Nonlinear Structures & Systems, Volume 1

Conference Proceedings of the Society for Experimental Mechanics Series
Gaetan Kerschen · Matthew R. W. Brake

Ludovic Renson Editors
Nonlinear Structures
& Systems, Volume 1
Proceedings of the 38th IMAC, A Conference
and Exposition on Structural Dynamics 2020
Conference Proceedings of the Society for Experimental Mechanics
Series
Series Editor
Kristin B. Zimmerman, Ph.D.
Society for Experimental Mechanics, Inc.,
Bethel, CT, USA
The Conference Proceedings of the Society for Experimental Mechanics Series presents early findings and case studies from
a wide range of fundamental and applied work across the broad range of fields that comprise Experimental Mechanics.
Series volumes follow the principle tracks or focus topics featured in each of the Society’s two annual conferences: IMAC,
A Conference and Exposition on Structural Dynamics, and the Society’s Annual Conference & Exposition and will address
critical areas of interest to researchers and design engineers working in all areas of Structural Dynamics, Solid Mechanics
and Materials Research.
More information about this series at http://www.springer.com/series/8922

Gaetan Kerschen • Matthew R. W. Brake • Ludovic Renson
Editors
Nonlinear Structures & Systems, Volume 1

Proceedings of the 38th IMAC, A Conference and Exposition on
Structural Dynamics 2020
Editors
Gaetan Kerschen Matthew R. W. Brake
Department of Aerospace Department of Mechanical Engineering
and Mechanical Engineering Rice University
University of Liège Houston, TX, USA
Liège, Belgium
Ludovic Renson
Merchant Venturers Building
University of Bristol
Bristol, UK
ISSN 2191-5644 ISSN 2191-5652 (electronic)

Conference Proceedings of the Society for Experimental Mechanics Series
ISBN 978-3-030-47625-0 ISBN 978-3-030-47626-7 (eBook)
https://doi.org/10.1007/978-3-030-47626-7
© The Society for Experimental Mechanics, Inc 2021

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights
of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific
statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date
of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional
affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
Nonlinear Structures & Systems represents one of eight volumes of technical papers presented at the 38th IMAC, A
Conference and Exposition on Structural Dynamics, organized by the Society for Experimental Mechanics, and held
in Houston, Texas, February 10–13, 2020. The full proceedings also include volumes on Dynamics of Civil Structures;
Model Validation and Uncertainty Quantification; Dynamic Substructures; Special Topics in Structural Dynamics &
Experimental Techniques; Rotating Machinery, Optical Methods & Scanning LDV Methods; Sensors and Instrumentation,
Aircraft/Aerospace, Energy Harvesting & Dynamic Environments Testing; and Topics in Modal Analysis &Testing.
Each collection presents early findings from experimental and computational investigations on an important area within
Structural Dynamics. Nonlinearity is one of these areas.
The vast majority of real engineering structures behave nonlinearly. Therefore, it is necessary to include nonlinear effects
in all the steps of the engineering design: in the experimental analysis tools (so that the nonlinear parameters can be correctly
identified) and in the mathematical and numerical models of the structure (in order to run accurate simulations). In so doing,
it will be possible to create a model representative of the reality which, once validated, can be used for better predictions.
Several nonlinear papers address theoretical and numerical aspects of nonlinear dynamics (covering rigorous theoretical
formulations and robust computational algorithms) as well as experimental techniques and analysis methods. There are also
papers dedicated to nonlinearity in practice where real-life examples of nonlinear structures will be discussed.
The organizers would like to thank the authors, presenters, session organizers, and session chairs for their participation in
this track.
Liège, Belgium G. Kerschen

Houston, TX, USA M. R. W. Brake
Bristol, UK L. Renson
v
Contents
1 Convergence Study on the C-Beam Using Joint Modes Based on Trial Vector Derivatives . . . . . . . . . . . . . . . . . . . . . 1
Florian Pichler and Wolfgang Witteveen
2 Semi Hyper-Reduction for Nonlinear Surface Loads on Finite Element Structures by the Use of Stress
Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Lukas Koller, Wolfgang Witteveen, and Florian Pichler
3 Parameter Identification of a Linear Substitution Model for Nonlinear Contact and Damping Inside
Lap Joints Using Distributed Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Stefan Pöchacker, Thomas Lauß, Florian Pichler, Stefan Oberpeilsteiner, and Wolfgang Witteveen
4 Flexible Multibody Dynamics: Which Terms of the Equations of Motion Have to Be Considered for
Which Problem? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Wolfgang Witteveen and Florian Pichler
5 Application of the Bouc-Wen Model to Bolted Joint Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Drithi Shetty, Matthew S. Allen, and Joseph D. Schoneman
6 Some Aspects of Using the Random Decrement Technique for Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Karsten K. Vesterholm, Rune Brincker, and Anders Brandt
7 Investigation of String Motions of Bowed String Instruments: A Finite Element Approach . . . . . . . . . . . . . . . . . . . 43
Özge Akar and Kai Willner
8 FreeDyn: A Free, Flexible and State of the Art Multibody Simulation Package for Education,
Research and Industrial Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Wolfgang Witteveen, Lauss Thomas, and Oberpeilsteiner Stefan
9 Force Probing to Access Potential Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Yawen Xu and Lawrence N. Virgin
10 Modeling and Experimental Validation of a Pylon Subassembly Mockup with Multiple Nonlinearities. . . . . . 59
Connor Ligeikis, Adam Bouma, Justin Shim, Simone Manzato, Robert J. Kuether, and Daniel R. Roettgen
11 Comparison Between Control-Based Continuation and Phase-Locked Loop Methods for the
Identification of Backbone Curves and Nonlinear Frequency Responses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Florian Müller, Gaëtan Abeloos, Erhan Ferhatoglu, Maren Scheel, Matthew R. W. Brake, Paolo Tiso, Ludovic
Renson, and Malte Krack
12 Implementing the Restoring Force Surface Method to Fit Experimentally Measured Modal Coupling
Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Benjamin Moldenhauer, Daniel R. Roettgen, and Benjamin Pacini
13 Identification of Backbone Curves and Nonlinear Frequency Responses using Control-based
Continuation and Local Gaussian Process Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Ludovic Renson
vii
viii Contents
14 Model Correlation to a Nonlinear Bolted Structure Using Quasi-Static Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . 87

Mitchell Wall, Seyed Iman Zare Estakhraji, and Matthew S. Allen
15 Numerical Continuation of Periodic Orbits for Harmonically Forced Nonlinear Systems
with Iwan Joints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Seyed Iman Zare Estakhraji, Matthew S. Allen, and Drithi Shetty
16 A Digital Absorber for Nonlinear Vibration Mitigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Ghislain Raze, Sylvain Guichaux, Andy Jadoul, Valery Broun, and Gaetan Kerschen
17 Control-Based Continuation of Nonlinear Structures Using Adaptive Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Gaëtan Abeloos, Ludovic Renson, Christophe Collette, and Gaetan Kerschen
18 Tuning of Finite Element Model Parameters to Match Nonlinear Reduced Order Models. . . . . . . . . . . . . . . . . . . . . 113
Kyusic Park and Matthew S. Allen
19 Towards an Understanding of the Transient Behavior of the Five-Parameter Iwan-Type Model . . . . . . . . . . . . . . 117
Robert M. Lacayo and Matthew S. Allen
20 On the Solution of Nonlinear Algebraic Equations Following Periodic Forced Response Analysis
of Nonlinear Structures Using Different Nonlinear Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
H. Sefa Kizilay and Ender Cigeroglu
21 Vibration-Based Bolt Tension Estimation for Multi-bolt Joints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
M. Brøns, A. Plaugmann, J. J. Thomsen, and A. Fidlin
22 Experimental Application of Control-Based-Continuation for Characterization of Isolated Modes on
Single- and Multiple-Degree-of-Freedom Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Gleb Kleyman, Martin Paehr, and Sebastian Tatzko
23 Nonlinear Function Selection and Parameter Estimation of Structures with Localised Nonlinearities,
Part 1: Numerical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Sina Safari and Julian M. Londono Monsalve
24 Tutorial on Nonlinear Reduced Order Modeling for Nominally Cyclic Symmetric Structures
and Rotating Machinery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Mainak Mitra, Andrea Lupini, Andrew Madden, Chiara Gastaldi, and Bogdan Epureanu
25 Excitation Techniques for Nonlinear Dynamic Systems: A Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Mahesh Nagesh, Akhil Sharma, Randall J. Allemang, and Allyn W. Phillips
26 Flutter of Double-Bay Panels with Finite Midbay Stiffness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
J. D. Schoneman
27 Nonlinear System Identification of a Jointed Structure Using Full-Field Data: Part II Analysis . . . . . . . . . . . . . . 185
Giancarlo Kosova, Mengshi Jin, Mattia Cenedese, Wei Chen, Aryan Singh, Debasish Jana,
Matthew R. W. Brake, Christoph W. Schwingshackl, Satish Nagarajaiah, Keegan J. Moore,
and Jean-Philippe Noël
28 Control Parameters in Non-linear Properties of Linear Guideway in Lateral Direction . . . . . . . . . . . . . . . . . . . . . . . 189
Ting-Yen Wu, Yi-Chun Lo, and Yum Ji Chan
29 Inferring Unstable Equilibrium Configurations from Observed Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Yawen Xu, Lawrence N. Virgin, and Richard Wiebe
30 Bolt Preload Loss Due to Modal Excitation of a C-Beam Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Max Miller, Chris Johnson, Noah Sonne, John Mersch, Robert J. Kuether, Jeff Smith, Jonel Ortiz,
Gustavo Castelluccio, and Keegan J. Moore
31 Nonlinear Dynamic Analysis of Bolted Joints: Detailed and Equivalent Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
N. Jamia, H. Jalali, J. Taghipour, M. I. Friswell, and H. H. Khodaparast
32 The Relevance of Nonlinear Normal Modes for Randomly Excited Nonlinear Mechanical Systems . . . . . . . . . . 223
Thomas Breunung and George Haller
Contents ix
33 Analysis of an Actuated Frictional Interface for Improved Dynamic Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

M. Lasen, Y. Sun, C. W. Schwingshackl, and D. Dini
34 Vibration Reduction of a Structure by Using Nonlinear Tuned Vibration Absorbers . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Muhammed Emin Dogan and Ender Cigeroglu
35 Simulation-Free Reduction Basis Interpolation to Reduce Parametrized Dynamic Models of
Geometrically Non-linear Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Christian H. Meyer and Daniel J. Rixen
36 Experimental Spectral Submanifold Reduced Order Models from Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . 249
Mattia Cenedese and George Haller
37 Development of an Experimental Rig for Emulating Undulatory Locomotion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
S. N. H. Syuhri, A. McCartney, and A. Cammarano
38 Identification and Modeling of a Variable Amplitude Fatigue Experiment Apparatus with Damaged
Beam Specimen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Hewenxuan Li and David Chelidze
39 Higher-Order Decompositions for Modal Identification and Model Order Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . 271
David Chelidze
40 Utilizing Modal Testing for Monitoring the Structural Health of Wind Tunnel Facility Hardware . . . . . . . . . . . 279
Kenneth Pederson and Vicente Suarez
41 An Efficient Coupled Modal Quasi-static Approach for Characterizing Non-linear Modal Properties
of Prestressed Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Nidish Narayanaa Balaji and Matthew R. W. Brake
42 An Assessment of the Applicability and Epistemic Uncertainties Inherent to Different Classes of
Friction Models for Modeling Bolted Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Justin H. Porter, Clayton R. Little, Nidish Narayanaa Balaji, and Matthew R. W. Brake
43 Hyper-Reduction Approaches for Contact Modeling with Small Tangential Displacements:
Applications for a Bolted Joint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Nidish Narayanaa Balaji, Tobias Dreher, Malte Krack, and Matthew R. W. Brake
44 Fatigue Damage of a Single-Edge Notched Beam Specimen Under Variable Amplitude Loading with
Similar Probabilistic and Cycle Counting Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
Chapter 1
Convergence Study on the C-Beam Using Joint Modes Based
on Trial Vector Derivatives
Florian Pichler and Wolfgang Witteveen
Abstract In the last years joint modes based on trial vector derivatives have been presented. These joint modes allow an
accurate consideration of nonlinear contact and friction forces inside jointed structures. If these joint modes are used for
model reduction, even high fidelity finite element models can be time-integrated with acceptable effort. In recent research on
joint nonlinearity and reduced order modelling a generic structure called “C-beam” or “S4 beam” was used. This structure
consists of two C-shaped beams assembled at their ends by two bolted joints. In this work, the previously mentioned model
reduction approach is applied to the C-Beam. A convergence study with respect to the number of additional joint modes for
different excitation levels is presented. In addition, it is investigated whether joint modes are able to reproduce well-known
damping phenomena due to friction inside joints at different preload levels. For this purpose, a detailed observation of local
sticking/slipping effects is given.
Keywords Joint contact · Friction · Damping · Model order reduction · Nonlinearity
1.1 Introduction
Complex mechanical structures are often an assembly of several substructures and hence are connected by some types of
mechanical joints like bolted joints. The nonlinear contact and friction forces and the resulting damping inside the joint can
significantly determine the dynamic response of such structures [1]. This is especially the case for widely distributed joints
where the stiffness and energy dissipation of the structure is mainly determined by the joint. On the other hand, the local
contact and friction forces influence the stress distribution inside the joint and therefor have significant influence on fatigue
and wear prediction. Nevertheless, these effects and their influence on dynamics are still an active field of research in terms
of experiments and numerical simulation.
Structures including contact and friction forces are commonly analyzed with the finite element method (FEM). This
method delivers accurate results in terms of contact stresses and local sticking/slipping effects. However, to ensure good
result quality a high mesh density in the joint region is required which leads to an unacceptable computational effort for
dynamic analyses. For the numerical time integration of linear structures model order reduction strategies like the Craig-
Bampton method [2] are commonly used. In the last years, several strategies have been presented to extend this method for
the computation of jointed structures. The most intuitive approach is to preserve the nonlinear degrees of freedom (DOF) at
the joint interface with static constraint modes. Depending on the size of the joint and the mesh density, this approach leads a
high number of DOF in the reduced system and hence, the advantage of model order reduction would be lost. Witteveen [3,
4] presented a specialized extension of the Craig-Bampton reduction base with so-called joint interface modes. Furthermore,
the authors presented in [5, 6] an extension strategy with joint modes based on the trial vector derivatives (also called joint
trial vectors JTVs) of the Craig-Bampton reduction base of the linear structure. The latter mentioned approach was integrated
via a C++ subroutine into the commercial available multibody software MSC.Adams. In this way, jointed structures can be
directly analyzed within a flexible multibody simulation involving realistic dynamic loads of the surrounding components.
The energy dissipation in joints is mainly caused by relative displacement of the contact pairs. A detailed experimental
investigation of dry friction inside joints has been performed by Gaul and his associates [1, 7]. Considering the entire joint
three types of motion can occur; sticking: the entire joint area remains stuck; micro-slip: the contact area is partially slipping
F. Pichler () · W. Witteveen

Department of Mechanical Engineering, School of Engineering, University of Applied Sciences Upper Austria, Wels, Austria
e-mail: [email protected]
© The Society for Experimental Mechanics, Inc 2021 1

G. Kerschen et al. (eds.), Nonlinear Structures & Systems, Volume 1, Conference Proceedings of the Society for
Experimental Mechanics Series, https://doi.org/10.1007/978-3-030-47626-7_1
2 F. Pichler and W. Witteveen
and sticking; macro-slip: the entire contact area is slipping. A numerical contact and friction model that enables gaping,
sticking or slipping can reproduce the characteristics of the experimental results. For an accurate representation of the micro-
slip and macro-slip friction, this numerical model has to be applied to a rather fine finite element (FE) mesh in the joint region.
With the mentioned approach of joint modes it is possible to perform dynamic simulations and get detailed information of
the contact and friction situation inside the joint. To analyze this information the authors implemented a postprocessing tool
which allows a graphical visualization of the contact situation.
A structure called “C-beam” or “S4 beam” was used several times in the last years for studies in the field of interface
reduction of bolted structures (see [8, 9]). This structure is used in this paper to present a detailed convergence study using
joint modes based on trail vector derivatives. Furthermore, the change in micro-slip and macro-slip regions inside the joint
is analyzed.
1.2 Background
In this theory section only a very brief review on joint modes and nonlinear joint forces in the context of multibody simulation
is given. General and detailed information on multibody simulation and model order reduction can be found in the literature,
for instance [10–12].
Extended reduction base with joint modes The joint modes used in this paper are implemented as presented in [6]. In terms
of model order reduction by projection the deformation state of a flexible structure x is approximated by r time invariant
trial vectors φ i , often also called modes. These modes are weighted with scaling factors qi , also called modal coordinates.
Mathematically this can be written as

r
x≈ φi qi = Φq (1.1)
i=1
where the matrix contains in its columns the trial vectors and the scaling factors are collected in the vector q. For linear
structures the component mode synthesis (CMS) based on the Craig-Bampton method [13] is well established. For this
reduction strategy, the matrix is a combination of constraint modes CM and normal modes NM . In [6] an extension of
this reduction base with preload modes PM which capture the deformation state due to preload forces and nonlinear contact
is presented. Additionally, the reduction base is extended with joint modes JM which allow an accurate computation of the
deformation inside the joint. The final reduction base can be given as

Φ = Φ CM Φ NM Φ PM Φ JM . (1.2)
Joint forces in the framework of multibody simulation The equations of motion (eom) for a flexible multibody system
including nonlinear joint forces in combination with constraint equations can be written as (see [5])
M̂ (q) q̈ + K̂ (q) + CTq (q) λ = Qv + Qe + Qnl

(1.3)
C (q) = 0.
In the latter equation, q denotes the vector of generalized coordinates, M̂, K̂ represent the mass and stiffness matrix of
the system. The vector Qv denotes the quadratic velocity vector, Qe the generalized external forces, and Qnl the generalized
nonlinear joint forces due to contact and friction. Furthermore, C and Cq represent the constraint equations and the constraint
Jacobian which is associated with the Lagrange multipliers λ. Partitioning the vector of generalized coordinates into
translational coordinates R, rotational coordinates , and flexible coordinates qf the eom for one jointed flexible body
can be written as
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
mRR mRΘ mRf R̈ CTR 0 0 0 R (Qv )R (Qe )R 0
⎣ ⎢ ⎥
mΘΘ mΘf ⎦ ⎣ Θ̈ ⎦ + ⎣ CTΘ ⎦ λ + ⎣ 0 0 ⎦ ⎣ Θ ⎦ = ⎣ (Qv )Θ ⎦ + ⎣ (Qe )Θ ⎦ + ⎣ 0 ⎦ .
T
sym mff q̈f Cf sym Ω 2 qf (Qv )f (Qe )f (Qnl )f
(1.4)
1 Convergence Study on the C-Beam Using Joint Modes Based on Trial Vector Derivatives 3
Fig. 1.1 Practical computation workflow
The generalized joint forces are therein computed via the nonlinear contact and friction forces on FE level fnl and the
extended reduction base defined by Eq. (1.2), resulting in
(Qnl )f = Φ T fnl . (1.5)
Practical computation workflow In Fig. 1.1 the practical computation workflow for the outlined theory is visualized. Based
on the FE model and the classical CMS reduction base a self-written graphical interface (Contact Definer) is used to
automatically compute the extended reduction base (Eq. (1.2)). Therefore, the contact pairs need to be determined on FE
level. The computation of contact and friction forces is performed by a user-written subroutine (Contact Analyzer) and the
vector of generalized joint forces is returned to the multibody simulation tool. After the simulation the Contact Definer can
be used for a detailed investigation of interesting contact values, for instance contact pressure or contact status (sticking,
slipping).
1.3 Model Description
Finite element model The analyzed structure was named “C-beam” or “S4 beam” in [8, 9]. The structure consists of two
C-shaped beams assembled at their ends by two bolted joints. The finite element model of the structure is depicted in Fig.
1.2 and the joint is marked with a red line. The model consists of 370450 nodes and 335010 linear hexaeder and pentaeder
solid elements. Inside the joint area approximately 6700 node to surface contact pairs are considered. In order to reproduce
the test bench described in [9] two nodes for a soft suspension (CSS1, CSS2) and one additional node for the application of
the force impulse (IA) are created. These nodes are coupled with RBE3 elements.
In Fig. 1.3 the representation of the screw in the FE model is depicted. The screw head is simplified using RBE3 coupling
elements. The screw shaft is modelled using beam elements. At the nodes SN1 and SN2 the bolt preload is applied. These
Fig. 1.2 Finite element model of the C-beam
Fig. 1.3 Representation of the screw
two nodes are connected with a soft beam element in order to avoid pure rigid body motion of the two C-shaped beams with
respect to each other.
Multibody model Representing the test bench of [9] in the multibody simulation the suspension of the structure with fishing
silk is realized with soft springs at the nodes CSS1 and CSS2. The excitation with the impact hammer on the test bench
was modeled with a Haversine force impulse with 1 ms. The bolt preload force was applied at the nodes SN1 and SN2
with 21.2 kN for both screws. The modal damping of the C-beam was set based on the eigenfrequencies to <2000 Hz: 1%;
2000–10,000 Hz 1–10% (smooth step); >10,000 Hz: 10%.
Reduction base The mode base for the convergence study of the C-beam was created as described in the previous section.
Based on the defined interface nodes of the FE model 22 constraint modes (2 × 6 due to CSS1 and CSS2; 6 due to IA; 4 due
to SN1 and SN2 for both screws) have been computed. Additionally 15 normal modes are considered in the reduction base.
The static deformation due to bolt preload of both screws is included with one preload mode. The number of joint modes is
varied in the next section in order to investigate their convergence behavior.
1.4 Convergence Study
Convergence of the preload state In a first step the results of the static preload state are compared to the results presented in
[9]. In the latter mentioned work, the same structure (FE model) was used and results for contact status and contact pressure
based on a nonlinear FE computation with Abaqus were presented. For the simulations in MSC.Adams a linear penalty model
Closed (Sticking)
Closed (Slipping)
Open
Abaqus presented in [9] MSC.Adams (20 joint modes)

Contact Pressure
3.41e+04
CPRESS 2.92e+04
+8.260e+03
+7.571e+03 2.44e+04
+6.883e+03
+6.195e+03
+5.506e+03 1.95e+04
+4.818e+03
+4.130e+03 1.46e+04
+3.442e+03
+2.753e+03 9.75e+03
+2.065e+03
+1.377e+03 4.87e+03
+6.883e+02
+0.000e+00
0.000
Abaqus presented in [9] MSC.Adams (20 joint modes)
Fig. 1.4 Comparison of the preload state; (a) contact status; (b) contact pressure
Fig. 1.5 Transient contact area for 15 N excitation force; left – absolute value; right – relative error
with a penalty stiffness of kN = 1.0e4 N/mm3 was used. In Fig. 1.4 the contact status and the contact pressure computed with
joint modes and a preload mode is compared with the Abaqus solution presented in [9].
The difference in the results of the contact pressures may be caused by the computation of the preload mode with
MSC.Nastran, which may be not the optimal choice for nonlinear FE computation. Nevertheless, in general a good
accordance with the results presented in [9] can be observed.
Convergence with respect to the excitation level In this section, the transient contact area is used as convergence criteria.
The contact area is defined by the area of nodes that are in contact divided by the entire joint area. The magnitude of the
Haversine impulse is varied for this study between 15 N (see Fig. 1.5) and 500 N (see Fig. 1.6). In both figures the absolute
value and a relative error with respect to a converged solution (200 joint modes). It can be seen that for both excitation levels
the relative error is below 5% already with 20 joint modes. For an excitation with 15 N the relative error is below 1% with
60 joint modes. For the excitation with 500 N this threshold is reached with 80 joint modes. It can been stated, that the
convergence is for both excitation levels similar and hence, not strongly depending on the excitation level.
Fig. 1.6 Transient contact area for 500 N excitation force; left – absolute value; right – relative error
×10–3
6
friction - full preload
friction - half preload
4
displacement [inch]
–2
–4
–6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time [s]
Fig. 1.7 Displacement of IA for full and half preload
Study with respect to the preload force For this investigation, the preload force was varied to determine the influence on
the friction. Therefore, dynamic simulations with a three parameter Coulomb friction model (see [5]) (friction coefficient
μ = 0.3; sticking stiffness kt1 = 3.0e4 N/mm3 ; slipping stiffness kt2 = 0.0 N/mm3 ) considering 60 joint modes were
performed. The simulations were performed with full preload (21.2 kN) and half preload (10.6 kN) over 4 s. In Fig. 1.7
the displacement of the force application point IA is plotted for full and half preload. It should be mentioned that the results
are shifted to zero, which means that the motion caused by the soft suspension is omitted. The figure shows that the damping
due to friction is higher in case of reduced preload. This has been expected since more slipping is possible. Furthermore, it
can be seen that the time range for meaningful slipping is longer for the case of half preload. For both variants, the curves
decrease with the same rate after approximately 3 s which means that the slipping is similar.
For the reduced preload the contact situation is investigated in detail for three different times in Fig. 1.8. In each row of
this figure the contact status (gapping, sticking, slipping) is visualized for approximately one oscillation of the first bending
mode. It can be clearly seen that at the begin of the simulation (t = 0.05 s) the slipping area is widely spread over the area in
contact which leads to a high damping due to friction. This slipping area is decreasing for the time instances at t = 0.5 s and
t = 4.0 s.
Fig. 1.8 Contact status for one oscillation of the first bending mode at different times: (a) 0.05 s, (b) 0.5 s, (c) 4.0 s; (blue – gapping, green –
sticking, red – slipping)
1.5 Conclusion
In this work an extended reduction base using a preload mode and joint modes as presented in [6] is investigated on the C-
beam. The theory of the extended reduction base and the consideration of nonlinear joint forces in the context of a multibody
simulation was shortly reviewed. Based on this, the practical simulation workflow was outlined.
A static convergence study showed that the preload state can be accurately reproduced by the preload mode compared with
a full nonlinear FEM computation. Dynamic investigations confirmed the excellent convergence of joint modes at different
excitation levels. The dynamic contact situation of 6700 node to surface contact pairs could be reproduced with 80 joint
modes. The preload force was varied in order to study the influence on damping and friction. The investigation showed that
more energy is dissipated in the case of reduced preload force. Furthermore, the dynamic contact situation inside the joint
was analyzed in detail with the Contact Definer, which enables a deeper understanding of the energy dissipation caused by
dry friction.
References
1. Gaul, L., Lenz, J.: Nonlinear dynamics of structures assembled by bolted joints. Acta Mech. 125(1–4), 169–181 (1997). https://doi.org/10.1007/
BF01177306
2. Craig, R.R., Bampton, M.C.C.: Coupling of substructures for dynamic analysis. AIAA J. 6(7), 1313–1319 (1968). https://doi.org/10.2514/
3.4741
3. Witteveen, W., Irschik, H.: Efficient mode-based computational approach for jointed structures: joint interface modes. AIAA J. 47(1), 252–263
(2009). https://doi.org/10.2514/1.38436
4. Witteveen, W., Sherif, K.: POD based computation of joint interface modes. In: Proulx, T. (ed.) Linking Models and Experiments, Volume 2:
Proceedings of the 29th IMAC, A Conference on Structural Dynamics. Springer, New York (2011)., ISBN 978-1-4419-9305-2:19–28, 2011
5. Pichler, F., Witteveen, W., Fischer, P.: A complete strategy for efficient and accurate multibody dynamics of flexible structures with large lap
joints considering contact and friction. Multibody Syst. Dyn. 40(4), 407–436 (2017). https://doi.org/10.1007/s11044-016-9555-2
6. Pichler, F., Witteveen, W., Fischer, P.: Reduced-order modeling of preloaded bolted structures in multibody systems by the use of trial vector
derivatives. J. Comput. Nonlinear Dyn. 12(5), 051032-051032-12 (2017). https://doi.org/10.1115/1.4036989
7. Lenz, J., Gaul, L.: The influence of microslip on the dynamic behavior of bolted joints. In: Proceedings of the IMAC XIII, Society of
Experimental Mechanics, pp. 248–254 (1995)
8. Hughes, P.J., Scott, W., Wu, W., Kuether, R.J., et al.: Interface reduction on Hurty/Craig-Bampton substructures with frictionless contact. In:
Kerschen, G. (ed.) Nonlinear Dynamics, Volume 1, Conference Proceedings of the Society for Experimental Mechanics Series, pp. 1–16.
Springer International Publishing, Cham, ISBN 978-3-319-74279-3 (2019)
9. Wall, M., Allen, M., Zare, I.: Predicting S4 beam joint nonlinearity using quasi-static modal analysis. In: Kerschen G., Brake M., Renson L.
(eds) Nonlinear Structures and Systems, Volume 1. Conference Proceedings of the Society for Experimental Mechanics Series, pp. 39–51.
Springer, Cham, ISBN 978-3-030-12390-1 (2020)
10. Shabana, A.A.: Flexible multibody dynamics: review of past and recent developments. Multibody Syst. Dyn. 1(2), 189–222 (1997). https://
doi.org/10.1023/A:1009773505418
11. Shabana, A.A.: Dynamics of Multibody Systems. Cambridge University Press, Cambridge, UK, ISBN 9780511610523 (2005)
12. Qu, Z.-Q.: Model Order Reduction Techniques with Applications in Finite Element Analysis. Springer, London, ISBN 978-1-4471-3827-3
(2004)
13. Craig, R.R.: A review of time-domain and frequency-domain component mode synthesis method. In: Combined Experimental/Analytical
Modeling of Dynamic Structural Systems; Proceedings of the Joint Mechanics Conference, pp. 1–30 (1985)
Chapter 2
Semi Hyper-Reduction for Nonlinear Surface Loads on Finite
Element Structures by the Use of Stress Modes
Lukas Koller, Wolfgang Witteveen, and Florian Pichler
Abstract The determination of nonlinear state-dependent surface loads, acting on finite element (FE) structures, represents
a computationally challenging and costly task in dynamic simulations. While for time integration an enormous reduction of
the FE models number of degrees of freedom (DOFs) can be achieved by subspace projection, the computation of nonlinear
surface loads usually depends on the non-reduced physical DOFs. In order to overcome this issue, so-called Hyper-Reduction
(HR) methods have been introduced. These methods try to compute the surface loads in a reduced subspace as well. In
this publication, an intermediate approach is proposed, which is called “Semi Hyper-Reduction” (SHR). The equations for
computing the surface loads are built up in the full space and then projected into a lower dimensional subspace via proper
force trial vectors. The required force trial vectors, called “stress modes”, thereby can be determined a priori without any
nonlinear computations using the full DOF model. As a numerical example, a 3D crank drive is used, where the piston and
the cylinder are separated by a hydrodynamic lubrication film, which is considered by Reynolds equation.
Keywords Model order reduction · Semi Hyper-Reduction · Multibody simulation · Reynolds equation ·
State-dependent nonlinear load
2.1 Introduction
For the computation of contact problems, within dynamic simulations, it is a common approach to describe the involved
contact partners by means of the finite element (FE) method. Depending on the complexity of the entire model, the number
of considered degrees of freedom (DOFs) may become dramatically high. In order to decrease the computational effort,
various model order reduction (MOR) methods were developed, which allow a DOF reduction. A common procedure is to
project the high-fidelity model with the full DOFs into a subspace where it can be described with much less equations [1].
If the considered problem shows a nonlinear behavior or nonlinear loads are acting, there may be a substantial loss in
the computational efficiency even if an MOR is applied. The reason for this behavior stems from the fact that such effects
usually are computed in the full DOF domain and must be projected into the reduced subspace afterwards. In literature,
efficient approaches for treating nonlinearities in the reduced subspace are known under the term of Hyper-Reduction (HR)
[2]. These methods aim at a description of nonlinear relationships in a reduced space as well. Different techniques, especially
for geometrical nonlinearities, are discussed in [2–7].
In several of the cited HR methods, training sets obtained from nonlinear full DOF computations are necessary to create a
proper subspace for both MOR and HR. In this contribution, the reduction procedure is based on deformation and stress
trail vectors, which are identified a priori from linear FE computations. The applied computational algorithm for their
determination is efficient and furthermore yields a linear relationship between the two reduction bases.
For applying the proposed method of SHR, the equations for describing the nonlinearity are constructed in the unreduced
domain, solely with information of the physical coordinates. This characteristic is the reason for the added term “Semi”, since
in a HR process the previous preliminary step is also done in a reduced manner. With a subspace spanned by the a priori
determined stress modes, the initial, full dimensional equations are reduced in a subsequent transformation. Consequently,
the effort for computing the nonlinearity is mainly influenced by the number of stress modes and is almost unaffected from
L. Koller () · W. Witteveen · F. Pichler

Study Program “Mechanical Engineering”, University of Applied Sciences Upper Austria, Wels, Austria

10 L. Koller et al.
the physical domain. The described intermediate reduction step also allows the implementation of the SHR approach into
nodal-based and physical DOF related routines with a reasonable programming effort.
2.2 Model Reduction and Description of the Nonlinear Term
For reducing the number of DOFs of an FE model, reduction via projection is used. Thereby, the equations of motion (EOM)
Mẍ + Kx = fLIN + fNL (x, ẋ) (2.1)
are transformed into a subspace which is spanned by m invariant trail vectors ϕi , collected in the (n × m) matrix . The
(n × 1) deformation state vector x is approximated by the weighted superposition
x ≈ ϕ1 q1 + · · · + ϕm qm = q (2.2)
with the corresponding scaling factors q1 , . . . , qm . Inserting (2.2) into (2.1) and a multiplication with from the left leads
to
Mq̈ + Kq = fLIN + fNL (q, q̇) ⇒ M̃q̈ + K̃q = f̃LIN + f̃NL (q, q̇) (2.3)
with the reduced (m × m) mass matrix M̃ and stiffness matrix K̃ instead of the full DOF (n × n) matrices M and K. This
projection procedure also results in the reduced (m × 1) vectors of linear loads f̃LIN and nonlinear, state-dependent loads
f̃NL . Since the nonlinear term still needs to be computed in the full DOF space, its determination remains as a bottleneck
in the numerical computation. In order to obtain a reduced description of the nonlinear surface loads, the concept of modal
stress recovery [8] is applied. This method allows reconstructing the stresses arising at the surface utilizing modal stress
modes. Since this contribution is restricted to contact situations with normal stresses, only those, which are evoked in normal
direction, are reconstructed with the modal normal stress modes ψni . For reasons of readability, these are now referred to as
stress modes. Mathematically, this can be described by the following scaled superposition with the l modal coordinates ri
σn ≈ ψn1 r1 + · · · + ψnl rl = ψn r, (2.4)
where the (c × l) matrix n holds the l modal normal stress modes. Thereby, the variable c denotes the number of normal
DOFs in the contact area. Based on the law of action-reaction, the fundamental concept of the proposed approach is to use
p = −σn ≈ −ψn r ⇒ p = p (r) (2.5)
for describing the pressures p at the c points in the contact area of the surface.
2.3 Computation of Stress Modes
The determination of a proper stress mode base n is based on the idea of modal stress recovery [8], which states, that for
n
each deformation mode ϕi a related stress mode ψi exists. If in (2.4) the stress modes based on the deformations modes
are used, the stress state can be reconstructed with the modal coordinates from (2.2). Thus, the formula from (2.4) can be
rewritten as
n n n
σn ≈ ψ1 q1 + · · · + ψm qm = ψ q, (2.6)
n
whereby the single stress modes are collected columnwise in the (c × m) matrix . However, a proper approximation of the
structures deformations x combined with a precise stress reconstruction demands two properties from : (1a) Description
of the structures global and (1b) local deformation behavior (especially in the contact area) and (2) the stress modes must
approximate the arising normal stresses accurately. An established mode base is those of Craig-Bampton (CB) CB , which
contains the (n × v) matrix v of vibrational modes and the (n × s) matrix s holding s deformation mode shapes stemming
2 Semi Hyper-Reduction for Nonlinear Surface Loads on Finite Element Structures by the Use of Stress Modes 11
from static load cases [1]. Since a classical CB mode base merely pictures the global deformations and thus is not suitable
for this task, an extension with the (n × l) matrix l is recommended in [9], leading to a (n × m) matrix
= [CB l ] , (2.7)
where m = v + s + l holds. Thereby, the modes in l are intended to enhance the properties in the contact area. For a detailed
information on l , see [9]. In order to obtain a simpler description of the reduced EOMs, the matrix is orthonormalized
n
with respect to the mass, resulting in the final mode base . According to the relationship from (2.6), the stress modes
are then determined with less effort out of . Since the global deformation modes does not evoke any surface stresses in
the contact area, a direct usage of the determined stress modes is not recommendable. For this reason, the method of proper
n
orthogonal decomposition (POD) [10] is applied to extract a significant subspace from yielding an approximation of the
modal normal stresses with
n
σn ≈ q ≈ n r. (2.8)
n
As demonstrated in [9], this subspace can be mathematically recognized in the singular values of . Since they indicate
a steep drop at the lth position, only the stress modes originating from l are expected to contain the relevant information.
2.4 Idea of Semi Hyper-Reduction
The general idea behind SHR is to build up the equations for nonlinear surface loads with the physical information of the c
DOFs, followed by a projection into a subspace with the relationship p = p(r) ≈ − n r. As a result, the dimension of the
describing mathematical problem decreases from c to l which leads to a reduced linear system of equations
g (p) = 0 ⇒ g (r) = 0 (2.9)
or a reduced minimization problem
min J (p) . . . min J (r)

s.t. ci (p) = 0 i = 1, . . . , nC1 ⇒ s.t. ci1 (r) = 0 i = 1, . . . , nC1
1
(2.10)
s.t. cj2 (p) ≥ 0 j = 1, . . . , nC2 .. . . . s.t. cj2 (r) ≥ 0 j = 1, . . . , nC2
for determining the nonlinearity, where g is a vectorial function and J stands for a scalar cost function. After solving (2.9)
or (2.10), the surface load must be computed with p = p(r) and a further multiplication with is necessary to obtain f̃NL
finally.
2.5 Application of Semi Hyper-Reduction to Reynolds Equation
The proposed method of SHR is applied onto Reynolds equation (RE)
∂ h3 ∂p ∂ h3 ∂p ∂h ∂h
+ = 6U + 12 (2.11)
∂x η ∂x ∂y η ∂y ∂x ∂t
where p stands for the pressure, h is the gap height, U is the velocity, x and y are the coordinates and t the time. Thereby,
only the velocity in the x-direction is relevant. The nodal-based finite difference (FD) method is a commonly used approach
for solving (2.11) and leads to a linear system of equations
Ap = b (2.12)
with the (c × c) matrix A, the (c × 1) vector b and the (c × 1) vector of nodal pressures p. If now SHR is applied, the
pressures p in (2.12) are described by (2.5) and a following multiplication with ( n ) from the left leads to
12 L. Koller et al.

− n A − n r = − n b ⇒ Ar = b. (2.13)
This operations result in the (l × l) matrix A, the (l × 1) vector b and the coefficients r are obtained as solution.
Consequently, the application of SHR yields a linear system of equations which dimensions are reduced from c to l. After
determining the coefficients r, the pressure distributions can be computed by
p ≈ − n r. (2.14)
Accordingly, solving RE by applying SHR requires an FD discretization and the additional projections to obtain A, b and
p. Compared to the FD method, the overall computational effort is reduced if solving the system of Eq. (2.13) and the three
transformations require less CPU time as the solution of (2.12).
2.6 Numerical Studies
The approach of SHR was applied to a 3D crank drive of a diesel engine, where the flexible piston and the flexible cylinder
liner are separated by a lubrication film based on RE. The mode base of the flexible piston consists of s = 6 static mode
shapes, v = 20 vibrational modes and l = 150 modes for reconstructing the surface pressures. For checking the result quality
of the SHR, also computations with the established FD method were performed. Thereby, the FD mesh consists of c = 1400
normal DOFs in the contact area. Since Guembel’s boundary condition [11] is applied, only positive pressures are regarded
in the solution. In Fig. 2.1, the obtained pressures are compared at 398.6 and 610.1 crank angle degrees (CADs). As can be
seen, the distributions match very well, which proves the capability of SHR. It is worth mentioning, that both methods are in
a good agreement for all other CADs too.
The efficiency of the SHR with respect to the FDM will enhance, if c increases compared to the number of stress modes
l. Since for the 3D model a comparative study with different meshes would be very time consuming, a 2D model is chosen
for this analysis. As shown in Fig. 2.2 and elaborated in [9], the number of nodes in the contact area and thus the number of
c is refined from 202 up to 3202 normal DOFs. The trend of the plotted curves reveals that the gap between the FDM and
the SHR clearly diverges with an increasing number of nodes. This discrepancy stems from the lower computational effort
required for solving the reduced system of Eq. (2.13), compared to the unreduced equations in (2.12). The slight increase for
the SHR approach can be traced back to the projections in (2.13) and (2.14), since their dimension correlates with c.
Fig. 2.1 Pressure distributions obtained from FDM (top) and SHR (bottom)
2 Semi Hyper-Reduction for Nonlinear Surface Loads on Finite Element Structures by the Use of Stress Modes 13
Fig. 2.2 Estimation of the computational time (2D) [9]
2.7 Conclusion
In this contribution, a novel method for the reduced computation of state-dependent and nonlinear surface loads based on
the use of stress modes is presented. The term SHR is chosen because it represents an intermediate approach between nodal-
and pure trial vector-based solution techniques. In a first step, the nonlinearity is described in the nodal domain, followed by
a subspace projection of the characterizing equations. For this reduction process, the a priori determined stress modes are
used as trail vectors. Since they can be computed without any dynamic simulations of the non-reduced nonlinearity, the costs
for their analysis remains relatively small. As the numerical results reveal, the result quality of the SHR is identical to the
FD method. Furthermore, due to the vector-based characteristic, a considerable efficiency gain is obtained if the number of
DOFs in the area of the nonlinearity increases.
References
1. Craig, R.R., Bampton, M.C.C.: Coupling of substructures for dynamic analyses. AIAA J. 6(7), 1313–1319 (1968)
2. Ryckelynck, D.: A priori hyperreduction method: an adaptive approach. J. Comput. Phys. 202(1), 346–366 (2005)
3. Barrault, M., Maday, Y., Nguyen, N.C., Patera, A.T.: An ‘empirical interpolation’ method: application to efficient reduced-basis discretization
of partial differential equations. C. R. Math. 339(9), 667–672 (2004)
4. Chaturantabut, S., Sorensen, D.C.: Nonlinear model reduction via discrete empirical interpolation. SIAM J. Sci. Comput. 32(5), 2737–2764
(2010)
5. Farhat, C., Avery, P., Chapman, T., Cortial, J.: Dimensional reduction of nonlinear finite element dynamic models with finite rotations and
energy-based mesh sampling and weighting for computational efficiency. Int. J. Numer. Methods Eng. 98(9), 625–662 (2014)
6. Hernández, J.A., Caicedo, M.A., Ferrer, A.: Dimensional hyper-reduction of nonlinear finite element models via empirical cubature. Comput.
Methods Appl. Mech. Eng. 313, 687–722 (2017)
7. Rutzmoser, J.B., Rixen, D.J.: A lean and efficient snapshot generation technique for the hyper-reduction of nonlinear structural dynamics.
Comput. Methods Appl. Mech. Eng. 325, 330–349 (2017)
8. Fischer, P., Witteveen, W.: Integrated MBS-FE-durability analysis of truck frame components by modal stresses. In: 15th ADAMS Users
Conference (2000)
9. Koller, L., Witteveen, W., Pichler, F., Fischer, P.: Semi hyper-reduction for finite element structures with nonlinear surface loads on the basis
of stress modes. J. Comput. Nonlin. Dyn (2020). https://doi.org/10.1115/1.4047334
10. Volkwein, S.: Model reduction using proper orthogonal decomposition. Available at: http://www.math.uni-konstanz.de/numerik/personen/
volkwein/teaching/POD-Vorlesung.pdf (2011). Accessed 15 Oct 2019
11. Bruce, R.W.: Handbook of Lubrication and Tribology, Volume II: Theory and Design. CRC Press, Boca Raton (2012)
Chapter 3
Parameter Identification of a Linear Substitution Model for
Nonlinear Contact and Damping Inside Lap Joints Using
Distributed Optimization
Study Program “Mechanical Engineering”
Stefan Pöchacker, Thomas Lauß, Florian Pichler, Stefan Oberpeilsteiner, and Wolfgang Witteveen
Abstract The numerical simulation of flexible system dynamics considering nonlinear contact and friction forces inside
joints is a very time-consuming process. In this work, a simple linear substitution model for the joint contact is proposed and
the parameters are identified. Subsequently, it is numerically examined whether this model is predictive and transferable. The
identified parameters are the tangential stiffness, normal stiffness, tangential damping and the contact area. All parameters
are assumed to be a function of the bolt preload. The reference is a high fidelity Finite Element (FE) model of two strips of
metal bolted together with three bolts. The solution of the reference model is calculated with distributed, nonlinear contact
and friction forces. During the optimization, the Fourier coefficient of the first mode was fitted. To speed up the large
number of time integrations a distributed optimization method is investigated. This decentralized method accelerates the
optimization process almost by the factor of the used computers. The determined parameters are used in a modified setting
and compared with a corresponding reference solution. The aim is to use the obtained substitute model for predictive and
tendential investigations of jointed structures.
Keywords Multibody system · Nonlinear joints · Numeric simulation · Parameter identification · Distributed
optimization · Linear substitution model · Fourier coefficient
3.1 Introduction and Motivation
In mechanical systems, joints between flexible bodies leads to nonlinear behavior and may have a strong effect on the
dynamic behavior (see [1] or [2] for example). This is caused by the nonlinear contact or friction forces. To simulate these
nonlinearities, several methods are available. The classical way is a nonlinear finite element computation. However, enormous
simulation times occur. Exact computation of high dynamics with finely meshed models is generally not feasible due to the
enormous computation times. Newer approaches result in a great time saving, while maintaining the same accuracy [3]. The
in-house tool JointAnalyser, which is based on such methods, allows calculations to be carried out with finely meshed, highly
dynamic models with moderate computing times.
Another modelling approach is the approximation of the essential behavior of a difficult nonlinear model with a linear
substitution model. Global properties can be predicted with sufficient accuracy, at least for smaller displacements. Such linear
models have the advantage of extremely short computation times. This contribution presents an investigation and parameter
identification of such a substitution model.
S. Pöchacker () · T. Lauß · F. Pichler · S. Oberpeilsteiner · W. Witteveen

University of Applied Sciences Upper Austria, Wels, Austria
e-mail: [email protected]; [email protected]; [email protected]; [email protected];
[email protected]

16 S. Pöchacker et al.
fast qualitative results

model-information
(preload, location, …) of different variants
variant 1
structure substitution
with model with
nonlinear linear bolted
joints variant 2
bolted joints
linear substitution variant 3

model generation
Fig. 3.1 Graphical representation of the research idea
3.2 Purpose of Research
Micro and macro slip can occur in a joint (e.g. a bolted connection). In general, macro slip should be avoided, as this indicate
a structure failure. However, micro slip is an important damping factor which occurs in addition to other types of damping,
and contributes significantly to the overall damping of the system [4].
Therefore, the goal is to find a purely linear substitution model of a bolt connection with physically interpretable
parameters. These parameters only depend on the preload [4, 5] and this makes the replacement model very easy and fast to
adapt to different joint situations. With the linear model, only the effect of slip in the microscopic range can be mapped. A
good description of micro and macro slip can be found in [6].
The idea of this research is illustrated in Fig. 3.1.
3.3 Mechanical Model of the Numerical Test Structure
3.3.1 Nonlinear High Fidelity FE-Model
To compute the dynamic behavior of a nonlinear structure, the following equations of motion has to be solved:
M ẍ + D ẋ + Cx = flin + fnl (3.1)
Thereby M is the mass, D the damping and C the stiffness matrix. x are the physical Degrees of Freedom (DoF’s), flin
the linear and fnl the nonlinear forces of the mechanical system. The single or double point above the DoF’s is the single or
double derivative with respect to time. With the JointAnalyser the real nonlinear conditions within a joint can be computed
very well. As the reference solution (kind of virtual experiment) a high fidelity Finite Element (FE) model with subsequent
Craig Bampton modal transformation [7] and additional joint modes [3] is used. The numerical test structure, a double beam
bolted three times with M8 bolts and clamped on one side, is sketched in Fig. 3.2. The bolts are not modeled. They are
included in the computation only as forces with corresponding constraint modes. In total, the reference solution is calculated
with 127 modes (80 joint modes and 47 Craig Bampton modes). Dynamic computations are simulated with the (Multi Body
Simulation) MBS program MSC ADAMS with linked JointAnalyser as a DLL for 1 [s] with a time increment of 1E-5. The
structure was loaded from the rest position with a stepforce at both tips in negative y-direction.
Each bolt is associated with an area of the beam. This can be seen in Fig. 3.3. Additionally, the dimensions are given in
the x-z plane. Figure 3.3 also shows the inside of the beam (dashed lines). In order to obtain approximately three identical
joints, no contact was defined for the first section, the clamping area.
3 Parameter Identification of a Linear Substitution Model for Nonlinear. . . 17
2 x 3mm
Preload 1 Bolt Head Seat D=14mm Preload 3
y Preload 2
x
z
Preload 1 Preload 2 Preload 3

2 x Stepforce
Fig. 3.2 Draft of the high fidelity model
Outside 525mm
75mm 75mm 2 x 75mm 2 x 75mm 75mm
30mm
x
z
No Contact Contact Area 450mm
Fig. 3.3 Outside of one beam with the acting area of the bolt head seats. The contact surface is divided in three equivalent sections (dashed lines)
3.3.2 Linear Substitution Model
The linear substitution model can be described by the following equation of motion in the physical space
M ẍ + D ẋ + Cx = flin (3.2)
or with the transformation x = q in the modal space
M̃ q̈ + D̃ q̇ + C̃q = f˜lin (3.3)
wherein the matrices and vectors are defined below.
x = q . . . DoF’s, (x are physical DoF’s, q are modalDoF’s) (3.4)
T M = M̃ = I . . . mass matrix (3.5)

⎡ ⎤
2η1 0
⎢ .. ⎥
T D = D̃ = ⎣ . ⎦ . . . damping matrix (3.6)
0 2ηn
⎡ 2 ⎤
1 0
⎢ .. ⎥
T C = C̃ = ⎣ . ⎦ . . . stiffness matrix (3.7)
0 2n
T flin = T (fc + fd + fl ) = f˜lin . . . force vector (3.8)

= ϕ1 · · · ϕn . . . matrix with trial vectors (3.9)
FE-Nodes
Interface- Contact Area

No Contact Node 225 x 2mm
z
Fig. 3.4 Schematic representation of the contact area partitioning
y
225 possible Nodepairs
z
Interface-Contact Node
dT
cT
cN
Fig. 3.5 Illustration of the linear contact model of an interface-contact node
The force vector flin consists of the parts: Force of interface nodes due to normal and tangential stiffness fc , force of
interface nodes due to tangential damping fd and force of the external load (tip load e.g.) fl . In order to vary the contact area
in the linear model, the potential contact area was divided into 225 strips, each 2[mm] wide. The corresponding FE nodes
are coupled with an interface node. So the range of the linear contact can be varied between 2 and 150[mm]. Now a linear
spring in tangential and normal direction as well as a viscous damping in tangential direction act between the corresponding
interface nodes of the upper and lower beam. The following drafts (Figs. 3.4 and 3.5) illustrates the situation of the linear
contact surface.
3.3.3 Used Model Parameters

High Fidelity FE-Model
The parameters
ofadhesion and sliding stiffness
were placed between the determined values of [4] and [5]. Sticking stiffness
kSt = 1000 mm3 , sliding stiffness kSl = 0 mm3 and friction coefficient μ = 0.3. The preload of the bolts has been varied
N N
between 16 and 3.5[kN]. The stepforce on the tips has an amount of 10[N] each.
Substitution Model
The mode basis of the linear substitution model has 26 eigenmodes with eigenfrequencies between 9.13 and 2719.29[Hz].
The modes 1 up to 18 are damped by 0.1% and the modes 19 up to 26 by 1%.
3.4 Distributed Optimization Strategy
There are direct methods (Nelder Mead method e.g.) to find an optimal set of parameters for a problem [8–10]. However,
these methods have a closed technique that can only be used in a sequential way. Within an optimization run, the integration
of the linear substitution model must be carried out very often. With several thousands calculations, this can take a few hours
of computation time.
3.4.1 Cost Function
The cost function

N
Ja = (y ti (pa ) − ȳ ti )2 (3.10)
i=1
is set up for this optimization task. Where y(pa ) are the amplitudes of the y-deflection of the two tips of the linear model in
frequency space with the parameter set pa . ȳ are the corresponding y-deflections of the reference model (Sect. 3.3.1) and N is
the number of frequencies. The scalar Ja is the quadratic error summed up over the optimization frequency (frequency-range
of the first mode). The detailed theory and procedure can be read in [11].
3.4.2 Space of Parameter
The linear model has a 4-dimensional parameter space (normal stiffness cN , tangential stiffness cT , tangential damping dT
and number of active interface-contact nodes n (Fig. 3.6)). The first three parameters are double values, but the number of
nodes is an integer. This would cause problems when using a simplex method. The simplest way of distribution would be
to optimize the double parameters on several computers at the same time for different but fixed integers for the activated
interface nodes.
In a design of experiments all parameters are discretized with the required step size. If a full factorial design is performed,
the cost function is computed with all parameter combinations. The results are stored in a table, the resulting costs for the
used parameter combination. In general, the number of simulations (grid points) is computed as a product based on the
individual discretizations of the parameters. If the parameters have to be discretized very finely, the number of grid points
increases very quickly. In order to reduce the number of necessary grid points, there are good methods of statistical design
of experiments [12–14].
Fig. 3.6 Illustration of the

4-dimensional parameter space
(cN . . . normal stiffness,
cT . . . tangential stiffness,
dT . . . tangential damping,
cT
n . . . number of active dT
interface-contact nodes)
cN
n dT
Ethernet/Internet
Main Computer
Ethernet/ Server
Internet Cluster e.g.
Sim. Nodes
…
Ethernet/Internet
Fig. 3.7 Schematic representation of the distributed computation resources
3.4.3 Distributed Resources
A number of simulation jobs are defined by a main computer and sent to a server. Every available simulation node asks the
server if a job has to be computed and sends the result back to the server after it is finished. As soon as all results have arrived
at the server, it reports back to the main computer with the collected results. On this computer the evaluation of the received
data can then be carried out (Fig. 3.7).
3.4.4 Determination of the Joint Parameters

Simplex Algorithm
The Simplex algorithm is a direct optimization method. It finds a minimum of the cost function by varying the parameters.
However, this method can only be distributed for different but fixed contact surfaces. For this purpose, simulation nodes are
parallel launched with one Simplex algorithm each for a fix defined contact surface. This method was used to determine the
precise joint parameters of the linear model for all preloads.
Distributed Optimization by Discretization of the Parameter Space
Once the computation results have been sent back to the main computer, they are available in table form. One row of the table
contains a parameter set and the associated resulted costs. This data can then be used to determine the optimal parameters.
The simplest method to determine the optimal joint parameters is to search for the minimum value of the cost function in this
table. The associated parameters represent a good set to simulate the global behavior of the nonlinear model with the linear
substitution model. To increase the accuracy, the parameter space can be discretized again near the first detected minimum.
Now the linear model can be simulated with the new parameter combinations again. This sequence can also be automated
very well and can be carried out several times until a defined variation limit is reached.
Another method to increase the accuracy is the multidimensional regression [15, 16]. With the help of this, a minimum of
the cost function can be determined and with the resulting parameters the joint can be mapped more precisely.
3.5 Results and Discussion
3.5.1 Virtual Testing with JointAnalyser
Figure 3.8 shows exemplary for some bolt preloads the y-deflection during the entire simulation time. Here, the reduction of
the oscillation mean value can be seen with lower preloads. A lower bolt preload results in a larger relative sliding movement
between the two sheets. This figure also shows a stiffening at higher preloads. The higher the preload of the bolts, the stiffer
the whole beam becomes and the higher the vibration frequency. Figure 3.9 shows the y-displacement curves enlarged in
the time range from 0.53 to 0.63 s. In addition, the small black boxes indicate the peak of the oscillation amplitude. This
illustrates the decrease of the amplitude and of the vibration frequency with a lower preload.
Figure 3.10 shows the magnitude of the Fast Fourier Transformation (FFT) of the y-deflection of the beam tip up to
450[Hz]. This shows that higher preloads leads to more significant contribution of higher modes.
Furthermore, the frequency range up to 30[Hz] was shown in Fig. 3.11. Here, the reduction of the eigenfrequency at lower
preload can be detected. In addition, the static deflection (mean value) of the beam tip was enlarged. An increase with a lower
preload can be seen here.
Fig. 3.8 y-deflection of the beam tip over the whole simulation time (preload PL in [N])
Fig. 3.9 y-deflection of the beam tip over time, zoomed in the time interval 0.53–0.63 [s] (preload PL in [N])
Fig. 3.10 Amplitude over the frequency (FFT(y-deflection)) of the ADAMS-simulation with different preloads. The first four Eigenfrequencies
are shown
PL=3.5[kN]
PL=16[kN] …
Fig. 3.11 Amplitude over the frequency (FFT(y-deflection)) of the ADAMS-simulation with different preloads. The first Eigenfrequency and the
mean value is shown
3.5.2 Findings of the Joint Parameters

Using the Simplex Algorithm (section “Simplex Algorithm”)
From the optimization results (Simplex algorithm), diagrams are created for each parameter. These are shown in the following
figures. The respective parameters are plotted according to the selected bolt preload. This procedure was performed with
different numbers of active interface-contact nodes, which is equivalent to different expansions of the joint contact. The
different curves show that the contact surface has an effect on the contact parameters. The beam has a width of 30[mm].
Hence a contact length of 20[mm] results in a contact area of 600[mm2 ], a length of 24[mm] results in a contact area of
720[mm2 ] and a length of 28[mm] results in a contact area of 840[mm2 ].
The graph of the tangential stiffness (Fig. 3.12) show in qualitative terms a physical behavior as it has already been
determined in measurements of other studies [4, 5]. The tangential stiffness increases only slightly with increasing preload.
An almost linear characteristic can clearly be noticed. In addition, the influence of the selected contact surface can be seen.
An increase of the contact surface causes a decrease of the stiffness and vice versa. However, the linearity and its gradient
remains the same.
Figure 3.13 shows the optimized normal stiffness over the preload. This simulation is also done for the areas as mentioned
before. With the data points of the 24[mm]-long contact area a regression line is computed and plotted additionally into the
diagram. The smallest quadratic error of the regression curve was obtained with a logarithmic trial function.
In [17] the property of the dynamic friction by bolted structural components is investigated. The results of the optimization
of the tangential damping (Fig. 3.14) shows a qualitatively good agreement to the results for higher preloads in the above
Tangential Stiffness [N/mm3]

12,0
10,0
8,0
6,0
4,0
2,0
0,0
0 2000 4000 6000 8000 10000 12000 14000 16000 18000
Preload [N]
28[mm] 24[mm] 20[mm] Linear (24[mm])
Fig. 3.12 Tangential stiffness over used preloads with three different lengths of the contact normalized to the acting contact area
Normal Stiffness [N/mm3]

3,00
2,50
2,00
1,50
1,00
0,50
0,00
0 2000 4000 6000 8000 10000 12000 14000 16000 18000
Preload [N]
28[mm] 24[mm] 20[mm] Log. (24[mm])
Fig. 3.13 Normal stiffness over used preloads with three different lengths of the contact normalized to the acting contact area
Tangential Damping [Ns/mm3]

1,00E-02
7,50E-03
5,00E-03
2,50E-03
0,00E+00
0 2000 4000 6000 8000 10000 12000 14000 16000 18000
Preload [N]
28[mm] 24[mm] 20[mm] Power (24[mm])

Fig. 3.14 Tangential damping over used preloads with three different lengths of the contact normalized to the acting contact area
Fig. 3.15 FFT of the beam tip deflections (lin. model and reference solution) over the frequency range of optimization
literature source. With the linear model, macroscopic displacements, which occurs at lower preloads, cannot be reproduced.
The tangential damping would have to increase significantly with lower preloads and decreases at zero preload towards zero
as well [17].
A disadvantage of the viscous damping concept is that, in the strict sense, the substitution model is only valid for one
frequency. If the fundamental oscillation has a significantly different frequency, it must be converted. As mentioned above,
a lower bolt preload results in a larger relative sliding movement between the two sheets of metal. This behavior cannot be
reproduced with the linear model. Therefore, the replacement model was only determined from a preload of 3500[N] up.
The smallest quadratic error of the regression curve was obtained with a power trial function.
Using the Distributed Optimization by Discretization of the Parameter Space (section “Distributed Optimization
by Discretization of the Parameter Space”)
In order to use the distributed resources even better, a parameter optimization with the method from section “Distributed
Optimization by Discretization of the Parameter Space” was carried out. The preload for each bolt was 10,000[N]. For
these optimization, the discretization of the individual parameters resulted in 15,600 different combinations. Theoretically,
the computation can be performed on as many computers as the number of combinations. However, due to the high
communication density of the server this is not stable and efficient. At the time of publication the environment architecture
was not yet so developed that so many computations could be processed. Therefore only a theoretical but realistic (smaller
simulation numbers were parallelized) time of about 2 [h] with 27 simulation nodes on 9 computers can be given. After
determining the cost function minimum, no smaller discretization around the minimum was performed. In contrast, the
sequential computation on one simulation node takes in approximately 31 h. The distributed optimization procedure on one
simulation node resulted in the following joint parameters:

N N Ns
cN = 0.8974 , cT = 4.6265 and dT = 1.2821E − 04 .
mm3 mm3 mm3
Figure 3.15 shows the FFT of the tip deflections (lin. model and reference solution) in the frequency range of optimization.
It can be seen that the amplitude of the first mode can already be mapped well.
Evaluation of the Used Optimization Methods
A big advantage of the Simplex method is that the parameters can be calculated very precisely without any intervention.
The cost function is scanned in a smaller range. However, it is not possible to determine whether the result is a global or
a local minimum. The distribution and thus the parallelizability is very limited by the sequential sequence. In this case, the
discretization of the contact surface could be parallelized.
Table 3.1 Overview of the determined parameters with the two different optimization methods

N N Ns
cN mm3
cT mm3
dT mm3
Method n[−]
Simplex alg. (section “Simplex Algorithm”) 0.8090 4.6476 3.4340E-04 6
Min. cost function (section “Distributed Optimization by Discretization of the Parameter Space”) 0.8974 4.6265 1.2821E-04 6
Table 3.2 Overview of the preload configurations for the verification of the linear substitution model. For the preload position see Fig. 3.2
Configuration Preload 1 [N] Preload 2 [N] Preload 3 [N]
1 10,000 5000 16,000
2 5000 16,000 10,000
3 16,000 10,000 5000
The method from section “Distributed Optimization by Discretization of the Parameter Space” can be parallelized very
well. In addition, the entire meaningful parameter space can be examined for a global minimum. Afterwards the accuracy
can be improved with further discretizations.
A comparison of the optimized parameters is shown in Table 3.1. The results of the damping do not look very similar at
first sight, but they agree best due to the discretization. Here a further optimization process with a finer discretization of the
parameters would yield a closer result.
3.5.3 Verification of the Substitution Model
In order to check the usability of the linear substitution model, simulations with three different preload configurations with
ADAMS are done. The resulting y-deflection of the reference model and the simulation result of the linear model are plotted.
For this computation of the linear model the optimized parameters of Sect. 3.5.2 (section “Simplex Algorithm”) are used.
The preload configurations are listed in Table 3.2. For the preload positions see Fig. 3.2.
The figures below show the y-deflection of the linear model and the reference model in the time domain. Additionally, the
FFT was carried out and plotted in the range of the first mode (range of optimization). The fundamental oscillation can be
reproduced well. This can be seen in the logarithmic diagrams where the FFT in the range of the first mode is plotted. From
the results of the performed verifications it can be concluded that this linear model with the used optimized parameters can
be used for a useful approximation of the global behavior of the nonlinear model. Thus, the presented method is suitable for
first estimations of the global linear dynamic behavior.
Preload Configuration 1 (Figs. 3.16 and 3.17):
Fig. 3.16 y-displacement of the beam tip (lin. model and reference solution) over the whole simulation time with preload configuration 1
Fig. 3.17 FFT of the beam tip deflections (lin. model and reference solution) over the frequency range of optimization (configuration 1)
3.6 Conclusion and Outlook
In this paper the determination and transferability of a linear substitution model for bolted joints is investigated. This
linear model has four parameters (normal and tangential stiffness, tangential damping and contact area) to map the physical
behavior. The parameters are identified as a function of the bolt preload force. It is only possible to approximate the nonlinear
model with a linear substitution model as long as the global structural behavior is kind of linear. The model parameters are
determined with a parameter identification in frequency domain (direct search method and distributed optimization). The
two tested optimization strategies were evaluated with regard to parallelizability and accuracy. The mechanical model are
two sheet metal strips bolted together and clamped on one side. The two sheets are connected with three bolts. After a
nonlinear dynamic reference simulation with the MBS software ADAMS and the connected JointAnalyser the linear model
is parameterized by means of the tip deflection. Therefore, the tip deflection over time was transformed with the FFT and the
amplitude of the first mode was fitted. Due to the large number of necessary calculations, these are distributed over several
simulation nodes and a significant amount of time might be saved. The result is a linear replacement model of a bolted
connection. This model can be adapted to the a required bolt preload and used for other structures. With the optimized joint
model a verification was carried out on the basis of three examples and resulted in a good agreement of the linear model to
the reference oscillation. From the results of the performed verifications it can be concluded that this linear model with the
used optimized parameters can be applied for the approximation of the global behavior of the nonlinear model. Thus, the
presented method is suitable for first estimations of the global linear dynamic behavior. In future it would also be possible to
process the calculated variations with the software toolbox HeuristicLab [18] in order to obtain the optimal joint parameters.
This toolbox can be used, for example, to find a mathematical model for the resulting data. With this a minimum of the cost
function and thus optimal parameters are determined afterwards. For example, the impact of each parameter is immediately
delivered.
Acknowledgments This research is part of the project Distributed Modelling and Simulation, DisMoSim, which is financed by the Forschungs
und Förderungs Gesellschaft, FFG
References
1. Gaul, L., Nitsche, R.: The role of friction in mechanical joints. Appl. Mech. Rev. 54(2), 93–106 (2001)
2. Reuss, P., Zeumer, B., Herrmann, J., Gaul, L.: Consideration of interface damping in dynamic substructuring. Conf. Proc. Soc. Exp. Mech.
Ser. 2, 81–88 (2012)
3. Pichler, F., Witteveen, W., Fischer, P.: Reduced-order modeling of preloaded bolted structures in multibody systems by the use of trial vector
derivatives. J. Comput. Nonlinear Dyn. 12(5), 051032-1–051032-12 (2017)
4. Lenz, J.: Strukturdynamik unter dem einfluß von mikro- und makroschlupf in fügestellen. Ph.D. thesis, Inst. A für Mechanik (1997)
5. Schwingshackl, C.W., Nowell, D.: The measurement of tangential contact stiffness for nonlinear dynamic analysis. In: Kerschen, G., Brake,
M.R.W., Renson, L. (eds.) Nonlinear Structures and Systems, vol. 1, pp. 165–167. Springer International Publishing, Basel, CH (2020)
6. Ehrlich, C., Schmidt, A., Gaul, L.: Microslip joint damping prediction using thin-layer elements. In: Allen, M., Mayes, R., Rixen, D. (eds.)
Dynamics of Coupled Structures, vol. 1, pp. 239–244. Springer International Publishing, Cham (2014)
7. Bampton, M.C.C., Craig, J.R.R.: Coupling of substructures for dynamic analyses. AIAA J. 6(7), 1313–1319 (1968)
8. Fletcher, R.: Practical Methods of Optimization, 2nd edn. Wiley, Chichester/New York (2010)
9. Nelder, J.A., Mead, R.: A simplex method for function minimization. Comput. J. 7(4), 308–313 (1965)
10. Nocedal, J., Wright, S.J.: Numerical Optimization, corr. 2. print edn. Springer Series in Operations Research. Springer, New York (2000)
11. Oberpeilsteiner, S., Lauß, T., Steiner, W., Nachbagauer, K.: A frequency domain approach for parameter identification in multibody dynamics.
Multibody Syst. Dyn. 43, 1–17 (2017)
12. Goos, P., Jones, B.: Optimal Design of Experiments: A Case Study Approach. Wiley, Hoboken (2011)
13. Kleijnen, J.P.C.: Design and Analysis of Simulation Experiments. International Series in Operations Research & Management Science, vol.
230, 2nd edn. Springer, Cham (2015)
14. Siebertz, K., van Bebber, D., Hochkirchen, T.: Statistische Versuchsplanung: Design of Experiments (DoE), 2nd edn. VDI-Buch Ser. Vieweg,
Berlin/Heidelberg (2017)
15. Bronshtein, I., Semendyayev, K., Musiol, G., Mühlig, H.: Handbook of Mathematics, 6th edn. Springer, Berlin/Heidelberg (2015)
16. Welc, J., Esquerdo, P.J.R.: Applied Regression Analysis for Business – Tools, Traps and Applications, 1st ed. 2018 edn. Springer,
Berlin/Heidelberg (2017)
17. Bournine, H., Wagg, D.J., Neild, S.A.: Vibration damping in bolted friction beam-columns. J. Sound Vib. 330(8), 1665–1679 (2011)
18. Wagner, S., Kronberger, G., Beham, A., Kommenda, M., Scheibenpflug, A., Pitzer, E., Vonolfen, S., Kofler, M., Winkler, S., Dorfer,
V., Affenzeller, M.: Architecture and design of the heuristicLab optimization environment. In: Advanced Methods and Applications in
Computational Intelligence. Topics in Intelligent Engineering and Informatics, vol. 6, pp. 197–261. Springer, Heidelberg (2014)
Chapter 4
Flexible Multibody Dynamics: Which Terms of the Equations
of Motion Have to Be Considered for Which Problem?
Wolfgang Witteveen and Florian Pichler
Abstract Scientists often face the task to carry out a “proof of concept” of a research idea. If this is related to multibody
simulation of flexible bodies, it is sometimes necessary to implement a simple and problem-oriented multibody simulation
code in programs like Matlab or Scilab. The question then arises, which terms of the equation of motion have to be considered
when the flexible body is modelled along the Floating Frame of Reference Formulation. The consideration of all effects leads
to very complex expressions and also require costly FE preprocessing. A remarkably simplified version can be obtained, by a
strict application of the small deformation assumption to the level of the kinetic energy. But what is the truth in the context of
a certain problem? Clear rules are provided, so that a researcher can decide for a particular mechanical system, which terms
are required and which ones can be neglected. This work reviews a paper that has been recently published by the authors
with respect to this question.
Keywords Flexible multibody dynamics · Flexible body · Floating frame of reference formulation · FFRF · Small
deformations
4.1 Introduction
A very common method for the mathematical description of elastic and small deformations of solid bodies in multibody
dynamics is the Floating Frame of Reference Formulation (FFRF), see [1]. The total movement of a body is subdivided into
a non-linear rigid body motion and a small elastic deformation. These deformations are expressed in the body fixed Floating
Frame and are approximated with trial functions, often also called modes. A full deviation of the equations of motion leads
to complex and CPU time consuming expressions. Experience shows that many terms are rarely necessary. Neglecting these
terms would simplify the programming of such a code and speed up its execution time. In addition, the complex terms contain
expressions that are not standard output of an FE program. If they could be neglected, then the usual output possibilities of
FE programs would be sufficient. Now the question arises when which terms have to be considered. In [2] a detailed answer
was given on that question. In this paper only the conclusions are recapitulated, without going into the details. It is intended
to be a kind of a “user manual” when which terms are recommended. The interested reader is referred to [2] for in-depth
background knowledge and numerical examples.
4.2 Starting Point: The Not and the at Most Simplified Equations of Motion
Although the FFRF can be formulated uniformly, there are some important details which influence the concrete appearance
of the equations of motion. The equations listed below are based on the following assumptions, see [2] Chap. 3 for more
informations.
• Rigid body rotation is parametrized via Euler Parameters
• A so called Buckens frame is used for the Floating Frame of Reference
W. Witteveen () · F. Pichler


32 W. Witteveen and F. Pichler
• The Floating Frame of Reference is fixed to the center of gravity of the undeformed body
• Mass normalized (pseudo) free surface modes are used for the former mentioned trial functions
• The initial orientation of the axis of the Floating Frame of Reference matches the principal axis of inertia of the
undeformed body.
A complete derivation based on the above assumptions leads to the following equations of motion for an undamped
flexible body without external loads:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
mI 0 0 R̈ 0 0 0 R 0
⎣ mΘΘ mΘf ⎦ ⎣ Θ̈ ⎦ + ⎣ 0 0 ⎦ ⎣ Θ⎦ = ⎣ (Qv )Θ ⎦
2
sym I q̈f sym Ω qf (Qv )f
T T
mΘΘ = G I0 + W1 Qf + Qf T W2 Qf G mΘf = G Qf T W3 (4.1)
˙ T I + W Q + Q T W Q ω − GT W Q̇ + 2Q T W Q̇ ω − 2G
(Qv )Θ = −2G ˙ T Q T W q̇
0 1 f f 2 f 1 f f 2 f f 3 f

(Qv )f = [ω1 I ω2 I ω3 I] 1
2 W1
T
+ W2 Qf ω − 2W3 T Q̇f ω
The body’s state is characterized by the vectors R, and qf , holding the translational, rotational and flexible coordinates.
The matrix I represents the identity matrix with the proper dimensions and the matrix 2 holds the eigenvalues of the modal
equations. The arrangement of the flexible coordinates qf in a somehow advantageous form delivers the matrix Qf . The
matrix G transforms the Euler Parameters into the angular velocity vector ω. The body’s inertia properties can be described
by the body’s mass m and the four matrices I0 , W1 , W2 and W3 . Those matrices can be computed via so called “shape
integrals” and contain basically information about the mass distribution of the undeformed body and the mode shapes. The
matrix I0 contains mass information of the undeformed body only and is equal to the inertia tensor of the undeformed
body. The other three matrixes contain mass information of the modes and describe the change in inertia due to the body’s
deformation. The symbol Qv is used for the so called quadratic velocity vector which contains the velocity depended inertia
forces.
In the literature a strict application of the small deformation assumption on the level of the kinetic energy is suggested,
see exemplarily [3]. This means, that in case of a sum of the body’s extension and its deformation, the deformation is always
neglected. This leads to
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
mI3 0 0 R̈ 0 0 0 R 0
⎢ ⎥ ⎢ ˙ TI ω⎥
G I0 G 0 ⎦ ⎣ Θ̈ ⎦ + ⎣ 0 0 ⎦ ⎣Θ ⎦
T
⎣ = ⎣ − 2G 0 ⎦
(4.2)
sym I q̈f sym Ω 2 qf 0
which is a remarkable simple set of equations. It can be seen, that (4.2) can be obtained out of (4.1) when all terms containing
the matrixes W1 , W2 and W3 are neglected.
As already mentioned, experience shows that (4.2) often leads to sufficiently accurate results. However, there are scenarios
where this is not the case.
4.3 Assumptions of Small Deformations and Its Implications
It has been assumed in [2] that the flexible deformation remains small at least against one dimension of the body. This is a
common restriction in the context of the FFRF. Based on this assumption it can be shown, that the norm of the matrix I0
is much higher than the norm of W1 , W2 and W3 . Moreover, the norm of W1 is higher as for the matrices W2 and W3 .
Furthermore it has been observed, that for the absolute value of the entries in W2 and W3 a limit of 2 can be given. Please
see [2] Chap. 6 for details on the former observations.
4 Flexible Multibody Dynamics: Which Terms of the Equations of Motion Have to Be Considered for Which Problem? 33
4.4 Deformation Depended Inertia
The term [I0 + W1 Qf + Qf T W2 Qf ] arises two times in the full set of Eq. (4.1). Once inside the mass matrix and a second
time inside the quadratic velocity vector Qv . In [2] it is shown that the state depended terms are just necessary in two
special (and rare) cases. The first one is a structure with a particular compliance against angular velocity and a special mass
distribution where the mass is concentrated at a certain distance of the rotational axes. The second case is a structure having a
dominant spatial extension in one direction like a beam. If one of those two situations takes place, an additional requirement
needs to be fulfilled, which is that the body is a free body or at least very softly connected to other bodies, so that its state is
mainly determined by its own inertia. Consequently, in most applications the state depended terms can be neglected. If not,
it turned out, that the linear and the quadratic term must be taken into account.
4.5 Inertia Coupling
Equation (4.1) shows a coupling in the mass matrix between the rotational and flexible degrees of freedom. This coupling is
examined more detail in [2]. It turns out that the conditions for the consideration of this coupling are the same as the once
for the state dependent inertia in the previous chapter.
4.6 Quadratic Velocity Vector
The quadratic velocity vector Qv is formed by centrifugal, Coriolis and Euler forces. The latter evaluate to zero when Euler
parameters are used. The gyroscopic forces of the undeformed body should be considered in any case. This is the remaining
term on the right hand side of Eq. (4.2). If rotor dynamic effects are not of interest but the body’s widening due to centrifugal
forces has to be considered, Eq. (4.1) can be simplified to
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎡ ⎤ ⎡ ⎤ 0
mI 0 0 R̈ 0 0 0 R
⎢ ⎥ ⎣ ⎦ ⎢ ˙TI ω ⎥
⎣
T
G I0 G 0 ⎦ Θ̈ + ⎣ 0 0 ⎦ ⎣ Θ⎦ = ⎢
⎣ − 2G
0
⎥
⎦ (4.3)
2 1
sym I q̈f sym Ω qf ω 1 I ω 2 I ω 3 I 2 W1 ω
T
Equation (4.3) implies that no special structure in the sense of the former two sections has to be considered. In the rare
case where this would not be the case, the state depended inertia and the mass coupling have to be considered as well as the
state depended term W2 Qf in (Qv )f .
Rotor dynamics effects are characterized by a strong coupling of the rotational and flexible degrees of freedom. If such
effects have to be considered, the quadratic velocity vector as given in Eq. (4.1) has to be considered. A slight optimization
is possible, when none of the two particular structures in the sense of the former two sections have to be considered. In such
a case, the term 2Qf T W2 Q̇f can be neglected.
4.7 Conclusion and Benefit
In this work the significance of all inertia related terms of a flexible body in the FFRF have been reviewed based on [2]. All
terms have been addressed and guidelines have been given when which term can be neglected or has to be considered.
The well-founded negligence of certain terms in the equations of motion leads to the following benefits: The first benefit
is obviously the reduction of the computational burden. The matrixes W1 , W2 and W3 are involved in matrix–matrix and
matrix–vector products which need to be computed several times at each iteration during the numerical time integration. A
second benefit is that the computation of W1 , W2 and W3 is not necessary at all when they are not needed. Consequently, the
eigenvalues and mode shapes are enough for flexible multibody dynamics. This is standard output of many FE codes, while
W1 , W2 and W3 are not. Finally, the set of guidelines given above removes uncertainty concerning the question of which
invariant needs to be considered and which not.
34 W. Witteveen and F. Pichler
References
1. Shabana, A.: Dynamics of Multibody Systems, 4th edn. Cambridge University Press, New York (2013)
2. Witteveen, W., Pichler, F.: On the relevance of inertia related terms in the equations of motion of a flexible body in the floating frame of reference
formulation. Multibody Syst. Dyn. 46, 77–105 (2019)
3. Geradin, M., Rixen, D.J.: Impulse-based substructuring in a floating frame to simulate high frequency dynamics in flexible multibody dynamics.
Multibody Syst. Dyn. 42, 47–77 (2018)
Chapter 5
Application of the Bouc-Wen Model to Bolted Joint Dynamics
Drithi Shetty, Matthew S. Allen, and Joseph D. Schoneman
Abstract Various numerical models have been developed to capture the dynamic, hysteretic behavior of different
mechanical systems. One such semi-physical model is the Bouc-Wen model, which relates the input displacement to the
output restoring force in a hysteretic way. This formulation is intended for any form of hysteresis and was originally applied
to force – deflection and flux – current diagrams of mechanical and ferromagnetic hysteresis. Built-up structures are also
known to show hysteretic behavior due to the slipping that occurs between interfaces bolted together. This paper tests how
effective the Bouc-Wen model is in capturing the power-law damping behavior observed in bolted joints by comparing it
with another commonly used numerical model – the Iwan model. While the Iwan element has been proven to be robust and
well-suited at capturing the power-law increase in energy dissipation and slow decrease in stiffness with vibration amplitude
exhibited by bolted interfaces, numerical integration of the same is currently computationally expensive. Time integration of
the Bouc-Wen model, on the other hand, is much more efficient, thus warranting the proposed study.
Keywords Non-linear dynamics · Bouc-Wen model · Bolted joints · Hysteresis · Parameter identification
5.1 Introduction
Mechanical fasteners have long been known to be an important source of stiffness and energy dissipation in built-up
structures. The slipping between two surfaces that are bolted together results in frictional dissipation of energy. As a result,
the restoring force versus displacement graph of systems consisting of bolted joints shows non-linear, hysteretic behavior. At
low force amplitudes, the edges of the contact patch slip, which is known as micro-slip. As the force amplitude increases, the
area of contact reduces, ultimately resulting in relative motion between the surfaces, also known as macro-slip. The change
in stiffness and damping that occurs due to bolted interfaces has also been found to be a function of the amplitude of the
response [1]. In order to capture this observed dynamics, lumped hysteretic models are used. One such model prevalent in
bolted joint dynamic analysis is the Iwan model [2], which consists of a parallel arrangement of spring-slider units known
as Jenkins elements. Solving the equation of motion of a system consisting of an Iwan model requires implicit integration
schemes that can be time consuming. Thus, alternative hysteretic models that would reduce the computational effort while
still providing the accuracy obtained with Iwan models must be considered.
A particular semi-physical, hysteretic model that has found applications in multiple areas, especially in civil and
mechanical engineering, is the Bouc-Wen model [3]. It consists of a non-linear, first-order ordinary differential equation
(ODE) that relates the input displacement to the output restoring force in a hysteretic way. This differential equation can be
easily integrated using explicit ODE solvers to obtain the non-linear restoring force. The existing literature predominantly
focuses on applying the Bouc-Wen model to capture the steady-state hysteretic behavior of various non-linear systems. In
D. Shetty ()
Department of Mechanical Engineering, UW-Madison, Madison, WI, USA
M. S. Allen
Department of Engineering Physics, UW-Madison, Madison, WI, USA
J. D. Schoneman
ATA Engineering, Inc., San Diego, CA, USA

36 D. Shetty et al.
bolted joints, however, the non-linearity is also characterized by power-law dissipation behavior – a behavior in which the
log of the energy dissipated (or effective modal damping ratio) increases linearly with the log of the vibration amplitude, and
a decrease in natural frequency of the system with increase in response amplitude [1]. This work analyses the effectiveness of
the Bouc-Wen model in capturing this power-law behavior by comparing it against Segalman’s four-parameter Iwan model
[4], which was derived specifically to simulate bolted joint dynamics.
The next section describes the mathematical representation of the Bouc-Wen model and the system identification method
adapted from [5] to estimate the Bouc-Wen parameters. The conference presentation will elaborate on the case study
presented here, in which the response of the Bouc-Wen model is compared against that of the Iwan model when an impulsive
force is applied on a mode of a built-up structure that exhibits non-linearity.
5.2 Theory
The equation of motion for a single degree-of-freedom (SDOF) system consisting of a hysteretic non-linear element can be
written as
mẍ + cẋ + kf x + fN L (x, z) = f (t) (5.1)
where m, c, and kf are the mass, linear damping, and linear stiffness of the system respectively. The non-linear restoring
force fN L (x, z) is a function of the displacement x and a hysteretic state variable z. The Bouc-Wen formulation for the
non-linear restoring force is given as
fN L = (1 − α)ki z (5.2)
ż = Aẋ − β|ẋ||z|n−1 z − γ ẋ|z|n (5.3)
k
where α = kfi is defined as the ratio of post-yield to pre-yield stiffness, which, for a bolted joint, corresponds to stiffness
at macro-slip (i.e. the linear stiffness of the system without the joint) to the stiffness when the joint is stuck. A, β and γ
are the Bouc-Wen parameters to be identified. An adaptation of the parametric identification method presented in [5] has
been used to identify the Bouc-Wen parameters for a given load-displacement curve. The system to be identified could be an
SDOF system or a non-linear mode of a structure. Assuming the modes to be uncoupled and the non-linearity to be weak, the
non-linear mode can be written as an SDOF system consisting of a non-linear element that is independent of other modes.
The system under consideration is excited by a harmonic force having fixed amplitude and frequency, and the steady-state
force-displacement curve is obtained. Because the Iwan joint is not velocity dependent, a very low forcing frequency can
be chosen so that the excitation is nearly quasi-static and the steady-state hysteresis curve is obtained within the first few
periods of integration. The same approach is used with the Bouc-Wen model. The amplitude of the force is chosen to be low
enough for the system to be in micro-slip.
Firstly, α and ki can be obtained from the low amplitude linear frequency, ω0 , and the high amplitude slip frequency, ω∞
using the following relations:
2
ω∞
α= ; ki = mω02 (5.4)
ω0
The parameters A, β and γ are then calculated using the method of least squares, given by Eq. 5.5,
⎡ ⎤
A
Fnl
= x −x|zi |n−1 zi −x|zi |n ⎣ β ⎦ (5.5)
(1 − α)ki
γ
where x is a vector of the difference between two consecutive displacements, fNL is the difference between consecutive
force value and ωi is the vector of hysteretic displacement calculated as wi = fN L /Kw for each time instant. Equation 5.5
is of the form Y = which can be solved for by taking the inverse of . To ensure that the matrix is well-
conditioned, each column of is normalized to its corresponding maximum value. The least-squares problem is solved
5 Application of the Bouc-Wen Model to Bolted Joint Dynamics 37
-4
10
0.6 213.95 0.11 2.5 50
Iwan
Bouc-Wen Iwan
0.105 45
Bouc-Wen
Non-linear restoring force (N)
% error
% error in natural frequency

0.4 213.9 % error
40
0.1
35
% error in damping
frequency Hz
0.2 213.85 0.095 2
30
damping
0.09
0 213.8 25
0.085
20
-0.2 213.75 0.08 1.5
15
0.075
10
-0.4 213.7
Iwan 0.07
Bouc-Wen
5
-0.6 213.65 0.065 1 0

-5 -4 -3 -2 -1 0 1 2 3 4 5 -5 -4 -3 -5 -4 -3
10 10 10 10 10 10
displacement (m) 10-6 Velocity amplitude (m/s) Velocity amplitude (m/s)
(a) Steady-state behavior, n = 2 (b) Frequency vs amplitude, n = 2 (c) Damping vs amplitude, n = 2

10-4
0.6 213.93 0.01 2.5 10
Iwan Iwan
Bouc-Wen 213.92 0.009 Bouc-Wen 9
Non-linear restoring force (N)
% error in natural frequency

0.4 % error
0.008 8
213.91
0.007 7
frequency Hz
% error in damping
0.2 213.9 2
0.006 6
damping
213.89
0 0.005 5
213.88
0.004 4
-0.2 213.87 1.5
0.003 3
213.86
Iwan 0.002 2
-0.4
Bouc-Wen
213.85 0.001 1
% error
-0.6 213.84 0 1 0
-5 -4 -3 -2 -1 0 1 2 3 4 5 10-5 10-4 10-3 10-5 10-4 10-3
displacement (m) 10-6 Velocity amplitude (m/s) Velocity amplitude (m/s)
(d) Steady-state behavior, n = 1.1 (e) Frequency vs amplitude, n = 1.1 (f) Damping vs amplitude, n = 1.1
Fig. 5.1 Comparing the quasi-static behavior and dynamic behavior obtained using two different values of the Bouc-Wen parameter n,
demonstrating the sensitivity of the ring-down response to n. (a) Steady-state behavior, n = 2. (b) Frequency vs amplitude, n = 2. (c) Damping
vs amplitude, n = 2. (d) Steady-state behavior, n = 1.1. (e) Frequency vs amplitude, n = 1.1. (f) Damping vs amplitude, n = 1.1
iteratively for different values of n. In most of the existing literature, an arbitrary value of n = 2 has been found to be good
enough. However, while the steady-state behavior of a system consisting of a Bouc-Wen element may not be sensitive to the
parameter n, the ring-down response is, as will be shown in the following numerical example.
5.3 Numerical Case Study: Sumali Beam
The identification method described in Sect. 5.2 was applied to an assembly of two beams bolted together, commonly referred
to as the Sumali beam. This structure consists of two thin, identical, stainless steel beams of length 508 mm, width 50.8 mm,
and thickness 6.35 mm, that overlap and are joined with four bolts. The first three elastic bending modes of this structure
are non-linear. The parameters of a modal Iwan model, which approximates each of these non-linear modes, were obtained
in [6], and have been used in this work as a benchmark against which the fitted Bouc-Wen model can be compared. The
second bending mode of the Sumali beam was considered. The response of a modal Iwan model for this non-linear mode
was numerically simulated for a harmonic force input using the Newmark-β integration method. The steady-state hysteresis
loop obtained was then used to calculate the Bouc-Wen parameters A, β and γ by iteratively solving the least-squares
problem, given by Eq. 5.5. The system was then excited by an impulsive fore and the time response obtained by integration
was post-processed using the Hilbert filter in order to obtain the amplitude-dependent damping and natural frequency.
Figure 5.1 shows the effect n has on the accuracy of the Bouc-Wen model. The hysteresis loops obtained using the Bouc-
Wen model show some deviation from the Iwan model, with the percentage error in the estimation of energy dissipated over
a cycle (obtained by calculating the area under the curve) being 3.46% when n = 2 and 7.86% when n = 1.1. However,
significant difference in accuracy can be observed in the prediction of the ring-down response. When n = 2, the Bouc-Wen
model gives a maximum error of 0.11% in the natural frequency estimate and a maximum error of 50% in the damping
estimate. When n is changed to 1.1 and the other Bouc-Wen parameters are re-calculated using the least-squares method, the
errors in both frequency and damping drop to a maximum of 0.01% and 9.3% respectively. There is a marginal improvement
38 D. Shetty et al.
in the agreement in natural frequency but a significant one in the prediction of damping. Thus, in order for the Bouc-Wen
model to be effectively used to capture the power-law dissipation observed in bolted joints, it appears the parameter n must
also be optimized.
Acknowledgments This material is based in part upon work supported by the National Science Foundation. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National
Science Foundation.
References
1. Segalman, D.J.: Modelling joint friction in structural dynamics. Struct. Control Health Monit. 13(1), 430–453 (2006)
2. Iwan, W.D.: A distributed-element model for hysteresis and its steady-state dynamic response. J. Appl. Mech. 33(4), 893–900 (1966)
3. Bouc, R.: A mathematical model for hysteresis. Acta Acustica united with Acustica, 24(1), 16–25 (1971)
4. Segalman, D.J.: A four-parameter Iwan model for lap-type joints. J. Appl. Mech. 72(5), 752–760 (2005)
5. Zhu, X., Lu, X.: Parametric identification of Bouc-Wen model and its application in mild steel damper modeling. Proc. Eng. 14, 318–324
(2011)
6. Lacayo, R.M., Deaner, B.J., Allen, M.S.: A numerical study on the limitations of modal Iwan models for impulsive excitations. J. Sound Vib.
390, 118–140 (2017)
Chapter 6
Some Aspects of Using the Random Decrement Technique
for Nonlinear Systems
Karsten K. Vesterholm, Rune Brincker, and Anders Brandt
Abstract The random decrement (RD) technique can be used to analyze the response signal from a system that has
amplitude dependent modal parameters. The implementation of RD for this purpose is usually done by simply applying
the technique at multiple amplitudes in the measured response signal. Modal parameters are then estimated based on RD
signatures using well known time domain modal parameter estimation methods. This analysis procedure originates from the
invention of the RD technique, and is described by several studies in the literature. However, the RD technique is developed
for linear systems, and caution must be exercised when applying it to nonlinear systems. In this study, several aspects of
applying of the RD technique on signals exhibiting nonlinear behavior are addressed. The principle of superposition does not
apply for a nonlinear system. This means the averaging process in RD can yield corrupted results. The benefit of a sufficiently
high sampling rate is described.
Keywords Random decrement · Nonlinear system · Random vibrations · Signal processing · System identification
6.1 Introduction
Modal parameter estimation performed on a structure during operation is called operational modal analysis (OMA). Here,
only the output is measured, the excitation force is unknown. OMA is based on an assumption of linearity, and can be
performed in the time domain or in the frequency domain, as described by Brincker and Ventura [1]. The methodology
in OMA is to measure the vibration of a structure, acceleration, velocity or displacement, then calculate the correlation
function matrix of the signals. From the correlation function matrix, modal parameters are estimated using well established
algorithms.
A method of estimating the correlation function is the random decrement (RD) technique. RD was invented in the late
1960’s and early 1970s by Henry A. Cole [2] as a method of obtaining a correlation function estimate of a signal, while
measuring the signal. This estimate became known as a RD signature. It was meant to be used to detect sudden changes in
the signal that would be difficult to observe directly from the signal. The mathematical foundation was defined by Vandiver
et al. [3], where the auto RD signature was related to the correlation function. Ibrahim [4] presented a method to perform the
RD calculation on multiple signals, yielding the cross RD signature. A further development came from Brincker et al. [5]
who generalized Vandiver et al.s result to the cross RD signature. A derivation of this result is presented by Asmussen [6].
The essential difference between the RD technique and a conventional correlation function estimate is that a triggering
condition is specified when using the RD technique. This means RD can be applied with the intent to analyze how the
system responds to a specific initial condition, consisting of an amplitude and a slope. For a linear system, the frequency and
rate of decay are independent of the initial conditions. This is not the case for a nonlinear system, because the principle of
superposition (PoS) does not apply.
K. K. Vesterholm () · A. Brandt

Department of Technology and Innovation, University of Southern Denmark, Odense, Denmark
e-mail: [email protected]; [email protected]
R. Brincker
Department of Civil Engineering, Technical University of Denmark, Lyngby, Denmark

40 K. K. Vesterholm et al.
During previous work of the authors investigating how the RD technique can be used to estimate amplitude dependent
modal parameters of nonlinear systems, aspects of the application of RD related to the PoS emerged. The present study
investigates these aspects in order to formulate good practices when applying RD to nonlinear systems.
6.2 Theory
The relationship between the RD signature and the correlation function and derivative of the correlation function is defined
as [6]
a2 b2
Rx,y (τ ) Ṙx,y (τ ) a1 ypy (y)dy b ẏpẏ (ẏ)d ẏ
Dx,y (τ ) = 2
· ã − 2
· b̃, ã = a2 , b̃ = 1b (6.1)
σy σẏ a1 py (y)dy
2
b pẏ (ẏ)d ẏ
1
The estimate of the RD signature is a conditional mean defined as [6]
1
N
D̂x,y (τ ) = x(ti + τ )|Ty(ti ) , Ty(ti ) = {a1 < y(t) < a2 , b1 < ẏ(t) < b2 } (6.2)
N
i=1
where x and y are vibration response signals, Ty(ti ) is the triggering condition (TC) where a1 and a2 determine the initial
amplitude, b1 and b2 define the initial slope, and N is the number of time samples in the response signal the satisfy the TC.
The time samples that satisfy the TC are called triggering points (TP). When implementing the RD technique, ã and b̃ are
the average value of the TPs in the signal and derivate signal respectively
1 1
N N
â = y(ti ), b̂ = ẏ(ti ) (6.3)
N N
i=1 i=1
When implementing a TC, it is important that these parameters are unbiased.
6.3 Analysis
When applying the RD technique to investigate nonlinear systems, a TC should be chosen that accommodates the fact that
PoS does not apply. This means, a triggering condition should not include a large range of amplitudes or slopes. An example
of such a TC, is the original and commonly used TC known as level crossing (LC) which is defined as
LC
Ty(t i)
= {a < y(t) < a + a, −∞ < ẏ(t) < −∞} (6.4)
where the amplitude is theoretically well defined, but the slope of TPs can have any value. An example where this causes
very large errors in modal parameter estimates of a nonlinear system can be found in a study by Vesterholm et al. [7]. Instead,
a TC with a specific amplitude and a specific slope should be chosen, such as local extremum (LE) [8]
LE
Ty(t i)
= {a1 < y(t) < a2 , ẏ(t) = 0} (6.5)
where a1 and a2 should define a narrow band in the response signal to ensure a well defined amplitude.
When using the RD technique to estimate amplitude dependent modal parameters, it is important for the implementation
of RD to be unbiased. The implementation of LC is done according to the definition by Brincker et al. [5]. In the present
study it is uncovered that this implementation of LC has a previously unknown amplitude bias, and a large standard deviation
of the signal value at the TP. This means implementing LC will yield a RD signature with a wrong initial amplitude, and
have TPs with a large range of amplitudes. The bias and large deviation are irrelevant for a linear system. However, they
are essential when investigating for amplitude dependent modal parameters, because a low deviation and accurate amplitude
are needed to accommodate the fact that PoS no longer applies. The amplitude bias and standard deviation of LC can be
eliminated by having a sufficiently high sampling rate as illustrated in Fig. 6.1. Here the random displacement response of a
single-degree-of-freedom (SDOF) system is simulated using a sampling rate 100 times the natural frequency. The signal is
6 Some Aspects of Using the Random Decrement Technique for Nonlinear Systems 41
Estimate of a Estimate of b Standard deviation of b estimate

0.5 0.3
20
True value
Level Crossing 10
0.4
0.2 Local Extremum 7
5
[m/s]
0.3 3
b [m/s]
a [m]
0.1
b
0.2 1
True value 0 0.5

0.1 Level Crossing Level Crossing
Local Extremum 0.2 Local Extremum
0 -0.1
4.5 12.1 30.3 100.0 227.4 4.5 12.1 30.3 100.0 227.4 4.5 12.1 30.3 100.0 227.4
Oversampling ratio fs/fr [-] Oversampling ratio fs/fr [-] Oversampling ratio fs/fr [-]
Fig. 6.1 Estimate of parameters ã and b̃ calculated using Equation (6.3) for an SDOF system at various sampling frequencies expressed as the
oversampling ratio. In the left pane, the a estimate is illustrated with an error bar illustrating the standard deviation. The middle pane shows the
estimate of b where the standard deviation of b is illustrated in the right pane
then decimated to 3 times the natural frequency, where, according to the sampling theorem, all information is still present.
The signal is then upsampled, and parameters ã and b̃ are estimated, for a number of oversampling ratios.
Applying LE, is with the purpose of identifying TPs with ẏ = 0. A measure of how well this is realized when implemented
can be investigated by estimating b̃ using Equation (6.3). Here, the estimate of b̃ is seen to have a significant standard
deviation, unless the signal is highly oversampled. A large standard deviation in this regard means that a large range of
slopes are identified as TPs, which is undesirable. LE is also shown to be very accurate in estimating ã.
6.4 Conclusion
The random decrement (RD) technique is a relevant tool for estimating modal parameters of nonlinear systems, because it can
be applied to investigate specific initial conditions of the system from a random response signal. Implementing a triggering
condition (TC) with this purpose, it is important to consider the principal of superposition, as this no longer applies. TCs that
defines specific initial conditions, amplitude and slope, should be used for nonlinear systems. When implementing TCs that
uses a well defined amplitude and slope, like local extremum, upsampling the response signal has the benefit of reducing the
standard deviation of the slopes of the triggering points. Upsampling the response signal also has the benefit of reducing a
discovered amplitude bias in level crossing, and the standard deviation of the signal amplitude at the triggering points.
Acknowledgments The authors acknowledge the funding received from Centre for Oil and Gas – DTU/Danish Hydrocarbon Research and
Technology Centre (DHRTC).
References
1. Brincker, R., Ventura, C.: Introduction to Operational Modal Analysis. Wiley, United Kingdom (2015)
2. Cole, H.A.: On-the-line analysis of random vibrations. In: AIAA/ASME 9th Structural Dynamics Materials Conference, Palm Springs (1968)
3. Vandiver, J.K., Dunwoody, A.B., Campbell, R.B., Cook, M.F.: A mathematical basis for the random decrement vibration signature analysis
technique. J. Mech. Des. 104(2), 307–313 (1982)
4. Ibrahim, S.R.: Random decrement technique for modal identification of structures. J. Spacecr. Rockets 14(11), 696–700 (1977)
5. Brincker, R., Krenk, S., Kirkegaard, P.H., Rytter, A.: Identication of dynamical properties from correlation function estimates, Bygningsstatiske
Meddelelser 63(1), 1–38 (1992)
6. Asmussen, J.C.: Modal analysis based on the random decrement technique: application to civil engineering structures. Ph.D. thesis, University
of Aalborg (1997)
7. Vesterholm, K.K., Brincker, R., Brandt, A.: Linearization of modal parameters in Duffing oscillator using the random decrement technique. In:
Proceedings of the International Conference on Noise and Vibration Engineering (ISMA) (2018)
8. Tamura, Y., Suganuma, S.-Y.: Evaluation of amplitude-dependent damping and natural frequency of buildings during strong winds. J. Wind
Eng. Ind. Aerodyn. 59(2), 115–130 (1996)
Chapter 7
Investigation of String Motions of Bowed String Instruments:
A Finite Element Approach
Özge Akar and Kai Willner
Abstract The bowed string motion of violin strings is examined using finite element analysis. The string is modeled as a bar
with a finite number of elements, whereas the bow is one single node. The cause of the sound is the stick-slip effect, which
occurs at steady bowing of a string. The contact area between the bow node and one string node is modeled as nodal contact.
This work aims at investigating the Schelleng Diagram as well as the Guettler diagram. The Schelleng diagram shows bow
force over bowing position and identifies the area of normal sound, i.e. the Helmholtz motion (von Helmholtz, Von den
Tonempfindungen als physiologische Grundlage für die Theorie der Musik. Vieweg, Braunschweig, 1863), overtone sound
and raucous sound, the latter depicts bow force over bow acceleration in order to find a proper parameter configuration for
“perfect bowing starts”. The influence of different friction curves of the rosin are modeled and simulated as well, which has
been tested and investigated extensively in the literature mainly experimentally, see e.g (Smith, Woodhouse, J Mech Phys
Solids 48(8):1633–1681, 2000).
An overall research of the finite-element modeling of bowed strings is started with this new simulation and modeling
set-up. The numerical results will be compared to experimental results in future work.
Keywords Helmholtz motion · Schelleng diagram · Guettler diagram · Finite element modeling · Stick-slip effect
7.1 Introduction
Bowed string instruments like the violin or the cello incorporate many physical phenomena in one single structure. One
occurring effect is the tribology of the rosin, which is the lubricant that the player puts on the bow in order to induce friction
and thereby sound. Furthermore, there is the contact between the bow and the string, the geometrical nonlinearity of the
string and the overall nonlinear dynamic behavior of the vibrating string. Hence, these instruments are versatile multiphysics
examples and thereby offer a vast field of exploration of physical effects that engineers and scientists can apply on many
other areas, e.g. industrial applications. However, bowed string instruments have been studied extensively since the nineteenth
century, starting with [3, 5]. The main goal was to find out how one can describe the bowed string motion. Since then, the
objectives of doing bowed string research have rather been finding out differences between good and excellent instruments
(in terms of sound radiation and playability) and creating virtual bowed string sound. Successful approaches are e.g. the
digital waveguide syn-thesis [3] or modal synthesis [4].
The idea of this work is exploring the aforementioned physical effects of bowed strings. For this purpose, the
“conventional” approaches to model the string vibration like digital waveguide synthesis [6] or modal synthesis [7] are not
suitable, due to the lack of physicality. The authors’ choice is a finite element (FE) approach, as it allows the consideration
of all relevant physical effects.
This paper starts with a short overview of the underlying theory of bowed string. Later, it introduces briefly two common
diagrams, the Schelleng diagram [1] that shows the dependency of bowing position and bow force, and the Guettler diagram
[2] depicting bowing force over bow acceleration in order to find perfect transients before steady-state string motion. The
simulation results are presented afterwards. A conclusion in the end adds up the insights of the results.
Ö. Akar () · K. Willner

Universität Erlangen-Nürnberg, Erlangen, Germany

44 Ö. Akar and K. Willner
7.2 Background
The stick-slip-effect This phenomenon is the cause for the string to radiate sound. The bow triggers a kink, which wanders
along the string from the contact point between bow and string to the end of the string (the nut) back to the upper string
end, the so-called bridge. During the sticking phase, the string moves constantly with the bow into one direction – the bow
deflects it and as soon as the string restoring force is large enough, it “breaks free” from the bow and snaps back. The kink
is between the contact point and the bridge. When the string force is small enough again, the bow grips the string again and
the process starts from the beginning. Figure 7.1 visualizes this effect. The drawing on the left is based on [1, 3], the image
on the right on [8]. For details refer also to [9].
The Eqs. (7.1) and (7.2) describe the radial return mapping [10], which is used to model the stick-slip effect:

FR = FRtrial
if FR ≤ μS · Fbow (7.1)
trial
KT = p

FR = μD · Fbow · sgn FRtrial
else (7.2)
KT = 0
Fbow is the applied bow force, while FR is the current string force. The test force FRtrial = FRold + p· (ustr − ubow ), with
FRold being the friction force in the previous step, is updated in each iteration step. KT is the tangential stiffness, p is the
penalty parameter. ustr and ubow are the string and the bow displacement, respectively. μD and μS are the dynamic and the
static friction coefficients.
The Schelleng diagram This diagram, shown in Fig. 7.2, is an empirically determined approach to find the “right” bowing
position with respect to the bowing force and vice versa. Both axes are in logarithmic scale. The x-axis shows the range in
which the player can move the bow. This area can be located between the bridge and the upper ending of the fingerboard.
The player’s intention is mainly to be in the green area, where the so-called Helmholtz-motion takes place. This motion
indicates normal string motion and hence normal sound. Bowing in the blue area leads to double or multislip motion and
thereby to “surface sound”, since higher harmonics are excited better than the fundamental string harmonic. In the red area,
the bow force is too high, so the resulting sound is rough and no periodic motion can occur. For details refer to [1, 9].
The Guettler diagram While the Schelleng diagram considers the steady-state period of the moving string, the Guettler
diagram looks at the transients. It plots the applied bow force over the bow acceleration and evaluates the duration, until
Helmholtz motion starts. If a transient lasts up to 0.05 s, it was a “perfect” bowing start. There is also a gray area in which
the start is considered mediocre. A bad transient is everything outside this region. Furthermore, for each bowing position,
there is a new plot, as the transients vary when the bowing position is changed. The Guettler diagram is explained in further
detail in [2].
Fig. 7.1 The stick-slip effect of a bowed string. The graph on the left shows the characteristic saw-tooth of the displacement over time in the upper
plot, the plot in the bottom the corresponding string velocity. The image on the right displays the kink moving along the string
7 Investigation of String Motions of Bowed String Instruments: A Finite Element Approach 45
Fig. 7.2 The Schelleng diagram based on [1]. The corresponding string motions are depicted around the main graph
Table 7.1 String and bow parameters and simulation setup values for the Schelleng diagram simulation
Parameters Value Unit Parameters Value Unit
String length 328 mm Damping coefficient 5 * 10−6 –
Number of elements 30 – Dynamic friction coefficient 0.3 –
String radius 0.41 mm Static friction coefficient 0.889 –
String cross section 0.528 mm2 Bow velocity 100 mm/s
String density 6.27 * 10−9 t/mm3 Simulation time 1 s
String pre-tension 105 N Time step 2.27 * 10−6 s
Penalty parameter 250 – Number of samples 4.41 * 105 –
Friction modeling To model the friction, usually Coulomb’s law (see Eq. (7.3)) is used. Many researchers in the context of
bowed strings attached a great value to the tribology of the rosin, so they conducted many experiments and found equations
to fit the measured friction curves, see e.g. [4] (Eq. (7.4)). In these models, the dynamic friction coefficient is velocity-
dependent, vstr being the string velocity and vbow the bow velocity. FN is the normal force.
Coulomb’s friction model
FR
μ= (7.3)
FN
Smith’s/Woodhouse’s model [4]
vstr − vbow vstr − vbow

μ = 0.4· exp + 0.45· exp + 0.35 (7.4)
0.01 0.1
7.3 Simulation
Schelleng diagram simulation The goal of this simulation is to find the “green area” of Fig. 7.2 for a fixed parameter setup
for a violin string G with a fundamental frequency of 196.89 Hz. The parameters in Table 7.1 are used. The string and bow
parameters are real parameters for aforementioned string, except the pre-tension. This value has to be adapted, until the
string frequency matches the real frequency of the string. The reason for that is the lack of accuracy in the stiffness and mass
matrices, as the string model consists of a finite number of one-dimensional bar elements.
The simulation results are depicted in Fig. 7.3. In the lower graph, the bowing position and the bow force vary. The x-axis
shows the bowing position (node 1 is near the bridge, node 6 is near the upper end of the fingerboard), the y-axis plots the
Fig. 7.3 Upper plots show the time integration results for one bowing position and one bow force. The colors correspond to the markers in the
lower graph, the Schelleng diagram. The green graph (a) is Helmholtz motion, blue plot (b) indicates multislip motion, red, (c), is for non-periodic
motion and black graph (d) shows ALF motion. The bow force varies from 0.3 N to 3 N and the bowing position (node 1 corresponds to the bridge,
node 6 is the upper end of the fingerboard). The markers depict the string motion
bow force from 0.3 N to 3.N. Both axes are in logarithmic scale. The different markers indicate the string motion. The blue
crosses symbolize multislip motion, which means that higher harmonics are excited. The red markers indicate rough sound.
The green circles show Helmholtz motion. At very high bow forces with very steady bowing, even subharmonics (anomalous
low frequencies, ALFs) can be excited, which are drawn as black circles. The upper plots show the time-integration results
for a constant bowing position (28 mm) and constant bow force. Here, the bow velocity, the damping and the friction
coefficients were changed in order to provoke clearer differences in the string motions. The detailed parameter setups and
simulation results will be shown in future work.
The Guettler diagram simulation For the simulation of the Guettler diagram, the parameters in Table 7.2 were used. 4
bowing positions were simulated starting from near the bridge at 18 mm going to the near-fingerboard region at 48 mm. The
applied force range is from 0.25 N to 2.25 N for each bowing position. The bow acceleration starts at 1000 mm/s2 and goes
up to 9000 mm/s2 with 9 steps in-between. In total, 324 simulations were run. Each step, i.e. one bow acceleration at one bow
force, takes 0.4 s. Figure 7.4 shows the simulation results, each plot displaying one bowing simulation. The crosses indicate
a very long transient before Helmholtz motion or even no Helmholtz motion at all. The squares show a mediocre transient,
where Helmholtz motion occurs (according the authors’ definition) after 0.1–0.15 s. Perfect transients are symbolized by
a circle. Here, Helmholtz motion starts after 0.05 s (by definition). On some conditions, i.e. at high bow forces near the
fingerboard, ALFs were produced. The triangles indicate this subfrequency motion.
7 Investigation of String Motions of Bowed String Instruments: A Finite Element Approach 47
Table 7.2 String and bow parameters and simulation setup values for the Guettler diagram simulation
Parameters Value Unit Parameters Value Unit
String length 328 mm Damping coefficient 5 * 10−6 –
Number of elements 30 – Dynamic friction coefficient 0.4 –
String radius 0.41 mm Static friction coefficient 0.889 –
String cross section 0.528 mm2 Bow velocity 100 mm/s
String density 6.27 * 10−9 t/mm3 Simulation time 0.4 s
String pre-tension 105 N Time step 1.13 * 10−6 s
Penalty parameter 250 – Number of samples 3.52 * 105 –
Fig. 7.4 Guettler diagram simulation results for different bowing positions starting from the bridge, plot (a), showing results of 2 positions in the
center (b, c), ending at (d), near the fingerboard
In total, the simulation results show that the closer the player bows at the bridge the more difficult it is to provoke
acceptable transients and the easier the transients are excited the more the player approaches the fingerboard. The results
correspond qualitatively to the measured and simulated results in [2].
The influence of friction The different friction models presented in this paper were applied on the string model. Figure
7.5 shows the simulation results for a Helmholtz motion simulation with different friction models, Coulomb’s model with
μD = 0.52 (the average value of Woodhouse’s model) and the 2 presented empirically constructed models ([4]). The
displacement is plotted over the simulation time. It is obvious that the simulation results do not differ a lot from each
other, the transient response differ from each other, the steady-state response varies only slightly for all 3 models. Bigger
differences can be seen at t = 0.75 s, where the string vibration starts fading.
It is the engineer’s task to reflect whether Coulomb’s law is not sufficient, since the other models are computationally
much more expensive. However, the presented bow-string model in this paper is extremely simplified, an increase in model
complexity could lead to a bigger impact of the friction model on the string motion behavior.
Fig. 7.5 Friction models applied on one-dimensional string model
7.4 Conclusion
Using the FE method in order to model the bowed string motion is a novel approach, which seems promising in the context
of studying multiphysics applications. In the beginning, a brief overview of the physics of bowed strings has been given,
explaining the stick-slip effect as well as the 2 main graphs to spot a “good” string vibration, the Schelleng diagram and the
Guettler diagram. The Schelleng diagram simulations in this paper show that by using FE, the authors were able to reproduce
familiar string motion behavior (e.g. Helmholtz and multislip motion). Furthermore, Guettler diagrams have been simulated.
The resulting graphs match the diagrams of the original paper [2] qualitatively. The simulation results of the different friction
models have shown that the resulting string motions differ in the transients from each other, but are similar in the steady-state
response. It has to be considered whether a friction model other than Coulomb’s model is needed, as the computation time
increases extensively using velocity-dependent friction models. However, increasing the bow-string FE model complexity
could also lead to a bigger influence of the friction. Further investigations will give more insights on this idea. In addition,
it has to be mentioned that using other time steps, the simulation results differ notably from each other. Hence, it is crucial
to compare the results with experimental data obtained by measurements of a real bowed string, what will be done in future
work.
References
1. Schelleng, J.C.: The bowed string and the player. J. Acoust. Soc. Am. 53(26), 26–41 (1973)
2. Guettler, K.: On the creation of the Helmholtz motion in bowed strings. Acta Acust. Acust. 88, 970–985 (2002)
3. von Helmholtz, H.: Von den Tonempfindungen als physiologische Grundlage für die Theorie der Musik. Vieweg, Braunschweig (1863)
4. Smith, J.H., Woodhouse, J.: The tribology of rosin. J. Mech. Phys. Solids. 48(8), 1633–1681 (2000)
5. Raman, C.V.: On the mechanical theory of the vibrations of bowed strings and of musical instruments of the violin family, with experimental
verification of the results-Part I. Bull. Indian Assoc. Cultiv. Sci. 15, 1–158 (1918)
6. McIntyre, M.E., Woodhouse, J.: On the fundamentals of bowed-string dynamics. Acta Acust. Acust. 43(2), 93–108 (1979)
7. Demoucron, M.: On the Control of Virtual Violins – Physical Modelling and Control of Bowed String Instruments. Royal Institute of
Technology, Stockholm (2009)
8. Cremer, L.: The Physics of the Violin. The MIT Press, Cambridge (1984)
9. Bennett, W.R.J.: The Science of Musical Sound – Volume 1: Stringed Instruments, Pipe Organs, and the Human Voice. Springer Nature
Switzerland AG, Cham (2018)
10. Giannokopoulos, A.E.: The return mapping method for the integration of friction constitutive relations. Comput. Struct. 32(1), 157–168 (1989)
Chapter 8
FreeDyn: A Free, Flexible and State of the Art Multibody
Simulation Package for Education, Research and Industrial
Applications
Wolfgang Witteveen, Lauss Thomas, and Oberpeilsteiner Stefan
Abstract FreeDyn (www.freedyn.at) is a free multibody simulation (MBS) package consisting of a graphical user interface
(GUI) and a C++ solver in which a modified version of the HHT time integration method is implemented. In order to
underline the solvers efficiency with respect to CPU time, comparative simulations of FreeDyn and a commercial software
product are presented in this paper. The main modelling elements of FreeDyn are rigid and flexible bodies, constraints,
forces and measures (angle, position, velocity . . . ). The FreeDyn standard force elements can be specified by user defined
expressions. However, if a more advanced force is required, the user can link a user written subroutine (DLL) to FreeDyn.
FreeDyn can be used as standalone software, which is controlled via the mentioned GUI. Alternatively, FreeDyn can be
linked via a C-Interface as a dynamic link library (DLL) to other software like MATLAB, Scilab or self-written codes.
The C-interface provides control functions for time integration and access to model data, which are normally hidden (e.g.
mass matrix, constraint forces, Jacobi matrices etc.). Beside industrial applications, FreeDyn can be used in education and
research.
In Education: Beside the classical teaching elements like the creation and simulation of MBS models, it is furthermore
possible, to look deeper inside the structure of the equations. This can be easily done by the use of the access routines of the
C-Interface.
In Research: FreeDyn enables quick implementations of particular research ideas in the context of MBS. The integration
of those ideas can be realized with a user written subroutine and/or alternatively, by the use of the C-Interface. A researcher
can focus on the research topic, while FreeDyn handles the MBS model. Moreover, due to the availability of very specific
information, the C-Interface of FreeDyn is suitable for extraordinary optimization problems, which is underlined by an
example and several literature citations.
The paper contains a brief review of the theoretical concepts on which FreeDyn is based. The previously mentioned
possibilities are briefly discussed, and underlined by illustrative and meaningful examples.
Keywords Multibody dynamics · Flexible multibody dynamic · Free software · Nonlineare mechanics · HHT time
integration
8.1 Introduction and Theoretical Background of FreeDyn
FreeDyn [1] is a freely available software package for the pre- and post processing and numerical time integration of
multibody dynamic systems. A graphical user interface (GUI) is available for pre- and post-processing. With this GUI
complex models can be built up with the support of typical graphic oriented methods. The results of the time integration can
be read in again, animated and plotted. The automatically generated equations of motion have the following characteristic:
M (q) q̈ + Cq T (q) λ = Q (q, q̇, t)

(8.1)
C (q, t) = 0
The vector q contains the degrees of freedom of the system and consists of n sub-vectors containing the degrees of freedom
of the n modeled bodies. The matrix M(q) is the state dependent mass matrix with block diagonal entries according to the
W. Witteveen () · L. Thomas · O. Stefan


50 W. Witteveen et al.
Fig. 8.1 CPU Time comparison with generic (left) and an industrial model (right)
mass matrices of the individual bodies. The vector Q contains the generalized forces and the quadratic velocity vector. The
constraint equations depend on the state vector q and the time t. They are collected in the vector C. For the consideration of
the constraint forces the transposed of the constraint Jacobian Cq T and the vector of the Lagrange multipliers λ are needed.
For the numerical time integration an HHT solver corresponding to [2] is implemented in C++. For the parameterization
of the rigid body rotation Euler parameters are chosen. The known problems concerning numerical damping in connection
with Euler parameters have been solved and implemented in the solver, see [3]. Flexible bodies are formulated according to
the Floating Frame of Reference Formulation (FFRF), see [4].
In order to underline FreeDyn’s efficiency a comparison with a commercial tool is given next. The first model is a more
generic model, see Fig. 8.1 on the left hand side. A four link mechanism is loaded via a force or a torque. Two identical
models have been build up in FreeDyn and in MSC.ADAMS [5]. The time integrations have been performed with the HHT
solver and comparable integration parameters. Depended on the actual set of parameters it turned out, that sometimes the
commercial code was faster (maximum 46%) and sometimes FreeDyn was faster (maximum 25%).
An image of the second example is given in Fig. 8.1 on the right hand side. An eight cylinder crank drive has been build
up. All connecting rods are modelled as flexible bodies with 21 Craig/Bampton modes. The crank drive is accelerated to
4200 rpm via a torque acting directly on the crank drive. The results of the commercial code MSC.ADAMS [5] and FreeDyn
are in good agreement. MSC.ADAMS [5] needed 20 seconds CPU time at an integrator step size of 4e-5 seconds while
FreeDyn needs 40 seconds. However, a much smaller integrator step size (e.g. 1e-5 seconds) increases the difference up a
factor of 5.
8.2 Standard Modelling Elements of FreeDyn
Simple rigid bodes like spheres are rods can be directly defined in the FreeDyn GUI. For more complex shaped rigid bodies
a STL-interface is available.
Flexible bodies are described along the FFRF. This formulation requires some special matrixes. At the moment, the
interface of MSC.ADAMS [5] is used.
Measures are virtual sensors to record the time history of certain quantities. With those measures displacements, velocities
and accelerations can be investigated.
FreeDyn supports currently three groups of constraints. One group locks the relative translation and rotations of certain
degrees of freedom of the involved bodies. With the second group of constraints motions between two degrees of freedom
can be enforced and finally a gear constraint is available. This enforces a certain ratio of displacement between to rotational
degrees of freedom.
Two approaches can be used to model forces inside FreeDyn. First, generic forces between two degrees of freedom can
be applied via the FreeDyn GUI. Those forces can depend on time and on arbitrary state variables (via measures). More
complex forces can be implemented using a dynamic linked library (dll). Such a dll can be programmed by any user and
linked to FreeDyn. In this way a Pacejka tire model has already been successfully implemented, see below.
8 FreeDyn: A Free, Flexible and State of the Art Multibody Simulation Package for Education, Research and Industrial Applications 51
Fig. 8.2 A 3D vehicle model in FreeDyn
Even contact forces between two arbitrarily shaped rigid bodies can be considered. Therefore, the approach and code
published by G. Hippmann has been linked to FreeDyn [6].
To demonstrate the capabilities of FreeDyn, some details of a successfully implemented vehicle model are given next, see
[7] and Fig. 8.2.
The model consists of 31 rigid bodies resulting in 217 generalized coordinates and 203 constraint equations. For a
meaningful investigation of the vehicle dynamics, the following model details are essential:
• A Pacejka tire model has been implanted via a dll.
• A fully detailed three-dimensional suspension kinematics using rigid bodies for control arms, uprights, push rods, rockers,
and tie rods is included in the model. For modeling the spring-damper element a non-linear force, defined by a state-
dependent spline, is used.
• The aerodynamical forces, drag and downforce, are given from a computational fluid dynamics simulation through
velocity-dependent coefficients. The resulting state-dependent drag force and drag torque are acting on the center of
mass of the chassis body. The downforce is split into a front axle load and a rear axle load.
• The steering angle is impressed by a time-dependent motion onto the steering rack body. The driving torques, as well as
the braking torques, are described by time-dependent spline torques. In a more advanced simulation a driver model has
been implemented, see Sect. 8.3 below.
• The location of the vehicle (reference point) can be given utilizing curvilinear coordinates. For that purpose, the vehicle
position is projected onto the center line (xc, yc) of the road. The coordinate s describes the arc length to the projected
point on the center line, whereas r denotes the normal distance from the center line. The transformation between the global
vehicle position x, y and r, s is given by a non-linear equations, see [7].
8.3 FreeDyn as C Library
FreeDyn’s C++ solver is compiled with a C-interface and can be downloaded as a dynamic linked library (dll). This gives
the possibility to link FreeDyn with software packages like Matlab [8] or Scilab [9]. Another scenario is to link FreeDyn
with a self written code. The user can focus on his own task (e.g. an optimization) and FreeDyn takes over the multibody
simulation, even of complex models.
The functionality of the interface can be roughly divided into three areas:
1. For each current time integration point, all kind of information can be requested from the solver. These are for example the
displacements, velocities, accelerations, measures, constraint forces, Lagrange multipliers, constraint Jacobian, Jacobian
of the full system and other quantities.
2. An existing model can be time-integrated in a certain interval. This interval can be very long, but also be very short, such
as one single integration time step.
52 W. Witteveen et al.
Fig. 8.3 A 3D vehicle controlled by a MATLAB [8] driver model
3. The interface can be used to modify model parameters. These are e.g. integration parameters, model states or forces.
In order to give a better idea, three functions of the interface are given as examples. For the sake of readability, not all
parameters are given, only some arbitrarily selected ones.
void getSystemState(double* oTime, … double oQ[], … double oL[]);

void solveEoM(… int* iKeepResultsInMemory …);
void modifyForce(int* iForceIndex, … double iDoubleData …);
After a successful simulation, the function getSystemState(...) can be used to read out the state variables and their time
derivatives as well as the Lagrange multipliers at a certain point in time. With the function solveEoM(...) a numerical
simulation can be started. The function modifyForce(...) can be used to change an external force.
The following tasks have been successfully processed with the help of FreeDyn’s C interface:
• Optimization task: Scilab [9] calls FreeDyn with a certain set of parameters and invokes a time integration. Afterwards
the results are read into Scilab [9] and a cost function is evaluated. Based on certain considerations model modifications
are triggered by Scilab [9] via the C interface. The time integration starts again and so on until an optimum is obtained.
An example of a similar strategy can be found in [10].
• Since only a single time step can be calculated in FreeDyn, a complex FreeDyn model can be used as a block inside
Matlab/Simulink [8] (or inside XCos in Scilab [9]). This enables the implementation of a controller in Matlab/Simulink
[8] for a multibody model.
Another scenario which can be imagined would be the co-simulation of several domains. For the multibody simulation
domain FreeDyn can be used. The co-simulation control and synchronization unit can access the multibody domain via the
C-interface in order to get and set the necessary data.
To illustrate the possibilities once again, some details of a successfully simulated example are given below. A complex
off-road vehicle model (40 bodies, 225 constraint equations) has been set up with in FreeDyn, see Fig. 8.3.
The sensitivity of a race track lap time has been investigated with respect to certain model parameters. The necessary
driver model has been implemented inside Matlab [8]. As a consequence Matlab [8] called the FreeDyn time integration
with the streering quantities obtained by the driver model implemented in Matlab [8].
8 FreeDyn: A Free, Flexible and State of the Art Multibody Simulation Package for Education, Research and Industrial Applications 53
8.4 FreeDyn in Industrial Applications
It should already be clear from the previous content of the paper that FreeDyn can be used for industrial problems with a
large number of bodies. The import of several complex and finely meshed FE structures is possible as well.
8.5 FreeDyn in Research
Researchers are frequently faced with the task of demonstrating or verifying their ideas by means of numerical examples.
In order to keep the effort manageable, sometimes minimal examples are constructed that have little to do with industrial
applications. In the context of multibody simulation, FreeDyn can be used to perform a meaningful proof of concept without
having to deal with the complexity of a full model multibody simulation. The development of a new tire model can be
imagined as an illustrative example. Instead of demonstrating the idea using a flat example with a few degrees of freedom,
FreeDyn can be used to implement the tire force either as a dll or via Matlab [8]. In both cases, the researcher can focus on
the actual research topic and nevertheless make a proof of concept with a meaningful model.
8.6 FreeDyn in Education
FreeDyn can be used to train the basic skills for a modern modelling and simulation of multibody systems. In addition,
the C-interface can also be used to get a deeper insight. It is conceivable for a lecture or a seminar that the effects of
different integration or model parameters can be investigated systematically. Another conceivable scenario would be the use
of FreeDyn in a control course. The C-interface can be utilized to design a controller for a (nonlinear) multibody system.
References
1. www.freedyn.at
2. Negrut, D., Rampalli, R., Ottarsson, G., Sajdak, A.: On an implementation of the Hilbert-Hughes-Taylor method in the context of index 3
differential-algebraic equations of multibody dynamics. J. Comput. Nonlinear Dyn. 2, 73–85 (2007). https://doi.org/10.1115/1.2389231
3. Sherif, K., Nachbagauer, K., Steiner, W., Lauß, T.: A modified HHT method for the numerical simulation of rigid body rotations with Euler
parameters. Multibody Syst. Dyn. 46(2), 181–202 (2019)
4. Shabana, A.: Dynamics of Multibody Systems, 4th edn. Cambridge University Press, New York (2013)
5. www.mscsoftware.com
6. Hippmann, G.: Modellierung von Kontakten komplex geformter Körper in der Mehrkörperdynamik. PhD Thesis, Institut fur Mechanik und
Mechatronik, TU Wien (2004) (http://pcm.hippmann.org/doc/hippmann_dissertation.pdf, cited 30.09.2019)
7. Oberpeilsteiner, S., Lauss, T., Eichmeir P., Steiner, W.: FreeDyn – a software for advanced multibody simulation: applications in the field of
vehicle dynamics. In: Proceedings of ECCOMAS Thematic Conference on Multibody Dynamics, July 15th–18th, 2019. Duisburg, Germany
8. https://de.mathworks.com/
9. www.scilab.org
10. Lauß, T., Oberpeilsteiner, S., Sherif, K., Steiner, W.: Inverse dynamics of an industrial robot using motion constraints. In: 2019 20th
International Conference on Research and Education in Mechatronics (REM), pp. 1–7. Wels, Austria (2019). https://doi.org/10.1109/
REM.2019.8744124
Chapter 9
Force Probing to Access Potential Energy
Yawen Xu and Lawrence N. Virgin
Abstract Perhaps the two most common forms of potential energy are those associated with gravitational and elastic forces.
In an experimental setting, if we can measure the force required to maintain equilibrium, then the extraction of potential
energy is relatively straightforward, since the force is the negative vector gradient of the potential. In this paper we apply this
approach to a small mass that is placed on various shapes under the action of gravity.
Keywords Potential energy · Conservative forces · Rolling ball · Load cells · Experiments
9.1 Introduction
In this paper we focus attention on a system comprised of a spherical ball that sits on a prescribed curve or surface, and
under the influence of gravity [1]. That is, a shape that entirely defines the form of potential energy [2]. By definition, at
points of horizontal tangency (zero slope) no force is needed to maintain the system in free equilibrium (both stable and
unstable). If the ball is moved away from a horizontal tangency such that it sits on a slope, then it requires some external
force (in the direction of the tangency) to keep it in that position: otherwise it will roll naturally downhill towards (away
from) a local potential energy minimum (maximum). For a system with a single degree of freedom it is relatively easy to
monitor the (pushing and pulling) force required to move the ball along an undulating curve, and this can be extended to 2D
corresponding to a surface.
We undertake a proof-of-concept approach in this paper, in which we accurately produce a number of physical curves
and surfaces, prescribed mathematically in one (curve) and two (surface) dimensions. The high resolution of these shapes
is achieved using additive manufacturing and CNC milling. An arrangement of load cells is used to measure the force
associated with maintaining the position of the ball on the shape at different locations. The resulting force map is then
integrated (numerically) to recover the potential energy of the system, i.e., the original shape. The process of numerical
integration is subject to the usual issues associated with finite data [3]. The distinction between linear and nonlinear is less
important than the role played by dimension, but in general the approach is well suited to situations in which a variety of
equilibria (including saddles) co-exist.
9.2 Examples
Three sample shapes are shown in Fig. 9.1. In part (a) a 1D curve exhibits a number of turning points; in part (b) a 1D shape
changes form as a function of a parameter B; and in part (c) we have a 2D surface. In each case, a steel ball-bearing is placed
on the prescribed curve/surface and the resulting horizontal forces measured as a function of position.
Y. Xu · L. N. Virgin ()
Department of Mechanical Engineering and Materials Science, Duke University, Durham, NC, USA

56 Y. Xu and L. N. Virgin
Fig. 9.1 (a) 1D curve, (b) a 1D curve parameterized by B, (c) a 2D surface
150
x(mm)
100
50
140 160 180 200 220 240 260

B(mm)
(a) (b)
Fig. 9.2 (a) the cage of load cells acting on the ball bearing, (b) contours of potential energy with equilibrium points superimposed (open data
points – unstable; closed – stable)
9.3 Experiments and Results
The arrangement of sensitive load cells as shown in Fig. 9.2a. The measured force is then integrated (numerically) to obtain
the potential energy and hence shape. An example is shown in Fig. 9.2b for the curve described in Fig. 9.1b. As a function
of B, the number of equilibria changes and this is the reflected in the contours of potential energy superimposed on the
parameter space.
In 2D, we measure forces in both the X and Y directions. This was done for the surface shown in Fig. 9.1c. The theoretical
contours are shown in Fig. 9.3a together with the measured equivalent in part (b).
9.4 Conclusions
This paper describes the measurement of forces to keep a ball in position on a variety of 1D and 2D shapes. From this
information, integration is performed on the force vector to obtain potential energy and hence the original shape on which
the ball sits. This approach has possible applications in force-probing strain energy in continuous elastic systems in which
less information is known a priori than the current case.
9 Force Probing to Access Potential Energy 57
100
80
60
40
20
y(mm)
0
-20
-40
-60
-80
-100
-100 -50 0 50 100
x(mm)
(a) (b)
Fig. 9.3 A 2D potential energy surface, (a) theoretical contours, (b) integrated from the measured forces. The contour scale is mm
References
1. Xu, Y., Virgin, L.N.: Probing the force field to identify potential energy. J. Appl. Mech. 86, 101008 (2019)
2. Feynman, R., Leighton, R.B., Sands, M.: The Feynman Lectures on Physics. Addison-Wesley, Reading (1964)
3. Davis, P.J., Rabinowitz, P.: Methods of Numerical Integration. Academic Press, Elsevier (1984)
Chapter 10
Modeling and Experimental Validation of a Pylon Subassembly
Mockup with Multiple Nonlinearities
Connor Ligeikis, Adam Bouma, Justin Shim, Simone Manzato, Robert J. Kuether, and Daniel R. Roettgen
Abstract The industrial approach to nonlinearities in structural dynamics is still very conservative, particularly from an
experimental point of view. A demo aluminum aircraft has been equipped with discrete nonlinear elements designed
to replicate real-world engine pylon subassemblies to increase awareness on the effects of nonlinearities in design, and
understand how these effects can be positively exploited, if properly understood. After some preliminary experiments aimed
at understanding the coupled behavior of the aircraft-pylon mockup, it became clear that more in-depth numerical and
experimental analyses are required on the pylon subassembly alone. For this paper, experimental data is collected to analyze
the nonlinear dynamic behavior of the pylon, leading to better understanding of the subassembly once it connects to the
aircraft. The pylon element has three main sources of nonlinearities: (1) geometric nonlinearities of the connecting beam, (2)
contact as the beam presses into the tapered block surface and (3) friction in the bolted connections. Backbone curves are
generated, which map the evolution of natural frequency and damping ratio with excitation amplitude. Using the experimental
data, a low-order nonlinear model is identified to replicate the backbone characteristics and response of the pylon.
Keywords Nonlinear dynamics · Backbone curves · Stepped sine testing · Free decay testing · Low-order modeling
10.1 Introduction
Nonlinearities are found in engineering structures and often originate from the design, presence of joints, material properties,
and geometric properties. Identifying the effect of these nonlinearities has been investigated by many researchers and various
methods can be used to measure or predict the response of a structure with nonlinearities [1]. The integration of experimental
nonlinear identification and finite element modeling of structures is helpful to the structural dynamics community and
is important for complex systems due to the common linear assumption of structures. A more accurate representation
of nonlinear behaviors in the finite element model of a structure or system is needed in order to obtain better response
predictions. Therefore, vibration testing exhibiting nonlinear phenomena can no longer be assumed as linear. While research
groups have been extensively working over the last decade to extend structural dynamics methods and approaches to deal with
nonlinearities, both numerically and experimentally, industry is still very conservative and slow to adapt to new techniques
into their engineering workflow. Cooper, DiMaio, and Ewins recently summarized an attempt to integrate finite element
models and test-based identification in nonlinear structural dynamics into a structured approach [2]. The objective of the
approach is to extend existing linear techniques with the addition of nonlinear analysis to construct a model which can
predict the nonlinear behavior of such identification methods and available nonlinear numerical algorithms to obtain a valid
FE model of the structure under consideration. Its attractiveness is that it does not rely on any specific method but uses a
collection of toolbox methods to obtain the final validated structural model.
C. Ligeikis · J. Shim
University of Michigan, Ann Arbor, MI, USA
A. Bouma ()
New Mexico State University, Las Cruces, NM, USA
S. Manzato
Siemens Industry Software, Leuven, Belgium
R. J. Kuether · D. R. Roettgen
Sandia National Laboratories, Albuquerque, NM, USA

60 C. Ligeikis et al.
Fig. 10.1 (a) Siemens aluminum airplane demo with attached pylons and (b) individual pylon
In an attempt to bridge the gap between academics and industrial communities, Siemens Industry Software has been
working on a demo aluminum aircraft, shown in Fig. 10.1a. This demo has been equipped with discrete nonlinear elements
(Fig. 10.1b) designed to replicate engine pylon subassemblies to increase awareness on the effects of nonlinearities in designs
and understand how these effects can be positively exploited if properly understood.
The approach proposed in [2] has been recently applied to the airplane-pylon assembly in Fig. 10.1a. From this initial
study [3], it became clear that a more in-depth analysis of the pylon behavior alone was required before a global model could
be developed and validated. In the airplane-pylon assembly, it was difficult to isolate the response of the pylon from other
phenomena in the assembly, and thus, the current investigation focuses on the pylon subassembly in isolation. A rigid box
fixture was designed to isolate the pylon dynamics so that a low-order nonlinear model could be created and validated with
experimental backbone data. The pylon subassembly is composed of two tapered blocks, a thin beam, and a square swinging
mass. Figure 10.1b shows an individual pylon that contains three possible sources of nonlinearity: (1) geometric nonlinearity
of the thin beam, (2) variable stiffness and damping as the beam increases its contact area with the tapered block, and (3)
friction in the bolted connections. To analyze the effect of these potential nonlinearities, two types of experimental methods
were utilized to characterize the pylon’s nonlinear behavior: stepped sine testing at different input forcing levels and free
decay testing. The CONCERTO method proposed by Carella and Ewins [4] was used to post-process the stepped sine data.
A Hilbert transform based method [5] and a zero-crossing method [6] were used to post-process the free decay data. The
three methods were used to generate frequency and damping backbone curves for the pylon’s first mode of vibration and
derive and calibrate a nonlinear single-degree-of-freedom (SDOF) model of the pylon subassembly.
The structure of this paper is as follows: Section 10.2 describes the theory and methods used for nonlinear system
identification. Section 10.3 describes the experimental setup and sensor location on the fixture and pylon assembly. In Sect.
10.4, the linear finite element model (FEM) is described leading to linear identification based on low-level modal testing
using a combination of hammer and pseudo-random shaker testing used to identify linear natural frequencies, damping,
and mode shapes with correlation and model calibration of the FEM. Section 10.5 describes nonlinear testing results and
provides a comparison of various methods used for nonlinear system identification. Section 10.6 details the development and
experimental validation of a nonlinear single degree of freedom model. Finally, conclusions and an outline of future work
are summarized in Sect. 10.7.
10.2 Theory
10.2.1 Concerto Method
The frequency response functions (FRF) from the stepped sine tests were analyzed using the CONCERTO (COde for
Nonlinear identifiCation from mEasured Response To vibratiOn) method proposed by Carella and Ewins [4]. In this
approach, two points on either side of the resonance peak with the same displacement response amplitude are used to
10 Modeling and Experimental Validation of a Pylon Subassembly Mockup with Multiple Nonlinearities 61
Fig. 10.2 Typical FRF used for CONCERTO method
determine the natural frequency and damping ratio by linearizing the system about that response amplitude. Given an FRF
relating the input force to an output displacement, the method proceeds as follows for an arbitrary response amplitude, Ax :
1. Choose an experimentally measured steady-state response on one side of the resonance peak. The FRF at this first point
is described by:
X (ω) (Ar + j Br )
H (ω) = = = R1 + j I1 (10.1)
F (ω) ω2n + ω21 + j ηr ω2n
where j is the imaginary number, ω1 is the corresponding frequency at the chosen amplitude, ωn is the natural frequency
of the system, ηr is the modal loss factor, Ar and Br are the real and imaginary parts of the modal constant, and R1 and I1
are the real and imaginary parts of the FRF.
2. Determine the frequency of the matching response amplitude on the opposite side of the resonance peak as shown in Fig.
10.2. The FRF at this second point is now defined as:
X (ω) (Ar + j Br )
H (ω) = = = R2 + j I2 (10.2)
F (ω) ω2n + ω22 + j ηr ω2n
where ω2 is the corresponding frequency at the interpolated point and R2 and I2 are the real and imaginary parts of the
FRF at the interpolated point.
3. By combining Eqs. (10.1) and (10.2) the following expressions for amplitude dependent natural frequency and damping
ratio are obtained:
Fig. 10.3 Potential issues with interpolation required for CONCERTO method

(R2 −R1 ) R2 ω22 −R1 ω12 +(I2 −I1 ) I2 ω22 −I1 ω12
ωn (Ax ) = (10.3)
(R2 −R1 ) +(I2 −I1 )
2 2

(I −I )
2 1 R2 ω22 −R1 ω21 +(R2 −R1 ) I2 ω22 −I1 ω21
ζ (Ax ) = −
(10.4)
2ω2n (R2 −R1 )2 +(I2 −I1 )2

4. Finally, by repeating the above steps at a variety of different response amplitudes, frequency and damping backbone
curves can be created by plotting the natural frequency and damping ratio with respect to amplitude.
One of the limitations of the CONCERTO method arises from the need to interpolate the frequency on the opposite side of
the resonance peak as illustrated in Fig. 10.3. It is especially problematic when a system with strong nonlinearities exhibits
multiple steady-state solutions at a given forcing frequency around a resonance. Experimental stepped sine tests are unable
to capture the unstable responses that would be needed to apply CONCERTO, and instead, exhibit the well-known jump
phenomena where the response of the system jumps from one stable steady-state solution to another lower energy response.
An improper interpolation of frequency at the jump phenomenon location leads to an incorrect estimation in the interpolated
forcing frequency and an under or over prediction of the calculated natural frequency or damping ratio from Eqs. (10.3) and
(10.4).
10.2.2 Hilbert Transform Method
One of the approaches utilized to analyze the free decay response data involved using the Hilbert transform with curve fitting
proposed by Sumali and Kellogg [5]; the theory is briefly reviewed here for completeness. The Hilbert transform is a linear
operator that shifts the frequency components of a signal by − π2 radians. The Hilbert transform can be used to find the
analytical signal, V(t), defined as:
∼
V (t) = v(t) + j v(t) (10.5)
∼
where v(t) is the original signal and v(t) is the Hilbert transform of the signal. The magnitude of the analytical signal
represents the envelope of the free decay response, and the derivative of the phase of the analytical signal represents
the damped oscillation frequency. Once the envelope decay time constant, C(t), and damped oscillation frequency are
determined, the damping ratio and natural frequency can be backed out as follows:

ωn (t) = ωd (t)2 + C(t)2 (10.6)
ζ (t) = − ωC(t)
n (t)
(10.7)
The raw Hilbert transform is especially sensitive to noise in the signal, so the curve fitting approach in [4] was used to
mitigate this issue. However, curve fitting may not be appropriate for strongly nonlinear solutions, which may require an
unrealistic nth order fit to capture the intricacies of the Hilbert transform. As an alternate to curve fitting, the use of a moving
average filter on the raw magnitude and phase of the Hilbert transform was also explored. The moving average filter took a
10-sample window around each point and averaged them, thus smoothing the curve of the raw data.
10.2.3 Zero-Crossing Method
The second approach to analyze the free decay response data utilized the zero-crossing method proposed by Londoño, Neild,
and Cooper [6]. In this approach, instantaneous frequency and damping ratios are estimated for each cycle of oscillation
during a ring down response measured from the system. Given a free decay signal x(t), the method proceeds as follows:
1. Determine the set of times, t0 , where x(t) crosses zero.
2. Estimate the instantaneous natural frequency for each zero-crossing time using the following equation:
−1
fˆ ti0 = ti+1
0 − t0
i−1
(10.8)
3. A moving average filter can be applied to the instantaneous frequency estimates to smooth out noisy data:

1 N −1 ˆ
f ti0 = N j =0 f
0
ti+j (10.9)
The choice of N here is dependent on the level of noise present in signal. For this study, we found that N = 10 produced
reasonable results.
4. Determine the maximum absolute value of x(t) and the corresponding time tia for each zero-crossing time interval using:
âi = max |x(t)| (10.10)

∀t| ti0 ≥t≥ti+1
0
tia = t ⇐⇒ aî = |x(t)| (10.11)


5. Interpolate tia , aî to obtain an amplitude envelope, Ax ti0 , with values that correspond to the zero-crossing times. For
this study, a simple first order interpolation was used.
6. Estimate the instantaneous damping ratio at each zero-crossing time using the log-decrement method:

ζ̂ ti0 = 01 ln (Ax (ti−1 )) − ln Ax ti+1
0
(10.12)
ω0 ti0 ti+1 −ti−1
0
7. A moving average filter can be used to smooth the instantaneous damping ratio estimates:

1 N −1
ζ ti0 = N j =0 ζ̂
0
ti+j (10.13)
Again, it was determined that N = 10 produced reasonable results.

The resulting frequencies, f ti0 , from Eq. (10.9), damping ratios, ζ ti0 , from Eq. (10.13) and the response amplitude,

Ax ti0 , all describe the backbone curves that depend on the amplitude of response.
10.3 Experimental Setup
The pylon subassembly was bolted to a rigid, aluminum test fixture, as shown in Fig. 10.4, to isolate the fundamental
vibration mode from the airplane assembly. The pylon was designed to behave rigidly in the frequency bandwidth around
the damper’s first bending mode at 7.2 Hz. The pylon could be bolted to the test fixture such that it either moved parallel or
perpendicular to the opening of the test fixture. The test fixture was bolted down to an optical breadboard which was glued
to the floor to ensure rigidity. Next to the test fixture, two granite slabs were glued to one another and the optical breadboard
to provide an appropriate height for the shaker. The shaker was glued to the granite slabs and was oriented for the stinger to
shake the test fixture perpendicular to the openings of the test fixture. The optimal location of the stinger on the test fixture
was determined using a series of hammer tests and finite element simulations to maximize the FRF of the pylon at its first
mode resonance.
Four 100 mv/g lightweight PCB uniaxial accelerometers were glued to the swinging mass of the pylon to monitor its
behavior. Also, a triaxial and uniaxial accelerometer were attached on opposite sides of the upper tapered blocks, and another
triaxial accelerometer was attached directly above the block on the test fixture to monitor any relative acceleration between
Fig. 10.4 Experimental setup

the pylon and test fixture. Finally, a triaxial accelerometer was attached directly opposite to the location of the load cell to
monitor the input point of the test fixture and to obtain drive point responses during shaker and hammer tests.
10.4 Linear Modal Testing and FEA Model Calibration
A detailed finite element model was generated using CUBIT [7] to discretize the mesh and Sierra Structural Dynamics finite
element codes [8] to simulate the modes and forced response of the pylon and the fixture-pylon assembly. The CAD geometry
of the assembly is shown in Fig. 10.5. An eigenvalue analysis was performed to obtain the natural frequencies and mode
shapes of the assembly with the base of the fixture having fixed boundary conditions. The model was used during the pre-test
analysis to determine the optimal shaker input location to maximize response of the pylon mass during nonlinear stepped
sine testing. The model here provided confirmation of the separation of the dynamics of the fixture and pylon and the modal
characteristics were validated with experimental measurements.
The first five mode shapes of the fixture-pylon are shown in Fig. 10.6. The first mode of vibration is a weak-axis bending
mode of the pylon with a natural frequency of 7.2 Hz. There is essentially no response in the fixture ensuring that the
dynamics of the pylon are adequately isolated. The second mode is a pylon torsional mode with a resonant frequency of
47.3 Hz. The fixture dynamics become important in the third vibration mode which corresponds to a bending mode in the
fixture at 80.5 Hz. The fourth mode corresponds to a higher order pylon bending mode, while the fifth and final mode
investigated is a ‘flapping’ mode where the top plate of the fixture bends and the pylon displaces perpendicular to the fixture
opening. The modal analysis on the fixture and pylon assembly confirm the fixture design achieved reasonable separation
between the fundamental bending mode (7.2 Hz) and the first bending mode of the fixture (80.5 Hz).
In order to validate and calibrate the finite element model of the pylon/fixture assembly, low level modal testing
was performed. Hammer and pseudorandom shaker tests were conducted to obtain FRFs relating input forces to output
accelerations at each of the accelerometer locations. The hammer impulses were applied at both the shaker drive point and
on the swinging mass of the pylon. These FRFs were then curve fit using the Siemens PLM Simcenter Testlab PolyMAX [9]
module to obtain experimental natural frequencies, mode shapes, and damping ratios. Using this data, the material properties
(i.e., elastic moduli and densities) and boundary conditions in the bolted connections of the pylon/fixture assembly were
updated to better match the experimental results. In general, the experimental frequencies agreed well with the FEA results
as reported in Table 10.1. The maximum frequency error of 3.66% occurred in the fourth mode of the assembly, however
the first bending mode of the pylon, which was of most interest in this investigation, was within 0.41% of the experiment.
In addition to the frequency comparison, there was excellent correlation between the experimental and FEA mode shapes
as shown by the Modal Assurance Criterion (MAC) plot in Fig. 10.7. Furthermore, the first mode had the least amount
Fig. 10.5 CAD model of pylon and fixture assembly

Fig. 10.6 Mode shapes obtained from FEA
Table 10.1 Experimental and Mode 1 2 3 4 5

FEA modal data
fexp (Hz) 7.25 45.79 78.07 96.30 134.75
fFEA (Hz) 7.23 47.28 80.48 99.89 134.73
% difference 0.41 3.20 3.04 3.66 0.015
ζ exp (%) 0.12 1.76 0.39 0.95 0.35
of damping while the second mode had the most damping. The lack of damping in the fundamental mode proved to be
challenging during the stepped sine testing, as described in the next section.
10.5 Nonlinear Testing
10.5.1 Stepped Sine Tests
To study the response of the first bending mode of the pylon at different forcing and response amplitudes, stepped sine tests
at a range of input forces were conducted near the first mode frequency around 7.2 Hz. Due to the low amount of damping
present in the first mode (0.12%), special care was taken to choose a measurement time that would limit the effect of leakage
from one frequency to another and ensure that the steady state response at each frequency interval was reached. Each test
took approximately 20–30 minutes to conduct, depending on the frequency range examined.
Fig. 10.7 MAC matrix
For strongly nonlinear solutions at large response amplitudes, it was expected that the stepped sine tests would be unable
to capture the true response of the system due to instability of the bent portion of the resonance peak in the strongly nonlinear
range. However, the jump phenomenon location reveals the approximate changes of the natural frequency. From the stepped
sine tests, response spectra (Fig. 10.8a) and FRF (Fig. 10.8b) data were collected. With increasing levels of force, the
resonance initially started to soften going from 0.5 to 7 N. Beyond this force level, the system started to stiffen as the steady-
state force amplitude increased from 7 to 20 N. The transition from softening to hardening behavior is attributed to the thin
beam contacting the tapered blocks. As the first vibration mode increases in response amplitude, the stiffness of the tapered
block increased the natural frequency of the nonlinear mode and significantly changed the dynamics of the system; this could
be verified visually from the experimental setup. The normalized force response in Fig. 10.8b shows that in the stiffening
regime, the ratio of output and input decreases at higher force levels.
10.5.2 Free Decay Tests
Free decay tests were performed by pulling back (i.e., applying an initial displacement) and releasing the pylon’s swinging
mass. This approach proved to be more repeatable than trying to apply an impulsive load with a modal hammer. The measured
accelerations were converted to displacements by numerically integrating twice using the cumtrapz MATLAB function.
Bandpass filtering was utilized both to prevent signal drifting and to isolate the first mode response of the oscillating pylon.
The filtfilt forward-backward filter function in MATLAB was used to prevent signal phase distortion. Figure 10.9 shows some
intermediate results obtained via the zero-crossing method. There was noticeable variation in the estimated natural frequency
and damping as the amplitude of vibration decreased.
10.5.3 Comparison of Backbone Curves
A comparison of the experimental pylon frequency and damping backbone curves is shown in Fig. 10.10. In general, the
frequency backbone curves were similar regardless of the testing or analysis method employed. There appears to be an initial
softening behavior where the natural frequency slightly decreases until reaching a displacement amplitude of approximately
0.0075 m. After this ‘elbow’, there is a sharp hardening where the natural frequency increases as the amplitude increases. This
‘elbow’ was observed to occur when the pylon’s thin beam makes full contact with the upper tapered blocks. The frequency
backbones obtained using the zero-crossing method, the Hilbert transform moving average method, and by peak-picking
the stepped sine data shown in Fig. 10.8a were in good agreement and were thought to provide the best approximation of
the true frequency backbone of the first pylon mode. In contrast to these reference data, the Hilbert transform with an 8th
(a )
(b)
Fig. 10.8 Stepped sine (a) response spectra and (b) FRFs
order polynomial curve fit appears to underestimate the location of the ‘elbow’ and overestimate the slope of the hardening
portion of the backbone. Furthermore, the CONCERTO method yielded a set of frequency backbones with tails that are not
consistent with the results obtained via the other analysis methods. These tails are due to the previously described issues
with the necessary interpolations required to apply the CONCERTO method and are thus not reliable estimates of the system
dynamics once the nonlinearity is very strongly excited.
The damping backbone curves were not as comparable as the frequency backbone curves. Both Hilbert tranform methods
gave results that seemed to overestimate the damping as compared to the zero-crossing method. In addition, the zero-crossing
method appeared to better capture possible intricacies in damping nonlinearity. More specifically, the damping ratio almost
linearly increased until an amplitude slighly less than 0.002 m. The damping ratio then remained constant until an amplitude
of 0.01 m and finally exhibited nonlinear behavior at larger amplitudes. However, through additional testing (which is
described in the next subsection), it was found that this high-amplitude nonlinear damping behavior is most likely caused by
the attached accelerometer cables. It is believed that the interpolations required for the CONCERTO method resulted in very
large errors in the estimation of damping ratios and those were not included in Fig. 10.10.
(a)
(b)
(c)
Fig. 10.9 Zero-crossing method: (a) free decay response and amplitude envelope, (b) instantaneous natural frequency, and (c) instantaneous log
decrement
10.5.4 Effect of Sensor Cables on Backbone Curves
Due to the low damping present in the system (<0.3% at small amplitudes) and the placement of the uniaxial accelerometers
on the pylon swinging mass, it was necessary to investigate the effects the orientation and the number of cables had on
the damping of the system. Thus, a series of free decay tests were conducted to study the impact of sensor cables on the
estimated frequency and damping backbone curves. These tests were performed by detaching differing numbers of cables
from the swinging mass of the pylon. The specific accelerometer cables studied are highlighted in Fig. 10.11a. It should be
noted that to limit mass loading changes, only the cables were removed and the accelerometers themselves remained glued
onto the pylon. The zero-crossing method was then used to derive the backbone curves shown in Fig. 10.11b. The legend
in this figure denotes which cables were attached for each test. It was discovered that the cables had very little impact on
the frequency backbone curve. In contrast, the amount of damping present in the pylon was significantly reduced as the
number of attached cables decreased. The cables also introduced additional nonlinear damping behavior at amplitudes larger
than 0.01 m. Finally, it was observed that certain cables had a larger effect on the pylon damping than others, in particular,
sensor s4, which seemed to introduce a “bump” when comparing the fairly linear blue curve and green curve. Because of the
non-negligible effects on damping, it is strongly recommended that future studies use non-contact methods, such as using a
laser vibrometer or DIC (digital image correlation), to characterize the pylon’s dynamics, in particular, the true damping of
the system.
10.6 Nonlinear SDOF Model Identification and Validation
10.6.1 Model Formulation
In order to most accurately capture the pylon’s dynamics, the backbone curves obtained with only one accelerometer cable
attached to the swinging mass (i.e., the blue curves in Fig. 10.11b) were utilized to develop a low-order nonlinear model of
the pylon. The pylon was modeled as a nonlinear SDOF system with the following equation of motion:
(a) (b)
Fig. 10.10 Comparison of experimental (a) frequency and (b) damping backbone curves using different post-processing techniques
Fig. 10.11 (a) Accelerometer labels and (b) effect of attached cables on backbone curves
mẍ(t) + cnl (x)ẋ(t) + knl (x)x(t) = f (t) (10.14)
where m is the pylon mass, cnl (x) is a nonlinear damping term that is a function of the displacement x, knl (x) is a nonlinear
stiffness term, and f (t) is the forcing term. It was found that the experimental damping backbone curve could be best
replicated using a nonlinear damping function with the following form:
Table 10.2 Optimal nonlinear SDOF model parameters

m c0 c1 c2 c3 k0 k1 k2 k3 kgap xgap
0.141 0.00564 0.753 −18.98 140.9 289.4 −957.9 90821 5113296 11440 0.00723
(kg) (Ns/m) (Ns/m1.5 ) (Ns/m2 ) (Ns/m2.5 ) (N/m) (N/m2 ) (N/m3 ) (N/m4 ) (N/m2 ) (m)
√
cnl (x) = c3 |x|1.5 + c2 |x| + c1 |x| + c0 (10.15)
where c0 , c1 , c2 , and c3 are nonlinear damping coefficients. Here the damping is essentially a polynomial function of the root
of the displacement. It was found that the stiffness backbone curve was best replicated using a piecewise function:

knl (x) = k3 |x|3 + k2 x 2 + k1 |x| + k0 + kgap |x| − xgap (10.16)
where k0 , k1 , k2 , and k3 are nonlinear stiffness coefficients, xgap is the displacement amplitude corresponding to the previously
described ‘elbow’ in the frequency backbone curve, and kgap is a nonlinear stiffness coefficient that is nonzero only when
|x| > xgap . Thus, the kgap (|x| − xgap ) term is zero when the pylon displacement is less than xgap and nonzero when the pylon
displacement is greater than xgap .
10.6.2 Model Parameter Identification
The nonlinear SDOF model was implemented in Simulink and had 10 unknown parameters: c0 , c1 , c2 , c3 , k0 , k1 , k2 , k3 ,
kgap , and xgap . The mass m was assumed to be equal to the pylon’s swinging mass. The unknown parameters were identified
using a two-phase optimization approach that minimized the difference between the experimental backbone curves and
the numerically simulated backbones. The numerical backbones were obtained by simulating a free decay test in Simulink
using a set of model parameters and then employing the zero-crossing method. The optimization was performed using the
fminsearch function in MATLAB.
First, the optimal stiffness parameters θ knl = k0 , k1 , k2 , k3 , kgap , xgap ] were obtained by minimizing the following
objective function:

θ k nl = argmin f exp (Ax ) − f sim (Ax ) (10.17)
θknl
where fexp (Ax ) is the experimental frequency backbone vector, fsim (Ax ) is the simulated frequency backbone vector, and
· denotes
the Euclidean norm. Next, with the optimal stiffness parameters held constant, the optimal damping parameters
θ cnl = c0 , c1 , c2 , c3 ] were identified by minimizing a similar objective function:

θ cnl = argmin ζ exp (Ax ) − ζ sim (Ax ) (10.18)
θcnl
where ζ exp (Ax ) is the experimental damping backbone vector and ζ sim (Ax ) is the simulated damping backbone vector.
This identification procedure yielded the nonlinear SDOF parameters listed in Table 10.2. The numerical backbone curves
obtained using these parameters matched well with the experimental backbone curves as shown in Fig. 10.12.
10.6.3 Experimental Validation
The identified nonlinear SDOF model was validated by comparing numerically simulated time histories with experimentally
measured time histories. The identified model was able to accurately capture the free decay response of the pylon as shown
in Fig. 10.13a. Next, a low-level hammer test was performed by striking the swinging mass of the pylon such that it did
not cause the thin beam to contact the upper tapered blocks. The results of this test are shown in Fig. 10.13b. It should
be noted that the experimental signals were low-pass filtered to obtain only the first mode response. Using the measured
Fig. 10.12 Comparison of experimental and simulated backbone curves
hammer force as the input, the model was able to accurately capture both the initial response to the impulse and the ring
down response. Finally, a high-level hammer test was performed to gauge how well the model captured the pylon’s behavior
when the thin beam made contact with the upper tapered blocks. However, in order to perform this high-level testing, it was
necessary to change the experimental configuration so a large enough hammer swing could be made without accidentally
contacting the fixture. This new configuration is shown in Fig. 10.14a (i.e. the pylon is essentially rotated 90◦ ). With this
new configuration, the model could capture the initial impulse response relatively well; however, there were slight errors in
both natural frequency shifting and damping during the ring down response.
To determine what caused these differences in the high-level results, experimental backbone curves were generated
for the new experimental configuration using the zero-crossing method. The backbones curves for the new and original
configurations were then overlaid as shown in Fig. 10.14b. While there were minimal differences in the frequency backbones,
there was noticeable variation in the damping backbones particularly at amplitudes below 0.008 m. These differences are
most likely due to effects of the orientation of the attached accelerometer cable.
10.7 Conclusions
This paper has presented a study investigating the nonlinearities observed during the experimental campaign of a pylon
subassembly in isolation from the operating assembly of a mock aircraft structure. The overall aim of the paper was to
demonstrate the application of several techniques for nonlinear system identification to derive a low order model of the
subsystem to eventually be coupled to the airplane model. This was achieved by two methods; stepped sine tests at different
forcing levels and free decay tests. The pylon exhibits a weak initial softening behavior and then a strong hardening behavior
originating from contact between the tapered blocks and thin beam of the subassembly. Of the methods used, the zero-
crossing approach seems to be most accurate and the easiest to implement in order to identify frequency and damping
versus amplitude backbone curves. Overall, it is difficult to determine where damping originates, whether it is structural,
friction, air, etc., but it has been confirmed experimentally that the sensor cables have a significant effect on pylon damping.
In future studies, this could be prevented by using laser Doppler vibrometry or digital image correlation to obtain better
damping estimates. Several future studies could follow-on from this work. Additional tests could be performed to evaluate
nonlinear SDOF model performance in other loading environments such as pseudorandom. Further investigations could
(a)
(b)
(c)
Fig. 10.13 Comparison of (a) free decay results, (b) low-level hammer input results, and (c) high-level hammer input results
explore possible nonlinearities in higher frequency modes or coupling between modes with the development of multi-DOF
models. The ultimate goal of this research is to eventually connect the pylon subassembly to the linear demo airplane model
in [3] to the nonlinear SDOF model and studying the combine system behavior. This would lead to a better understanding as
to whether simplified and isolated tests of the nonlinearity in a complex system can be identified and included into the full
system.
Acknowledgements This research was conducted at the 2019 Nonlinear Mechanics and Dynamics (NOMAD) Research Institute supported by
Sandia National Laboratories and hosted by the University of New Mexico. Sandia National Laboratories is a multimission laboratory managed
and operated by National Technology and Engineering Solutions of Sandia, LLC., a wholly owned subsidiary of Honeywell International, Inc.,
for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA-0003525. The authors would also like to
thank Bill Flynn from Siemens Industry Software NV for supplying the data acquisition systems used to collect the experimental measurements
presented throughout this work. SAND2019-12265 C.
Fig. 10.14 (a) Experimental configuration for high-level hammer test and (b) comparison of experimental backbone curves for original and new
experimental configurations
References
1. Noel, J.P., Kerschen, G.: Nonlinear system identification in structural dynamic: 10 more years of progress. Mech. Syst. Signal Process. 83, 2–35
(2017)
2. Cooper, S.B., DiMaio, D., Ewins, D.J.: Integration of system identification and finite element modelling of nonlinear vibrating structures. Mech.
Syst. Signal Process. 102, 401–430 (2017)
3. Cooper, S.B., Manzato, S., Borzacchiello, A., Bregant, L., Peeters, B.: Investigating nonlinearities in a demo aircraft structure under sine
excitation. In: Model Validation and Uncertainty Quantification, vol. 3, pp. 41–57. Springer, Cham (2020)
4. Carrella, A., Ewins, D.J.: Identifying and quantifying structural nonlinearities in engineering applications from measured frequency response
functions. Mech. Syst. Signal Process. 25(3), 1011–1027 (2011)
5. Sumali, H., Kellogg, R.A.: Calculating Damping from Ring-Down Using Hilbert Transform and Curve Fitting. (No. SAND2011-1960C). Sandia
National Lab. (SNL-NM), Albuquerque (2011)
6. Londoño, J.M., Neild, S.A., Cooper, J.E.: Identification of backbone curves of nonlinear systems from resonance decay responses. J. Sound Vib.
348, 224–238 (2015)
7. CUBIT Development Team. CUBIT 15.4 User Documentation. SAND2019-3478 W
8. Sierra Structural Dynamics Development Team (March 2019). Sierra/SD – User’s Notes – 4.52. SAND2019-3063
9. Peeters, B., Van der Auweraer, H., Guillaume, P., Leuridan, J.: The PolyMAX frequency-domain method: a new standard for modal parameter
estimation. Shock Vib. 11(3, 4), 395–409 (2004)
Chapter 11
Comparison Between Control-Based Continuation
and Phase-Locked Loop Methods for the Identification
of Backbone Curves and Nonlinear Frequency Responses
Florian Müller, Gaëtan Abeloos, Erhan Ferhatoglu, Maren Scheel, Matthew R. W. Brake, Paolo Tiso,
Ludovic Renson, and Malte Krack
Abstract Control-based continuation (CBC) and phase-locked loops (PLL) are two experimental testing methods that have
demonstrated great potential for the non-parametric identification of key nonlinear dynamic features such as nonlinear
frequency responses and backbone curves. Both CBC and PLL exploit stabilizing feedback control to steer the dynamics
of the tested system towards the responses of interest and overcome important difficulties experienced when applying
conventional testing methods such as sine sweeps to nonlinear systems. For instance, if properly designed, the feedback
controller can prevent the system from exhibiting untimely transitions between coexisting responses or even losing stability
due to bifurcations. This contribution aims to highlight the similarities that exist between CBC and PLL and present the
first thorough comparison of their capabilities. Comparisons are supported by numerical simulations as well as experimental
data collected on a conceptually simple nonlinear structure primarily composed of a thin curved beam. The beam is doubly
clamped and exhibits nonlinear geometric effects for moderate excitation amplitudes.
Keywords Control-based continuation · Phase-locked loop · Backbone curves · Nonlinear dynamics · Experimental
identification
11.1 Introduction
Experimental characterization of the dynamics of nonlinear structures is challenging for multiple reasons. The characteristics
of the structural response can be heavily dependent on the level of excitation of the structure. For instance, the frequency at
which the structure’s displacement and the force applied to it are in phase quadrature can vary with the response amplitude.
This relationship is called the backbone curve of the structure. Additionally, several periodic orbits can exist for identical
excitation signals. This can be illustrated by the frequency response curves, the relationship between the structural response
and the frequency of an harmonic excitation of constant amplitude. Consequently, classical methods used to characterize
F. Müller · M. Scheel · M. Krack

Institute of Aircraft Propulsion Systems, University of Stuttgart, Stuttgart, Germany
e-mail: [email protected]; [email protected]; [email protected]
G. Abeloos
Department of Aerospace and Mechanical Engineering, University of Liège, Liège, Belgium
E. Ferhatoglu
Department of Mechanical and Aerospace Engineering, Politecnico di Torino, Torino, Italy
M. R. W. Brake
Department of Mechanical Engineering, Rice University, Houston, TX, USA
P. Tiso
Department of Mechanical and Process Engineering, ETH Zürich, Zurich, Switzerland
L. Renson ()
Merchant Venturers Building, University of Bristol, Bristol, UK

76 F. Müller et al.
linear structures encounter difficulties. For instance, performing a sine sweep on a nonlinear structure can result in sudden
jumps from one periodic orbit to another and might not unveil all possible limit states (including isolated and unstable ones).
In recent years, new methods for experimental characterization using feedback control were introduced in the literature.
The present work consists in performing for the first time the comparison between the phase-locked loop method (PLL) [1]
and the control-based continuation (CBC) [2]. Both method can directly obtain the backbone curve in a single experiment.
The consistency between the methods is studied as well as their capabilities.
11.2 Characterization Methods
The phase-locked loop (PLL) method imposes a phase lag v ∗ (t) between the structural response r(t) and the excitation
signal fexc (t) using feedback control. Figure 11.1a shows the structure of the PLL feedback loop. The phase lag v(t) is
detected between the measured response signal r(t) and the oscillating signal u(t). The phase lag detector relies on the fact
that the product of two harmonic signals possesses a constant part whose value depends on the phase lag. This constant
part is isolated using a low-pass filter whose cut-off frequency is below the first harmonic of the signals. Subtracting v(t)
from the phase lag target v ∗ (t) gives the phase error e(t) = v ∗ − v(t) which is fed into a PID controller. The purpose of
the controller is to cancel e(t) completely. The resulting signal y(t) has frequency units and can be seen as the frequency
difference between the linear resonance frequency w and the excitation frequency f (t) = w + y(t). The time integral of the
excitation frequency is the excitation phase p(t). The cosine of p(t) gives the excitation signal u(t). In order to control the
excitation level, u(t) is multiplied by a gain F̂ creating the signal fexc (t) which feeds the actuator connected to the structure.
Control-based continuation (CBC) imposes the response amplitude using feedback control as well. Figure 11.1b shows
the feedback loop of CBC. The response r(t) is multiharmonic due to the nonlinearities in the experiment. r(t) is directly
compared to a response target r ∗ (t) to create the error signal e(t), fed into a PD controller. The resulting excitation signal
y(t) is fed into the shaker connected to the structure. Contrary to the PLL, the control error e(t) does not need to be cancelled
completely as the target signal acts only as a proxy for the system states. In this case, y(t) must be a non-zero monoharmonic
signal, and so must e(t). The main challenge of the method is to find an adequate r ∗ (t). The first harmonic of r ∗ (t) is the
parameter controlling the excitation level. However, if r ∗ (t) was monoharmonic, the error signal e(t) would contain the
harmonics of r(t) and the controller would be invasive, i.e. the excitation signal y(t) would be multiharmonic whereas a
(a)
(b)
Fig. 11.1 Diagrams of the PLL and CBC methods, both using different feedback loops to impose respectively the excitation-response phase lag
and the response amplitude; the experiment comprises an actuator, sensors, and the structure. (a) The signals are the following: r(t) the measured
response, p(t) the excitation phase, v(t) the measured phase lag, v ∗ (t) the target phase lag, e(t) the error, y(t) the PID controller output with
units of frequency, w the linear resonance frequency, f (t) the excitation frequency, and the excitation signal u(t) multiplied by the gain F̂ to
create the signal fexc (t) fed into the actuator of the experiment. (b) The signals are the following: r(t) the measured response, r ∗ (t) the target
response, e(t) the error, y(t) the PD controller output with units of voltage which is fed into the shaker; iterations are made on r ∗ (t) to cancel the
non-fundamental harmonics of e(t)
11 Comparison Between Control-Based Continuation and Phase-Locked. . . 77
monoharmonic excitation is desired. After reaching steady-state, the higher harmonics of r ∗ (t) can be replaced by the ones
of r(t). This is repeated until the higher harmonics are identical in both signals. At this point, e(t) is monoharmonic and so
is y(t): the controller is non-invasive.
11.3 Experimental Results
Both methods are applied on the structure shown in Fig. 11.2. It is a slightly curved beam bolted to a base which exhibits
softening-hardening behavior. The base is mounted vertically on a shaker with an impedance head at their connection to
measure the force and the acceleration at this point. The velocity of the beam is measured by a laser vibrometer (see Fig. 11.2).
Backbones measured successively by CBC and PLL are shown and compared in Fig. 11.3 and present a clear softening-
hardening behavior. They are very consistent with one another, both methods measured the same frequency for the fold
bifurcation in the backbone curve. Figure 11.4 shows the frequency response curves (FRCs) obtained by moth methods.
Figure 11.4a shows the FRCs directly measured by the PLL method. Although FRCs could have also been measured with
the CBC method, they were interpolated from the data collected during amplitude sweeps [2]. These interpolated FRCs are
Fig. 11.2 Base-excited experiment comprising a shaker actuating the structure, itself composed of a rigid base and a thin curved beam; sensors
consist in a laser vibrometer and an impedance head measuring the force and the acceleration at the shaker-base connection
317
316
Frequency [Hz]
315
314
313
PLL
CBC
312
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Response amplitude [m/s]
Fig. 11.3 Comparison between the backbone curve directly measured with the PLL method and with the CBC
78 F. Müller et al.
0.4 0.4
Backbone (measured)
0.5 FRCs (measured)
0.35 0.35
0.45
5N


0.3 0.3
0.4
6N 2.7 N
0.35 0.25 0.25
0.3 1N
2.7 N 0.2 0.2
0.25
0.15 0.15
0.2
0.15 0.3 N 0.1 0.1

0.1 Backbone (PLL, measured)
0.05 FRC (PLL, measured)
0.05
0.05 Backbone (CBC, measured) Backbone (measured)
FRC (CBC, interpolated) FRCs (interpolated)
0 0 0
312 314 316 318 312 313 314 315 312 313 314 315
Frequency [Hz] Frequency [Hz] Frequency [Hz]
(a) (b) (c)
Fig. 11.4 Frequency response curves either measured directly with PLL or interpolated from data measured with CBC. (a) FRCs directly measured
with PLL. (b) Comparison between the FRCs measured with PLL and iterpolated from data measured with CBC at a forcing amplitude of 2.7 N.
(c) FRCs interpolated from data measured with CBC
shown in Fig. 11.4c. Corresponding FRCs obtained by both methods are compared in Fig. 11.4b. Because the experiments
were not made successively, the backbones are not entirely consistent, but both them and the FRCs show high consistency
between both methods.
11.4 Conclusion
The phase-locked loop (PLL) and control-based continuation (CBC) were directly compared when applied to the same
softening-hardening structure. Both methods are capable of directly measuring the backbone curve of the structure and give
out consistent results. The PLL method can directly measure the frequency response curves experimentally. The CBC method
can in theory do the same but in this work, it has been used to measure the full manifold of the structural response, from
which the frequency response curves has been interpolated. The methods give consistent results for the backbones and the
frequency response curves.
Acknowledgments G.A. is funded by the FRIA grant of the Fonds National de la Recherche Scientifique (FNRS), L.R. is funded by a Research
Fellowship from the Royal Academy of Engineering (RF1516/15/11). They gratefully acknowledge the financial support of the Royal Academy
of Engineering and the FNRS.
References
1. Scheel, M., Peter, S., Leine, R.I., Krack, M.: A phase resonance approach for modal testing of structures with nonlinear dissipation. J. Sound
Vib. 435, 56–73 (2018)
2. Renson, L., Gonzalez-Buelga, A., Barton, D.A.W., Neild, S.: Robust identification of backbone curves using control-based continuation. J.
Sound Vib. 367, 145–158 (2016)
Chapter 12
Implementing the Restoring Force Surface Method to Fit
Experimentally Measured Modal Coupling Effects
Benjamin Moldenhauer, Daniel R. Roettgen, and Benjamin Pacini
Abstract In complex structures, particularly those with jointed interfaces, the dynamic response of individual modes can
behave nonlinearly. To simplify analyzing and modeling this response, it is typically assumed that modes are uncoupled,
in that each responds independently of the excitation level of other modes. This assumption is derived from the belief that,
while modal coupling generally exists in physical structures, its effects are relatively small and negligible. This practice is
reinforced by the fact that the actual causes of modal coupling are poorly understood and difficult to model. To that end, this
work attempts to isolate and fit a model to the effects of modal coupling in experimental data from a nonlinear structure. After
performing a low-level test to determine the linear natural frequencies and damping ratios of several modes, sine beat testing
is used to individually excite each mode and record its nonlinear dynamic response. The Restoring Force Surface (RFS)
method is then implemented to fit a nonlinear model to each isolated modal response. Sine beats are then done on multiple
modes simultaneously, in which the response is assumed to be a combination of the nonlinear models of each isolated mode
and some coupling term between them. As the terms modeling the individual modes are known, the only unknown is the
coupling term. This procedure is performed on several mode pairs and excitation levels to evaluate the effectiveness of all
proposed coupling models and gauge the significance of modal coupling in the structure.
Keywords Modal coupling · Restoring force surface · Nonlinear system identification · Hilbert transform · Sine beat
12.1 Introduction
Creating models to accurately predict the dynamic response of complex structures becomes increasingly difficult as sources
of nonlinearity become more significant. This is especially true in designs that contain many bolted joints, as even seemingly
minute motion through the interface can cause the opposing faces to slip relative to one another. This motion, called microslip,
can be a considerable source of damping in the structure due to the frictional forces that develop in the joint. A side effect
of this is that the dynamic modes of the structure can exhibit some degree of coupling, where the response of one is altered
by the excitation of another. These phenomena are poorly understood and have proven quite difficult to accurately measure
and model. Consequently, it is usually assumed that the modes are only weakly coupled and any effects due to this are
insignificant and may be neglected. While this approach has thus far yielded acceptable results, future designs may not
allow for such simplifications and require an accurate description of modal coupling. This work explores the applicability
of implementing the Restoring Force Surface (RFS) method of nonlinear system identification to detect and model modal
coupling in experimental data from a physical structure. This is done by first characterizing the linear modal properties of
the structure, the mass-normalized modeshapes, natural frequencies, and damping ratios of each mode. A modal shaker is
Sandia National Laboratories is a multimission laboratory managed and operated by National Technology and Engineering Solutions of Sandia,
LLC., a wholly owned subsidiary of Honeywell International, Inc., for the U.S. Department of Energy’s National Nuclear Security Administration
under contract DE-NA-0003525. This paper describes objective technical results and analysis. Any subjective views or opinions that might be
expressed in the paper do not necessarily represent the views of the U.S. Department of Energy or the United States Government.
B. Moldenhauer ()
University of Wisconsin – Madison, Madison, WI, USA
D. R. Roettgen · B. Pacini

80 B. Moldenhauer et al.
then used to excite individual modes to high amplitude, such that a RFS model can be fit to the nonlinear response of each
mode. The shaker is then used to excite two modes simultaneously and, using the linear parameters and the nonlinear model
of each isolated mode as knowns, RFS is used to fit a model to any residual modal coupling effects.
12.2 Theory
As proposed by Masri and Caughey [1], RFS is a method of nonlinear system identification that leverages the fundamental
equation of motion (EOM) of a system to isolate its nonlinearity. A single degree-of-freedom modal EOM with general
nonlinearity is given in Eq. (12.1), where q̈ is acceleration, ζ is the linear damping ratio, ωn is the linear natural frequency, q̇
is velocity, q is displacement, fNL (q, q̇) is some internal nonlinear force that is a function of the displacement and velocity,
and fext is an external force acting upon the system.
q̈ + 2ζ ωn q̇ + ωn2 q + fNL (q, q̇) = fext (12.1)
By transferring the linear terms to the RHS, as in Eq. (12.2), one obtains an expression for the cumulative nonlinearity in
the system, or the restoring force, in terms of easily measured quantities.
fNL (q, q̇) = fext − q̈ − 2ζ ωn q̇ − ωn2 q (12.2)
A model for the restoring force can be formed as in Eq. (12.3), where it is assumed that the nonlinearity can be
characterized by polynomial expressions that are an even and odd function of the displacement and velocity raised to some
power.

fNL (q, q̇) = kq q i + k|q| |q|j + cq̇ q̇ k + c|q̇| |q̇|l (12.3)
After choosing the number and order of these terms, their coefficients are found via a least-squares solution as in Eq.
(12.4), where the restoring force is premultiplied by the pseudoinverse of the nonlinear terms. A more detailed explanation
of these computations is given in [2].
T †
kq k|q| cq̇ c|q̇| = qi |q|j q̇k |q̇|l fext − q̈ − 2ζ ωn q̇ − ωn2 q (12.4)
Coupling between two modes can be modeled as in Eq. (12.5), where the restoring force is now taken to be a function of
responses from both modes and nonlinear terms are formed from combinations of those responses; exponents are neglected
here for brevity.

fNL,C (q1 , q2 , q̇1 , q̇2 ) = σq1 q2 q1 q2 + σq1 q̇2 q1 q̇2 + σq̇1 q2 q̇1 q2 + σq̇1 q̇2 q̇1 q̇2 (12.5)
In general, the order of any of these nonlinear terms is not necessarily an integer number; fractional powers may be
considered. While this allows for very flexible selection of parameters when choosing the model form, there is consequently
an infinite number of models that could be fit to the restoring force, with widely varying levels of success.
12.3 Test Case
The structure analyzed in this work is an aluminum cylinder-plate-beam (CPB) assembly that has been extensively studied in
previous works, such as [2, 3]. The test setup was used to collect linear data from low-level burst random input from a shaker,
and high-level nonlinear responses from sine beat shaker excitations; [3] explores these testing methods in great detail. These
processes are implemented to characterize the first three elastic modes of the CPB, which are the first cantilever modes of
the beam in the soft and stiff directions and the first drum mode of the plate. Coupling between these modes is investigated
by exciting the possible pairings between them; 1–2, 1–3, and 2–3.
12 Implementing the Restoring Force Surface Method to Fit Experimentally Measured Modal Coupling Effects 81
Table 12.1 RFS model terms and coefficients for CPB elastic mode 1
Term qi |q|j q̇ k |q̇|l
Exponent 2.3, 2.6, 2.7, 2.8 1.5, 1.7, 2.4, 4 1.7, 2, 2.5, 3.2 1.7, 2.6, 2.9, 3.2
Modal Acceleratio
RFS NL Fit 1000 Exp
Magnitude
100 Sim
0
-1000
0 200 400 600 800 1000 0 0.5 1 1.5 2
Frequency[Hz] Time[s]
Damping Ratio [-]

10-3
Frequency [Hz]
130
10 Exp
125 Sim
8 Linear
Exp
120 Sim 6
Linear
4
115
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Time[s] Time[s]
Fig. 12.1 Top left – Experimental restoring force for mode 1 and the least squares fit of the nonlinear terms; Top right – Experimental vs Simulation
modal acceleration; Bottom left: Instantaneous Natural Frequency; Bottom right: Instantaneous Damping Ratio
To form the restoring force, the system velocity and displacement are found by integrating the measured acceleration. This
was done in the frequency domain by dividing the acceleration by iω and −ω2 for velocity and displacement respectively,
and then applying a high pass filter to eliminate low frequency drift. The model form fit to this nonlinear restoring force was
determined through a simple combination of trial and error and Monte Carlo simulations, where a specified number of terms
were given a random exponent from a pool of possible values. With coefficients from the least squares solution between
the restoring force and the nonlinear terms, the response of the resultant model is computed by numerically integrating its
EOM. The accuracy of the model is determined by comparing the instantaneous natural frequency and damping ratio of the
integrated response to those of the measured experimental response. For the first soft cantilever beam mode of the CPB, four
of each of the nonlinear displacement and velocity terms were given random powers between 1.1 and 4, at 0.1 increments,
until a sufficiently accurate result was found; the final set of exponents are given in Table 12.1.
The restoring force and the resultant least squares fit of the nonlinear terms is shown in the top left plot of Fig. 12.1; there
is good agreement throughout the frequency range. From numerically integrating the model, where the measured force was
interpolated and used as the applied force in the EOM, the response is shown with the measured experimental acceleration
in the top right of Fig. 12.1, and the instantaneous natural frequency and damping ratio are given in the bottom left and right
plots, respectively. While the frequency of the response is very close to the truth data, the damping ratio is less accurate. As
this is the best model that could be found from any number of terms and range of exponents, it would seem that damping
is difficult to model regardless of how well the restoring force is fit. It was also observed that, even if the restoring force is
accurately modeled, the resultant simulation can deviate greatly from the measured response. The best fit models of the other
modes, and the coupling between them, will be explored in the accompanying presentation.
Acknowledgment This manuscript has been authored by National Technology and Engineering Solutions of Sandia, LLC under Contract No.
DE-NA0003525 with the U.S. Department of Energy/National Nuclear Security Administration. The United States Government retains and the
publisher, by accepting the article for publication, acknowledges that the United States Government retains a non-exclusive, paid-up, irrevocable,
world-wide license to publish or reproduce the published form of this manuscript, or allow others to do so, for United States Government purposes.
82 B. Moldenhauer et al.
References
1. Masri, S.F., Caughey, T.K.: A nonparametric identification technique for nonlinear dynamic problems. J. Appl. Mech. 46, 433–447 (1979)
2. Mayes, R.L., Pacini, B.R., Roettgen, D.R.: A modal model to simulate typical structural dynamic nonlinearity. In: Proceedings of the 34th
IMAC, Orlando (2016)
3. Roettgen, D.R., Pacini, B.R., Mayes, R.L.: Techniques for nonlinear identification and maximizing modal response. In Proceedings of the 37th
IMAC, Orlando (2019)
Chapter 13
Identification of Backbone Curves and Nonlinear Frequency
Responses using Control-based Continuation and Local Gaussian
Process Regression
Ludovic Renson
Abstract Control-based continuation (CBC) is a general and systematic method to probe the dynamics of nonlinear
experiments. In this paper, CBC is combined with a novel continuation algorithm that is robust to experimental noise and
enables the tracking of important nonlinear dynamic features such as backbone and nonlinear frequency response curves.
The method uses Gaussian process regression to create a local model of the response surface on which standard numerical
continuation algorithms can be applied. The local model evolves as continuation explores the experimental parameter space,
exploiting previously captured data to actively select the next data points to collect such that they maximise the potential
information gain about the feature of interest. The method is demonstrated experimentally on a nonlinear structure featuring
harmonically-coupled modes. The regression model is also exploited to estimate the uncertainty of the identified features.
Keywords Nonlinear vibrations · Experimental testing · Control-based continuation · Backbone curves · Nonlinear
frequency response · Gaussian process regression · Uncertainty quantification
13.1 Introduction
Without the need for a mathematical model, control-based continuation (CBC) is a means to apply the principles of numerical
continuation directly to a physical system and experimentally identify key dynamic features such as nonlinear frequency
response curves (NLFR) [1], backbone curves [2] and isola [3]. Initially proposed by Sieber and Krauskopf [4], CBC has
been experimentally demonstrated on a range of mechanical systems, including a parametrically-excited pendulum [5], an
impact oscillator [6] and a cantilever beam with a nonlinear mechanism at its free tip [3].
The fundamental idea of CBC is to use feedback control to stabilise the dynamics of the experiment whilst making the
control system non-invasive such that it does not modify the position in parameter space of the response of the open-loop
experiment of interest. This non-invasiveness requirement defines a zero-problem whose solutions can be found and tracked
in the experiment using path-following methods. Depending on the dynamic features of interest, this zero-problem is often
extended to include additional conditions constraining, for instance, an excitation amplitude or a phase difference.
Effectively and reliably tracking the solutions of such zero-problems during tests is challenging. Most existing path-
following methods are ideal only in a numerical context where the solution path is smooth and derivatives can be evaluated
to high precision. This is not easily achievable in experiments where solutions and derivative estimates are corrupted by
measurement noise. In this paper, I investigate the use of an algorithm combining CBC and Gaussian process regression
(GPR) to significantly improve the reliability of the continuation process. This algorithm builds on earlier work [7] and was
exploited to trace out limit-point bifurcation curves [8]. This paper shows how the same algorithm can also be used to identify
backbone and NLFR curves.
L. Renson ()

84 L. Renson
13.2 Experimental Set-Up
The set-up considered for the experimental demonstration of the algorithm is a steel cantilever beam whose free end is
attached to two linear springs arranged to give rise to geometric nonlinearity at large vibration amplitudes [8]. The length
of the beam as well as the pre-tension in the springs were carefully adjusted such that a 3:1 modal interaction between
the first two bending modes of structure exist. The structure is excited near the clamp using an electrodynamic shaker. The
force applied to the structure is measured using a PCB208C03 force transducer and beam tip vibrations are measured using
an Omron ZX2-LD100 displacement laser sensor. All measurements are made at 1 kHz sampling with no filtering. The
continuation algorithm run on a laptop computer directly connected to the real-time controller (RTC) box via a USB cable.
A third-order feedback control law is used to stabilise the dynamics of the system [8].
13.3 Continuation Principles
GPR provides local models that are smooth and cheap to evaluate and thus enables the use of established numerical
continuation techniques. The GPR models used here possess two inputs which correspond to two experimental adjustable
parameters: the excitation frequency (ω) and one of the fundamental Fourier coefficient of the control target signal (A).
To find NLFR curves, GPR is used to model the force amplitude as a function of both input parameters. Tracking NLFRs
corresponds to finding the input parameters (ω, A) such that the applied force is constant. For backbone curves, GPR models
the phase difference between the beam response and the applied force as a function of both input parameters. Tracking the
backbone curves correspond to finding the input parameters (ω, A) such that the phase difference is π/2. The constant
applied force and the phase quadrature conditions define the zero-problems solved by the continuation algorithm when
tracking NLFR and backbone curves, respectively.
The algorithm starts by collection an initial set of data points for different values of the input parameters ω and A. Based
on the data points captured, a GPR model is built and a solution of the zero problem is found. New data points are then
collected to guarantee that the solution found is not an artifact of the model and does not critically depend on the current
data. To decide where to collect data in the input parameter space, the current GPR model is exploited to numerically assess
the effect of additional data points on the zero-problem. The data point which is found to have the largest effect on the value of
the zero-problem is considered to be the most informative and hence is experimentally collected and then added to the model.
The procedure is repeated until the effect of any new data point on the zero-problem is below a user-defined threshold. At
this stage, the solution of the continuation problem is said to be robust to new data and the continuation algorithm continues
to progress in parameter space. The procedure is repeated for every new solution of the zero problem.
13.4 Results
Data points collected using the online continuation algorithm are all included in one GPR model representing the force
amplitude as a function of both input parameters. The NLFR curves obtained from this model are shown in solid black in
Fig. 13.1a. Two values of the force amplitude are considered (1 N and 2 N). A similar model is built with the data collected
during the backbone continuation. The backbone curve extracted from this model is shown in Fig. 13.1b. The collection of
data points not only at the dynamical feature of interest but also around it enables us to analyse the sensitivity of this feature
to experimentally collected data points and highlight regions of greater uncertainty. To visualize this uncertainty, 300 fold
curves were numerically computed for different data sets. Each set is based on all the data points collected using the online
algorithm but with 10% of the points, selected randomly, removed. NLFR (backbone) curves obtained with these smaller
data sets are shown in gray in Fig. 13.1. For the NLFR, regions of larger result variability, i.e. where the overall curve appears
thicker, are located around the resonance. For the backbone curve, larger result variability appears in the softening region at
low amplitude. Outside the region where data is available, NLFR and backbone curves are found to diverge very quickly.
13 Identification of Backbone Curves and Nonlinear Frequency Responses. . . 85
Fig. 13.1 (a) (—) NLFR curves obtained at 1N and 2N using all the data points collected. (—) NLFR curves obtained with 10% of the points,
selected randomly, removed. (b) (—) Backbone curve obtained using all the data points collected. (—) Backbone curves obtained with 10% of the
points, selected randomly, removed
13.5 Conclusions
In this contribution, I show that the algorithm combining control-based continuation with Gaussian process regression
introduced in [8] to track fold bifurcation curves can also be exploited track nonlinear frequency response and backbone
curves. The variability of the identified features with respect to collected data points can also be assessed.
Acknowledgments The author gratefully acknowledges the financial support of the Royal Academy of Engineering (Research Fellowship
#RF1516/15/11).
References
1. Barton, D.A.W., Sieber, J.: Systematic experimental exploration of bifurcations with noninvasive control. Phys. Rev. E 87(5), 052916 (2013)
2. Renson, L., Gonzalez-Buelga, A., Barton, D.A.W., Neild, S.A.: Robust identification of backbone curves using control-based continuation. J.
Sound Vib. 367, 145–158 (2016)
3. Renson, L., Shaw, A.D., Barton, D.A.W., Neild, S.A.: Application of control-based continuation to a nonlinear structure with harmonically
coupled modes. Mech. Syst. Signal Process. 120, 449–464 (2019)
4. Sieber, J., Krauskopf, B.: Control based bifurcation analysis for experiments. Nonlinear Dynamics 51(3), 365–377 (2008)
5. Sieber, J., Krauskopf, B., Wagg, D., Gonzalez-Buelga, A., Neild, S.: Control-based continuation of unstable periodic orbits. J. Comput. Nonlinear
Dyn. 6(1), 011005 (2010)
6. Bureau, E., Schilder, F., Ferreira Santos, I., Thomsen, J.J., Starke, J.: Experimental bifurcation analysis of an impact oscillator – tuning a
non-invasive control scheme. J. Sound Vib. 332(22), 5883–5897 (2013)
7. Renson, L., Barton, D.A.W., Neild, S.A.: Experimental tracking of limit-point bifurcations and backbone curves using control-based
continuation. Int. J. Bifurcation Chaos 27(01), 1730002 (2017)
8. Renson, L., Sieber, J., Barton, D.A.W., Shaw, A.D., Neild, S.A.: Numerical continuation in nonlinear experiments using local gaussian process
regression. Nonlinear Dynamics 98, 2811–2826 (2019)
Chapter 14
Model Correlation to a Nonlinear Bolted Structure Using
Quasi-Static Modal Analysis
Mitchell Wall, Seyed Iman Zare Estakhraji, and Matthew S. Allen
Abstract Bolted joints are common in many engineering structures, yet they introduce complexity when the interest is
predicting the dynamic response of a system. Under large load amplitudes, the joint contact interfaces will slip and cause
the response to be nonlinear. A method proposed by Festjens et al., and later elaborated upon by the authors and dubbed
Quasi-Static Modal Analysis (QSMA), has made it feasible to model the contact between structures in detail and thus
predict the nonlinear dynamic behavior. Prior works have shown that this is feasible (Wall et al, Predicting S4 beam joint
nonlinearity using quasi-static modal analysis. In: IMAC XXXVII; Jewell et al, J Sound Vib 479:115376, 2020, yet) no
work has rigorously correlated such a finite element model with experimental measurements. This work takes a step in that
direction by quantifying the effect of various features in the FEM to the amplitude dependent damping and natural frequency
predicted by QSMA. Specifically, the effect of the friction coefficient and the curvature of the contact interface on the QSMA
predictions is found and quantified. The results so far show that some of these effects could improve model agreement.
Keywords Quasi-Static Modal Analysis · Nonlinearity · FEA · Damping · Natural frequency
14.1 Introduction
Most structures utilize bolted joints to connect critical load bearing components. These bolted joints introduce many
complications when modeling the dynamics of a structure because under certain conditions a bolted joint can exhibit
nonlinearity with respect to load amplitude. The mechanism causing the nonlinearity is the relative slip that occurs at
the contact interfaces between two clamped components. Slip at the contact interface will dissipate energy and reduce the
effective stiffness of the structure with increasing load amplitude. This may occur in two regimes. In the micro-slip regime
the edges of the contact patch will begin to slip relative to each other, but at least some parts of the joint remains stuck and
the joint shows only weak to moderate nonlinearity. The other regime is macro-slip, where the entire interface begins to slide,
exhibiting a significant increase in nonlinearity. The focus of this work is to model the nonlinearity restricted to micro-slip.
One of the greatest challenges to solving these problems is the computational cost. Modeling macroscale structural
geometry must be balanced with modeling the microscale parameters that dictate the frictional energy dissipation. To address
this, Festjens et al. [3] proposed a method that broke up the bolted structure into linear regions away from the joints and
nonlinear regions near the joints. Then the mode shapes of the linearized structure could be modeled with a coarse finite
element mesh and used as boundary conditions for a much more refined nonlinear model of the joint. Mode shapes would
be calculated for the linearized structure, which would be valid at low force amplitudes, and then they could be updated as
the mode shapes changed at higher amplitudes. The computational cost was further improved by applying Masing’s rules,
described in [4, 5], which assumes that the initial loading curve, or backbone curve, is simply a scaled version of the full
hysteresis curve the structure would undergo in a full vibration cycle. Using the backbone curve, one can also calculate
the instantaneous natural frequency of the structure, and using the area inside the hysteresis curve, one can calculate the
instantaneous damping ratio.
A simplified version of this method [3], was eventually developed in [6]. This method, here referred to as Quasi-Static
Modal Analysis (QSMA), included the difference that the structure was not separated into linear and nonlinear regions, so
that the nonlinear quasi-static problem could be solved only once on the whole model, and the approach was more amenable
M. Wall () · S. I. Z. Estakhraji · M. S. Allen

Department of Engineering Physics, University of Wisconsin – Madison, Madison, WI, USA

88 M. Wall et al.
to commercial FEA software. Furthermore, in [6] the QSMA method was first used together with model updating to obtain
a model that correlated well with experimental data.
QSMA was further developed as a means to create a predictive FEM for damping nonlinearity in [1, 2]. Where various
contact formulations and experimental parameters were explored in both 2-D and 3-D models using Coulomb friction.
Physically reasonable results were found in both cases, although they fell short in being truly predictive of the experimentally
observed nonlinearity for an experimental benchmark structure called the S4 beam. This work hopes to expand on this
progress by creating a high fidelity FEM that uses more accurate measurements of the geometry of the contact surface to
create a more accurate model for the S4 beam. Furthermore, since the sub-millimeters scale surface topography is uncertain,
parametric studies are used to explore the sensitivity of the modal frequency and damping to the surface curvature.
14.2 Preliminary Results
Experimental measurements [7] showed that two modes of the S4 beam exhibited significant nonlinearity, the second and
the sixth. The second mode is the first in-phase bending mode and the sixth mode is the first shearing mode. Shown in Fig.
14.1, the two modes load the contact interface differently; the second begins to slip at the end of the interfaces closest to
the center of the beam due to transverse shears from the bending, and the sixth begins to slip in a more circular pattern as
the joint is loaded by transverse shear forces and torsional forces. In the cases presented here only mode two is considered,
although both modes are being correlated in order to have a thorough validation of the FEM.
To reach agreement between the model and the experimental measurements, two parameters are considered here: The
coefficient of friction between the two clamped members and the curvature of the contact surfaces. The coefficient of friction
is varied over a range of possible values. A cylindrical convex curvature is applied in the direction perpendicular to the length
of the beam. Curvatures with a radius of 1.27 m and 2.54 m are considered here, which produce a difference in height of
about 508 microns (20 mil) and microns 254 (10 mil) across the width of the S4 Beam. This is in the range of the non-flatness
that was measured in the S4 Beam that was tested. Cylindrical curvature was chosen in order to determine how curvature in
different directions effected the response.
Figures 14.2 and 14.3 show two parameter studies compared with experimental data. No linear damping mechanism
was included in the FEM so the magenta line indicates the approximation made for the amount of linear damping in the
experimental system. This value was subtracted from the raw experimental damping curve to calculate the nonlinear damping
curve in green. The approximation for the linear damping value was made by choosing the value that made the nonlinear
curve match the expected power-law form in micro-slip [8]. The bottom plot shows the difference between the linear natural
frequency and that computed at each amplitude, revealing how much the natural frequency of the structure would change as
the modal amplitude increases. Modal velocity has been expressed as the velocity of the point on the structure that has the
largest velocity in the mode in question.
Figure 14.2 shows how the curvature of the interface changes the structure. The flat model distributes the bolt preload
over a larger surface, so it has a lower pressure in the joint, it also shows the most nonlinearity. This agrees with previous
findings in [1] that show that lower joint pressure increases nonlinearity, presumably because there is a lower threshold to
slip. The models tend to over predict the change in damping and natural frequency and go completely into macroslip well
before the experimental data, which does not allow us to compare the models for higher amplitudes.
Fig. 14.1 Joint interface slip contours and mode shapes for mode 2 (left) and mode 6 (right)
14 Model Correlation to a Nonlinear Bolted Structure Using Quasi-Static Modal Analysis 89
Fig. 14.2 Effective natural frequency and damping of flat and curved interfaces compared to experimental data
Fig. 14.3 Various coefficients of friction for a given cylindrical curvature of 1.27 m
Figure 14.3 shows the amplitude dependent damping and natural frequency curves for various coefficients of friction and
a single cylindrical curvature of 1.27 m. The coefficient of friction has less effect than the curvature but driving it higher does
tend to push the predicted damping curve slightly closer to the experimental results.
14.3 Moving Forward
The cylindrically curved surfaces used here were implemented in order to obtain insight into how sensitive the structure is to
surface curvature. The actual surface is more spherical or elliptical, or may even have islands that deviate from the otherwise
flat or curved surfaces by 10’s to 100’s of microns (up to a few thousandths of an inch). Findings emphasize the need to
accurately predict the preload and joint pressure distribution. To simulate the real surface, precise surface roughness and
surface topology measurements are being acquired and will be used to update the FEM. To make a definitive conclusion
on the ability of this FEM to replicate the measured behavior, subsequent results will use this more accurate surface
representation of the structure.
90 M. Wall et al.
Acknowledgements This material is based in part upon work supported by the National Science Foundation under Grant Number CMMI-
1561810. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily
reflect the views of the National Science Foundation.
References
1. Wall, M., Allen, M.S., Zare, I.: Predicting S4 beam joint nonlinearity using quasi-static modal analysis. In: Presented at the 37th International
Modal Analysis Conference (IMAC XXXVII) (2019)
2. Jewell, E., Allen, M.S., Zare, I., Wall, M.: Application of quasi-static modal analysis to a finite element model with experimental correlation. J.
Sound Vib. 479, 115376 (2020)
3. Festjens, H., Chevallier, G., Dion, J.-L.: A numerical tool for the design of assembled structures under dynamic loads. Int. J. Mech. Sci. 75,
170–177 (2013)
5. Masing, G.: Eigenspannungen und verfestigung beim messing (self stretching and hardening for brass). Presented at the 2nd Int. Congress for
Applied Mechanics, pp. 332–335, 1926
6. Lacayo, R.M., Allen, M.S.: Updating structural models containing nonlinear Iwan joints using quasi-static modal analysis. Mech. Syst. Signal
Process. 118(1), 133–157 (2019)
7. Singh, A., et al.: Experimental characterization of a new benchmark structure for prediction of damping nonlinearity. Presented at the 36th
International Modal Analysis Conference (IMAC XXXVI), 2018
8. Segalman, D.J., et al.: Handbook on Dynamics of Jointed Structures. Sandia National Laboratories, Albuquerque (2009)
Chapter 15
Numerical Continuation of Periodic Orbits for Harmonically
Forced Nonlinear Systems with Iwan Joints
Seyed Iman Zare Estakhraji, Matthew S. Allen, and Drithi Shetty
Abstract A common numerical model for bolted or riveted joints is the Iwan model, which uses a number of discrete Jenkins
elements to capture the hysteretic behaviour of the system. Previously, the Iwan model has been primarily implemented
within time-domain simulations using the Newmark integrator. However, in many applications it is desirable to predict the
nonlinear Frequency Response Functions (FRFs) of a structure that contains joints. In order to do so, the steady-state response
must be estimated over a range of frequencies, and it is very time consuming to simply compute the time response until
steady-state is reached. Furthermore, the implicit nature of the state variables (i.e. the appearance of “hidden” state variables
in the form of slider displacements) makes it non-trivial to use continuation to compute the frequency response using already
established techniques such as the shooting method. This paper presents a novel method to numerically compute the non-
linear FRFs of a single degree-of-freedom (SDOF) system with an Iwan element. The maximum displacement over the
response period is included as a state variable, along with the initial displacement and velocity, and we demonstrate that
the position of all sliders can be calculated using these states. The shooting method is modified to account for the added
state variable. The method has been tested by computing the FRFs of a SDOF system containing of an Iwan element at
multiple force amplitudes. The results show that the proposed method is able to compute the steady-state response even at
large force amplitudes, when the system behaves quite non-linearly.
Keywords Iwan model · FRF · Newton-Raphson · Pseudo-arclength continuation · Periodic orbits · Steady-state
response
15.1 Introduction
Built-up structures are typically modeled using linear solvers with springs approximating the joints and using proportional
damping to account for the energy dissipation. While this may be an acceptable approximation for certain practical cases, a
more accurate estimation of the damping is essential in many applications. Gaul and Lenz [1] showed, through experimental
and numerical analyses, that there is significant non-linear influence of bolted joints on the damping and frequency
characteristics of a system. Mechanical fasteners, unlike welded joints, allow some play between the contact surfaces along
the direction tangential to the surface. This results in frictional energy dissipation and a change in stiffness. Both the stiffness
and energy dissipation also show amplitude dependent behavior (i.e. nonlinearity). At low force amplitudes, only the edges of
the contact interface slip while a majority of the joint surface remains stuck. This is called micro-slip. As the force amplitude
increases, the contact patch shrinks until there is relative motion between the surfaces, referred to as macro-slip.
Modeling the above-explained interface interactions can be very computationally expensive, as observed by Jewell et al.
[2]. As an alternative, lumped, hysteretic models are used, that simulate the non-linear dynamics of the system to match what
is experimentally observed. One such hysteretic model is the Iwan model [3], initially introduced for metal elasto-plasticity,
it has also been shown to be effective in capturing joint behavior [4]. It comprises of a parallel system of spring-slider units
known as Jenkins elements. The most commonly used adaptation of the Iwan model is Segalman’s four parameter model [5],
with the four parameters accounting for the joint stiffness, the macro-slip force, the transition to macro-slip, and the power
law energy dissipation that many joints have been found to exhibit in micro-slip.
S. I. Z. Estakhraji () · M. S. Allen · D. Shetty

University of Wisconsin – Madison, Madison, UW, USA
e-mail: [email protected]; [email protected]; [email protected]; [email protected]

92 S. I. Z. Estakhraji et al.
An important quantitative measure of the dynamic behavior of a system is the Frequency Response Function (FRF).
FRFs give information about the frequencies at which a system exhibits resonant behavior. This must be accounted for
when designing structures in order to avoid damage and failure. While the FRF of a linear system can be easily computed
using closed-form expressions, obtaining the FRF of a non-linear system is not a trivial task. One approach to do so is
the harmonic balance method [6], which approximates an analytical solution to the equation of motion, i.e. a non-linear
differential equation. However, in order to use the harmonic balance method, one must be able to write the restoring
force as a nonlinear function of the states. This cannot be done with a Jenkins or Coulomb friction element, as the force
depends not only on the state but also on the past history of the states, and in a discontinuous way. An Iwan joint typically
contains 30–100 Coulomb slider elements, compounding the difficulty if an approach such as regularization is used [7]. It is,
however, interesting to note that Segalman [5] essentially used a one-term harmonic balance approach to derive closed-form
expressions for the effective forces in an Iwan element as a function of vibration amplitude, but these are only applicable in
the micro-slip regime. Beyond the harmonic-balance type methods, another approach is to use numerical integration coupled
with a shooting algorithm to find periodic solutions and a continuation method to find a prediction for the next solution along
a branch. Peeters et al. [8] presented an efficient continuation scheme, based on shooting and pseudo-arc length continuation,
but their approach was specialized to undamped, unforced systems. Sracic and Allen [9] implemented a similar approach, but
using finite differences rather than analytical gradients, and elaborated on how this could be used to compute the nonlinear
frequency response of a damped, harmonically forced non-linear system. This paper adapts the method developed in [9] to
be applicable to a non-linear system containing Iwan elements. To do so, the slider positions must be calculated in order to
correctly estimate the steady-state displacement and velocity. In this work we propose to address this by adding the maximum
imposed displacement as one of the state variables, as will be elaborated subsequently.
The following section provides an overview of Iwan elements and the shooting method, followed by an explanation of the
algorithm used to numerically develop the FRF for a single DOF system consisting of a spring, mass and an Iwan element.
The results of testing the proposed method on a model SDOF system containing an Iwan element are presented, with FRFs
being obtained for different force amplitudes. The results show that the method performs well even when the non-linearity
is significant.
15.2 Theoretical Background
This section will review the theoretical framework surrounding the dynamics associated with bolted joints and the general
continuation procedure presented in [8].
15.2.1 Overview of the Iwan Model
The Iwan model is commonly used to represent the interface interactions in bolted joints. An Iwan element is a one-
dimensional, parallel arrangement of Jenkins elements. Each Jenkins element consists of a linear spring in series with a
friction slider. For a SDOF system with a single Iwan element, the equation of motion, given by Eq. 15.1, consists of linear
terms and a non-linear force term, Fnl .
mẍ + Clin ẋ + Klin x + Fnl (x, t, y(t, φ)) = Fext (15.1)
where x(t) is the system displacement, y(t, φ) is the current displacement of sliders, Clin is the linear damping co-efficient,
Klin is the linear stiffness, and Fnl is the non-linear force due to the Iwan joint, given by Eq. 15.2, which was derived in [3].
∞
Fnl (x, t, y) = kρ(φ)[x(t) − y(t, φ)]dφ (15.2)
0
In the above equation, k is the stiffness common to all Jenkins elements, and ρ(φ) is the population density of the sliders
of strength φ. A slider starts to slip when the system displacement is greater than the strength φ of that slider. The population
density plays an important role in the analysis of an Iwan joint and specifies the state of sliders based on the position,
velocity and strength of the sliders. The equation of motion, Eq. 15.1, is a non-linear second order differential equation.
The response is history dependent and hence, the differential equation is solved using implicit integration schemes, for
15 Numerical Continuation of Periodic Orbits for Harmonically Forced Nonlinear Systems with Iwan Joints 93
example the Newmark-β method, explained in [10]. As seen in Eq. 15.2, the distribution of sliders is theoretically continuous.
However, to numerically evaluate the differential equation, a discrete number of sliders are assumed, with 30–100 sliders
typically being a good approximation.
The objective of this paper is to obtain the FRF of a SDOF system consisting of an Iwan element type non-linearity. The
FRF, by definition, is a transfer function that models the steady-state output of a system to a harmonic input. The equation
of motion of a vibrating system is a second order differential equation and hence, the steady-state output is equal to the
particular solution of this differential equation. The FRF for linear systems can be obtained analytically using the method of
undetermined coefficients according to which the particular solution corresponding to a harmonic excitation is also harmonic
with it’s frequency equal to the excitation frequency. However, Eq. 15.1 is non-linear and does not obey the same relation.
As one can imagine, integrating the non-linear ordinary differential equation until steady-state is reached, and doing so for
varying harmonic loads of increasing frequencies, would be extremely computationally expensive and hence impractical.
A possible alternative is to use a continuation method to numerically calculate the steady-state displacement and velocity of
the non-linear system for a harmonic force input. The FRF can then be obtained by plotting the amplitude of steady-state
response normalized to the force amplitude versus the corresponding forcing frequency. Section 15.2.2 briefly explains the
general continuation procedure that has been derived in [8, 9]. What makes the non-linearity in the Iwan model different
and more complicated is the added state variables, i.e. the displacements of the sliders. In order for the response of the Iwan
model to a harmonic excitation to reach steady-state, each slider must also show steady-state behavior, yet each slips at a
different load, given by the distribution function, ρ(φ). Thus, the continuation procedure must be modified to include the
state of the sliders. Adding the state of each of the 100 sliders would make the computations cumbersome. Fortunately,
it was found that this can be avoided for an Iwan joint by instead using the maximum response displacement as the third
state variable. Then, as will be further elaborated in Sect. 15.3, the position of all sliders can be calculated if this maximum
displacement is known.
15.2.2 General Continuation Procedure
Numerical continuation can be used to obtain the FRFs of a nonlinear system. To do so, a shooting function is defined
to solve a boundary value problem at a particular forcing frequency, which is then coupled with a continuation method to
march ahead in frequency. For a response to be considered steady-state, the displacement and velocity must be periodic.
Theoretically, the difference between the response after a period of oscillation and at the beginning of the oscillation must be
zero. This concept forms the backbone of the shooting function, H (z0 , t, T ), which is defined to quantify whether a certain
set of initial conditions, (z0 , T ), produce a periodic response. Equation 15.3 gives the expression for the shooting function,
H (z0 , T ) = zT (z0 , T ) − z0 (15.3)
where zT = z(z0 , t = T ) is the state vector found by integrating with initial condition z0 over one period, i.e. until t = T
and T is the time period of the force, which is equal to the reciprocal of the forcing frequency. Ideally, the shooting function
must equal zero. However, in practice, a convergence criterion is implemented, as shown in Eq. 15.4,
||H (z0 , T )||

< (15.4)
||z0 ||
where is a user-defined small value. For the cases presented here, = 1×10−6 was found to provide good results. A starting
guess needs to be provided to begin the continuation procedure. A frequency value well below the natural frequency of the
system can be chosen as the starting point, and the FRF for the linearized (i.e. low amplitude) system at that frequency can be
used as the starting guess. The equation of motion is then integrated over the assumed period and convergence is evaluated
using Eq. 15.4. When the convergence criterion is not satisfied, a Newton-Raphson correction method (described in [9] as
NRCM1) is used to update the initial conditions, z0 and the period, T . The equations for the Newton-Raphson correction
method, referred to in [9] as NCRM1, are not repeated here for brevity; they can be found in [9]. It must be noted here that
at each iteration, the period of the external force must also be updated to match the corrected period.
Once an initial solution is obtained, the algorithm must march forward in frequency in order to estimate the entire FRF. To
do so, a pseudo-arclength continuation method [11] is employed that consists of a prediction step followed by a correction
step. The predictor step involves calculating a tangent vector to the solution branch using Eq. 15.5,

{0}
∂z0 |z0,(j ) ,T(j ) ∂T |z0,(j ) ,T(j )
∂H ∂H
{P(j ) } = (15.5a)
0
{P(j ) } = [Ps,(j
T
) PT ,(j ) ]T (15.5b)
where {P } is the desired tangent vector. Then, the first guess for the initial conditions of the next periodic solution is
calculated using Eqs. 15.6 and 15.7.
z0,pr,(j +1) = z0,j + sj Pz,(j ) (15.6)
Tpr,(j +1) = Tj + sj PT ,(j ) (15.7)
In the above equations, z0,pr,(j +1) and Tpr,(j +1) are the predictions for the (j + 1)th solution and sj is the j th step size.
While a constant small step size can be applied, a more computationally efficient approach is to implement an adaptive step
size (as used in [8]). That has been used here; the first step size is set by the user and subsequent step sizes are calculated
using Eq. 15.8.
K∗
sj = sign(sj −1 {Pj }T {Pj −1 }) |sj −1 | (15.8)
Kj −1
where Kj −1 is the integer number of iterates that were required to update the previous shooting function solution and K ∗ is
an optimal number of shooting iterates, supplied by the user.
The correction step is then used to obtain the steady state solution near this prediction. The Newton-Raphson method
used earlier with the shooting function is now modified to enforce the condition that the correction must be orthogonal to the
tangent vector, as expressed in Eq. 15.9.
! ! ⎧−H (z ⎫
(k)
z0,(j +1) ⎨ , T(j ) )⎬
∂z0 |z0,(j ) ,T(j ) ∂T |z0,(j ) ,T(j )
∂H ∂H 0,(j )
= 0 (15.9)
T
Ps,(j ) P T ,(j ) T(j(k)+1) ⎩ ⎭
0
This two-step continuation procedure is implemented until all of the solutions in the frequency range of interest have
been computed. The amplitude of steady-state response normalized to the force amplitude versus the corresponding forcing
frequency then gives the nonlinear FRF.
15.3 Modified Continuation Method for the Iwan Model
The general continuation procedure explained in Sect. 15.2.2 can be adapted to estimate the periodic solution of a system
consisting of an Iwan element. Time integration over one period of oscillation is performed to calculate x(T ), ẋ(T ) for a
particular initial guess x(0), ẋ(0), which form the input to the shooting function. The Newmark-beta integration scheme is
used and the non-linear force term in the equation of motion, Eq. 15.1, is calculated as a function of the current displacement
and slider position. When the system begins at rest, i.e. initial conditions are zero, the displacements of the sliders are all
zero. However, to accurately determine the non-linear force for a non-zero initial displacement and/or velocity, the initial
slider positions must also be provided. Section 15.3.1 explains the variables needed in order to determine the state of all of
the sliders at any instant and Sect. 15.3.2 gives the modifications required in the general continuation procedure in order to
incorporate the slider states.
15.3.1 State of Sliders
Figure 15.1 shows the steady-state behavior of two sliders over four periods for a harmonically forced SDOF system with
an Iwan joint, which is modeled with ten Jenkins elements. The time histories of two of the sliders are shown, the first slider
and last slider, which correspond to the weakest (first to slip) and strongest sliders respectively. The first slider shows nearly
10 First slider
y 1(t)
0
-10
0 0.5 1 1.5 2 2.5 3 3.5 4
Time/Period
10
Force at the joint
Force(N)
-10
0.5 1 1.5 2 2.5 3 3.5 4
Time/Period
2
Last slider
y 100(t)
-2
0.5 1 1.5 2 2.5 3 3.5 4
Time/Period
Fig. 15.1 Periodic behavior of the sliders and the force of the joint
continuous slip over a cycle of vibration whereas the last slider sticks and slips periodically. Each slider begins the cycle
at a different displacement. In order to implement the shooting method in a basic sense, one would need to add all of the
slider displacements as state variables (i.e. in the shooting function). That would be extremely expensive because often 100
to 250 sliders are needed to accurately model an Iwan joint in micro-slip, which means that many extra state variables would
be added to the problem, and the Jacobian of the shooting function would need to be computed with respect to all of these
variables.
In order to explain the idea behind the algorithm introduced in this paper (and which is shown in Fig. 15.3), it is necessary
to understand how the states of the sliders evolve over a typical vibration cycle. Therefore, Fig. 15.2 which shows the state
of the sliders in six different cases, which are discussed in this section. In Fig. 15.2, the red vertical line represents the Iwan
joint and the sliders are shown with small circles connected to springs. Slipping sliders are marked with horizontal red lines
while the stuck sliders are shown with black lines. When a Jenkins element slips, the point at which it is attached to ground
moves or the slider position yi (t) is said to change; the points yi (t) are shown by dots connected by a blue vertical line in
Fig. 15.2. The red vertical line is the instantaneous position of the whole Iwan element x and the yellow and purple lines
show the maximum extent of the displacement over the periodic response.
The point at which each Jenkins element slips can be specified using either the force, fi , as done in [3], or based on the
maximum extension of that slider’s spring. The maximum extension of the ith spring can be defined as [5]:
fi
φi = (15.10)
ki
where fi and ki are the maximum allowable force in and the stiffness of the ith Jenkins element respectively. Therefore,
a slider starts to slip when the displacement is larger than φ of that element. In all cases shown in Fig. 15.2, the 9th and
10th sliders do not slip. since their respective strengths, φ are greater than the maximum system displacement, xmax , i.e.
φ9 , φ10 > xmax . These sliders have been dubbed as “stuck at all times” and their features will be explained below.
For the time t=0.28377 T and =0.99128 For the time t=0.39413 T and =0.99128
n n
12 12
10 10
Joint Number
Joint Number
8 8
6 6
4 4
2 2
0 0
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
Displacement Displacement
(a) When 0 < x < xmax and ẋ > 0, the 1th to 7th sliders are (b) When x = xmax with positive velocity ẋ > 0. The 1th to
sliding, the 8th slider is stuck at some initial location. The 9th and 8th sliders are sliding. The sliders 9th and 10th remain stuck at
10th sliders are stuck at equilibrium. equilibrium.
For the time t=0.39415 T and =0.99128 n
12 12
10 10
Joint Number
Joint Number
8 8
6 6
4 4
2 2
0 0
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
(c) At the turning point, where x = xmax and, ẋ < 0. The 1st to (d) As the system starts to move in the negative x direction, the 1st
8th sliders are now stuck at (xmax − φ) and the 9th and 10th slider slider (weakest slider) begins to slip, while the 2nd to 8th sliders
remain stuck at equilibrium. remain stuck at (xmax − φ).
12 12
10 10
Joint Number
Joint Number
8 8
6 6
4 4
2 2
0 0
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
(e) Additional sliders slip the displacement progresses. At the in- (f) When the displacement reaches xmin , its largest negative value,
stant shown the 1st to 4th sliders are slipping and the 5th to 8th the velocity again goes to zero and now the 1st to 8th sliders are
sliders are still stuck at (xmax − φ). stuck at (xmin + φ).
Fig. 15.2 State of Iwan sliders at different time instants during a cycle of vibration. (a) When 0 < x < xmax and ẋ > 0, the 1th to 7th sliders are
sliding, the 8th slider is stuck at some initial location. The 9th and 10th sliders are stuck at equilibrium. (b) When x = xmax with positive velocity
ẋ > 0. The 1th to 8th sliders are sliding. The sliders 9th and 10th remain stuck at equilibrium. (c) At the turning point, where x = xmax and, ẋ < 0.
The 1st to 8th sliders are now stuck at (xmax − φ) and the 9th and 10th slider remain stuck at equilibrium. (d) As the system starts to move in the
negative x direction, the 1st slider (weakest slider) begins to slip, while the 2nd to 8th sliders remain stuck at (xmax − φ). (e) Additional sliders slip
the displacement progresses. At the instant shown the 1st to 4th sliders are slipping and the 5th to 8th sliders are still stuck at (xmax − φ). (f) When
the displacement reaches xmin , its largest negative value, the velocity again goes to zero and now the 1st to 8th sliders are stuck at (xmin + φ)
Stuck at All Time Instants
This is the case for the slider number i if φi > xmax as explained above. So, in this case the slider position would be zero
all the time (i.e. yi = 0). One can imagine that the normal force in these Jenkins elements is large enough so that they can
stretch to xmax without slipping.
For the case of other eight slipping sliders, the maximum extension that can be supported by each of their springs is less
than xmax . Therefore, the sliders will begin to slip at some point during the motion. Since slipping happens after the spring
has extended to its maximum value φ, it can be concluded that the slider is following the joint, with an offset in displacement
equal to φ from the joint, as is shown in Fig. 15.2a, and with velocity equal to that of the joint. The value of φ depends on
the distribution function selected for the joint. In order to represent the power-law type dissipation observed in bolted joints
in the micro-slip region, a power law distribution function was chosen in [5], given by Eq. 15.11,
ρ(φ) = Rφ χ [H (φ) − H (φ − φmax )] + Sδ(φ − φmax ), (15.11)
where φmax is the displacement at which all sliders slip, S is the slope of the force-displacement curve just before macro-slip,
and R and χ define the strength and slope of the power-law energy dissipation.
There are some moments, here called “reversal moments,” in which the joint changes its direction (i.e. the velocity goes
to zero). The position of the joint in those moments will be referred to here as a “reversal point”. The reversal points occur
at the peaks in the displacement versus time, and that they dictate the state of sliders as shown in Fig. 15.2c. As can be seen
in Fig. 15.2b, f, the sliders continue to slip until the joint reaches xmax or xmin . A moment after that, the sliders are stuck
because of the fact that the joint has changed its direction and starts to compress the spring as is shown in Fig. 15.2c. At this
instant, Fig. 15.2c, all the sliders are stuck. The sliders that never slipped are at the equilibrium position while other sliders
are at a distance of φ from the joint. Hence, the position of all of the sliders can be determined by using xmax and φ at this
moment. This is the main idea that the algorithm presented in this work uses, in order to use the peak displacement at the
reversal points as a state variable.
At the instant that the joint has reached to xmax and is changing its direction, the springs of slipped sliders are already
extended to their max value φ. At this moment, the joint starts to compress the sliders in the other direction as is shown in
Fig. 15.2d. The joint has to move by a value of φ to compress the springs to their equilibrium position while the sliders are
still stuck. Then, the joint needs to move by an extra value of φ to extend the springs in order to make the sliders slip as is
shown in Fig. 15.2e which means the joint has to move 2φ to make the sliders slip. That shows the importance of 2φ as a
criteria to evaluation of initial state of sliders. In this way, two different states have been defined for the joint as “When ||x0 ||
belongs to ||xmax − 2φi ||” and “When ||x0 || doesn’t belong to ||xmax − 2φi ||” which are shown in Fig. 15.3 and also are
explained below:
When ||x0 || Belongs to ||xmax − 2φi ||
The sliders that have φi < xmax slip during the integration period, but there are some intervals in which they stick. Simply
put, the slider sticks each time the velocity reverses (or |x(t)| is maximum). So, it can be shown that if x0 and ẋ0 have
different signs and also ||x0 || > ||xmax − 2φi ||, for a certain slider, then it should be stuck. This is the case for the 2nd to 8th
sliders shown in Fig. 15.2d. Otherwise, if x0 and ẋ0 have the same sign, the slider should be slipping and so its velocity will
match that of the joint as is the case for the 1st to 7th sliders in Fig. 15.2a.
When ||x0 || Doesn’t Belong to ||xmax − 2φi ||
Any slider for which φi < xmax , if ||x0 || < ||xmax − 2φi || has exceeded its slip threshold, and so they slip and the velocity
of the slider ẏi (t) = ẋ(t). These sliders are following the joint and keeping a fixed distance of φi from the joint, as shown in
Fig. 15.2e for the 1st to 4th sliders.
Thus, by adding the maximum displacement over the period of oscillation, xmax , or the reversal point, as a state, one can
determine the locations of all of the sliders yi . This will make it possible to use continuation with the Iwan joint, as explained
in more detail in Sect. 15.3.2.
15.3.2 Adding the State of Sliders to the Continuation Method

The basic shooting method works by assigning an initial condition z0 = x0 ẋ0 , integrating the system over one period
(using the Newmark method for example) to obtain x(T ) and ẋ(T ), and then comparing those with x0 and ẋ0 to check
periodicity. If the convergence criterion is not satisfied, then Newton’s method is used to find an update to the initial
conditions. In the proposed algorithm, xmax is added as a state variable and so we must also develop a means for correcting
xmax in each iteration and also predicting it for the next point along the FRF branch. Once the initial displacement and xmax
are known, one can find the initial state of each slider, as explained in the schematic in Fig. 15.3. Then one can compute the
initial nonlinear force fj0 in the Iwan joint and also the initial acceleration and supply these to the Newmark integrator as an
initial guess.

ẍ0 = M −1 Fext0 − fj0 − C x˙0 − Kx0 (15.12)
Once the additional state variable, xmax , is added to the state vector, the initial conditions used in the shooting function
are defined as:
⎡ ⎤
x0
z0 = ⎣ ẋ0 ⎦ (15.13)
xmax
To accommodate this additional state variable, the Jacobian matrices are modified. The variables used in Newton-Raphson
method will then be x0 , ẋ0 , T , xmax , and so the correction equation is now,
Guess and Use Iwan_static Make
Initializing
Yes No Yes
No
No Yes
Fig. 15.3 The algorithm used in this paper to find the initial state of sliders
⎡ ∂x ⎤⎡ x0
⎤
⎡ ⎤
∂x − 1
∂x ∂x ∂x
∂ ẋ0 ∂xmax ∂T ⎢ ⎥ x0 (T , x0 , ẋ0 , xmax ) − x0
⎢ 0∂ ẋ ∂ ẋ ⎥ ⎢ ẋ0 ⎥ ⎣
⎣ ∂x0 ∂ ẋ0 − 1 ∂xmax
∂ ẋ ∂ ẋ
∂T ⎦ ⎣ = ẋ0 (T , x0 , ẋ0 , xmax ) − ẋ0 ⎦ . (15.14)
xmax ⎦
∂xmax
∂x0
∂xmax ∂xmax
∂ ẋ0
∂xmax
∂xmax − 1 ∂T
xmax (T , x0 , ẋ0 , xmax ) − xmax
T
Based on Eq. 15.14, all of the state variables have an effect on the response and the periodicity. As xmax is changed, it
changes the initial state of the sliders, which affects the non-linear force and subsequently, the system acceleration at each
time instant, leading to changes in x(t), ẋ(t). Similar to [9], the Finite Difference Method has been used here to calculate the
Jacobian used in Eq. 15.14. For the Jacobian with respect to the state variables, a pre-defined step-size, z = 1 × 10−3 is used
and the equation of motion is solved for the incremented initial conditions, z0 + z . The perturbed state-vector obtained, zp
is then used in the Finite Difference equation, given as
∂H z p − z0
= (15.15)
∂z z
A similar procedure is followed to obtain the Jacobian with respect to the period T with two changes – a smaller step-size,
T = 1 × 10−8 is chosen, and the time period of the harmonic force is also increased by T when the period is perturbed
by T . Using a smaller step size increases the computational cost, but was found to make the finite difference calculations
more robust.
Once a solution has been found, the pseudo-arclength method can be used to find a prediction for the next solution using
a tangent to the curve. Mathematically, we define a prediction P and force it to be perpendicular to the Jacobian in Eq. 15.14
as follows:
⎡ ∂x ⎤⎡Px0
⎤
⎡ ⎤
−1 ∂x ∂x ∂x
0
⎢
∂x0 ∂ ẋ0 ∂xmax ∂T
⎥ ⎢ ⎥
⎢ Pẋ0 ⎥ = ⎣0⎦
∂ ẋ0 − 1
∂ ẋ ∂ ẋ ∂ ẋ ∂ ẋ
⎣ ∂x0 ∂xmax ∂T ⎦ ⎣P ⎦ (15.16)
∂xmax ∂xmax ∂xmax ∂xmax xmax
∂xmax − 1 ∂T
∂x0 ∂ ẋ0
0
PT
The shooting function remains the same, as shown below, although the dependence on xmax is made explicit. We also
require xmax to satisfy the shooting function. Rather than combine all of the variables in one shooting function, two separate
shooting functions are used, as given below,
Hz0 (z0 , t, T , xmax ) = zT (z0 , t, T , xmax ) − z0 (15.17a)
Hxmax (z0 , t, T , xmax ) = max(x(t))0≤t≤T − xmax 0 (15.17b)
where zT (z0 , t, T , xmax ) is the state vector found by integrating with initial condition z0 and xmax until t = T . Moreover,
the criteria for considering the solution converged has changed as below:
||Hz0 (z0 , T , xmax )||

< (15.18a)
||z0 ||
||Hxmax (z0 , T , xmax )||

< (15.18b)
||xmax ||
where is considered 10−6 in the results shown in this paper.
15.4 Results and Discussion
In order to test the method, a SDOF system consisting of an Iwan element was considered. The system parameters can be
found in Table 15.1, where Fs , Kt , χ and β are the Iwan parameters as described in [5].
The algorithm described in Sect. 15.3.2 was used to find the periodic orbit for a range of different forcing frequencies, and
the process was repeated for various force amplitudes. While a linear FRF has a shape that is independent of the amplitude
Table 15.1 Linear and Parameter Value

non-linear parameters of the
SDOF system studied Klin (N/m) 9
ζlin 2.78 × 10−5
M(kg) 10
Fs (N ) 10.0
Kt (N/m) 1
χ −0.5
β 3.0
Amplitude = 0.001N
Amplitude = 0.01N
Amplitude = 0.1N
Amplitude = 0.4N
103
Amplitude = 1N
Amplitude = 2N
Linear FRF
|X|/F (s2/kg)
102
101
100
0.156 0.1565 0.157 0.1575 0.158 0.1585 0.159 0.1595 0.16
Frequency (Hz)
Fig. 15.4 Non-linear FRFs for different force amplitudes compared against the linear FRF
of force applied, the shape of the non-linear FRFs of the Iwan model shown in Fig. 15.4 are observed to change as the
amplitude, and hence the amount of non-linearity, increases.
The lowering of the peaks as the force amplitude increases is indicative of the increase in the effective damping due to the
Iwan element. This is expected as the number of sliders that slip increases as the amplitude of the force increases. Hence, the
amount of energy dissipated, and consequently the damping of the system, increases. It can also be observed that the Iwan
element, in the test case presented here, behaves as a softening non-linear spring since the natural frequency of the system
decreases, i.e. the peak shifts to the left, as the force amplitude increases. On comparing it with the linear FRF, represented
by the red dashed line in Fig. 15.4, it can be seen that the Iwan element not only increases the system damping but also
changes the resonant frequency, although only slightly. For example, at a force of amplitude 0.1 N, the peak in the FRF shifts
downward by 0.56%. Even then, if a harmonic force was applied at this amplitude and at the peak frequency of 0.1583 Hz,
the amplitude of vibration would be about five times higher than that predicted by the linear FRF, due to the shift in resonant
frequency.
Figure 15.5 shows the response over two periods for a few points on the FRFs, which are marked with red dots in Fig. 15.4.
The first row shows the response of the joint, x(t), which is normalized by the maximum value of displacement max(x(t)). In
this figure, more phase lag can be observed for the higher load amplitudes. Although there is no visible harmonic distortion,
the number of sliders slipping does change quite a bit for different force amplitudes as is shown in the second row of Fig. 15.5.
It can be seen that higher load amplitudes results in a larger number of sliders slipping. For the case of force amplitude =
1N, the response shows macro-slip with all of the sliders slipping while for lower force amplitude, some sliders remain in
Fig. 15.5 Comparing the displacements, fraction of slipped sliders, non-linear restoring fore and the position of the 75th slider for different force
amplitudes
micro-slip. Moreover, the case shown and explained in Fig. 15.2d can be observed as the sharp vertical lines in the second
row of Fig. 15.5. The sharp vertical lines show the instance right after the joint has changed its direction when only the
weakest sliders have slipped.
The third row of Fig. 15.5 shows the non-linear force in the joint fnl in Eq. 15.1 normalized by FS from Table 15.1. First,
it can be observed that as the force amplitude increases, the non-linear force becomes closer to a “cosine” force, i.e. π/2
phase lag, while the external force is a “sine”. Moreover, as expected, there is a larger non-linear force for the higher force
amplitudes. In the case of Amp = 1 N, the non-linear force is equal to the force required to cause macro-slip, FS while for
Amp = 0.001 N, it is almost two percent of FS (i.e. close to the linear response), and for Amp = 0.1 N, the non-linear force is
about 49% of FS , which case the joint is exhibiting micro-slip.
The 75th slider is chosen as a sample to depict the response of sliders over time and its normalized response
y75 (t)/max(y75 (t)) is shown in the fourth row of Fig. 15.5. The horizontal part of the slider response (more visible for
the case of Amp = 0.001N) is related to the interval in which the slider is stuck when the joint moves 2φ from one reversal
point towards another (i.e. from xmax to xmin or vice versa). In this case, as the force amplitude increases and the response of
10-6
5
Error of initial displacement
Percent Error
4
1
1 2 3 4 5 6 7 8 9
t/T
10-4
0
Percent Error
Error of initial velocity
-1
-2
1 2 3 4 5 6 7 8 9
t/T
Fig. 15.6 The Error of the values of xnT and ẋnT for n = 1, 2, . . . , 9 for the force amplitude of 2 and ω = 0.989ωn
the slider becomes more continuous because the ratio of x2φ max
becomes smaller for the larger force amplitude. Furthermore,
it can be seen that the phase lags for the response of the slider are different for each load case; the phase lag seen for the
response of the joint is a consequence of different phase lags in each of the sliders.
In order to check if the results obtained from the algorithm truly have reached steady state (and hence have perfect
periodicity), consider the point with high force amplitude (i.e. Amp = 2 N) and a frequency close to the natural frequency
(i.e. ω = 0.989ωn ), near the peak of the cyan curve in Fig. 15.4. The error in the values of xnT and ẋnT for n = 1, 2, . . . , 9
compared to the initial values is shown in Fig. 15.6. It can be seen that error is negligible even and does not grow very quickly
even as the response is integrated over several periods.
A slight drawback of the proposed algorithm is its computation time. The run time varies depending on the force amplitude
but on an average, it takes about 17 min to compute a single FRF. At each forcing frequency, the non-linear equation of motion
needs to be integrated multiple times, for the different iterations of initial conditions, until the shooting function satisfies the
convergence criterion. The Newmark-β integration scheme used is computationally expensive, contributing to about 80% of
the total run time. The circular markers on each FRF in Fig. 15.4 indicate the frequencies at which the response amplitude
was computed, based on the size of the step taken by the algorithm. It can be seen that in all of the FRFs, the density of the
markers increases around the peak, implying that the continuation procedure requires a smaller prediction step in order to
follow the bend in the peak, due to the sudden change in the tangent vector in the prediction step. This has been indicated
in the zoom-in on the peak of the FRF corresponding to a force of 2 N in Fig. 15.4. For the same reason, at a higher force
amplitude more time is required to compute the FRF since the distortion of the peak is more prominent.
15.5 Conclusion
This paper demonstrates that the shooting/continuation procedure in [8] can be adapted for a SDOF system that includes an
Iwan element, without having to include all slider positions as state variables. Instead, the maximum response displacement
can be used to find the position of all sliders. This is an interesting feature of Iwan joints that is perhaps not well understood.
The nonlinear Frequency Response Functions obtained for the SDOF system studied here were shown to deviate from
a linear FRF in the expected way, in that it shows an increase in damping and a shift in the resonant frequency as the
forcing amplitude increases. Furthermore, the results obtained are truly periodic solutions to the equation of motion, as by
construction the algorithm verifies this at every step. The approach is somewhat computationally expensive, as the implicit
Newmark integrator must integrate the non-linear equation of motion over the forcing period several times per solution point.
However, the authors are not aware of another algorithm that has succeeded in computing the NLFRFs of a system with an
Iwan joint, and so this approach can provide valuable information. If a faster method can be devised in the future (e.g. using
multi-harmonic balance) then this method could serve as a benchmark against which those approaches can be tested, to verify
that any assumptions made do not degrade the accuracy significantly.
This paper reports only on the application of the algorithm to an SDOF system. The authors are working to use similar
concepts to extend this method to MDOF systems. Efforts are also underway to reduce the computation time, possibly by
improving the prediction and/or correction methods so as to reduce the number of iterations required for convergence.
Acknowledgments This material is based in part upon work supported by the National Science Foundation. Any opinions, findings, and
conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National
Science Foundation.
References
1. Gaul, L., Lenz, J.: Nonlinear dynamics of structures assembled by bolted joints. Acta Mech. 125(1), 169–181 (1997)
2. Jewell, E.A., Allen, M.S., Lacayo, R.: Predicting damping of a cantilever beam with a bolted joint using quasi-static modal analysis. In: Volume
8: 29th Conference on Mechanical Vibration and Noise, page V008T12A019, Cleveland, Aug (2017). ASME
3. Iwan. W.D.: A distributed-element model for hysteresis and its steady-state dynamic response. J. Appl. Mech. 33(4), 893–900 (1966)
4. Segalman, D.J.: Modelling joint friction in structural dynamics. Struct. Control. Health Monit. 13(1), 430–453 (2006)
5. Segalman, D.J.: A four-parameter iwan model for lap-type joints. J. App. Mech. 72(5), 752–760 (2005)
6. Krylov, N.M., Bogoliubov, N.N.: Introduction to non-linear mechanics, vol. 11. Princeton University Press (1949)
7. Krack, M., Gross, J.: Harmonic Balance for Nonlinear Vibration Problems. Mathematical Engineering. Springer International Publishing,
Cham (2019)
8. Peeters, M., Viguié, R., Sérandour, G., Kerschen, G., Golinval, J-C.: Nonlinear normal modes, Part II: toward a practical computation using
numerical continuation techniques. Mech. Syst. Signal Process. 23(1), 195–216 (2009)
9. Sracic, M.W.: Numerical continuation of periodic orbits for harmonically forced nonlinear systems. In: Civil Engineering Topics, vol. 4, pp.
51–69. Springer, New York (2011)
10. Cook, R.D., Malkus, D.S., Plesha, M.E., Witt, R.J.: Concepts and applications of finite element analysis, vol. 4. Wiley, New York (1974)
11. Seydel, R.: Practical bifurcation and stability analysis. In: Interdisciplinary Applied Mathematics, 3 edn. Springer, New York (2010)
Chapter 16
A Digital Absorber for Nonlinear Vibration Mitigation
Ghislain Raze, Sylvain Guichaux, Andy Jadoul, Valery Broun, and Gaetan Kerschen
Abstract In this study, a digital impedance is used to realize both a linear and a nonlinear piezoelectric tuned vibration
absorber in order to mitigate the vibrations of a nonlinear structure. The digital processing unit enables the synthesis
of impedances with arbitrary functional forms, thereby easing the implementation of nonlinear absorbers. The superior
performance of the nonlinear absorber over its linear counterpart is demonstrated experimentally. Various nonlinear
functional forms are also tested in the absorber and illustrate the relevance of the principle of similarity (i.e. the same
nonlinear functional form as that in the host structure should be used in the absorber).
Keywords Piezoelectric vibration absorber · Synthetic impedance · Nonlinear vibrations · Nonlinear vibration
absorber · Principle of similarity
16.1 Introduction
Resonant piezoelectric shunt damping as proposed by Hagood and von Flotow [1] is a popular approach for passive vibration
control. A piezoelectric transducer bonded to a vibrating structure converts part of its mechanical energy into electrical
energy. This energy may in turn be dissipated in a so-called shunt circuit. Efficient vibration mitigation by resonant shunt
circuits is conditioned upon a precise tuning of the shunt circuit to the frequency of the host structure.
Real-world structures exhibit a nonlinear behavior to some extent. This can be detrimental to tuned absorbers, since
the frequency of the host structure may vary with the forcing amplitude and the absorber may get detuned, leading to
high-amplitude vibrations. Agnes and Inman [2] observed that including nonlinearities in the absorber as well may have a
beneficial effect on performance. Habib and Kerschen [3] showed how these nonlinearities may be tuned so as to increase the
range of forcing amplitudes over which the absorber is efficient. A central idea in their approach is the principle of similarity,
i.e., the nonlinear absorber should possess the same nonlinearity type as that of the host structure. Soltani and Kerschen [4]
extended this principle to piezoelectric vibration absorbers using a capacitor with a nonlinear capacitance.
The theoretical developments in [4] were experimentally validated by Lossouarn et al. [5] using saturable inductors instead
of a nonlinear capacitance. By contrast, digital piezoelectric absorbers, as proposed by Fleming et al. [6], are used in this
work. A digital absorber comprises a digital unit that can be programmed to follow arbitrary input-output relations. This unit
is coupled to an analog board containing a current source and electronic circuitry to make the interface between the digital
unit and the piezoelectric transducer. The absorber is connected to the transducer and acts as an impedance from its point of
view. The advantage of using a digital absorber comes from the flexibility provided to synthesize virtually any impedance;
this however comes at the expense of energy consumption.
G. Raze () · G. Kerschen

S. Guichaux · A. Jadoul · V. Broun
Engineering Department, Haute Ecole de la Province de Liège, Liège, Belgium

106 G. Raze et al.
16.2 Piezoelectric Shunt Damping
The structure to be controlled is modelled as a single-degree-of-freedom (SDOF) Duffing oscillator coupled to a piezoelectric
stack connected to a shunt circuit, as shown in Fig. 16.1. The equations relating the displacement x of the SDOF oscillator,
the external forcing f and the piezoelectric charge q read
(
mẍ + cẋ + koc x + k3 x 3 − θ q = f
, (16.1)
Lq̈ + R q̇ + C1ε q + Cn |q|n−1 q − θ x = 0
p
where m, c, k and k3 represent the mass, damping coefficient, open-circuit stiffness and nonlinear stiffness coefficient of
the SDOF oscillator, respectively, θ is a piezoelectric coupling coefficient, L, R and Cn are the inductance, resistance
and nonlinear capacitance coefficient (Cn = 0 for a linear shunt circuit) of the shunt circuit, respectively, and Cpε is the
piezoelectric capacitance at constant strain. The values of L and R may be tuned such that the frequency response of the
underlying linear system exhibits two peaks of equal amplitude in place of the original high-amplitude resonance [7], as
depicted in Fig. 16.1. When the structure behaves nonlinearly, detuning of the linear absorber may occur. The purposeful use
of nonlinearities in the absorber can mitigate this undesirable phenomenon. In line with [3], the value of n should be chosen
according to the principle of similarity, i.e. n = 3. The value of the coefficient Cn may then be found from [4].
The host structure considered for the experiments is a cantilever beam covered with two arrays of piezoelectric patches.
A clamped thin lamina is attached to its free end, and is responsible for an overall hardening behavior. The structure can
accurately be modelled as a Duffing oscillator when considering its first bending mode [5]. A digital absorber is connected
to the structure through the patches in order to mitigate its vibrations. Figure 16.2 compares the performance of the absorber
when a linear or a nonlinear (with cubic capacitance) shunt circuit is synthesized. The detuning phenomenon described in the
previous section is clearly visible for the linear absorber: the rightmost peak grows in a high-amplitude response. Conversely,
the nonlinear absorber maintains its performance (i.e., equal peaks are still observed) over the tested forcing amplitude range.
Owing to the flexibility provided by the digital absorber, other mathematical functions can easily be synthesized. To
experimentally test the abovementioned principle of similarity, quadratic (n = 2) and quintic (n = 5) capacitances were
considered in the absorber. The corresponding frequency responses are shown in Fig. 16.3. Clearly, the amplitude-dependent
character of these responses illustrates the relevance of the principle of similarity.
(a) (b)
Fig. 16.1 Schematics of the system (a) and frequency response function of the underlying linear system (b) when the piezoelectric transducer is
short-circuited (▬ ), open-circuited (▬ ) and connected to an optimally tuned RL shunt circuit (▬ )
16 A Digital Absorber for Nonlinear Vibration Mitigation 107
(a) (b)
Fig. 16.2 Experimental frequency response of the structure controlled by a linear (a) and cubic (b) digital absorber under an external forcing
amplitude of 0.2 N (▬ ), 0.4 N (▬ ), 0.6 N (▬ ) and 0.8 N (▬ )
(a) (b)
Fig. 16.3 Experimental frequency response of the structure controlled by a quadratic (a) and quintic (b) digital absorber under an external forcing
amplitude of 0.2 N (▬ ), 0.4 N (▬ ), 0.6 N (▬ ), 0.8 N (▬ ) and 1 N (▬ )
16.4 Conclusion
The digital impedance eases the implementation of various types of piezoelectric absorbers. Taking advantage of this
flexibility, this work compared experimentally the vibration mitigation performance on a nonlinear host structure of a linear
and a nonlinear piezoelectric shunt circuit. The superiority of the latter over the former was demonstrated. The significance
of the principle of similarity was also validated.
Acknowledgements G. Raze and G. Kerschen would like to acknowledge the financial support of the SPW (WALInnov grant 1610122).
References
1. Hagood, N.W., von Flotow, A.: Damping of structural vibrations with piezoelectric materials and passive electrical networks. J. Sound Vib.
146(2), 243–268 (1991)
2. Agnes, G.S., Inman, D.J.: Nonlinear piezoelectric vibration absorbers. Smart Mater. Struct. 5(5), 704–714 (1996)
3. Habib, G., Kerschen, G.: A principle of similarity for nonlinear vibration absorbers. Phys. D Nonlinear Phenom. 332, 1–8 (2016)
4. Soltani, P., Kerschen, G.: The nonlinear piezoelectric tuned vibration absorber. Smart Mater. Struct. 24(7), 075015 (2015)
5. Lossouarn, B., Deü, J.-F., Kerschen, G.: A fully passive nonlinear piezoelectric vibration absorber. Philos. Trans. R. Soc. A Math. Phys. Eng.
Sci. 376(2127), 20170142 (2018)
108 G. Raze et al.
6. Fleming, A.J., Behrens, S., Moheimani, S.O.R.: Synthetic impedance for implementation of piezoelectric shunt-damping circuits. Electron. Lett.
36(18), 1525 (2000)
7. Soltani, P., Kerschen, G., Tondreau, G., Deraemaeker, A.: Piezoelectric vibration damping using resonant shunt circuits: an exact solution. Smart
Mater. Struct. 23(12), 125014 (2014)
Chapter 17
Control-Based Continuation of Nonlinear Structures Using
Adaptive Filtering
Gaëtan Abeloos, Ludovic Renson, Christophe Collette, and Gaetan Kerschen
Abstract Control-Based Continuation uses feedback control to follow stable and unstable branches of periodic orbits of a
nonlinear system without the need for advanced post-processing of experimental data. CBC relies on an iterative scheme
to modify the harmonic content of the control reference and obtain a non-invasive control signal. This scheme currently
requires to wait for the experiment to settle down to steady-state and hence runs offline (i.e. at a much lower frequency
than the feedback controller). This paper proposes to replace this conventional iterative scheme by adaptive filters. Adaptive
filters can directly synthesize either the excitation or the control reference adequately and can operate online (i.e. at the
same frequency as the feedback controller). This novel approach is found to significantly accelerate convergence to non-
invasive steady-state responses to the extend that the structure response can be characterized in a nearly-continuous amplitude
sweep. Furthermore, the stability of the controller does not appear to be affected.
Keywords Control-based continuation · Active filtering · Non-invasive control · Nonlinear vibrations · Experimental
characterization
17.1 State of the Art of Control-Based Continuation
The characterization of a nonlinear structure consists in determining its behavior under harmonic excitation. Nonlinear
structures can reach different periodic orbits under identical excitation, each one with its own response amplitude and
stability. Some of the responses are unstable and cannot be observed experimentally without diverging towards another
periodic orbit. Control-Based Continuation (CBC) is an experimental method that stabilizes the structure using feedback
control to reach these unstable response branches [1, 2]. Characterizing the unstable branches is useful to uncover potential
hidden branches that, even though stable, cannot be reached by performing uncontrolled frequency sweeps [3].
Nonlinear structures respond to monoharmonic excitation with multiharmonic responses. These multiple harmonics are
fed back through the controller into the excitation signal, making it multiharmonic. As such, the response of the structure is
different than under monoharmonic excitation and the control is considered invasive [4]. An important part of experimental
continuation procedures is to ensure that the controller is non-invasive and that the total excitation signal (control action
included) is monoharmonic, i.e. that its non-fundamental harmonics are canceled.
One way to render the controller non-invasive is to ensure that the non-fundamental harmonics of the target x ∗ cancel the
ones of x. In this way, the difference d = x ∗ − x fed into the controller is monoharmonic and so is the controller output:
the excitation is monoharmonic and the controller is non-invasive. Figure 17.1 shows the current implementation of CBC
[4–6]. The method consists in two loops. One online feedback loop with a controller generating the excitation signal u from
G. Abeloos () · G. Kerschen

C. Collette
Department of Bio, Electro and Mechanical Systems, Université libre de Bruxelles, Brussels, Belgium
L. Renson

110 G. Abeloos et al.
Fig. 17.1 Implementation of

Control-Based Continuation with
the online feedback loop and the
offline iteration loop: the
excitation u is the output of the
controller being fed with the
difference d between the target
response x ∗ and the measured
response x; the target comprises a
monoharmonic part x1∗
determined by the user and
non-fundamental harmonics x= ∗
1
resulting from Picard iterations
on x
(a) (b)
Fig. 17.2 Two different architectures using adaptive filters to remove the offline iteration loop, giving fully online systems; the diagonal arrows
behind the filters H1 and H=1 mark the reference signals. (a) The adaptive filter H1 synthesizes a monoharmonic excitation signal f using the
multiharmonic controller output u as a reference; the input q1 contains the fundamental harmonic. (b) The adaptive filter H=1 synthesizes the
non-fundamental harmonics x= ∗ of x ∗ using x as a reference; the input q
1 =1 contains every harmonic except the fundamental
the difference d. The other loop adds to the target’s fundamental harmonic x1∗ the non-fundamental harmonics x= ∗ . These
1
are currently defined offline, for instance with Picard iterations. This is repeated until the non-fundamental harmonics of x ∗
cancel the ones of x, making d and u monoharmonic. The continuation parameter is the amplitude of x1∗ .
17.2 Control-Based Continuation with Adaptive Filtering
Adaptive filters have been used for online harmonic elimination in the literature [7]. The contrary is done in this work:
they are used for harmonic selection in order to replace the offline loop shown in Fig. 17.1 and have a fully online system.
The filter uses input signals q, comprising orthogonal sine waves of specific harmonics’ frequencies that are combined to
synthesize an output signal. The RLS algorithm minimizes the squared error between this output signal and a reference
signal. This algorithm combines good performance with the ability to function with transient signals.
Figure 17.2a shows a system where an adaptive filter is used to synthesize a monoharmonic excitation signal f using
the multiharmonic controller output u as reference. In this way, d and u can be multiharmonic while the system is excited
monoharmonically. The response target x1∗ does not have to cancel the non-fundamental harmonics of x and thus can be
monoharmonic. Figure 17.2b shows a fully online system using the same strategy as the one shown in Fig. 17.1. The adaptive
∗ of the target using x as a reference. Consequently, x ∗ cancels
filter is used to synthesize the non-fundamental harmonics x=1 =1
the non-fundamental harmonics of x.
With these methods, no offline iteration is needed on the control target to ensure controller non-invasiveness, avoiding
waiting for steady-state between each iteration. Traditional continuation by increasing the amplitude of x1∗ with finite steps
17 Control-Based Continuation of Nonlinear Structures Using Adaptive Filtering 111
is still possible but not necessary anymore. The amplitude of x1∗ can be swept continuously and all responses of the structure
to constant frequency harmonic excitation will be explored in full transient, removing the need to wait for steady-state
altogether.
The system shown in Fig. 17.2a was tested on a softening-hardening single degree-of-freedom energy harvester [8]. Control-
based continuation was used to stabilize the system at an unstable periodic orbit to show that the filter has not significantly
affected controller stabilization. Figure 17.3a shows rich harmonic content in the excitation signal f fed into the shaker
before the filter is activated. After 11 s, the filter is activated and the excitation becomes monoharmonic. The structure stays
Excitation [V]
141.2 1e-2
88.9 3e-3
Frequency [Hz]
56 1e-3
35.3
3e-4
22.2
1e-4
14
3e-5
8.8
1e-5
2 4 6 8 10 12 14 16 18 20
Time [s]
(a) The non-fundamental harmonics of the excitation signal are reduced by an
order of magnitude after activation of the adaptive filter at 11 s
Response [m]
141.2
1e-4
88.9
Frequency [Hz]
56 3e-5
35.3
1e-5
22.2
14 3e-6
8.8
1e-6
2 4 6 8 10 12 14 16 18 20
Time [s]
(b) The response signal is unchanged after activation of the adaptive filter at 11 s
Fig. 17.3 A softening-hardening structure is stabilized at an unstable periodic orbit with CBC; before activation of the filter at 11 s, the controller
is invasive and the excitation is multiharmonic; after filter activation, the controller becomes non-invasive without destabilizing the structure from
its periodic orbit. (a) The non-fundamental harmonics of the excitation signal are reduced by an order of magnitude after activation of the adaptive
filter at 11 s. (b) The response signal is unchanged after activation of the adaptive filter at 11 s
112 G. Abeloos et al.
at the same unstable periodic orbit but the controller becomes non-invasive, as shown in Fig. 17.3b. With the functionality
of the filtering being shown, the full response manifold of the structure can be drawn by successive full-transient sweeps of
response amplitude at constant frequency.
17.4 Conclusion
A method for online canceling of the feedback controller invasiveness has been introduced. Adaptive filters are used to either
synthesize a monoharmonic excitation signal, or to synthesize a control target canceling the non-fundamental harmonics fed
into the controller. This method removes the need to make offline iterations on the control target while performing control-
based continuation. It opens the possibility of exploring the structural responses with full-transient sweeps of the control
target amplitude rather than finite continuations steps, accelerating the structural characterization procedure.
Acknowledgments G.A. is funded by the FRIA grant of the Fonds National de la Recherche Scientifique (FNRS), L.R. is funded by a Research
Fellowship from the Royal Academy of Engineering (RF1516/15/11). They gratefully acknowledge the financial support of the Royal Academy
of Engineering and the FNRS.
References
1. Sieber, J., Krauskopf, B.: Control based bifurcation analysis for experiments. Nonlinear Dynamics 51, 356–377 (2008)
2. Sieber, J., Gonzalez-Buelga, A., Neild, S.A., Wagg, D.J., Krauskopf, B.: Experimental continuation of periodic orbits through a fold. Phys. Rev.
Lett. 100, 244101 (2008)
3. Renson, L., Shaw, A.D., Barton, D.A.W., Neild, S.A.: Application of control-based continuation to a nonlinear structure with harmonically
coupled modes. Mech. Syst. Signal Process. 120, 449–464 (2019)
4. Barton, D.A.W., Sieber, J.: Systematic experimental exploration of bifurcations with noninvasive control. Phys. Rev. E 87, 052916 (2013)
5. Renson, L., Gonzalez-Buelga, A., Barton, D.A.W., Neild, S.: Robust identification of backbone curves using control-based continuation. J.
Sound Vib. 367, 145–158 (2016)
6. Renson, L., Barton, D.A.W., Neild, S.: Experimental tracking of limit-point bifurcations and backbone curves using control-based continuation.
Int. J. Bifurcation Chaos 27, 1730002 (2017)
7. Blasko, V.: A novel method for selective harmonic elimination in power electronic equipment. IEEE Trans. Power Electron. 22, 223–228 (2007)
8. Barton, D.A.W., Burrow, S.G.: Numerical continuation in a physical experiment: investigation of a nonlinear energy harvester. J. Comput.
Nonlinear Dyn. 6, 011010 (2001)
Chapter 18
Tuning of Finite Element Model Parameters to Match Nonlinear
Reduced Order Models
Kyusic Park and Matthew S. Allen
Abstract There has been a growing demand for nonlinear model updating procedures in the structural dynamics community
as advanced vehicles have started to incorporate nonlinearity into their designs. Finite element (FE) model updating is
difficult for a nonlinear structure for several reasons: there may be many unknown parameters in the FE model so that
multiple solutions may exist, and it is very expensive to compute the structure’s nonlinear normal modes from a full FE
model that are an excellent metric to use for nonlinear updating. A recent work showed that some of these challenges can be
overcome by updating a nonlinear reduced order model (ROM) rather than the full FE model. The updated ROM can be used
to compute response statistics, stresses and ultimately life of the structure. However, one drawback of this approach is that
one does not gain physical insight into which parameters in the FE model were in error, and so it is difficult to transfer the
lessons learned to future FE models. This work explores the feasibility of updating a FE model to correlate with a ROM with
a set of known parameters. The target ROM parameters consist of the linear stiffness and the nonlinear stiffness coefficients
which have been identified previously by updating the ROM to correlate with experimental measurements. In this process an
optimization routine is setup in which the free parameters are those of the FE model, such as boundary stiffness springs,
imperfections, pre-stress, etc. The optimization procedure is wrapped around a ROM creation algorithm that iteratively tunes
the FE parameters, creates a ROM, and evaluates the ROM parameters with respect to the target ROM in order to minimize
the objective function. This work will demonstrate the procedure on a numerical example of a curved beam in order to verify
its effectiveness on the nonlinear model updating.
Keywords Nonlinear dynamics · Nonlinear model updating · Reduced order models · Nonlinear normal modes ·
Gradient-based optimization
18.1 Introduction
This work demonstrates a new approach to model updating in which a finite element (FE) model is updated to correlate
with a reduced order model (ROM), which is already known to correlate well with measurements. The framework optimizes
the FE model parameters by a gradient based method so that the ROM of the FE model matches a target ROM, which
has presumably been derived experimentally [1]. The following section briefly outlines the reduced order modeling and an
overview of the FE parameter tuning framework. Then, a numerical example of a curved beam is presented and discussed.
K. Park ()
Department of Mechanical Engineering, University of Wisconsin – Madison, Madison, UW, USA
M. S. Allen
Department of Engineering Physics, University of Wisconsin – Madison, Madison, UW, USA

114 K. Park and M. S. Allen
18.2 Theory
The n DOF FE model of a geometrically nonlinear structure can be expressed as
Mẍ + Cẋ + Kx + fnl (x) = f(t) (18.1)
where M, C, and K are the n × n mass, damping, and linear stiffness matrices respectively. The n × 1 nonlinear restoring
force fnl is a function of the displacement vector x. The full equations of motion of the FE model in a physical space can be
projected onto a modal subspace by using x(t) = m q(t), where m is the n × m mass normalized mode shape matrix and
q is the m × 1 modal displacement. The reduced DOF of modal coordinates can be significantly smaller than the DOF of
original physical coordinates (m n). Then, the rth nonlinear modal equation becomes
q̈r + cr q̇r + ωr2 qr + θr (q1 , q2 , . . . , qm ) = Tr f(t) (18.2)
where the nonlinear restoring force θr is a function of the modal displacements as θr (q) = Tr fnl (m q). The nonlinear
restoring force that captures the geometric nonlinearity of a linear elastic system can be well approximated as

m
m
m
m
m
θr (q1 , q2 , . . . qm ) = Ar (i, j )qi qj + Br (i, j, k)qi qj qk (18.3)
i=1 j =i i=1 j =i k=j
where Ar and Br are the quadratic and cubic nonlinear stiffness terms, respectively. This work uses the implicit condensation
and expansion (ICE) method in order to find the nonlinear stiffness terms, which was presented in detail in [2].
The proposed procedure is illustrated in Fig. 18.1a, which follows the yellow path of the flowchart. It uses an original FE
model and a target ROM that was previously updated toward an experimental system. In fact, instead of directly updating
the FE model parameters to those of an actual system (red path in the flowchart), their mismatches were incorporated into a
ROM so that its resulting nonlinear stiffness terms could be updated toward the experimental data at a significantly reduced
computational cost (green path in the flowchart). Then, our framework updates the original FE model to match with the
updated ROM by tuning a set of design variables that can be chosen from the FE model parameters, such as imperfections
in geometry, pre-stress distribution, boundary conditions for stiffness, etc. The optimization involves a loop in which the FE
model parameters are tuned, a ROM is created by the ICE method, and the ROM coefficients are evaluated with respect to the
target ROM coefficients until their discrepancies are minimized (Fig. 18.1b). The optimization routine finds the minimum
objective function based on a gradient based method, i.e. the interior-point algorithm [3], which is implemented in the
MATLAB R
function, fmincon.
Fig. 18.1 Representation of the finite element model tuning framework: (a) an overall flow diagram and (b) an optimization diagram
18 Tuning of Finite Element Model Parameters to Match Nonlinear Reduced Order Models 115
18.3 Numerical Example: Curved Beam
In this work, the proposed procedure is tested numerically on the FE model of the curved beam with axial boundary springs in
Fig. 18.2. The beam had length of 9 in., width of 0.5 in., thickness of 0.031 in., and radius of curvature of 80 in.. The Young’s
modulus was 29,700 ksi, Poisson’s ratio was 0.28 and mass density was 7.36 ×104 lb-s2 /in.4 , in order to approximate steel.
The stiffness of axial springs was 1.60 × 104 lbf/in., approximating relatively soft boundary conditions. The beam model
was meshed with 40 two node beam elements to have 117 free DOF. Then, two design variables for the optimization were
chosen among the FE model parameters, i.e. the radius of curvature that can capture the imperfection of curvature of the
beam and the axial spring stiffness that can identify the boundary conditions, based on an assumption that these two physical
aspects are uncertain. Meanwhile, a target ROM of the beam model was generated from a reference FE model that has
small differences in design parameters compared to the original beam model. In this example, the reference FE model was
identical to the beam model except for the radius of curvature, R = 100 in. (i.e. 125% of the original beam model) and the
axial boundary stiffness, Kx = 2.00 × 104 lbf/in. (i.e. 125% of the original beam model).
In this example, using a single mode ROM composed of the first bending mode was sufficient to accurately update the
design variables of the beam model. The linear and nonlinear coefficients of a single mode ROM of the initial beam model
as well as those of the target ROM are presented in Table 18.1. The optimization took 13 iterations to minimize the gap
between the coefficients of the two ROMs as shown in Fig. 18.3. The converged curves of two design variables in Fig. 18.4
also demonstrates that the tuning is accurately performed with relatively small number of iterations. The percentage errors
of the updated design variables with respect to the reference values were below 0.0001% at 8th iteration.
Fig. 18.2 Schematic of a beam with uncertain radius of curvature R and axial stiffness terms Kx
Table 18.1 Linear and nonlinear stiffness coefficients of the optimized ROM and the target ROM
Coefficients Initial value Final value Target value
ωr 907.8337 833.9752 833.9752
A1 (1, 1) −1.3458 × 109 −1.2828 × 109 −1.2828 × 109
B1 (1, 1, 1) 6.1813 × 1011 7.5551 × 1011 7.5551 × 1011
100
Normalized Cost Function Value
10-5
10-10
10-15
10-20
0 5 10 15
Number of Iterations
Fig. 18.3 Convergence history of cost function value normalized by the target ROM coefficients
116 K. Park and M. S. Allen
Radius of Curvature 104

105 Axial Spring Stiffness
2.1
Axial Spring Stiffness (lbf/in)

Radius of Curvature (in)
100 2
1.9
95
1.8
90
1.7
85
1.6
80 1.5
0 5 10 15
Number of Iterations
Fig. 18.4 Convergence history of the design variables
It is important to mention that the design parameters should be properly chosen in order to effectively account for the
nonlinear stiffness terms of the updating ROM. For example, the beam thickness captures only the linear behaviors of the
beam model, so that the parameter could never contribute to updating the nonlinear terms. On the other hand, the radius of
curvature of the beam and the axial spring stiffness were able to account for both linear and nonlinear stiffnesses, which
resulted in an accurate tuning. Also, note that the gradient-based optimization used in this work provides only a local
minimum. Thus, it is possible that the updated design variables may not be the true values, although it was not the case in the
presented example. More complex examples that can further explore these issues will be discussed during the presentation.
Acknowledgments This work was supported by the Air Force Office of Scientific Research, Award # FA9550-17-1-0009, under the Multi-Scale
Structural Mechanics and Prognosis program managed by Dr. Jaimie Tiley. The authors would also like to thank Christopher I. Van Damme who
contributed to our numerical study.
References
1. I Van Damme, C.: Model correlation and updating of geometrically nonlinear structural models using nonlinear normal modes and the multi-
harmonic balance method. Ph.D dissertation, University of Wisconsin-Madison (2019)
2. Kuether, R.J., Deaner, B.J., Hollkamp, J.J., Allen, M.S.: Evaluation of geometrically nonlinear reduced-order models with nonlinear normal
modes. AIAA J. 53(11), 3273–3285 (2015)
3. Byrd, R.H., Gilbert, J.C., Nocedal, J.: A trust region method based on interior point techniques for nonlinear programming. Math. Program.
89(1), 149–185 (2000)
Chapter 19
Towards an Understanding of the Transient Behavior
of the Five-Parameter Iwan-Type Model
Robert M. Lacayo and Matthew S. Allen
Abstract Mignolet (J Sound Vib 349:289–298, 2015) demonstrated mathematically that his proposed five-parameter Iwan-
type model weakens the coupling between the change in effective stiffness and change in effective damping of the model as
vibration amplitude changes. Several experimental studies on bolted-joint structures have experienced difficulty in fitting
a traditional Iwan model to match measurements, so this advantage is sorely needed. However, Mignolet’s work only
considered steady-state harmonic motion, and the stiction behavior of the internal sliders was not studied in great detail.
In this work, the force-constitutive formulation of the five-parameter Iwan-type model is implemented computationally and
then examined to understand its behavior in more general transient scenarios. The five-parameter model is found to have
a complicated dependence on its displacement history. If the model reaches a maximum steady-state displacement after a
ring-up response (such as occurs when the system is excited at resonance by a shaker), the stiffness and damping it exhibits
is consistent with those formulated by Mignolet. When the vibration decays below the maximum-achieved displacement,
however, the effective stiffness and damping revert to power-law behavior with amplitude. The power-law behavior is
functionally similar to that of the four-parameter Iwan model (on which the five-parameter model is based), so the advantage
of the weakened coupling between the stiffness and damping is lost in a ring-down response.
Keywords Iwan model · Microslip · Hysteresis · Friction · Nonlinear damping
19.1 Introduction
A well-known model for microslip is the parallel-series distributed-element model developed by Iwan [1]. This model
consists of a parallel arrangement of a large (ideally infinite) number of Jenkins elements (a spring and Coulomb slider in
series), where each Jenkins element has the same spring stiffness and slips at a different force. As the Iwan model displaces
from equilibrium, each of the sliders begins to slip one after another. Segalman later defined a four-parameter distribution
function that gives the force at which each Jenkins element slips, such that the energy dissipated per cycle by the Iwan model
undergoing harmonic motion follows a power-law trend with the loading amplitude [2], as observed in many experiments
[3–8].
The four-parameter Iwan model strongly couples the stiffness with the dissipation, and this sometimes makes it
challenging to fit the model to experimental measurements. Mignolet [7] investigated this problem and proposed allowing
each slider to have different static and dynamic friction coefficients. This allows each slider to exhibit stiction: a large static
friction force that decreases when the slider begins to move. Mignolet assigned a fifth parameter to govern this decrease in the
friction force, and he used his new five-parameter Iwan-type formulation to derive closed-form expressions for amplitude-
dependent stiffness and dissipation during steady-state harmonic motion. With these expressions, he was able to fit his model
R. M. Lacayo ()
M. S. Allen
University of Wisconsin – Madison, Madison, WI, USA

118 R. M. Lacayo and M. S. Allen
to stiffness and dissipation measurements successfully. However, Mignolet never investigated how the five-parameter Iwan-
type model could be implemented for transient vibration, so the simplified expressions that he developed have not yet been
verified, nor has anyone studied the implications of the fifth parameter on the transient response of a system with an Iwan
joint.
This work presents an implementation of the five-parameter Iwan-type model and documents an investigation on the
amplitude-dependent stiffness and dissipation of that element in more general transient settings. The results show that
the five-parameter model has a complicated dependence on the past history of its displacement, especially on the largest
displacement the joint has experienced. As a result, the behaviors that Mignolet discussed would only be observed if a virgin
structure were excited monotonically up to the steady-state conditions that he assumed. In contrast, in a response where the
amplitude decreases with time, the power-law behavior becomes functionally equivalent to that of the four-parameter model,
so the advantage of weakened coupling between the stiffness and damping is lost. This study highlights how difficult it is to
account for stiction when studying the vibration of joints and the additional implications that this presents, which should be
addressed by future models.
19.2 Overview of the Five-Parameter Iwan-Type Model
An Iwan element consists of an infinite number of sliders that all slip at different force levels, as described by a distribution
function, ρ(φ). The force in the Iwan element is a summation over these sliders [2] as follows,
∞
F (t) = (t, φ)ρ(φ)dφ, (19.1)
0
where
(
u(t) − x(t) for |u(t) − x(t)| < φ (sticking),
(t, φ) = (19.2)
φ otherwise (slipping),
tracks the current deflection, x, of each spring. The distribution function, ρ(φ), defined in [2], is chosen so that the Iwan
element dissipates energy according to some power of the displacement amplitude, i.e. the log damping vs. log vibration
amplitude is linear with a slope corresponding to the power-law exponent.
If the displacement of the five-parameter Iwan-type model is defined with a sinusoidal function, as in u(t) = u0 sin(t),
then Mignolet shows that the force exerted by the model while in microslip (u0 < φmax ) reaches a peak amplitude of [7]
χ +2
F0 (u0 ) = KT u0 − u0 , (19.3)
χ +1
Rφmax R [χ + 2 − θ (χ + 1)]
KT = + S, = . (19.4)
(χ + 1) (χ + 1)(χ + 2)
These expressions were then used to predict the tangent stiffness and energy dissipation of his five parameter Iwan element,
resulting in the equations given in [7]. However, in that work a computational element was never developed, so the transient
response of Mignolet’s five-parameter model was never explored.
To investigate the transient behavior of a five-parameter Iwan-type model, a computational element embodying its force-
constitutive relationship was created. Here the constitutive model is approximated by using a finite number of sliders, as
done in [2], and assigning a slip threshold to each Jenkins element so that the integral in Eq. (19.1) becomes a summation.
Also, an additional variable had to be added to track the state of each slider, whether stuck or slipping. Further details will
be provided in the presentation.
19 Towards an Understanding of the Transient Behavior of the Five-Parameter Iwan-Type Model 119
19.3 The Transient Behavior of the Five-Parameter Iwan-Type Model
The hysteresis of the Iwan-type element was observed in three different cases. The first case considered is the steady-
state response of the element after virgin loading. This case matches that used by Mignolet when deriving Eq. (19.3). The
remaining two cases examine the hysteresis in response to ring-up and ring-down deflection.
19.3.1 Steady-State Behavior After Virgin Loading
An Iwan element with parameters χ = −0.6, φmax = 6 × 10−4 , R = 7 × 105 , S = 9 × 103 and θ = 0.3 was used. The
element was discretized into a set of N =10,000 sliders (many more than typical to ensure a smooth FJ approximating a
continuous Iwan-type element). The discrete element was subjected to a sinusoidal deflection of the form
u(t) = u0 sin(2π t) (19.5)
for one and one-quarter cycles (t ∈ [0, 1.25]) sampled at step size 0.001. The element was displaced to an amplitude u0 =
5×10−4 . The slider states were first initialized to zero, i.e. a virgin structure. The resulting force-deflection curve is plotted
in Fig. 19.1a. The first quarter cycle creates an initial loading curve up to u0 , and the remaining full cycle forms the reverse
and forward curves of a hysteresis bounded by ±u0 .
While not elaborated here, one can use this result to show that if the Iwan-type element is loaded monotonically up to a
certain displacement and then undergoes steady-state vibration at that amplitude, then the hysteresis it exhibits has a secant
slope and enclosed area that is consistent with Mignolet’s expressions for amplitude-dependent stiffness and dissipation.
A hysteresis that exhibits this property is referred to as a “Mignolet hysteresis” in the rest of this work.
19.3.2 Ring-Up and Ring-Down Behavior
Two different sinusoidal displacement signals were constructed: one for the ring-up and the other for the ring-down. The ring-
up signal was designed so that the amplitude increases linearly from a low boundary, uL = 2 × 10−4 , to a high boundary,
uH = 5.5 × 10−4 , in two cycles, and then completes another cycle at the high boundary. The third cycle is meant to create a
complete hysteresis loop for comparison with the Mignolet hysteresis at the final amplitude. Similarly, the amplitude of the
ring-down signal was designed to change from uH to uL in two cycles followed by one cycle at uL . The slider states in the
two cases were initialized to account for the element having already displaced from zero to its starting location on the virgin
loading curve.
(a) (b) (c)

10 A
D
5
FJ (t)
-5
E
B
-10
C
-5 0 5
u(t) × 10-4
Fig. 19.1 Force-deflection curves for the discrete Iwan-type element in the case of (a) sinusoidal displacement, (b) ring-up, and (c) ring-down
force-displacement curves for θ = 0.3
120 R. M. Lacayo and M. S. Allen
The resulting force-deflection curves are shown in Fig. 19.1. For the ring-up curve (Fig. 19.1b), each forward and reverse
path starts on the initial loading curve (point A) and proceeds to the opposite displacement (point B) by following the path of
a Mignolet hysteresis. After reaching that displacement, the path of initial loading is followed until load reversal (point C).
The third cycle (D to E to D) exactly traces the Mignolet hysteresis loop associated with the highest-achieved amplitude.
The ring-down force-deflection curve (Fig. 19.1c), however, results in a deviation from the Mignolet hysteresis. In the first
half-cycle (point A to point B), the steady-state hysteresis path is followed as usual, but load reversal occurs well before
reaching the initial loading curve. The ring-down curve also reverses early after the second half-cycle (point C), and it
is more apparent that the regime of linear force versus displacement (the path with constant slope just before point C) is
shortening. This constant slope regime ultimately disappears, and the shape of the hysteresis in the third full-cycle (E to F
to E) is completely different from the Mignolet hysteresis. As a result, in the ring down case shown here the secant slope
(effective stiffness) would be in error by −14.8% and the enclosed area (energy dissipated per cycle) would be in error by
90.3% of the values calculated from Mignolet’s equations.
19.4 Conclusion
The results of Sect. 19.3.2 reveal an important caveat regarding Mignolet’s steady-state assumptions for the five-parameter
Iwan-type model. By allowing the kinetic friction force to be less than the maximum static friction force, the five-parameter
model exhibits a form of history-dependence in that the shape of the hysteresis at lower amplitudes is controlled by that of the
highest-achieved amplitude. The ring-down cycles in a general five-parameter model hardly resemble those of the Mignolet
hysteresis, so, therefore, Mignolet’s expressions for steady-state stiffness and dissipation cannot be used to estimate the
same properties in a ring-down response. This demonstrates that Mignolet’s steady-state stiffness and dissipation equations
implicitly assume that the steady state was reached via a monotonically-increasing displacement amplitude.
In the conference presentation, further case studies will be presented and elaborated. Furthermore, dynamic simulations on
the transient (ring-down) response of four- and five-parameter Iwan models will be presented, showing that the response of
a system with a five-parameter model can be exactly represented with a four-parameter model after the amplitude decreases
below the hysteretic regime of constant slope.
References
1. Iwan, W.D.: A distributed-element model for hysteresis and its steady-state dynamic response. J. Appl. Mech. 33(4), 893–900 (1966)
Process. 118, 133–157 (2019)
4. Deaner, B.J., Allen, M.S., Starr, M.J., Segalman, D.J., Sumali, H.: Application of viscous and Iwan modal damping models to experimental
measurements from bolted structures. J. Vib. Acoust. 137(2), 021012 (2015)
5. Roettgen, D.R., Allen, M.S.: Nonlinear characterization of a bolted, industrial structure using a modal framework. Mech. Syst. Signal Process.
84(Part B), 152–170 (2017)
6. Lacayo, R.M., Deaner, B.J., Allen, M.S.: A numerical study on the limitations of modal Iwan models for impulsive excitations. J. Sound Vib.
390, 118–140 (2017)
7. Mignolet, M.P., Song, P., Wang, X.Q.: A stochastic Iwan-type model for joint behavior variability modeling. J. Sound Vib. 349, 289–298 (2015)
8. Lacayo, R., Pesaresi, L., Gross, J., Fochler, D., Armand, J., Salles, L., Schwingshackl, C., Allen, M., Brake, M.: Nonlinear modeling of structures
with bolted joints: a comparison of two approaches based on a time-domain and frequency-domain solver. Mech. Syst. Signal Process. 114,
413–438 (2019)
Chapter 20
On the Solution of Nonlinear Algebraic Equations Following
Periodic Forced Response Analysis of Nonlinear Structures Using
Different Nonlinear Solvers
H. Sefa Kizilay and Ender Cigeroglu
Abstract In periodic forced response analysis of nonlinear structures, most of the time analytical solutions cannot be
obtained due to the complex behavior of the nonlinearity and/or due to the number of nonlinear equations to be solved.
Therefore, numerical methods are widely used. For periodic forced response analysis of nonlinear systems, generally
Harmonic Balance Method (HBM) or Describing Function Method (DFM), which transform the nonlinear differential
equations into a set of nonlinear algebraic equations, are used. In the literature, there exist several nonlinear algebraic equation
solvers based on Newton’s method which have different convergence properties and computational expense. In this paper,
comparison of computational performance of different nonlinear algebraic equation solvers are studied where, solvers with
different convergence order are selected based on the number of Jacobian matrix and vector function evaluations. In order to
compare the performance of these selected nonlinear solvers, a lumped parameter model with cubic stiffness nonlinearity is
considered. Several case studies are performed and nonlinear solvers are compared to each other in terms of solution time
based on error tolerance used.
Keywords Newton’s method · Newton-like solvers · Nonlinear algebraic equations · Numerical solution · Nonlinear
vibrations · Harmonic balance method
20.1 Introduction
Presently, there exists many engineering applications where the effects of nonlinearities cannot be neglected. Some examples
of such nonlinear systems occur in the design of gas turbine engine blades [1, 2], bolted joints [3, 4] and large deformations
in structures [5]. For those systems, where nonlinearity changes the dynamic characteristics significantly, many problems are
encountered in the process of obtaining reliable and accurate solutions. In the literature, many researchers study to develop
new solution methods for system of nonlinear algebraic equations in order to increase the accuracy and stability of the
solution process and reduce the computational effort.
In periodic forced response analysis of nonlinear structures, nonlinear differential equations of motion are converted into
a set of nonlinear algebraic equations by using Harmonic Balance Method (HBM) [1–4] or Describing Function Method
(DFM) [6, 7]. Application of HBM or DFM results in a set of nonlinear algebraic equations which needs to be solved for
the unknown displacement vector by a nonlinear algebraic equation solver. The most common nonlinear solver utilized in
periodic forced response analysis of nonlinear structures is Newton’s method with arc-length continuation [8–13].
In this paper, different Newton-like solvers [14, 15] are compared in terms of total iteration number and computational
time for the determination of forced response of nonlinear structures. Convergence order, number of Jacobian matrix
calculations and number of vector function evaluations are considered in the selection process of these solvers. A lumped
parameter system with cubic stiffness nonlinearity is considered as the case study in order to compare the performance of
the solvers.
H. S. Kizilay · E. Cigeroglu ()

Department of Mechanical Engineering, Middle East Technical University, Ankara, Turkey

122 H. S. Kizilay and E. Cigeroglu
20.2 Mathematical Modeling
In this part, methods used in the determination of periodic forced response analysis of nonlinear systems are presented.
Equations of motion of a nonlinear mechanical system can written in the following form
Mẍ(t) + Cẋ(t) + Kx(t) + fN (t) = f(t), (20.1)
where x is the displacement vector and dot represents differentiation with respect to time. Here, M, C and K represent the
mass, viscous damping and stiffness matrices of the system, respectively. fN is the vector of internal nonlinear forcing and f
is the external harmonic excitation force vector.
20.2.1 Harmonic Balance Method
Frequency domain methods are computationally more efficient for the determination of steady-state response of nonlinear
structures compared to time domain methods; hence, they are widely preferred for forced response analysis of nonlinear
systems [1–4]. Since, periodic forced response of nonlinear systems is of interest, in this study, HBM which is one of the
most widely used frequency domain methods, is utilized in the analysis of nonlinear systems.
For harmonic excitation, response of the nonlinear system is also assumed as periodic. Therefore, displacement, nonlinear
internal forcing and excitation force vectors can be expressed by Fourier series as follows by using only a single harmonic
term
x(t) = xs sin (ωt) + xc cos (ωt) , (20.2)
fN (t) = fsN sin (ωt) + fcN cos (ωt) , (20.3)
f(t) = fs sin (ωt) + fc cos (ωt) . (20.4)
Substituting Eqs. (20.2) to (20.4) into Eq. (20.1), the following set of nonlinear algebraic equations is obtained
s s
K − ω2 M −ωC fN f
q + = c . (20.5)
ωC K − ω2 M fcN f
where, q = {xs xc }T which contains the sine and cosine coefficients of unknown displacement vector.
This nonlinear algebraic equation set can be written in a residual vector as
s s
K − ω2 M −ωC f f
R (q, ω) = q+ N − c = 0, (20.6)
ωC K−ω M
2 fNc f
which needs to be solved by using a nonlinear solver.
20.3 Solution of Nonlinear Algebraic Equations
In this section, Newton’s method and Newton-like nonlinear solvers used for the solution of Eq. (20.6) are briefly presented.
A single iteration of Newton’s method can be given as follows
−1
qk+1 = qk − J qk · R qk , ω , (20.7)
20 On the Solution of Nonlinear Algebraic Equations Following Periodic. . . 123

where, J qk = ∂R (q, ω) /∂q|q=qk is the Jacobian matrix and k is the iteration number. Newton’s method is an order 2
and a single step method which requires 1 Jacobian and 2 vector function evaluations at each step and denoted as M1 in
the paper. In the literature, there exists different method which follows a similar procedure and have different convergence
orders. In the solutions of nonlinear algebraic equations, computationally the most expensive part is the calculation of the
Jacobian matrix. Therefore, in the current study, the methods considered are selected from the widely used methods and in
such a way that they require at most 2 Jacobian matrix evaluations per step of the nonlinear solver [14–21].
The method developed by Darvish and Barati [14] is a 3rd order newton type two-step method which requires 1 Jacobian
and 2 vector function evaluations at each step and it is denoted by M2. Cordero et al. [16] derived a new method to design
predictor-corrector methods for systems of nonlinear equations. It requires 2 Jacobian and 2 vector function evaluations at
every step. This three-step method (M3) is 6th order. Sharma et al. [17] developed a Newton-like three-step method for
systems of nonlinear equations which requires 1 Jacobian and 3 vector function evaluations at each step. This method (M4)
is 5th order. Sharma et al. [18] suggested a 5th order three-step method (M5) for solving systems of nonlinear equations.
At every iteration step, 2 Jacobian matrices and 2 vector functions need to be calculated. Liu et al. [19] suggested a 3rd
order two-step method (M6) for solving systems of nonlinear Eq. (20.2) Jacobian matrices, 1 vector function evaluations are
required at each step. Noor et al. [20] suggested a 3rd order two-step method for solving systems of nonlinear equations.
This method (M7) requires 2 Jacobian and 1 vector function evaluations at each step. Cordero et al. [21] suggested a two-
step variant of Newton’s method (M8) to solve functions of several variables. 2 Jacobian matrices and 1 vector function
evaluations are required at each step. Chord [22] suggested a new approach to modify Newton’s method. The main aim of
this method (M9) is to reduce the number of Jacobian matrix calculation. Jacobian matrix is calculated at the first iteration
and then it is used in Chord’s method. Shamanski [22] method (M10) is applied in such a way that Jacobian matrix is
calculated for every 2 steps.
20.4 Case Study
In this part, the 4-degree of freedom (DOF) lumped parameter model with cubic stiffness nonlinearity given in Fig. 20.1 is
studied to compare the performance of the nonlinear solvers in Table 20.1. Paremeters of this lumped parameter model are
m1, 2, 3, 4 = 1 kg, k1, 2, 3, 4, 5 = 2500 N/m, kc = 1000 N/m3 . Structural damping with a loss factor of 0.01 is considered in the
system and a sinusoidal force with an amplitude of 120 N is applied on the second mass.
Fig. 20.1 4 DOF lumped parameter model with cubic stiffness nonlinearity
Table 20.1 Specific calculation aspects of the solvers

Number of Jacobian calculations Number of function evaluations Order of convergence
M1 1 1 2
M2 1 2 3
M3 2 2 6
M4 1 3 5
M5 2 2 5
M6 2 1 3
M7 2 1 3
M8 2 1 3
M9 1 1 1≤q<2
M10 1 1 1<q<2
Fig. 20.2 Displacement receptance of linear and nonlinear 4 DOF lumped parameter model
Table 20.2 Performance comparison of the solvers for constant arc-length parameter (0.04) and different error criteria
Error criterion [1e−6] Error criterion [1e−8] Error criterion [1e−10]
Total number of iterations Normalized time Total number of iterations Normalized time Total number of iterations Normalized time
M1 9494 1.18 10,608 1.22 12,512 1.30
M2 8278 1.11 8622 1.15 9721 1.20
M3 8173 1.65 8206 1.62 8236 1.67
M4 9694 1.28 10,767 1.36 13,081 1.57
M5 8178 1.62 8225 1.61 8265 1.71
M6 9276 2.26 10,503 2.47 12,242 2.89
M7 9187 2.25 10,429 2.46 12,084 2.84
M8 8266 1.58 9299 1.69 10,809 1.93
M9 9616 1.02 11,147 1.10 14,011 1.13
M10 9553 1.00 10,976 1.08 13,662 1.19
Response amplitude of the second mass normalized with respect the applied forcing amplitude is given Fig. 20.2 for
linear and nonlinear systems. It should be noted that since cubic stiffness is a continuous nonlinearity, it is effective in all
frequency range throughout analysis. Investigating Fig. 20.2, it can be concluded that the system is highly nonlinear which
makes convergence of the nonlinear solvers harder. Therefore performance of the solvers are studied on this highly nonlinear
case.
In order to understand the effect of error criterion and arc-length parameter on the performance of the nonlinear solvers,
different stopping error criteria (1e−6, 1e−8, and 1e−10) and different arc-length parameters (0.32, 0.16, 0.12, 0.08, 0.04,
0.02, and 0.01) are investigated in the solution process.
Firstly, the solvers are investigated by keeping the arc-length parameter constant at 0.04 and changing error criterion
1e−6, 1e−8 and 1e−10. Total number of iterations and normalized computational time are given in Table 20.2 and Fig.
20.3. Computational time are normalized according to solver with the minimum computational time which turns out to be the
Shamanski method (M10). It can be concluded that decreasing error criteria increases total number of iterations and hence,
computational time. It can be understood that method M6 and M7 which require 2 Jacobian matrix evaluations at each step
has the worst performance compared to the other solvers. In addition, M2, Chord (M9) and Shamanski (M10) methods are
promising methods compared to Newton’s method (M1) for this specific arc-length parameter and error criteria. In addition,
M3 and M5 methods have minimum total iteration numbers for all error criteria; however due to high computational load of
M3 and M5, they are not competitive with M2, M9 and M10.
Secondly, the solvers are investigated by keeping error criteria constant at 1e−8 and changing arc-length parameters (0.32,
0.16, 0.12, 0.08, 0.04, 0.02, and 0.01). Computational performance of the solvers are given in Table 20.3 and Fig. 20.4.
Cross symbol “x” in Table 20.3 indicates that nonlinear solver used cannot converge for this specific case. Computational
Fig. 20.3 Normalized computational time comparison of the solvers for constant arc-length parameter (0.04) and different error criteria
Table 20.3 Performance comparison of the solvers for constant error criteria (1e−8) and different arc-length parameters
Arc-length parameter
0.32 0.16 0.12 0.08 0.04 0.02 0.01
Normalized time M1 1.29 2.31 3.01 4.26 8.04 15.11 29.44
M2 x 2.10 2.75 4.03 7.58 14.48 29.44
M3 1.45 2.76 3.69 5.56 10.68 21.21 44.12
M4 x x x 4.93 8.96 18.02 33.77
M5 x x x 5.43 10.61 21.67 43.48
M6 2.86 4.95 6.34 8.70 16.28 32.71 57.62
M7 2.81 4.89 6.23 8.65 16.21 32.13 57.34
M8 1.62 3.08 4.04 5.78 11.14 22.60 43.03
M9 x x x 3.52 7.25 13.73 27.00
M10 1.00 1.95 2.55 3.74 7.12 13.94 27.70
time values are normalized according to solver that gives the minimum computational time. In this case, Shamanski method
(M9) resulted in the minimum computational time. It can be realized that decreasing arc-length parameter increases total
number of iterations and computational time. In addition, non-convergent methods (M2-M4-M5-M6) start to converge at
lower arc-length parameters. It can be conlcuded that convergence domain of those solvers are narrower than the other ones.
Moreover, methods M6 and M7 have the worst performance compared to the other solvers. Shamanski (M10) and method M2
are promising methods for higher arc-length parameters compared to Newton’s method (M1). Chord (M9) and Shamanski
(M10) methods can be advantageous for smaller arc-length parameters.
20.5 Conclusion
In this paper, computational performance of different nonlinear algebraic equation solvers are investigated for periodic
forced response analysis of nonlinear structures. Ten different solvers are chosen based on different convergence order
and the number of Jacobian matrix and vector function evaluations. In performance comparison study, a 4-DOF lumped
parameter model with cubic stiffness nonlinearity is considered. Computational performances of the selected solvers are
tested for different values of the stopping error criterion and arc-length parameter. It is observed that M2, Shamanski (M9)
and Chord (M10) methods can be promising methods to decrease computational time, whereas, higher order methods which
requires more Jacobian and/or vector function evaluations are not suitable for periodic forced response analysis of nonlinear
structures.
Fig. 20.4 Normalized computational time comparison of the solvers for constant error criteria (1e−8) and different arc-length parameters
References
1. Sanliturk, K.Y., Imregun, M., Ewins, D.J.: Harmonic balance vibration analysis of turbine blades with friction dampers. J. Vib. Acoust. 119,
96–103 (1997)
2. Von Groll, G., Ewins, D.J.: The harmonic balance method with arc-length continuation in rotor/stator contact problems. J. Sound Vib. 241,
223–233 (2001)
3. Jaumouillé, V., Sinou, J.J., Petitjean, B.: An adaptive harmonic balance method for predicting the nonlinear dynamic responses of mechanical
systems—application to bolted structures. J. Sound Vib. 329(19), 048–4067 (2010)
4. Karapistik, G., Cigeroglu, E.: Reduced order modeling of bolted joints in frequency domain. In: Nonlinear Structures and Systems, Volume 1
Conference Proceedings of the Society for Experimental Mechanics Series, pp. 235–238 (2019) doi: https://doi.org/10.1007/978-3-030-12391-
8_33
5. Lewandowski, R.: Computational formulation for periodic vibration of geometrically nonlinear structures—part 2: numerical strategy and
examples. Int. J. Solids Struct. 34(15), 1949–1964 (1997)
6. Budak, E., Özgüven, H.N.: Iterative receptance method for determining harmonic response of structures with symmetrical non-linearities.
Mech. Syst. Signal Process. 7(1), 75–87 (1993)
7. Tanrikulu, O., Kuran, B., Özgüven, H.N., Imregun, M.: Forced harmonic response analysis of nonlinear structures using describing functions.
AIAA J. 31(7), 1313–1320 (1993)
8. Cigeroglu, E., An, N., Menq, C.H.: A microslip friction model with normal load variation induced by normal motion. Nonlinear Dyn. 50(3),
609 (2007)
9. Nayfeh, A.H., Balachandran, B.: Applied nonlinear dynamics: analytical, computational, and experimental methods. Agawan, MA: Wiley-
VCH Verlag GmbH. (1995)
10. Sarrouy, E., Grolet, A., Thouverez, F.: Global and bifurcation analysis of a structure with cyclic symmetry. Int. J. Non-Linear Mech. 46(5),
727–737 (2011)
11. Sundararajan, P., Noah, S.T.: Dynamics of forced nonlinear systems using shooting/arc-length continuation method—application to rotor
systems. J. Vib. Acoust. 119(1), 9–20 (1997). https://doi.org/10.1115/1.2889694
12. Ferreira, J.V., Serpa, A.L.: Application of the arc-length method in nonlinear frequency response. J. Sound Vib. 284(1–2), 133–149 (2005).
https://doi.org/10.1016/j.jsv.2004.06.025
13. Cigeroglu, E., Samandari, H.: Nonlinear free vibration of double walled carbon nanotubes by using describing function method with multiple
trial functions. Physica E. 46, 160–173 (2012). https://doi.org/10.1016/j.physe.2012.09.016
14. Darvish, M.T., Barati, A.: A third-order newton-type method to solve systems of nonlinear equations. Appl. Math. Comput. 187, 630–635
(2007). https://doi.org/10.1016/j.amc.2006.08.080
15. Decker, D.W., Kelley, C.T.: Sublinear convergence of the Chord method at singular points. Numer. Math. 42(2), 147–154 (1983). https://
doi.org/10.1007/bf01395307
16. Cordero, A., Torregrosa, J.R., Vassileva, M.P.: Pseudocomposition: a technique to design predictor–corrector methods for systems of nonlinear
equations. Appl. Math. Comput. 218(23), 11496–11504 (2012). https://doi.org/10.1016/j.amc.2012.04.081
17. Sharma, J.R., Guha, R.K.: Simple yet efficient Newton-like method for systems of nonlinear equations. Calcolo. 53(3), 1–23 (2015)
18. Sharma, J.R., Gupta, P.: An efficient fifth order method for solving systems of nonlinear equations. Comput. Math. Appl. 67(3), 591–601
(2014). https://doi.org/10.1016/j.camwa.2013.12.004
19. Liu, Z., Fang, Q.: A new Newton-type method with third-order for solving systems of nonlinear equations. J. Appl. Math. Phys. 3(10), 1256
(2015)
20. Noor, M.A., Waseem, M.: Some iterative methods for solving a system of nonlinear equations. Comput. Math. Appl. 57(1), 101–106 (2009)
21. Cordero, A., Torregrosa, J.R.: Variants of Newton’s method for functions of several variables. Appl. Math. Comput. 183(1), 199–208 (2006)
22. Shamanski, V.E.: A modification of Newton’s method. Ukrain Mat. Z. 19, 133–138 (1967)
Chapter 21
Vibration-Based Bolt Tension Estimation for Multi-bolt Joints
M. Brøns, A. Plaugmann, J. J. Thomsen, and A. Fidlin
Abstract Critical bolted connections exist in many engineering structures, from pressurized pipelines to wind turbines.
Often there are legal demands for maintaining necessary bolt tension in such joints to prevent failure. The available tools for
tightening makes it challenging to obtain the correct tension in a bolt, as well as subsequently checking if a bolt is in fact
tightened to the correct level. Recent work proposes to use vibrations for estimating tension in a bolt. Estimation is possible
by measuring and analysing transverse natural frequencies and damping ratios induced by e.g. a transversal or longitudinal
hammer impact on the bolt itself. The work so far has focused on a single bolt. Most bolted joints consist of many bolts, e.g.
a flange connection often has a ring of almost identical bolts. Identical bolts, with almost the same tension, will also have
very similar boundary conditions and thus almost the same natural frequencies. If there is only very light damping between
two adjoining bolts, a frequency response measured on one bolt after an impact might include the vibrational response of
both bolts (the coupling might even be so strong that the two bolts cannot be viewed as entities on their own) leading to the
question: How to separate which frequencies belong to which bolt?
Keywords Bolted joints · Multiple bolts · Tension estimation · Transient vibrations · Coupled vibrations ·
Double-beam model
21.1 Introduction
It is of interest to be able to check the tension in critical bolted joints, i.e. joints that if loosened cause critical failures, such as
wind turbine collapses, pressurized gas pipe leaks etc. Current techniques include torque control, turn control, and ultrasonic
methods [1–3]. A newly proposed technique is to impulse hammer impact a bolt and estimate the tension by analysis of
the recorded dynamic response [4–5]. It is advantageous, as control can be done without employing heavy equipment,
and tightening will then only be performed on bolts that are in fact loose. This can potentially reduce maintenance costs
significantly. The new technique tested on a single bolt show that: at low tension the squared natural frequency of the first
bending mode increases strongly with tension, as the bolt is tightened the squared natural frequency changes more weakly
and approximately linearly with tension.
However, for application purpose it is not enough that this approach works for a single bolt. To use transverse vibrations
for estimation of tension, it has to be investigated if the behavior observed for a single bolt is the same in a multi-bolt
structure. This work addresses the issue of which frequency in a measured response belongs to which bolt, in a multi-bolt
joint. Earlier work [6] showed that it is possible to excite bending natural frequencies in a real multi-bolt structure, but also
that identification of peaks is not trivial. To provide initial insight, a double Bernoulli-Euler beam model coupled with springs
at the boundaries is suggested. The boundary stiffness of each bolt couples to bolt tension by an analytical expression [4],
which previously proved able to reproduce experimental results for a single bolt accurately. For comparison, a 3D COMSOL
model is employed.
M. Brøns () · A. Plaugmann · J. J. Thomsen

Department of Mechanical Engineering, Technical University of Denmark, Lyngby, Denmark
A. Fidlin
Institute of Engineering Mechanics, Karlsruhe Institute of Technology, Karlsruhe, Germany

130 M. Brøns et al.
21.2 Theoretical and Experimental Modeling and Analysis
A two-beam model representing two bolts tightened in the same structure is set up, as shown in Fig. 21.1. An analytical
expression for correlating bolt tension to boundary stiffness is applied, following the procedure of [4]:
c1 tanh (c2 p)
kn = , (21.1)
1 + c3 e−c4 p
where kn is the axial stiffness, c1–4 are positive constants determined by fitting to experimental data (from a single bolt
setup), and p is nondimensional tension (p = Nl2 /EI). The axial stiffness relates to translational (Kw,1–2 ) and rotational
stiffness (Kθ,1–2 ) by constants defined by geometrical properties given in [4]. The coupling springs are roughly modeled as
the geometric mean between the bolts boundary stiffness (more advanced coupling modes are currently investigated):

kw,c (p1 , p2 ) = cw kw,1 kw,2 , k,,c (p1 , p2 ) = c, k,,1 k,,2 , (21.2)
where cw and cθ are unknown constants that need to be fitted. By applying these relations, setting up the equations of motion,
and using mode shape expansion, frequency responses for the two-beam model can be obtained.
For comparison, adapted from [4], Fig. 21.2 shows experimental reference results for a single M12x250 mm bolt tightened
in a single solid structure. Figure 21.2a shows the squared first transverse natural frequency as a function of tension, and Fig.
21.2b the corresponding damping ratio. The yield stress of the bolt is σ y = 640 MPa and σ is the actual axial stress.
For tensions exceeding about 20% of the yield stress, the experimentally obtained squared natural frequency starts to
increase linearly with tension. The damping ratio decreases with tension. Figure 21.3a shows the experimental setup for the
two-bolt test. Two M12 bolts in an aluminum structure, both mounted with accelerometers and excited by transverse hammer
impact at the other end of the bolt. A force washer on each bolt measures the tensions. Two cases are tested: when the two
bolts have close to equal tension, and when one is tight and the other looser. Figure 21.3b shows a 3D solid COMSOL-model
of the experimental setup; it can provide additional insights compared to the simple beam model. The same relation between
boundary stiffness and tension is used here as in the beam model, Eqs. (21.1) and (21.2), but the coupling due to the structure
is much closer to reality.
Fig. 21.1 Two-beam model, with boundary springs, axial tensions N1 and N2 and coupling springs
21 Vibration-Based Bolt Tension Estimation for Multi-bolt Joints 131
Fig. 21.2 (a) Measured first natural frequency squared as a function of tension for reference measurement. Three different test series indicated by
◦, , ♦. Asterisk markers: new measurements from two-bolt setup (cf. Table 21.1). (b) Damping ratio ζ as a function of tension
Fig. 21.3 (a) Experimental setup: Two bolts in an aluminum structure. Accelerometers mounted transversely on the bolt heads. Transverse hammer
hit on threaded bolt end. (b) For comparison: COMSOL 3D solid model. It is still a rough interpretation of the real setup, but a geometry much
more detailed than the beam model
21.3 Results
Figure 21.4 shows auto spectrums for two cases: one where the two bolts’ tensions are almost identical (upper spectrums,
47% of yield stress), one where one bolt is tight, and the other looser (lower spectrums, 70% and 49% of yield stress).
Left figures are response measured on Bolt 1, the one given the impact, the right figures are measured on Bolt 2, which
is only vibrating as a secondary effect. The beam model and COMSOL model is not tuned in the attempt to capture the
experimental response quantitatively, but is meant to be used just for qualitative comparison. Firstly, the figures illustrate
how the identification of the natural frequency corresponding to that of the impacted and measured bolt is not trivial, as
several peaks of the same order of magnitude appear in the same response. The multiple peaks arise from the response of
the other bolt, and since the bolts are cross sectionally symmetric, there can be peaks for each transverse direction. In this
case, the assumption is that the frequency with the largest magnitude is that of the impacted bolt (in the impact direction);
both models support this assumption (although the difference is not big). Table 21.1 shows the identified natural frequencies
and damping ratios. The damping ratio is generally small and the tighter the bolts the smaller the damping ratio, as for the
case of a single bolt. For the case where the bolts are far apart in tension, the first natural frequencies align with those in Fig.
21.2. As experimental mode shapes are not available, as they were in [5], those obtained by COMSOL and from the beam
model are used to decide on modes. COMSOL mode shapes predicts, for the case of almost equally tension in the bolts, that
the modes couple, which means that instead of two individual bending modes of each bolt, there are two coupled modes
(for each transverse direction): one with the bolts bending co-phase and one counter-phase. The first natural frequency, the
counter-phase mode, aligns with the results in Fig. 21.2. As the clamped structure is asymmetric, the co-phase motion has
132 M. Brøns et al.
Fig. 21.4 FRF for two bolts. Left: Measurement on Bolt 1 (the bolt hit); Right: Measurement on Bolt 2 (the bolt not hit). Upper: N1 = 25.5 kN,
N2 = 25.6 kN; Lower: N1 = 37.8 kN, N2 = 26.6 kN. Dashed line: experimental FRF; solid line: two-beam model; Dashed-dotted line: COMSOL
prediction
Table 21.1 Experimental natural frequencies and damping ratios for the two bolts at two different combinations of tension. There are two
transverse frequencies, since the bolts are axissymmetric; it is that in the direction of the hit. The hammer impact is on Bolt 1. CP/CTP: Co-
phase/counter-phase. Modes shapes are assumptions for the experimental frequencies on basis on the COMSOL predicted mode shapes
Case Bolt no. 1st natural frequency [Hz] Damping ratio [%] 2nd natural frequency [Hz] Damping ratio [%]
N1 = 25.5 kNN2 = 25.6 kN 1&2 946 CP1090 CTP 0.170.16 2539 CP2858 CTP 0.090.12
N1 = 37.8 kNN2 = 26.6 kN 1 1006 0.13 2630 0.08
2 963 0.14 2535 0.09
a significantly different frequency than the counter-phase. The beam model does not take account of this asymmetry, so the
co-and counter-phase mode predictions are the same.
21.4 Conclusion
Experiments performed with two bolts in a structure served to illustrate the challenges in estimating bolt tension from
vibrations with multiple bolts. A two-beam mathematical model showed that by simple modelling it is possible to capture
some of the tendencies found in experiments. The model cannot fit both first and second natural frequency with the
same physical parameters, so there are some significant features neglected. The COMSOL-model captures some of the
asymmetrical behavior, but a drawback of the model is that it is specific to that particular setup. For application purpose, it
is important that it is possible to identify the frequencies, and determine if each bolt has its own bending mode and unique
frequency. From these experiments alone it cannot be determined if the case with two bolts of almost equal tension have
coupled mode shapes; i.e. a mode shape includes bending of both bolts. The frequency of the assumed counter-phase mode
aligns with results for a single bolt, and if that is the true mode shape, it means that both bolts have the same frequency and
tension. To check that, measurements of mode shapes were advantageous. Furthermore, the coupling could also be sensitive
to the geometry and damping of the structure. The aluminum structure is only inducing very little damping; thus vibrations
can easily transmit from one bolt to the next. However, when the two bolts’ tensions are a little further apart, each bolt has
its own bending frequency, and it follows the same relation between tension and frequency as a single bolt in a structure.
Acknowledgement This work is financially supported by the Danish Council for Independent Research, grant DFF-6111-00385.
References
1. Bickford, J.H.: Introduction to the Design and Behavior of Bolted Joints – Non-Gasketed Joints, 4th edn. CRC Press (2007)
2. Joshi, S.G., Pathare, R.G.: Ultrasonic instrument bolt stress. Ultrasonics, (November), pp 270–274 (1984)
3. Nassar, S.A., Veeram, A.B.: Ultrasonic control of fastener tightening using varying wave speed. J. Press. Vessel. Technol. 128(3), 427 (2006)
21 Vibration-Based Bolt Tension Estimation for Multi-bolt Joints 133
4. Sah, S.M., Thomsen, J.J., Brøns, M., Fidlin, A., Tcherniak, D.: Estimating bolt tightness using transverse natural frequencies. J. Sound Vib. 431,
137–149 (2018). https://doi.org/10.1016/j.jsv.2018.05.040
5. Brøns, M., Thomsen, J.J., Fidlin, A., Tcherniak, D., Sah, S. M.: Modal impact testing for estimating bolted joint tightness. In Proceedings of
the International Conference on Noise and Vibration Engineering (ISMA2018), Sep. 17–19, 2018, Leuven, Belgium, 8.pp.
6. Brøns, M., Thomsen, J.J., Sah, S.M., Tcherniak, D., Fidlin, A.: Analysis of transient vibrations for estimating bolted joint tightness. In
Proceedings of International Modal Analysis Conference - IMAC 2019 (IMAC XXXVII), January 28–31, Orlando, Florida, USA, 1.pp.
Chapter 22
Experimental Application of Control-Based-Continuation
for Characterization of Isolated Modes on Single- and
Multiple-Degree-of-Freedom Systems
Gleb Kleyman, Martin Paehr, and Sebastian Tatzko
Abstract In this study, two dynamic systems are experimentally investigated using control-based-continuation (CBC) near
their resonant frequencies. One system is a single mass oscillator with a stiffness nonlinearity, the other system is a pair
of slim beams with a friction element in between and thus significant nonlinear damping. The focus of this study is on
the implementation of the CBC-method and the interpretation of the experimental results. The dynamic load-displacement
characteristics of both systems (s-curves) at different frequencies and the backbone-curves are presented.
Keywords Control-cased-continuation · Backbone-curve · Nonlinear damping · Nonlinear frequency response · s-curve
22.1 Introduction
Control-based-continuation (CBC), a recently published approach for the identification of nonlinear dynamic systems [1],
has aroused great interest in the experimental community. The key aspect of CBC is the characterization of the dynamic
load-displacement characteristic of nonlinear systems. However, the amplitude-dependent resonance frequency curves,
also known as backbone-curves, can also be determined experimentally with the CBC if they are combined with suitable
phase control [2]. In addition to amplitude-dependent damping curves and mode shapes, backbone-curves are used to
characterize nonlinear-normal modes, which is an extension of linear mode theory to nonlinear systems [3]. The CBC-
method is closely related to path following techniques for solving of nonlinear problems [4]. Details on the method and
different experimental applications are described in several papers e.g. [5–7]. A friction type nonlinearity has so far only
been investigated numerically [8]. However, friction damping has many applications, as for example in turbomachinery. The
dynamic properties of friction-damped systems can become particularly exciting if, in addition to nonlinear damping, there
is also a stiffness nonlinearity. Therefore, this study examines two systems using the CBC method: a weakly damped system
with pure stiffness nonlinearity and a system with a non-ideal friction contact between two slim beams.
22.2 Theoretical Background of Control Based Continuation
Nonlinear mechanical systems may have an ambiguous dynamic load-displacement characteristics. Due to their typical shape
in systems with stiffness nonlinearity, they are often referred to as s-curves. Figure 22.1 shows a schematic of such a s-curve,
with the unstable area highlighted in gray. In this region, the system will respond to a certain harmonic excitation force
by adjusting to either the value of the upper or lower branch (black), depending on the previous time history and initial
values. The red section marks unstable periodic orbits. The CBC-method exploits the s-curves experimentally in order to
characterize nonlinear systems near their resonance frequency by a two step procedure: first the system is stabilized by a
feedback-control as shown in Fig. 22.1 and then the control is made noninvasive. The CBC-method within this study is a
simplification of the full CBC-method. Therefore, only the amplitude of the reference displacement xref (t) is parameterized.
For details on this simplified CBC-method, see [5–7]. For a detailed description of the full method see [4]. Here we limit
ourselves to a brief presentation of the experimental results.
G. Kleyman () · M. Paehr · S. Tatzko

Institute of Dynamics and Vibration Research, Department of Mechanical Engineering, Leibniz University Hannover, Hannover, Germany
e-mail: [email protected]; [email protected]; [email protected]

136 G. Kleyman et al.
controller shaker + system
(a) (b)
Fig. 22.1 (a) Exemplary relationship between the amplitudes of the excitation force and displacement of a nonlinear system (fundamental
harmonic). (b) Feedback-control loop of the CBC-experiment
3 1
4
2
(a) (b)
Fig. 22.2 (a) Experimental setups of the duffing-oscillator. (b) S-curve (black), backbone-curve (red) and section plane along 0.12 A current
amplitude (gray)
22.3 Experiments
This study examined two different experimental setups. System (A) is a weakly damped single mass oscillator with Duffing-
type stiffness nonlinearity, see Fig. 22.2. System (B) consists of a pair of beams with a shared support and an asymmetrical
friction element (a so-called underplatform damper) in between the beams, see Fig. 22.3. The friction element is pressed
against the triangular platforms of the beams with constant normal force.
For the experiments, the CBC-method was implemented on a dSpace system. A voice-coil actuator was used for the excitation
of system (A) and a shaker for system (B). The displacement was measured with a laser triangulation sensor. For system
(B) the first out-of-phase mode was investigated. The s-curves (black dots) are measured step-by-step at fixed excitation
frequencies applying the CBC-procedure. Of course, sufficient time must be allowed between the single iteration steps for
the system to settle down. However, for the backbone-curves (red crosses) the frequency is set as a variable parameter. Using
a zero-point search (in our case the regula falsi method), the frequency is varied until a phase angle of π/2 between the
fundamental harmonics of excitation force and displacement is reached. The CBC-algorithm is run each time between the
frequency variations in order to get a noninvasive solution.
22 Experimental Application of Control-Based-Continuation for. . . 137
3 2
(a) (b)
Fig. 22.3 (a) Experimental setups of the friction-damped beams. (b) S-curve (black) and backbone-curve (red) measurements
System (A) shows the expected and already well known behavior. With increasing displacement amplitude the backbone-
curve is bent towards higher frequencies, as it is typical for a Duffing-oscillator. The gray section plane in Fig. 22.2 visualizes
the nonlinear frequency response at a constant forcing amplitude. It can be seen, that the stiffening of the system leads to
an overhanging frequency response, where the lower branch usually is unstable. With the feedback-control, however, every
single operating point can be monitored. The s-curve measurements of system (B) show that in the transition between sticking
and slipping motion there are oscillation regimes with an ambiguity in the load-displacement curve. Based on our experience
with system (A), we interpret this as an additional stiffness nonlinearity. After closer examination of the contact area, we
assume that the effect is caused by a clearance between the friction element and one of the platforms, which is leading to a
stiffening at higher displacement levels. However, the exact nature of this phenomenon is still under investigation. For this
reason, further tests with higher displacement resolution and different normal forces are carried out. Anyway, this example
shows that due to unforeseen contact conditions the real dynamics of friction-damped systems can become so complex that
an analysis with a stabilizing control such as CBC is absolutely essential.
22.4 Conclusion
We have successfully applied the CBC-method to two systems with different nonlinearities. The friction-damped system
surprised with a load-displacement characteristic which shows ambiguities. Such ambiguities usually mark unstable periodic
orbits, as shown by the example of the Duffing-type oscillator. They therefore indicate an additional stiffness nonlinearity.
This is probably caused by a clearance in the contact between the friction element and the beam. The investigations will be
continued.
References
1. Sieber, J., Krauskopf, B., Wagg, D., Neild, N., Gonzalez-Buelga, A.: Control-based continuation of unstable periodic orbits. J. Comput.
Nonlinear Dyn. 6(1), 011005 (2010)
2. Renson, L., Gonzalez-Buelga, A., Barton, D., Neild, S.: Robust identification of backbone curves using control-based continuation. J. Sound
Vib. 367, 145–158 (2016)
3. Kerschen, G., Peeters, M., Golinval, J., Vakakis, A.: Nonlinear normal modes, part I: a usefull framework for the structural dynamicist. Mech.
4. Schilder, F., Bureau, E., Santos, I.F., Thomsen, J.J., Starke, J.: Experimental bifurcation analysis – continuation for noise-contaminated zero
problems. J. Sound Vib. 358, 251–266 (2015)
138 G. Kleyman et al.
5. Barton, D.: Control-based continuation: bifurcation and stability analysis for physical experiments. Mech. Syst. Signal Process. 84(Part B),
54–64 (2017)
6. Renson, L., Shaw, A., Barton, D., Neild, S.: Application of control-based continuation to a nonlinear structure with harmonically coupled modes.
Mech. Syst. Signal Process. 120, 449–464 (2019)
7. Bureau, E., Schilder, F., Elmegard, M., Santos, I., Thomsen, J., Starke, J.: Experimental bifurcation analysis of an impact oscillator – determining
stability. J. Sound Vib. 333(21), 5464–5474 (2014)
8. Sieber, J., Krauskopf, B.: Control based bifurcation analysis for experiments. Nonlinear Dyn. 51, 365–377 (2008)
Chapter 23
Nonlinear Function Selection and Parameter Estimation
of Structures with Localised Nonlinearities, Part 1: Numerical
Analysis
Sina Safari and Julian M. Londono Monsalve
Abstract Although techniques for system identification of linear dynamical systems are well developed, for nonlinear
systems, it is still an open research question. Particularly, finding the correct form of the nonlinear functions (e.g. polynomial,
multi-linear) and their parameters for dynamic systems with unknown nonlinearities still represents a crucial aspect to be
addressed. Many methods have been presented in the literature to characterize and estimate the parameters of nonlinear
dynamical systems; yet, most of them require pre-assumption of the form of the nonlinear function. In order to overcome this
issue, in this study a data driven nonlinear autoregressive with exogenous inputs (NARX) model is considered together with
a polynomial basis of model terms. We present here the mathematical background and numerical simulations of the proposed
methodology. We use Forward Regression Orthogonal Least Square (FROLS) algorithm to select the terms in the NARX
model considered. FROLS is a recursive algorithm able to find the more significant model terms based on the Error Reduction
Ratio (ERR) criterion. We also explore ways to improve the selection of nonlinear terms using an optimization procedure
during which models corresponding to various candidate structures are estimated based on Prediction Error Method (PEM),
and the one providing the best fitness to the simulation data is selected. The procedure is motivated by the desire to quantify
the degree of nonlinearity of a system, with the ultimate goal of updating a finite-element or other mathematical model
to capture the nonlinear effects accurately. Single and multi-degrees of freedom systems that include a wide variety of
nonlinearities are considered in this study. The results show that selecting the nonlinear function based on ERR is not
affected by noise while the coefficient estimation depends on signal-to-noise ratio (SNR) value and the combination of
different nonlinear functions. Discussions are made on sufficiency and efficiency of various optimization tools dealing with
the coefficient estimation for MDOF system with localised nonlinearities.
Keywords Nonlinear system identification · Data driven model · Nonlinearity characterization · Nonlinearity
quantification · NARX model
23.1 Introduction
Mathematical modelling of structure is a main topic in the field of structural vibration. Reliable and accurate models would
enable the structural dynamist to control and optimize the performance of critical structures. In particular, predicting the
behaviour of structures containing nonlinearities has gained more interest in recent years. Nonlinear system identification
(NSI) tools have been developed focusing on detection, localization, characterization and quantification of nonlinearities in
structural systems [1]. NSI methods offer a better understanding of nonlinear systems, discover the type of the nonlinearities
and estimates their model.
Nonlinear function selection and parameter estimation for systems with complex nonlinear behaviour is still challenging
issue. In most of the NSI methods like conditioned reverse path (CRP) method [2], time domain nonlinear subspace
identification (TNSI) method [3], an ansatz is made as an educated guess for nonlinear function before doing the parameter
quantification that is verified later by examining it based on experimental results. Furthermore, backbone (BCBN) curve
estimated based on Resonance Decay Method (RDM) employs a curve-fitting approach to find out the coefficient using
Harmonic Balance (HB) method [4–6]. NSI requires an accurate characterisation of the non-linear forces present in a
system before any parameter estimation can be attempted. Nonlinear function selection and parameter estimation can be done
S. Safari · J. M. Londono Monsalve ()

College of Engineering, Mathematics and Physical Sciences, University of Exeter, Exeter, UK

140 S. Safari and J. M. Londono Monsalve
using various optimization algorithms which finds a proper describing function based on simulation data. As an alternative
approach in this study, we employed Forward Regression Orthogonal Least Squares (FROLS) algorithm to find the dominant
term at each consecutive step.
Unlike the RDM which uses the free decay response obtained after releasing the structure from resonance condition,
in this research we used the forced response of the system to estimate the nonlinear terms and coefficient that best fit
the system measured response. To do this, FROLS algorithm followed by NARX data-driven method [7] is used as an
input-output identification technique. The method is examined using various polynomial nonlinearities on a single-degree-
of-freedom system followed by its application to Multi-Degree-of-Freedom (MDOF) system. The objectives in this paper
are: (i) assessing the effects of SNR and various optimization approach for coefficient estimation (ii) finding out the unknown
nonlinearities from the library of terms (iii) implementing the DDM for MDOF systems.
23.2 Methodology
The NSI method considered in this study is the so-called nonlinear autoregressive models with exogenous inputs (NARX).
The NARX models are a special case of the more general class of nonlinear autoregressive moving average models with
exogenous inputs (NARMAX) [7] and they are defined in discrete time. Accordingly, the output y(k) on time instant k of a
NARX model is given by a nonlinear function F of the following expressions:

y(k) = F y (k − 1) , y (k − 1) , . . . , y k − ny ,
(23.1)
y (k − nk ) , y (k − nk − 1) , . . . , y (k − nk − nu ) + e(k)
In practice, several choices of the lag numbers ny , nu , and nk are tested to find a good model candidate for past outputs y
and past input u. Furthermore, the noise e in Eq. (23.1) is independently and identically distributed (i.i.d.) with zero mean,
finite variance and purely additive to the model. For the polynomial case the generic linear-in-parameters representation can
be used:
M
y(k) = θi pi (k) + e(k) (23.2)
i=1
where pi are the regressors formed by some combination of predetermined model terms which can be either static or
dynamical models [7] and θi are the coefficient for M regressors.
The current state-of-the-art method of identifying a NARX model, starting from input and output measurements, uses an
Orthogonal Least Squares (OLS) algorithm instead of Least Squares (LS) algorithm to choose which nonlinear terms should
be added in the model. Choosing some terms to add in the models, while leaving others out, is important to keep the NARX
model manageable and parsimonious. The OLS algorithm creates a set of orthogonal basis functions w1 , . . . ,wi , using
Gram-Schmidt algorithm and then chooses the orthogonal monomial wi which maximizes the so-called Error Reduction
Ratio (EER):
(y, wi )
EER i = (23.3)
(y, y)2 (wi , wi )2
This continues by adding more terms to the model from term library until a small enough model error is achieved. This
general search is called Forward Regression, and is the state-of-the-art NARX modelling technique known as FROLS [7].
The FROLS identification scheme yields usable NARX models, which often are sparse. The sum of error reduction ratios
(SERR) can be used as cut-off criteria for the term selection:
M
SEER = EER i (23.4)
i=1
The coefficient estimation can be carried out through the FROLS algorithm, however, the estimation accuracy will be poor
for noisy signals. Alternative methods used to enhance the estimation accuracy using prediction error minimization (PEM)
or simulation error minimization (SEM) in which the coefficients can be obtained using optimization algorithms. PEM uses
one step ahead prediction in the fitness function based on measured data. Whereas, SEM uses the simulated output signal in
the fitness function.
23 Nonlinear Function Selection and Parameter Estimation of Structures with Localised Nonlinearities, Part 1: Numerical Analysis 141
23.3 Numerical Analysis
The equation of motion in physical space for a dynamic system including stiffness non-linearity can be written as:
n
M q̈ + C q̇ + Kq + ρT f (ρi q) = F (23.5)
i=1 i
In this equation, q̈, q̇ and q are the acceleration, velocity and displacement vectors; M, C and K are the matrices of mass,
damping and stiffness, F is the force vector and ρ is a location vector of each n nonlinear element. f is the nonlinear function
which can be termed as describing function. Nonlinear term library assumed polynomials up to 7 order in this study: q|q|,
q3 , |q|q3 , q5 , |q|q5 , q7 .
23.3.1 SDOF System
Single Degree of Freedom (SDOF) numerical simulations with polynomial stiffness nonlinearities are considered to study
the capabilities of FROLS in NARX algorithm for selection and estimation of nonlinear parameters included in the model.
To address this, NSI is carried out by using a Data Driven Method (DDM) based on the best discretization of the Nonlinear
Differential Equation (NDE) model of the system. Xiaobiao Ge et al. [9] theoretically proved that the best discretization of
the NDE model to be used in the NARX is mixed difference approach i.e. backward-forward difference (BFD) and forward
backward difference (FBD). In the present study, we considered FBD approach.
q(ti )−q(ti −Δt)

q̇ = Δt
q(ti +Δt)−2q(ti )+q(ti −Δt)
(23.6)
q̈ = Δt 2
Here we compare different approaches to identify the nonlinear SDOF system. Firstly, we consider that the external force,
acceleration, velocity and displacement of the system are all available and the coefficients of NDE can be evaluated through
the curve-fitting process. Secondly, the NDE in Eq. (23.5) divided to the Linear Force (LF) and Nonlinear Restoring Force
(NRF) and tried to find the best static nonlinear terms for the NRF part using FROLS algorithm and curve-fitting used for
coefficient estimation. Thirdly, the capability of NARX in estimation of dynamic characteristics is assessed considering both
linear and nonlinear discretised parts.
The characteristics of SDOF examples (Fig. 23.1) considered here for numerical study is shown in Table 23.1 which
includes quadratic, cubic and quintic nonlinear terms.
Case 1: Cubic stiffness The first example used in this study is a cubic SDOF system characterized in [5]. The parameters of
the underlying linear system and nonlinearities included in Table 23.1 (Case 1). The response of Case 1 is considered to be
noise-free under harmonic excitation at the resonance condition. The sum of error reduction ratios (SERR) by adding each
nonlinear term to the model indicates that the cubic term is selected and provides a perfect description of the nonlinearity in
the system. The calculation is iterated for moving buffer through signals used for the identification.
Fig. 23.1 Single degree of

freedom system with stiffness
nonlinearity
Table 23.1 Characteristics of single degree of freedom dynamical systems

NL
Case m (kg) ξ K (N/m) f (Hz) k2 (N/m2 ) k3 (N/m3 ) k5 (N/m5 )
1 1.5 0.0063 6000 10.066 0 7e6 0
2 0.6874 0.0045 3.3e+04 34.87 −9.05e5 1.05e9 0
3 0.6874 0.0045 3.3e+04 34.87 0 0 1.8e13
4 0.6874 0.0045 3.3e+04 34.87 −9.05e5 1.05e9 −3.1e13
(a) (b) (c)
Fig. 23.2 Identified SDOF cubic example (case 1) (a) nonlinear force-displacement curve (b) backbone curve (c) stepped sine response at low
vibration level F = 0.2 N
Table 23.2 Identification results for SDOF system case 1: cubic stiffness
Identification results
System parameters RD method – NARX method
– – Value Error (%) Value Error (%)
M (kg) 1.5 1.5 0 – 1.505448 0.363175
C (N s/m) 0.8 0.8 0 0.804776 0.597028
K (N/m) 6000 6001.051 0.017512 6019.519 0.325322
K3 (N/m3 ) 7000000 6983302 0.238545 6979713 0.289818
Figure 23.2 presents the results of identification for SDOF cubic example using the NARX and RDM methods. The
results show a comparable and satisfactory performance of both procedures in terms of the estimation of the nonlinear
force-displacement response, backbone curve and stepped sine simulation. Only minor differences can be observed for the
backbone curve obtained from NARX estimation (Fig. 23.2b) which mostly present a good approximation of backbone curve
trend. The identification results also presented in Table 23.2 show that all the dynamic parameters of cubic system properly
identified using NARX method with small estimation error around 0.3%. In this case the RDM gives a relatively lower
estimation error for stiffness coefficients in comparison with NARX method.
Case 2: Quadratic and cubic stiffness The second case correspond to a SDOF system with a combination of quadratic and
cubic stiffness nonlinearities. Stepped-sine sweep simulation results is presented for different levels of excitation (F = {0.01,
0.05, 0.1, 0.25, 0.5, 1} N) in Fig. 23.3. The backbone curve extracted using RDM [5] is also superimposed onto the stepped-
sine simulation responses. For the sake of comparison, we examined many signals obtained from different frequency and
force levels in the stepped-sine responses to find out the terms describing the nonlinearity in the system. NRF calculated after
identifying the linear system and is used with corresponding displacement signal as inputs for the FROLS algorithm.
SEER and root mean square (RMS) selection criteria are presented in Fig. 23.4 for three piece of signals selected from
a pool of examined signals. Iterations done for buffers of signals in time and the order of selected terms is also present in
the Fig. 23.4. It can be seen that the free decay signal properly introduces the nonlinear terms, however, SEER indicates that
there is a minor contribution of another harmonic |(y − 1)|(y − 1)5 in some part of the time domain signal. Another signal
is used from high force level (A) responses as shown in Fig. 23.3 indicates that there is four term contributing to the NRF
signal: |(y − 1)|(y − 1)3 , (y − 1)7 , |(y − 1)|(y − 1), (y − 1)3 . Further, a response from medium force level (F = 0.25 N and
f = 34.8 Hz) is considered for term selection which provides a better representation of the nonlinear terms. Therefore, by
adding each term to the prediction process SEER value convergence to 1.
Furthermore, a sweep sine simulation is carried out for F = 1 N and f = 34–38 Hz. Sweep sine displacement response
is presented in Fig. 23.5 with the selected terms for every buffer of response. As it can be seen, the terms selected for
lower displacement amplitude are the representative true nonlinear terms and as the amplitude grows due to the nonlinear
normal mode resonance the higher order polynomial terms are ranked based on EER. Therefore, it should be noted that the
nonlinearities activated in the lower amplitude should be taken into account carefully.
Following the terms selection process, NARX algorithm employed for coefficient estimation. Figure 23.6a presents the
results of backbone curves for identified systems using different selected terms. It can be seen that the nonlinear terms
Fig. 23.3 Stepped-sine sweep simulation results for case 2: quadratic and cubic stiffness (A-high force level: F = 1 N and f = 36.8 Hz, B-medium
force level: F = 0.25 N and f = 34.8 Hz)
Fig. 23.4 SEER and RMS values for term selection using free decay, high force level (A) and medium force level (B) signals for case 2: quadratic
and cubic stiffness
selected and estimated based on the signal obtained from higher force level at the resonance condition is satisfactory only
for that part of the backbone curve. On the other hand, the terms reflected based on the signal from medium force level are
well-compatible with the true backbone curve. The stepped-sine simulations of the identified system based on medium force
Fig. 23.5 Terms selection based on sweep sine response for SDOF – case 2
Fig. 23.6 Comparison of the identified systems for case 2 (a) Backbone (BCBN) curves (b) Stepped-sine simulation
level is shown Fig. 23.6b in which a good agreement observed. Base on the different signals examined in this part, a response
signal of a medium force level at the resonance condition for a specific system results in a better term selection based on
FROLS algorithm.
To discuss the effects of noise on the term selection and coefficient estimation, three levels of SNR = 40, 80 and 100 dB
considered in this study. It has been observed that term selection is not effected by noise and FROLS properly delivers
the nonlinear terms. As it can be seen from Fig. 23.7, simple curve-fitting process [4] used for coefficient estimation of all
parameters in nonlinear dynamic system is failed when we add some noise to the output responses. It should be noted that the
normalized signals based on each buffer amplitude is used for coefficient estimation and Fig. 23.7 presents the normalized
coefficients. Accordingly, NARX method used for coefficient estimation of nonlinear dynamic system using PEM and SEM
algorithms. The estimation process initiates with PEM and proceeds with SEM to minimize the error for noisy signals.
The results are reported for noise-free and noisy signals obtained for case 2 example in Table 23.3. It can be seen that the
estimation error for the noise free signal is less than 1% and for signals with SNR = 40 and 80 dB the error is in the range
of 29% and 25% respectively for quadratic term. It is evident that NARX method based on SEM demonstrated superior
performance in coefficient estimation of polynomial nonlinear systems.
The optimization algorithms embedded in the optimization toolbox of MATLAB package is considered herein for
coefficient estimation. Basically, a weighted sum of squares of the errors is used as a loss function. This may be minimized
using a quasi-newton algorithm such as the Gauss-Newton or Levenberg-Marquardt algorithm [10]. Regardless of whether a
classical quasi-Newton algorithm, or a separable least square based approach is used, the result is a non-convex optimization
which must be solved iteratively. To achieve a satisfactory estimation, two practical point should be noted: (a) the input
Fig. 23.7 Normalized parameter estimation for noisy case based on approach A (a) SNR: 40 dB (b) SNR: 80 dB (c) SNR: 100 dB
Table 23.3 Identification results for SDOF system case 2: quadratic and cubic stiffness
System parameters Noise free – SNR: 40 dB – SNR: 80 dB
– – Value Error (%) Value Error (%) Value Error (%)
M (kg) 0.6874 0.6884 0.15 – 0.6972 1.43 – 0.6839 0.5
C (N s/m) 1.35551 1.3665 0.81 1.3377 1.32 1.3376 1.32
K (N/m) 3.3e+04 33033 0.1 33604 1.83 32972.83 0.08
K2 (N/m2 ) −9.05e5 −899754 0.58 −1.17E+6 29.25 −1.136e+6 25.523
K3 (N/m3 ) 1.05e9 1E+09 0.57 1.144E+9 9.02 1.1165e+9 6.3328
Fig. 23.8 Stepped-sine sweep simulation results for case 3: quintic stiffness with estimated backbone (BCBN) curves
signal should be like that the dominant nonlinearities in the system overall response for various operational conditions (low
amplitude and extreme operational condition) sufficiently contributed (b) avoid any sudden change like jumps in the response
or divide them to various operating conditions for coefficient estimation.
Case 3: Quintic stiffness The third example used in this study for SDOF system includes a quintic stiffness nonlinearities.
The stepped-sine sweep simulation result for force levels F = {0.05, 0.1, 0.25, 0.5, 1} N is shown in Fig. 23.8. The higher
order nonlinear terms can be considered for the systems with a sudden change in the frequency in a specific deformation
amplitude. Accordingly, the terms selection carried out using a forced response for medium force level and also sweep
sine response. The quantic term is perfectly selected as the only polynomial term contributing to the system nonlinear
Fig. 23.9 SEER and RMS values for term selection in case 4: quadratic, cubic and quintic stiffness (medium force level: F = 1 N and f = 35.5 Hz)
(a) (b)
Fig. 23.10 Identification results for case 4: quadratic, cubic and quintic stiffness nonlinearities (a) displacement response (b) estimated backbone
curves and stepped sine simulation
response. The estimated backbone curve extracted using RDM and NARX is demonstrated in Fig. 23.8 for the validation of
identification.
Case 4: Quadratic, cubic and quantic stiffness The forth example used in this study for SDOF system contains quadratic,
cubic and quintic stiffness nonlinearities. A harmonic medium force level excitation is used for term selection in case 4 SDOF
system. Both of the SEER and RMS measures presented in Fig. 23.9 indicates a proper selection of polynomial nonlinearities
(y − 1)3 , |(y − 1)|(y − 1), (y − 1)5 .
Sweep sine excitation between 33 and 38 Hz with amplitude of 2 N is used for this example to generate the most useful
information of nonlinearity in the operational boundary of system. Estimated coefficients using NARX method are used to
compare the displacement response of simulated (true) and identified SDOF system (case 4) in Fig. 23.10a. The results show
that the fitted model is able to predict the response of the true structural model with reasonable accuracy as the goodness-
of-fit reported for the identified system is 98.52%. To further validate the identified system, the stepped sine simulation for
several force levels F = {0.01, 0.05, 0.1, 0.25, 0.5, 1, 2} N is overlaid with estimated backbone curves in Fig. 23.10b. The
estimated backbone curves closely represent the nonlinear behaviour, nevertheless, the minor discrepancy can be attributed to
the errors in the coefficient estimation of nonlinear terms especially quadratic term error which is about 15%. The estimated
coefficients for the identified SDOF system are also reported for SNR = 40 and 80 dB values in Table 23.4.
Table 23.4 Identification results for SDOF system case 4: quadratic, cubic and quintic stiffness
System parameters Noise free – SNR: 40 dB – SNR: 80 dB
– – Value Error (%) Value Error (%) Value Error (%)
M (kg) 0.6874 0.6853 0.3 – 0.618142 10.1 – 0.6853 0.3
C (N s/m) 1.35551 1.3781 1.7 1.455183 7.4 1.3781 1.7
K (N/m) 3.3e+04 32821.88 0.54 29442.62 10.8 32819.87 0.5
K2 (N/m2 ) −9.05e5 −7.7E+05 14.95 −556955 38.5 −7.7E+05 15.1
K3 (N/m3 ) 1.05e9 9.95E+08 5.22 9.91E+08 5.7 9.9E+08 5.3
K5 (N/m3 ) −3.1e13 −2.9E+13 5.16 −3.4E+13 11.2 −2.94E+13 5.2
Fig. 23.11 Three DOF systems. Four cases consider different locations of the nonlinear elements
Table 23.5 Characteristics of three degree of freedom nonlinear dynamical systems

Nonlinear coefficients
Case m (kg) ξ k (N/m) f (Hz) Location vectors (ρ i ) k2 (N/m2 ) k3 (N/m3 )
1 0.6801 0.0045 1.6386e+04 f1 = 18.9072 [1 0–1] −9.05e5 1.05e9
2 0.6801 0.0045 1.6386e+04 f2 = 34.9359 [0 1 0] −9.05e5 1.05e9
3 0.6801 0.0045 1.6386e+04 f3 = 45.6460 [1 0 0] −9.05e5 1.05e9
4 0.6801 0.0045 1.6386e+04 [1 0 0; 1–1 0;0 1–1; 0 0 1] −9.05e5 1.05e9
23.3.2 Application to MDOF System
A nonlinear MDOF system is now examined. The aim is to show how the procedure introduced above can be extended to
system identification of nonlinear MDOF systems. This can be accomplished by applying the proposed procedure either
to responses of individual masses in physical coordinates or to individual modal coordinates after projecting the system
responses into a linear modal space.
For proportional damping the modal mass and linear modal stiffness terms are uncoupled and the linear modes may be
coupled nonlinearly by additional terms in the system equations [8]. It was considered that the use of a model based up on
modal (or possibly mixed physical/modal) space is arguably a candidate for being an identification algorithm [4–6].
A 3DOF lightly damped system is considered in this study with nonlinear elements localised in different positions as
shown in Fig. 23.11. Quadratic and cubic stiffness nonlinearity of the form k2 q|q| + k3 q3 , is considered in all four cases
presented below. Case 1, Case 2 and Case 3 consider one nonlinear element while Case 4 considers four identical nonlinear
elements. The characteristics of MDOF examples in this study are presented in Table 23.5.
Equation (23.7) presents the equation of motion for 3DOF example. The damping matrix is calculated assuming modal
damping of 0.0045 in all modes.
⎡ ⎤⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
100 q̈1 q̇1 2 −1 0 q1 F1
⎣
m 010 ⎦ ⎣ ⎦ ⎣ ⎦ ⎣
q̈2 + C q̇2 + k − 1 2 −1 ⎦ ⎣ q2 + N RF = F2 ⎦
⎦ ⎣ (23.7)
001 q̈3 q̇3 0 −1 2 q3 F3
If the equation of motion is expressed in the modal coordinates, ui , using the matrix of mode shapes, Φ, we can write:
Fig. 23.12 SEER and RMS values for term selection using medium force level (F = 0.25, f = 34.936 Hz) signal
Fig. 23.13 Stepped-sine sweep simulation results obtained from driving point for 3DOF system with quadratic + cubic stiffness nonlinearities
ü1 + 2ζ1 ωn1 u̇ + ωn21 u1 + fnl1 (u1 , u2 , u3 ) = φ1 T F

ü2 + 2ζ2 ωn2 u̇2 + ωn22 u2 + fnl2 (u1 , u2 , u3 ) = φ2 T F (23.8)
ü3 + 2ζ2 ωn2 u̇3 + ωn23 u3 + fnl3 (u1 , u2 , u3 ) = φ3 T F
here, ωni is the i-th natural frequency of the underlying linear system, ξ i is the damping ratio and fnli is the nonlinear function
in term of the interacting modal coordinates at the i-th resonance. The nonlinearity is assumed to be polynomial of the form
fnli = ai (c1 u1 + c2 u2 + c3 u3 ) |(c1 u1 + c2 u2 + c3 u3 )| + bi (c1 u1 + c2 u2 + c3 u3 )3 + . . . (23.9)
where ai , bi are the coefficients of the i-th nonlinear function to be estimated and ci are available from ρ i Φ.
The term selection process is carried out for the 3DOF examples using a harmonic force at medium force level (F = 0.25,
f2 = 34.936 Hz). The RMS values presented in Fig. 23.12 for 3DOF cases which shows that the quadratic and cubic terms
fulfil the nonlinearity in the system.
The results of stepped-sine sweep simulation for 3DOF cases are presented in different levels of excitation (F = {0.125,
0.25, 0.5, 1} N) in Fig. 23.13. The localized nonlinearities in the MDOF cases are considered to generate nonlinear effects
at different resonances. For instance, Case 1 with one nonlinear element between masses 1 and 3 generates nonlinear effects
in the mode 2 only. Sweep up and down carried out and the backbone curves extracted from RDM (blue dashed lines) and
estimated using NARX (purple dotted lines) is also shown against stepped-sine simulations. It can be seen that the predicted
model based on NARX is able to represent the nonlinearity effects on the frequency response functions (FRFs). For further
comparison, Table 23.6 shows the estimated coefficient using NARX method. It should be noted that less than 5% error is
observed for estimated nonlinear coefficients.
Table 23.6 Estimated nonlinear coefficients for 3DOF cases

System
parameters Identification results
Case 1 – Case 2 – Case 3 – Case 4
– – Value Error (%) Value Error (%) Value Error (%) Value Error (%)
K2 (N/m2 ) −9.05e5 −8.5E+05 6.25 – −8.8E+05 2.3 – −9.50E+05 4.9 – −8.79E+05 2.9
K3 (N/m3 ) 1.05e9 1.02E+09 2.54 1.0E+09 1.0 1.05E+09 0.2 9.90E+08 5.7
23.4 Conclusion
This paper discusses a method for the dynamic identification of structures containing localised nonlinear stiffnesses.
The approach discussed herein comprise two steps: nonlinearity characterization based on the FROLS algorithm and
quantification based on NARX method respectively. Numerical examples for SDOF and MDOF systems with several
polynomial terms is considered in this study. The structure excited using harmonic force in a single frequency or sweep-
sine simulation, at a time, enabling measurements to be made in larger displacement regimes where nonlinearities are more
likely to be observed. Time domain data alongside modal equation of motion, that include nonlinear modal interaction terms,
enables identification of both linear and nonlinear parameters. We have discussed the use of appropriate signal to be used
for nonlinear term selection and coefficient estimation. It was observed that the nonlinear terms selected based on EER and
RMS criteria using medium harmonic force level and free decay response from resonance condition provides most accurate
information of nonlinear terms. From the practical point of view, terms selection is not affected by reasonable amount
of output noise and the optimization algorithms used for coefficient estimation performance well in case the excitation used
activates nonlinearities in different operating levels. A sweep sine excitation for medium to high force levels is recommended
for more efficient and accurate coefficient estimation.
Acknowledgment Mr. Safari is support by a PhD scholarship from the College of Engineering, Mathematics, and Physical Sciences, University
of Exeter which is gratefully acknowledged.
References
1. Noël, J.P., Kerschen, G.: Nonlinear system identification in structural dynamics: 10 more years of progress. Mech. Syst. Signal Process. 83,
2–35 (2017)
2. Richards, C.M., Singh, R.: Identification of multi-degree-of-freedom non-linear systems under random excitation by the “reverse path” spectral
method. J. Sound Vib. 213(4), 673–708 (1998)
3. Marchesiello, S., Garibaldi, L.: A time domain approach for identifying nonlinear vibrating structures by subspace methods. Mech. Syst. Signal
Process. 22, 81–101 (2008)
4. Platten, M.F., Wright, J.R., Dimitriadis, G., Cooper, J.E.: Identification of multi-degree of freedom non-linear systems using an extended modal
space model. Mech. Syst. Signal Process. 23, 8–29 (2009)
5. Londoño, J.M., Neild, S., Cooper, J.E.: Identification of backbone curves of nonlinear systems from resonance decay responses. J. Sound Vib.
348, 224–238 (2015). https://doi.org/10.1016/j.jsv.2015.03.015
6. Londoño, J.M., Cooper, J.E., Neild, S.A.: Identification of systems containing nonlinear stiffnesses using backbone curves. Mech. Syst. Signal
Process. 84, 116–135 (2017)
7. Billings, S.: Nonlinear System Identification: NARMAX Methods in the Time, Frequency and Spatio-Temporal Domains. Wiley, Chichester
(2013)
8. Al-Hadid, M.A., Wright, J.R.: Developments in the force-state mapping technique for non-linear systems and the extension to the location of
non-linear elements in a lumped-parameter system. Mech. Syst. Signal Process. 3(3), 269–290 (1989)
9. Xiaobiao, G., Luo, Z., Ma, Y., Liu, H., Zhu, Y.: A novel data-driven model based parameter estimation of nonlinear systems. J Sound Vib. 453,
188–200 (2019)
10. Ljung, L.: System Identification: Theory for the User, 2nd edn. Prentice Hall, Upper Saddle River (1999)
Chapter 24
Tutorial on Nonlinear Reduced Order Modeling for Nominally
Cyclic Symmetric Structures and Rotating Machinery
Mainak Mitra, Andrea Lupini, Andrew Madden, Chiara Gastaldi, and Bogdan Epureanu
Abstract Many important engineering structures such as rotating machinery, including turbine bladed disks, gears,
flywheels and satellites are comprised of repeated (nominally identical) substructures arranged circumferentially with cyclic
symmetry. Due to this unique arrangement, the system matrices and consequently the dynamics of such structures exhibit
specific characteristics (Mitra and Epureanu, ASME Appl Mech Rev. https://doi.org/10.1115/1.4043083, 2019; Olson et al.
ASME Appl Mech Rev, 66(4):040803, 2014). Extensive scientific study and analysis has been conducted on this topic in
recent decades. Of particular interest is the change in dynamic behavior when there are deviations in substructures from their
nominal, even to a small extent. Colloquially termed mistuning, such deviations are practically impossible to avoid. They
manifest as material or geometric differences due to causes such as manufacturing tolerances, wear and differential operation
conditions (Castanier and Pierre, J Propuls Power 22/2:384, 2006). Mistuning can lead to strain energy localization, higher
system responses and reduction of the operational life cycle and should therefore be carefully considered in the design and
analysis of structures. The current industrial practice is to use Monte Carlo simulations to characterize mistuning effects
using randomly generated deviations in substructures of the nominal design (Mitra and Epureanu, ASME Appl Mech Rev.
https://doi.org/10.1115/1.4043083, 2019; Castanier and Pierre, J Propuls Power 22/2:384, 2006). Since thousands of dynamic
simulations might be required to characterize a single design, full order high fidelity models remain prohibitively expensive.
For such tasks, reduced order models (ROMs) are employed instead (Castanier and Pierre, J Propuls Power 22/2:384, 2006;
Baek and Epureanu, ASME J Vib Acoust, 139(6):061011, 2017). However, obtaining fast and accurate ROMs for cyclic
structures with nonlinearities (Mitra and Epureanu, ASME Appl Mech Rev. https://doi.org/10.1115/1.4043083, 2019; Baek
and Epureanu, ASME J Vib Acoust, 139(6):061011, 2017; Zucca and Firrone, J Sound Vib, 333:916–926, 2014) remains a
challenging task. This tutorial aims at summarizing and highlighting some of the most relevant techniques that have been
proposed to date, with a specific focus on nonlinear ROMs including contact nonlinearities.
Keywords Cyclic symmetry · Reduced order models · Mistuning · Nonlinear dynamics · Rotordynamics
24.1 Cyclic Symmetry and Constraint Equations
The dynamics of a cyclic symmetric structure may be completely captured by establishing displacement constraint equations
between degrees of freedom of high (H) and low edges (L) of the structure as shown in Fig. 24.1 (a). Due to symmetry, the
displacements at the edges are related by a phase difference, and the familiar dynamic equation corresponding to the full
M. Mitra ()
ANSYS Inc, Canonsburg, PA, USA
A. Lupini
A. Madden
ANSYS Inc, Ann Arbor, MI, USA
C. Gastaldi
Politecnico di Torino, Torino, Italy
B. Epureanu

152 M. Mitra et al.
(a) (b) (c)

Fig. 24.1 (a) Finite element models (b) Frequency vs Nodal Diameter plot for a turbine blisk (c) Campbell diagram (generated using ANSYS® )
model in Eq. (24.1) may be reduced to the order of number of DOFs in a single sector via a transformation [1, 2] to the cyclic
domain as shown in Eq. (24.2):
Mq̈ + Cq̇ + Kq = f; q = [q1 , q2 , . . . qN ]T (24.1)
2π T
qH = Re qL eipα ; α = ; Ms q̈s + Cs q̇s + Ks qs = fs ; qs = qL , qI (24.2)
N
where M, C and K represent system mass, damping and stiffness matrices respectively. q and f represent the generalized
coordinates and excitation force vectors. N is the number of sectors, α is the sector angle. Subscript s represents the sector
level constrained cyclic DOFs (L, I). The phase difference imposed by the cyclic constraints arises from a spatial Fourier
decomposition of harmonic p. Since the dynamics represented by each spatial harmonic is independent, the full (N×DOFs )
system reduces to N independent systems of size DOFs . Moreover, simulating one or two of these independent systems
corresponding to the dominant harmonic in the excitation is typically sufficient to obtain predictions when linear effects
are predominant in the structure. This is especially true in rotating stages of turbines whose impinging pressure wave
has a dominant spatial harmonic imparted by the upstream static stage (stator) comprised of cyclically symmetric vanes.
Modal analysis may also be carried out in the cyclic domain. The effect of the Fourier space spanned by the modes of a
cyclic symmetric structure are evident in the modes, which may exhibit a certain number of nodal diameters along which
displacements are zero. Modes of spatial harmonics (except for p = 0, N/2 when applicable) are commonly paired by identical
natural frequencies leading to characteristic Frequency vs. Nodal Diameter plots of several high-density modal regions as
shown in Fig. 24.1(b). Rotating structures which tend to be cyclic symmetric are commonly affected by spin softening,
Coriolis and rotational prestress effects. Consequently, the natural frequencies are dependent on the rotational velocity, an
effect usually analyzed using Campbell diagrams shown in Fig. 24.1(c).
24.2 Mistuning and Probabilistic Simulations
The breaking of perfect cyclic symmetry, most commonly due to geometric or structural variations between sectors, is
referred to as mistuning [1, 3]. Mistuning is judged to be ‘small’ when the deviation from the cyclic baseline is significant
for natural frequencies but not for the modes. When modes are affected significantly, the mistuning is judged to be ‘large’.
Even small mistuning can affect the system response significantly as it couples the various spatial harmonics and a linear
response is a superposition of multiple modes in a narrow frequency region. The uncertainty in the underlying parametric
variations make exact response predictions difficult and Monte Carlo probabilistic simulations are used to characterize the
expected response amplification factor (AF) over the nominal cyclic case as shown in Fig. 24.2(a). Further time savings
may be realized in probabilistic analysis by fitting available simulation response data to the expected Weibull distribution as
shown in Fig. 24.2(b), thereby predicting the statistics with relatively fewer simulations [1].
24 Tutorial on Nonlinear Reduced Order Modeling for Nominally Cyclic Symmetric Structures and Rotating Machinery 153
(a) (b)
Fig. 24.2 (a) Nonlinear Response for tuned and mistuned analysis of a shrouded blisk (b) Fitting AFs from 50 mistuned responses to a Weibull
distribution (compared to distribution from 1000 cases)
24.3 Nonlinearities
Friction and contacts are the most commonly encountered nonlinearities in cyclic structures such as bladed disks,
although their effects vary significantly based on their location and the components affected [4]. For instance, the relative
displacements and normal loads may be different by orders of magnitude for a blade and disk joint when compared to
contact between a rotor and its casing. Usually, steady state nonlinear amplitude responses are of interest and are modeled by
applying harmonic balance (HB) to the time-domain equations to yield a multi-harmonic frequency domain representation
as follows:
h h
−h2 ω2 M + ihωC + K qh = fE + fN (24.3)
h
where [.] represents the hth temporal harmonic in the frequency domain and subscripts E and N represent the applied
excitation and nonlinear forces respectively. Contact forces which depend on relative displacements and internal slip variables
cannot be calculated a-priori or expressed directly in the frequency domain. Consequently, an alternating frequency time
h
(AFT) approach is used. At each simulation iteration, fN is determined by applying an inverse Fourier transform to its
time domain periodic solution, which in turn is obtained from time-domain displacements calculated by applying forward
Fourier transforms to system displacements. Other nonlinearities such as material nonlinearities (coatings), and multi-physics
effects such as contact heat generation and temperature effects are also commonly observed, especially in gas turbine rotors.
However, these are beyond the scope of this tutorial whose focus is on contact-related nonlinearities.
24.4 Reduced Order Modeling
Even when linear effects are dominant, reduction of cyclic systems is challenging due to mistuning. Substructuring using
methods such as component mode synthesis (CMS) [1, 3] is a strategy which reduces a system by partitioning it into
components and using the interface DOFs between these components to apply appropriate constraints, while reducing
internal DOFs of the components, typically via modal truncation [1]. A CMS reduction followed by HB analysis is often used
for harmonic analyses of bladed disks with localized nonlinearities such as contact interfaces as illustrated in Fig. 24.3(a).
Component mode mistuning (CMM) [1, 3] is a method which adapts CMS to mistuned cyclic structures, treating the baseline
tuned structure and the mistuning (often parametrized as deviations in stiffness or mass in blades) as individual components
as illustrated in Fig. 24.3(b). For small mistuning, the span of tuned and mistuned modes is assumed to be identical. This
allows reduction and projection of the mistuned quantities back onto the Fourier space of the tuned blisk, resulting in
computationally efficient simulations.
The nonlinearities representing friction do not have a closed form representation and thus are not conducive to techniques
such as polynomial expansion [1, 4]. Other general techniques for generating nonlinear ROMs, such as proper orthogonal
modes, nonlinear normal modes prove to be too computationally expensive for practical applications. Instead, nonlinear
154 M. Mitra et al.
(a) (b)
Fig. 24.3 (a) CMS followed by HBM for simulating nonlinear dynamics of a shrouded blisk (b) Schematic showing system partitioning for CMM
ROMs for bladed disks commonly focus on approximating nonlinear subspaces representing dynamics specific to the nature
of nonlinearity (stick-slip contact, frictionless chatter, amplitude-dependent coatings). Linear reductions are often leveraged
in addition to nonlinear reductions, since usually only a portion of the structure such as the blade, shrouds or coatings have
significant nonlinearities. Some notable nonlinear ROMs discussed in this tutorial are based on bilinear modes (BLMs),
piecewise linear modes (PLMs) [1], adaptive microslip projections (AMP) [5] and equivalent energy single harmonic
approximations [6].
24.5 Conclusions
The aim of this tutorial is to present the ideas and techniques commonly employed in dynamic modeling and simulation of
cyclic symmetric structures, leading up to a discussion of the state-of-the-art reduced order models used for fast-generation of
nonlinear ROMs designed for probabilistic analyses. The tutorial assumes basic knowledge of dynamics, but no prior specific
knowledge regarding cyclic symmetric structures. It introduces the participants to concepts unique to the field such as engine
order excitation, frequency-nodal diameter plots, block-circulant matrices and amplification factor. It also discusses dynamic
reduction techniques for various types of nonlinearities in these structures such as Coulomb friction and amplitude-dependent
nonlinearities in the time-domain (harmonic-balance) as well as the spatial domain (projection-based ROMs).
References
1. Mitra, M., Epureanu, B.I.: Dynamic modeling and projection-based reduction methods for bladed disks with nonlinear frictional and intermittent
contact interfaces. ASME. Appl. Mech. Rev. (2019). https://doi.org/10.1115/1.4043083
2. Olson, B.J., Shaw, S.W., Shi, C., Pierre, C., Parker, R.G.: Circulant matrices and their application to vibration analysis. ASME. Appl. Mech.
Rev. 66(4), 040803 (2014). https://doi.org/10.1115/1.4027722
3. Castanier, M.P., Pierre, C.: Modeling and analysis of mistuned bladed disk vibration: current status and emerging directions. J. Propuls. Power.
22/2, 384–396 (2006)
4. Zucca, S., Firrone, C.M.: Nonlinear dynamics of mechanical systems with friction contacts: coupled static and dynamic multi-harmonic balance
method and multiple solutions. J. Sound Vib. 333, 916–926 (2014)
24 Tutorial on Nonlinear Reduced Order Modeling for Nominally Cyclic Symmetric Structures and Rotating Machinery 155
5. Mitra, M., Zucca, S. Epureanu, B.I.: Adaptive microslip projection for reduction of frictional and contact nonlinearities in shrouded blisks.
Journal of Computational and Nonlinear Dynamics (2016). https://doi.org/10.1115/1.4033003
6. Baek, S., Epureanu, B.: Reduced-order modeling of bladed disks with friction ring dampers. ASME. J. Vib. Acoust. 139(6), 061011 (2017).
https://doi.org/10.1115/1.4036952
Chapter 25
Excitation Techniques for Nonlinear Dynamic Systems:
A Summary
Mahesh Nagesh, Akhil Sharma, Randall J. Allemang, and Allyn W. Phillips
Abstract Experimental analysis of nonlinear dynamic mechanical systems is a challenging task. Although well established
and widely accepted techniques for Modal Parameter Estimation (MPE) of structures within the linearity assumption exist,
several challenges arise once these assumptions cease to be valid i.e. nonlinear systems. Foremost of these challenges in
analyzing nonlinear systems is the excitation technique used. While several traditional and non-traditional techniques are
available for exciting these systems, each has its own merits and demerits during experimentation and post-processing. This
paper summarizes some of these techniques applied to a physical structure and provides a comprehensive discussion on
the results obtained. The excitation techniques include impact hammer, electrodynamic shakers, pneumatic system and step
relaxation.
Keywords Nonlinear systems · MIMO excitation techniques · Shaker excitation · Pneumatic excitation · Backbone
curves
25.1 Introduction
Detection, identification and characterization of nonlinear dynamics in structures is very challenging. Although well-
established Modal Parameters Estimation (MPE) [1, 3, 7] techniques exists for structures behaving under a linear assumption,
nonlinear systems have no such generalized methodologies that can be universally applied [8–10, 12]. The greatest among
challenges in experimental analysis of nonlinear systems is in controlling the various input parameters that widely affect the
response of a system. In a linear MPE process, effects of test setup/equipment on structures, other transients and variances
that affect the test structures are considered negligible; situations are often encountered where such effects have profound
consequences thus emphasizing the study of nonlinear systems in great detail.
Excitation techniques, i.e. hardware, associated controls and data processing techniques are important for nonlinear
systems. MPE on linear structures are usually performed using several well-established excitation techniques often in a
Multiple Input Multiple Output (MIMO) [2, 3] configuration. Linear MPE using a MIMO configuration is possible due
to small amplitudes of vibrations and Maxwell’s reciprocity among several assumptions being adherent to the systems.
When large amplitudes of vibrations are encountered, such as those discussed in this paper, one or more of these
assumptions of linearity in behavior are violated. Such structures require special and careful considerations during testing
and characterization. In addition, one must be aware of general erroneous results that can be encountered when several
techniques established for linear structural systems are applied to structures that exhibit nonlinear behavior. This paper
describes some popular and novel excitation techniques for nonlinear dynamic systems and typical associated response
characteristics, and particular interest in methods that can be applied to MDOF and MIMO situations rather than just SDOF
and SISO situations since MIMO/MDOF approaches have proven very advantageous in evaluating linear modal parameters
and similar advantages are expected when evaluating nonlinear systems.
M. Nagesh () · A. Sharma · R. J. Allemang · A. W. Phillips

Department of Mechanical and Materials Engineering, College of Engineering and Applied Sciences, University of Cincinnati, Cincinnati, OH,
USA

158 M. Nagesh et al.
25.2 Test Structure
A test structure shown in Fig. 25.1 was setup at the University of Cincinnati, Structural Dynamics Research Laboratory
(UC-SDRL) [6, 11] to perform the various tests detailed in this paper. This structure comprises of a cantilever beam and
a double-clamped beam of properties gives in Table 25.1. The free end of the cantilever beam is rigidly attached to the
double-clamped beam at its center such that the combined structure behaves as a single entity. It must be noted that the
cantilever beam has a very large flexural rigidity than the double clamped beam. For small loads and associated deflections,
the system behaves similar to a linear structure but exhibits a cubic hardening stiffness type nonlinearity for very large loads.
A static load versus displacement profile is obtained and a polynomial fit performed in both directions of loading as shown
in Fig. 25.2. A slight bilinearity (difference in slopes) is observed for the forward and backward loading profiles that is
attributed to imperfections in the rigid joint connecting the two beams. Since the magnitude of this difference in slopes is
much smaller than the other terms contributing to the stiffness at the center, the system can be largely classified as exhibiting
cubic hardening nonlinearity. The system can also be approximated as a Single Degree of Freedom (SDOF) type system
with a cubic nonlinear spring. Discussions in this paper are made using both SDOF system analysis as well as a continuum
(and MDOF). The beam structures are mounted on a rigid frame whose natural frequencies and other characteristics do
not interfere with the beam natural frequencies and characteristics under study. Instrumentation of the test setup comprises
of nine PCB Piezotronics miniature ‘Wheat Grain’ type single axis accelerometers (PCB Model 352C23) mounted on the
beams and four general purpose ‘Tear Drop’ type accelerometers (PCB Model TLD352A56) mounted on the frame.
Several excitation techniques are available for the dynamic response analysis of this structures. One or more of the
techniques use the same hardware but vary entirely in the actual excitation provided to the systems and consequently each
technique has its own merits and demerits. Some of these techniques may be unsuitable in the presence of nonlinearities and
may even provide erroneous results that can be incorrectly diagnosed as a linear system. A new excitation technique using a
Fig. 25.1 (a) Schematic of test structure with instrumentation details (b) First Natural Frequency of the test tetup at 40.809 Hz and zeta 1.9535%
Table 25.1 Test Setup Beam Properties

Beam Type Length (m) Width (m) Height (m) Young’s Modulus (GPa) Density (kg/m3 )
Double clamped beam 0.4572 0.0127 1.5875X10−3 200 7850
Cantilever beam 0.381 0.0127 7.9375X10−3
25 Excitation Techniques for Nonlinear Dynamic Systems: A Summary 159
Fig. 25.2 Static displacement characteristics of test setup showing predominantly cubic nonlinearity
Table 25.2 Digital Signal Processing (DSP) Parameters for Various Excitation Techniques
Excitation Technique Fmaximum (Hz) # Frequency Lines Window Fsamp (Hz)
Impact 800 1000 Force-expo 2048
Shaker (all tests) 800 8000 Hanning
Pneumatic piston 800 8000 Hanning
pneumatic single acting piston is studied here as a novel idea that can possibly address demerits of some of the traditional
excitation techniques mentioned in Table 25.3. Specifications for the various hardware used is provided in Table 25.4.
25.3 Modal Analysis – Linear Response
A Multiple Reference Impact Test (MRIT) with 13 inputs and 45 outputs was performed on the test structure with digital
signal processing (DSP) characteristics given in Table 25.2. The FRFs were estimated using a Hv technique and linear modal
parameter estimation is performed using UC-SDRL’s X-Modal software package. F The modes obtained for the given system
are detailed in Fig. 25.3. The first natural frequency (∼
=40 Hz) is considered for study throughout this paper. The mode shape
is shown in Fig. 25.1 and is a combined first mode for both the cantilever and the double clamped beam. The other modes
can be classified as dominated by one of the two beams due to their varying flexural rigidities. Since the MRIT is limited
in its frequency resolution due to very small impact response durations, the z domain Rational Fraction Polynomial (RFP-z)
algorithm [13], a high order complex frequency domain algorithm available in UC SDRL’s X-Modal software package was
used to establish the first natural frequency of the structure as 40.809 Hz and synthesize corresponding frequency response
functions (FRF) for further analysis.
A low excitation of force of 1 N rms is applied to the system with shaker (Modal Shop Model 2060E) and pneumatic
systems using various types of input excitation signals in both Single Input Multiple Output (SIMO) and MIMO configuration
as described in Table 25.4. From the force-displacement profile in Fig. 25.2, it can be concluded that at such small levels of
excitations, the structure predominantly behaves linearly. A comparison of the various linearized tests is shown in Fig. 25.4.
To establish linearity and independence of the test system from any of the excitation system characteristics, results from each
of the excitation techniques are compared against each other.
Table 25.3 Comparison of Some Common Excitation Techniques

Steady State Sine Pure Random Pseudo Random Random Periodic Chirp Impact Burst Random
Minimize leakage No No Yes Yes Yes Yes Yes
Signal-to-noise ratio Very high Fair Fair Fair High Low Fair
RMS-to-peak ratio High Fair Fair Fair High Low Fair
Test measurement time Very long Good Very short Fair Fair Very short Very short
Controlled frequency content Yes Yes* Yes* Yes* Yes* No No
Controlled amplitude content Yes No Yes* No Yes* No No
Removes distortion No Yes No Yes No No Yes
Characterize nonlinearity Yes No No No Yes No No
*special hardware required
Table 25.4 Case wise Description of Excitation Techniques and Natural Frequencies Estimated (RFP-z)
Case Excitation Technique Typical Hardware First Natural Freq Estimate (Hz) Deviation (%)
1 MRIT Impact hammer 40.809 Ref.
2 SIMO random Electrodynamic shaker 40.172 1.559
3 SIMO random w/cyclic Avg 41.947 2.788
4 SIMO burst random 39.071 4.257
5 SIMO periodic random 40.723 0.210
6 SIMO Pseudo random 40.300 1.246
7 MIMO random 42.723 4.691
8 Pneumatic random Single acting pneumatic piston 37.522 8.082
9 Pneumatic random w/cyclic 37.090 9.111
Figure 25.4 shows the comparison of various test methods against each other for a low excitation level. The MRIT raw
and synthesized FRF are shown along with raw FRF data estimated using a Hv estimation technique for all other excitation
techniques, except pneumatic excitation where H2 estimation technique is used. From Fig. 25.4 it is observed that the SIMO
shaker excitations with different types of excitation frequency fall within 2% of the first natural frequency obtained from the
impact test. This confirms that the testing equipment associated with shaker tests themselves do not affect the characteristics
of the system. This however is not true for the MIMO shaker excitation with random input since the same natural frequency
is now shifted to 42 Hz. In this case, a second shaker is attached to the thin double clamped beam that has very low flexural
rigidity. Any equipment addition to such lightweight members adds further stiffness to the system which affects the overall
characteristics and increases the natural frequency as seen in Fig. 25.4. In addition, an MPE was obtained for all these
excitation techniques to estimate the first natural frequency. The results of the same is given in Table 25.4.
Figure 25.5 shows the Modal Assurance Criterion (MAC) for the first mode between the different methods of excitation.
Since the first mode shown in Fig. 25.1 is from an MRIT, a reduced set of eigen vectors are used to compare and obtain a
MAC since other excitation methods provide a limited spatial basis for comparison. It is evident from the MAC that the first
mode obtained from various excitation techniques are essential the same spatially since MAC values in most cases are close
to 1. An exception to this is the MIMO configuration random excitation technique that has a largely altered first mode. This
can again be attributed to the second shaker being attached to the thin double clamped beam of very low flexural rigidity and
thereby altering the system characteristics drastically. It must also be noted that the MRIT and SIMO shaker excitations have
a very high MAC value thereby confirming independence of the excitation techniques from the system characteristics. The
pneumatic excitation, although have a lower frequency for the first mode, spatially provides the same modal parameters.
25.4 Nonlinear Anaysis – Shaker Excitation
Commonly used excitation techniques mentioned in Table 25.3 are applied to the test setup at higher force levels and their
responses are processed in various forms. All nonlinear analysis performed using the shaker are in SIMO configuration
(unless mentioned otherwise) where the excitation is applied to the test setup as shown in Fig. 25.1. The responses are for
the accelerometer at the rigid junction between the two beams.
Fig. 25.3 Complex Mode Indicator Function (CMIF) plot for the test setup from MRIT. Type 1 modes indicate predominantly cantilever beam
type mode and type 2 indicates predominantly double clamped beam type mode
Fig. 25.4 FRF magnitude comparison between various excitation techniques for Mode 1
Fig. 25.5 Modal Assurance Criterion (MAC) for the first mode of the test setup obtained from various excitation techniques
25.4.1 Stepped Steady State Sine Excitation Using Shakers
A steady state sine excitation is the best excitation technique to characterize and understand the nonlinearities present in a
given system. Since the given mode of interest is the first mode, a narrow band from 38 Hz to 43 Hz with frequency resolution
of 0.1 Hz is selected and at each discrete frequency, the test setup is excited at various force levels (RMS) as shown in Fig.
25.6. The response at the rigid junction between the two beams is analyzed since this location corresponds to maximum
deflection. The results from this plot, collectively known as a Backbone Curve shows a cubic hardening type nonlinearity.
Commensurately, the response corresponding to 1 N at these frequencies is of the order of responses given in the linear FRFs
as shown in Fig. 25.4.
Stepped Sine (Steady Sine Response) at discrete frequencies gives the most accurate characterization of nonlinearities
in a given structure. This method however is time consuming and requires special hardware and controls that can provide
adequate excitation since the excitations and responses are at/near natural frequencies. Excitations at these frequencies are
accompanied by large responses and sustaining a fixed force level is hence hard. In addition, many shakers are incapable
of operating at such frequencies since the stroke available is limited and many shakers are not designed to provide such
large output forces required to obtain a sustained sine wave at these frequencies. This technique is applied in a SIMO test
configuration and MIMO type responses cannot be obtained. Uneven step sizes in frequency is another issue since the post
processing using FFT techniques can be challenging.
25.4.2 Random Excitation Using Shakers
Random excitations are the most common type excitations used in structural dynamics. At low excitation levels (linear
response), sufficient averages and Hanning window, random excitation is used to obtain FRFs with fairly undistorted peaks
that can be used in MPE. However, irrespective of the FRF estimation technique (H1 , H2 , Hv ), for large force levels, random
excitation provides extremely distorted data and distorted peaks at harmonics, indicating the presence of nonlinearities. The
distorted harmonics have another disadvantage of obscuring other modes that may be close to these harmonics.
Figure 25.7 shows wideband characteristics of random excitation applied to the test setup which indicate the harmonics
and other distortions with a Hv type FRF estimation technique and 40 power spectrum averages at all force levels. In a
broader sense, random excitations provide an overall perspective of any nonlinearities present in a structure and can only be
used to detect their presence and fairly identify the type but can provide negligible characterization of nonlinearities. Also,
random excitations are not very helpful in detecting very weak nonlinearities, since the effects of these nonlinearities gets
neutralized due to averaging.
Fig. 25.6 Backbone curve for test setup
Fig. 25.7 (a) Random excitation FRF at various force levels (b)Random excitation showing large distortions around the peak frequency as the
nonlinearity in responses increases due to large forcing levels [Random with 40 PowerSpectrum Averages]
Figure 25.8 shows random excitation with cyclic averaging (20 power spectrum averages with 4 cycles per average).
Cyclic averaging reduces leakage. While cyclic averaging does have a lot of benefits in enhancing FRF estimations for
linear responses, there does not seem to be significant improvements when nonlinear responses are encountered. However, it
removes the information between the frequency bins which may be beneficial.
Fig. 25.8 Random excitation with cyclic averaging showing similar trends [Random with 20 Power Spectrum Averages with 4 cyclic averages
each]
25.4.3 Periodic Excitation Using Shakers
Unlike random excitation techniques, periodic excitation contains energy only in increments of f or at discrete frequencies.
Two commonly used periodic excitation techniques are the periodic random and pseudo random. The former has random
amplitude and random phase at the discrete frequencies while the latter has a constant amplitude and random phase at discrete
frequencies. Nonlinear systems typically show periodic content since at several discrete frequencies energy is added to the
system which may coincide with harmonics corresponding to the type of nonlinearity at a given excitation level. However,
this type of excitation may fail to detect nonlinearities if the frequency resolution f is not sufficiently finite and can thereby
miss any discrete frequencies where nonlinearities can be highlighted.
Figures 25.9 and 25.10 show pseudo random and periodic random response of the test setup for various forcing levels
(RMS). Periodic excitations generally minimize leakage which is evident from the coherence values at 1 N. Increased levels
of excitation in linear systems typically do not show changes in the coherence; increase in excitation will increase the
coherence at anti-resonances due to increased signal to noise ratio. However, the coherence drop at higher excitation levels
in Figs. 25.9 and 25.10 indicates nonlinearity.
25.4.4 Transient Excitation Using Shakers
The burst random technique is a transient technique performed using a uniform window (or exponential window) where a
random input force is applied for a certain duration and the rest of the time block has no external excitation thereby allowing
the system to decay freely. The advantage of this technique over other transient techniques such as impact testing lies in the
control of excitation levels. The initial burst length can be controlled in a similar manner as random excitations. However,
damping of the test structure plays a crucial role in determining the type of response and its ability to capture nonlinearities.
Nonlinearities may be better detected in very lightly damped structures with extremely slow decay If the structure has larger
damping and thereby faster decay rates, the FRF generally is devoid of any nonlinear content and the natural frequencies
may tend to push towards the linear modes. Figure 25.11 shows the burst random technique applied to the test setup and
several demerits of the burst random technique are highlighted from the FRFs obtained.
Fig. 25.9 Pseudo Random excitation FRF and Coherence
Fig. 25.10 Periodic Random excitation FRF and Coherence
25.5 Nonlinear Analysis – Pneumatic Excitation
Pneumatic excitation is a hybrid excitation method combining features from impact and shaker excitation. In this method,
multiple impacts are used to excite a structure to elicit a random response from it. The principal idea behind this method is
that any random signal can be theoretically represented by successive impacts with random magnitude. This input produces
Fig. 25.11 Burst Random excitation
a response like the one obtained by applying a random force with a shaker. The primary advantage of pneumatic excitation,
other than its minimal cost, is not have the excitation system attached to the structure. This eliminates any shaker impedance
from the estimated FRFs. In addition, a multiple input configuration is quite easy to achieve. Typical information about
the Pneumatic excitation setup can be found in [14]. The FTE (Force Transducer on Exciter) configuration is used for this
experiment as shown in Fig. 25.12.
In this method, pneumatic cylinders powered by compressed air are used to excite a structure. DC solenoid valves regulate
the air flow into the cylinders. The solenoid valves are operated through a switching circuit and thus, control the duration
and interval between successive impacts. A typical random pulse train (digital time series) with 5 V magnitude is given as
an input to the circuit as shown in Fig. 25.13. The magnitude of impacts is mostly dependent on air pressure. However, the
pulse width and time delay (delay between successive pulses) also have a minor impact on the magnitude of impacts.
For this experiment, the structure (non-linear rig) had to be excited at different RMS force levels to observe its behavior
as it goes into the non-linear regime. The minimum RMS force level was 1 N and the maximum RMS force level was 20 N.
Achieving such high magnitude RMS force levels with random pulse train was difficult as all forward strokes of the exciter
would not hit the structure due to randomness of pulse width and time delay. The need to avoid hard seating of the cylinder
also puts a constraint on the values of pulse width and time delay. Although a random pulse train was used by Chawla [6] to
excite the same structure, he did not use such high RMS force levels for it. Therefore, a pulse train with constant pulse width
and time delay is used as input for this experiment as shown in Figs. 25.14.
A typical input force and response produced by a constant pulse train is shown in. Using this type of input and various
values of pulse width, time delay and pressure (‘pneumatic settings’), input force (RMS) levels of 1 N, 2 N, 5 N, 10 N, 15 N
and 20 N were generated. Details of these pneumatic settings are given in Table 25.5. All the DSP settings were identical to
the shaker tests. The driving point FRF for all these force levels using both power spectrum and cyclic averaging is shown in
Fig. 25.16.
The peaks in all the FRFs (even at 1 N) are somewhat distorted. The distortion at lower RMS force levels seems to be
a function of the excitation as it was not observed while conducting shaker tests with any form of excitation. Figure 25.16
shows that the peak frequency for 5 N, 10 N, 15 N, 20 N RMS is greater than 1 N RMS force level. Unfortunately, for higher
force levels (greater than or equal to 5 N) the FRFs are too distorted to differentiate between their peak frequencies. Another
important thing to notice from these FRFs is that cyclic averaging is not able to improve the quality of the FRFs. Observing
these results, it can be concluded that although a constant pulse train input is able to generate required RMS force levels,
peaks of the FRFs obtained are distorted even at low force (linear region) levels. The reasons behind these distortions and
ways to improve the FRF measurements are currently under investigation.
Fig. 25.12 Pneumatic Cylinder Schematic
Fig. 25.13 Random Pulse Width Excitation

Fig. 25.14 Constant Pulse Width Excitation
Fig. 25.15 Input Force and Response from Pneumatic Excitation
25.6 Nonlinear Analysis – Step Relaxation
Step relaxation is a broadband excitation technique involving a quick release of a static tension allowing the system to decay
freely from the amplitude of static force applied. It is similar to impact testing when a sudden change in force is applied
at a given location. Mathematically, such systems correspond to initial value problems where the system is at rest and a
Table 25.5 Pneumatic Excitation Parameters

Target RMS force (N) Generated RMS force(N) Pressure (Psi) Pulse Width Time Delay
1 1.02 20 60 88
2 2.01 13 74 127
5 4.87 20 65 128
10 10.25 28 70 220
15 15.10 44 70 320
20 20.05 56 70 410
Fig. 25.16 (a) FRF with 40 Power Spectrum Averages (b) FRF with 20 Power Spectrum Averages with 4 Cyclic Averages each
large initial condition is applied and the decay is analyzed for system characteristics. Historically, such techniques have
been applied to structures where other forms of excitations are not possible such as bridges and wind turbines. A significant
advantage of this technique is the large force levels that can be applied as initial loads, since dynamic loading through shakers
is absent. For the test setup considered in this paper, this excitation was used at almost 3 times static load as compared to the
maximum dynamic load of 20 N applied using shakers.
The response of the structure to these large initial loads is processed in split time blocks. Since the decay characteristics
contain responses at large and small amplitudes, block corresponding to the large amplitudes are typically nonlinear in nature
and as the response further decays, the smaller amplitudes can be approximated to linear responses. It must be noted that
oscillations in the linear regions correspond to linear natural frequencies obtained from the various linear tests mentioned in
the earlier sections. Figure 25.17 shows a typical step relaxation analysis depicting large amplitude responses at the start and
the decay toward smaller amplitudes. Furthermore, the magnitude of frequency domain content is relatively higher at low
frequencies as the step relaxation method excites low frequencies better than higher frequencies.
Figures 25.18, 25.19, 25.20, and 25.21 show the responses at select time instants during the decay response for various
static loading ranging from 22.2 N through 89 N at approximately 22.2 N increments. It is important to note that damping
of the test structure plays a crucial role in the decay characteristics and the feasibility of this excitation technique. For very
lightly damped structures, the decay is slow and hence variation in amplitudes is more evident. Such responses may not be
obtained in structures with large damping. The static loading on the structure must be within its yield limits to prevent any
permanent damage due to large static loads applied to the structure. This technique has been utilized on large towers and
trusses, since the lower frequency response is well excited while the high frequency response is not excited for this type of
forcing function.
Fig. 25.17 Step Relaxation Response for 89 N static load describing shifting in frequencies and decay toward the linear modes
Fig. 25.18 Accelerometer Auto Power Spectrum at discrete time intervals for 22.2 N static load step relaxation excitation
25.7 Non-Contact Excitation Sources
Small structures with very low damping cannot be excited using most of the excitation techniques discussed in the previous
sections. This is due to mass and stiffness effects from addition of any instrumentation added to the system. The miniature
accelerometers are still significantly heavy and can distort the response characteristics of such small structures. To overcome
these difficulties, Baver et.al [4, 5] used a magnetic excitation method using a series a electromagnets. A pair of such
electromagnets on either side of the test specimen is connected is opposite polarity such that a constant magnetic field is
available throughout the region the specimen is excited without making contact with the electromagnets. This setup was used
to detect both bending and torsional modes, and the results were satisfactorily correlated with a Finite Element Model (FEM).
The response of the specimen is made using a Laser Doppler Vibrometer (LDV), a non-contact measurement technique. Such
Fig. 25.21 Accelerometer Auto Power Spectrum at discrete time intervals for 89 N static load step relaxation excitation
excitation techniques are beneficial for nonlinear structures since many lightly damped small structures exhibit nonlinear
responses and the technique is currently under investigation.
25.8 Conclusion
This paper summarizes the importance of excitation techniques including hardware, data processing and representation for
a test structure with known nonlinear response characteristics. While several excitation techniques are commonly applied in
both linear and nonlinear analysis of dynamic systems, seldom have the instrumentation and structural effects been studied
separately. Through the linear modal analysis, independence between the test instrumentation and hardware, and the system
response was established thereby ensuring that any nonlinear responses obtained are purely due to the dynamic characteristics
of the system. Various well-established linear excitation techniques and their responses are visited for a system having a
nonlinear response and typical characteristics, capabilities along with merits/demerits of the excitation technique is observed
here. A combined shaker-impact type technique using a pneumatic piston is also studied and various issues with the same
have been highlighted. However, an investigation is underway to address these issues. Using electromagnetic shakers as
an excitation technique is another area requiring further study. It is also observed that while traditional shaker equipment
may have limitations on the total dynamic load applied, some very large responses can be studied effectively using a step
relaxation approach provided the structure is lightly damped.
References
1. Allemang, R.J.: Investigation of some Multiple Input/Output Frequency Response Function Experimental Modal Analysis Techniques, p. 358.
University of Cincinnati, PhD Dissertation (1980)
2. Allemang, R.J., Brown, D.L.: Correlation coefficient for modal vector analysis. Proc Int Modal Anal Conf Exhib. 110–116 (1982)
3. Allemang RJ, Phillips AW (2004) The unified matrix polynomial approach to understanding modal parameter estimation: an update. Proc 2004
Int Conf noise Vib Eng ISMA 2373–2401
4. Baver, B.: Property Identification of Viscoelastic Coatings through Non-Contact Experimental Modal Analysis, 57pp. University of Cincinnati,
MS Thesis (2016)
5. Baver BC, Phillips AW, Allemang RJ, Kim J (2016) Magnetic excitation and the effects on modal frequency and damping. In: Mains M (ed)
Topics in Modal Analysis & Testing, Volume 10. Springer, Cham, pp 347–354
6. Chawla, R.D.: Investigation of Asymmetric Cubic Nonlinearity Using Broadband Excitation, p. 76. University of Cincinnati, MS Thesis (2019)
7. Ewins, D.J.: Modal Testing: Theory, Practice and Application, 2nd edn. Wiley (2006)
8. Kerschen, G., Peeters, M., Golinval, J.C., Stéphan, C.: Nonlinear modal analysis of a full-scale aircraft. J. Aircr. 50, 1409–1419 (2013). https:/
/doi.org/10.2514/1.C031918
9. Kerschen, G., Worden, K., Vakakis, A.F., Golinval, J.C.: Past, present and future of nonlinear system identification in structural dynamics.
Mech. Syst. Signal Process. 20, 505–592 (2006). https://doi.org/10.1016/j.ymssp.2005.04.008
10. Noël JP, Kerschen G (2016) 10 years of advances in nonlinear system identification in structural dynamics: a review. Proc ISMA 2016 - Int
Conf noise Vib Eng USD2016- Int Conf uncertain Struct Dyn 2709–2745
11. Pandiya, N.: Design and Validation of a MIMO Nonlinear Vibration Test Rig with Hardening Stiffness Characteristics in Multiple Degrees of
Freedom, p. 133. University of Cincinnati, MS Thesis (2017)
12. Phillips, A.W.: Investigation of an Experimental Dynamics Test Simulation Methodology, p. 138. University of Cincinnati, PhD Dissertation
(1991)
13. Richardson, M.H., Formenti, D.L.: Parameter estimation from frequency response measurements using rational fraction polynomials. Proc Int
Modal Anal Conf Exhib. 167–181 (1982)
14. Sharma, A.: A New Multiple Input Random Excitation Technique Utilizing Pneumatic Cylinders, p. 81. University of Cincinnati, MS Thesis
(2016)
Chapter 26
Flutter of Double-Bay Panels with Finite Midbay Stiffness
J. D. Schoneman
Abstract Partitioned or multibay panel designs are a common aerostructural configuration and a potential source of concern
for dynamic response and flutter analysis. Fortunately, past work has shown that the flutter boundary of an equally partitioned
multibay panel is substantially similar to that of a single, isolated panel—identical, in fact, for the double-bay case. A related
question concerns the flutter of multibay panels when the intermediate supports are of only finite stiffness. Preliminary
investigation on this front reveals that when the intermediate support is not sufficiently stiff, a very large drop in the flutter
boundary can occur due to frequency coalescence between even and odd structural modes. This finding has possible design
ramifications but may be of even more interest on the experimental front, providing experimental aeroelasticians an additional
parameter with which to investigate flutter and postflutter behavior of panels.
Keywords Aeroelasticity · Panel flutter · Elastic boundary conditions
26.1 Introduction
Multibay panel configurations are commonly used by aerospace designers to maintain minimum skin thickness over
large surface areas while keeping dynamic response and flutter boundaries within acceptable limits. Contrary to initial
expectations, work by Dowell [1] demonstrated that multibay panels with equal partitioning in the streamwise direction
exhibit flutter boundaries either identical or substantially similar to the well-known flutter boundary of a single panel. While
the flutter dynamic pressure does eventually drop off for a large number of bays, the value for double- and single-bay panels
is identical.
This finding is of particular interest as it pertains to the design of a test article for an upcoming aerothermoelastic (ATE)
validation campaign. The test hardware and plan is under development by ATA Engineering, Inc., (ATA) in conjunction
with the Air Force Research Laboratory (AFRL) Structural Sciences Center (SSC). The test article itself, referred to as the
Discovery Experiment Panel (DEP), is depicted in Fig. 26.1. The basic geometry is an asymmetric double-bay box panel
with a nominal thickness of 50 mil (1.3 mm). With measurements of 20 in the streamwise direction and 12 in. spanwise,
basic calculations1 at the Arnold Engineering Development Center (AEDC) Tunnel C facility reveal that the flat panel is
quite far from its flutter boundary under achievable test conditions: Mach 4.0, with stagnation pressures and temperatures of
100 psia and 1600 R corresponding roughly to the true-altitude values for Mach 4 flight at 70,000 ft. [2].
Multiple factors complicate this type of preliminary flutter analysis: most obviously the effects of static pressure
differential, cavity acoustics, and large temperature excursions in the panel that cause almost immediate buckling. Given
the vast literature on panel flutter, any number of prior works may be selected to better understand these topic. For instance,
Dowell studied curvature and static preload effects [3]; Nydick, Friedmann, and Zhong discussed thermal loading effects
on curved panels [4]; and acoustic theories are quite well developed and overviewed by Dowell [5]. Some amount of
1 Consideringa titanium panel, for instance, the nondimensional dynamic pressure values are λ = 43 when considering the length of the larger bay
and λ = 148 when considering the full 20 in. length of the DEP, both well below the standard reference value of λ ≈ 343 for an infinite, simply
supported panel.
J. D. Schoneman ()

176 J. D. Schoneman
Fig. 26.1 DEP test article for ATE validation, designed by ATA Engineering in conjunction with the AFRL SSC
experimental data on the topic of aerothermoelastic panel flutter also exists due to Dixon et al. [6, 7]. Together, these works
will help ATA/AFRL verify and possibly validate the flutter analysis procedures used to design the DEP and its experimental
program.
A consideration which, to the author’s knowledge, has not been given detailed coverage is the effect of finite stiffness
bay supports on the flutter boundary. The crossrib height of the DEP is a possible design parameter and can be adjusted for
future testing if analytical results indicate that interesting behavior may be observed. To that end, this study builds on the
multibay flutter analysis of Dowell [1] by introducing a variable stiffness vertical restraint at the centerpoint of a double-bay
configuration. The work also serves as initial verification of the finite element model (FEM)–based panel flutter code, which
ATA has developed to study the ATE flutter problem, with numerical predictions matching the theory quite closely.
Results of this preliminary study indicate that, for certain values of the restraint stiffness, the flutter boundary drops
dramatically, reaching nearly zero dynamic pressure in some cases. The mechanism for this behavior is the shift in the
structure’s odd bending-mode frequencies, which must cross over the even bending-mode frequencies that are unaffected by
changes in the restraint. Although further analytical study on the topic is needed, the explanation for the observed behavior
appears to be fundamentally sound and corresponds to similar behavior in curved panels [3]. These findings have possible
design ramifications, but the more interesting result may relate to the design of aeroelastic experiments, particularly the DEP
with its adjustable crossrib design parameter. Future work on this topic will focus on examining the finite-stiffness midpoint
restraint in the context of initial curvature, thermal buckling, and three-dimensional panel models.
26.2 Aeroelastic Models
The development below follows a classical panel flutter formulation previous research on aeroelastic models [1]. A
nominally flat, infinitely wide, symmetric double-bay panel is considered; each bay is of length L for a total panel length
of 2L. Deflections are restricted to the linear regime, and aerodynamics are modeled using the first-order piston theory
approximation, which is a quasi-steady theory yielding a time-invariant aeroelastic system with no explicit frequency
dependence. The partial differential equations of bending motion wn for panel n are

∂ 4 wn ∂ 2 wn 2q∞ ∂wn 1 ∂wn
D + ρm t + + =0 (26.1)
∂x 4 ∂t 2 M∞ ∂x U∞ ∂t
The panel’s additional physical properties are its flexural rigidity D, density ρ m , and thickness t. Relevant flow properties
are the dynamic pressure q∞ , Mach number M∞ , and freestream speed U∞ . For the double-bay with pinned boundaries, these
equations must satisfy zero-displacement and zero-moment conditions at the ends, with slope and moment compatibility at
the intermediate support.
∂ 2 w1 (0) ∂ 2 w2 (L)
w1 (0) = w2 (L) = = =0 (26.2a)
∂x 2 ∂x 2
w1 (L) = w2 (0) = 0 (26.2b)

26 Flutter of Double-Bay Panels with Finite Midbay Stiffness 177
∂w1 (L) ∂w2 (0)

= (26.2c)
∂x ∂x
∂ 2 w1 (L) ∂ 2 w2 (0)
= (26.2d)
∂x 2 ∂x 2
After invoking the typical assumption of a periodic response at frequency ω and nondimensionalizing, eq. (26.1) can be
represented as
∂ 4 ŵn ∂ ŵn
4
+λ − π 4 Z ŵn = 0 (26.3)
∂ξ ∂ξ
2
The key parameters are nondimensional dynamic pressure λ = 2q∞ L3 /M∞ D and the analytical eigenvalues Z = ωω0 −
)
jgT ωω0 , with imaginary unit j and fundamental frequency ω0 = π 2 D/ ρm tL4 . Differentiation takes place with respect
to the nondimensional distance ξ = x/L. The total system damping, denoted by gT , is in general a combination of aerodynamic
damping and viscous damping within the structure. In this work, only aerodynamic damping is considered, so that

12 ρ∞ a∞ L 2
gT = g A = 1 − ν 2 (26.4)
π4 ρm am t
To find the eigenvalues, a spatial solution of the form

4
ŵn = Ck,n eξpk (26.5)
k=1
is specified; the values pk are roots of the quartic equation p4 + λp − π 4 Z = 0. This solution satisfies the differential eqs.
(26.3) and boundary conditions for a proper choice of coefficients Ck, n . Using eq. (26.5) to impose the boundary conditions
leads to a homogenous linear system Ac = 0; as usual, nontrivial solutions exist only when det A = 0. Once the eigenvalues
are known for a given λ, the response frequencies are given by R {Z}, and the flutter boundary itself corresponds to the
{Z}
condition gA = √IR {Z}
.
In this work, the eigenvalues Z(λ) were located by specifying λ and numerically minimizing the matrix determinant for
a complex Z. This approach is inelegant and somewhat unwieldy; additionally, it has particular trouble in correctly tracking
the “branch” of each eigenvalue for higher numbers of panels. Thus, only the exact analytical solution for a single panel is
demonstrated in the results below. Since the flutter boundary for the double-bay panel is all but identical to the single-bay
case, omission of the double-bay analytical results is of little consequence.
An alternate approach to analyze eq. (26.1) involves a discretization using the finite element method. This type of approach
is being used by ATA to analyze the flutter behavior of the DEP. Finite element discretization results in a fairly large system
of linear, second-order ordinary differential equations,

2q∞ 2q∞
Mẅ + Kẇ ẇ + K + Kw w = 0 (26.6)
U∞ M ∞ M∞
The matrices M and K are the mass and stiffness matrices, while Kẇ and Kw are the aerodynamic damping and
stiffness matrices. The quasi-steady nature of piston theory means that these latter two matrices can be constructed either
by elementary geometric considerations or straightforward application of the finite element method. In this work, all finite
element and stability calculations were performed in MATLAB, although generally the structural matrices may be obtained
from commercial finite element software such as Abaqus.
Stability can be analyzed at full order if desired, but a modal reduction is typically more convenient. Solving the structural
eigenvalue problem allows one to retain m modeshapes φi such that φTi Kφi = ωi,0 2 (the square of the ith “dry” natural
frequency) and φTi Mφi = 1. All other terms in the transformed mass and stiffness matrices are zero, due to orthogonality of
the modeshapes. Displacements are represented using a reduced set of modal coordinates q, with w = [φ1 φ2 . . . φm ]q = q.
The reduced aeroelastic system is
178 J. D. Schoneman

2q∞ 2q∞
q̈ + Kq̇ q̇ + Λ + Kq q = 0 (26.7)
U∞ M ∞ M∞
The aerodynamic matrices are obtained via routine coordinate transformation, and the modal stiffness matrix is diagonal
with Λii = ωi,0
2 . To assess the stability of a given flow condition, the system can be recast into state-space form as a first
order ordinary differential equation, yielding

!
q̇ 0m Im q q
= 2q∞ 2q∞ =A (26.8)
q̈ − U∞ M∞ Kq̇ − Λ+ M∞ Kq q̇ q̇
All oscillatory eigenvalues of the system take the form λi, i + m = ni ± jωi , with ωi the damped natural frequency and ni
the total damping of each mode. Asymptotic stability is assured as long as R {λi } < 0 for all i. The flutter boundary itself is
given by R {λi } = 0 for any i.
An important distinction between the FEM/modal (hereafter “numerical”) method and the analytical procedure is the
treatment of damping. Using the state-space formulation, aerodynamic damping for a given flow condition is implicitly
included. This approach has the advantage of determing a specific flutter boundary for the flow condition of interest, but it
does not yield the flutter boundary as a function of arbitrary values of aerodynamic and viscous damping, so is somewhat
less convenient for design purposes. On the other hand, once the modal reduction is performed, eigenvalues for a variety of
flow conditions can be quickly computed, mitigating this inconvenience.
Analytical and numerical results are compared below for the double-bay, infinite-width case of interest.
26.3 Fem and Verification of Numerical Approach
Analytical flutter behavior can only be obtained through consideration of nondimesional parameters. The FEM, on the
other hand, must carry along physically representative properties, with results nondimensionalized after the fact. Structural,
material, flow, and mesh properties are summarized in Table 26.1. The finite element mesh was constructed from four-node
shells based on a simple plate/membrane formulation; to avoid shear locking, bending terms were fully integrated while
shear terms were evaluated via reduced integration as suggested by previous research [8]. Only a small spanwise strip of the
panel was modeled, so the infinite-width condition was enforced by adding restrictions on spanwise motion and streamwise
rotation to the edges of the FEM. Twenty modes of the FEM were retained for flutter computations.
Computations were performed for both a single and double-bay panel. Evolution of frequency as a function of λ for each
case is compared to the theoretical solution in Fig. 26.2. Prior to frequency coalescence, the curves match almost precisely.
After coalescence, a divergence between the analytical and numerical result is evident. It is not yet clear to the author whether
this corresponds to a true difference in prediction or simply a difference in the quantity plotted—further study is required.
The flutter boundary itself is shown in Fig. 26.3—as expected, flutter boundaries predicted by the single- and double-bay
finite element models were indistinguishable. The analytical result provides the flutter boundary for an arbitrary amount of
aerodynamic damping, with the black dashed line corresponding to the actual value of aerodynamic damping for the present
flow conditions, which varies as a function of λ. The numerical method predicts a single flutter value of λ corresponding to
the selected flow conditions. There is an observable offset between the point at which the analytical flutter boundary intersects
the aerodynamic damping line and the numerical prediction of flutter dynamic pressure; however, the resulting discrepancy is
below 1% in terms of λ. This amount of error is not currently a major concern—further analysis at a variety of aerodynamic
damping conditions will help to better assess the accuracy of the numerical procedure. For the small amount of aerodynamic
damping present at these conditions, λflutter ≈ 343, the dynamic pressure corresponding to frequency coalescence.
Table 26.1 Summary of structural, flow, and FEM properties used for numerical computation
D [N · m] ρ m [kg/m2 ] t [m] L [m]
Structure 6.4 2700 0.001 0.5
Flow M∞ p∞ [Pa] U∞ [m/s]
4.0 0–5000 1152
FEM Streamwise elements Spanwise elements Degrees of freedom
100 4 3030
Fig. 26.2 Frequency evolution comparisons between an exact, analytical solution for the single bay dynamics and the results predicted by
numerical models of a single and double-bay panel
Fig. 26.3 Analytical flutter boundary (black, solid) along with the aerodynamic damping for present conditions (black, dotted) and the flutter
dynamic pressure predicted by both single- and double-bay FEMs (blue, dashed)
26.4 Examination of Partially Restrained Intermediate Support
With some confidence available in the numerical method, the effect of a partially restrained intermediate support on the
double-bay panel’s flutter boundary can be considered. The midpoint displacement constraint of equation (26.2b) is amended
to retain the equality requirement but remove the zero-value specification, and a spring stiffness ks is associated with the
midpoint displacement. This additional stiffness is straightforward to add to the finite element stiffness matrix from equation
(26.6), but modification of the nondimensional analytical formulation of equation (26.3) is not so obvious and requires further
consideration. As such, only results from the numerical approach are presented here.
180 J. D. Schoneman
Fig. 26.4 Frequency evolution with dynamic pressure λ for selected values of the midpoint restraint stiffness ks . Single-bay frequencies are plotted
in each pane for reference
For reference, denote the flutter dynamic pressure for the infinitely stiff support as λ∗ = 3432 . When ks ≈ 0, the double-bay
panel will behave as a single panel of length 2L, and its flutter dynamic pressure will drop to λ∗ /8 ≈ 43. As ks increases, one
might expect some type of monotonic transition between λ∗ /8 and λ∗ . Figure 26.4 depicts the evolution of frequency/dynamic
pressure for several values of ks , (normalized by a stiffness parameter D/L3 ), demonstrating that no such well-behaved
transition occurs. Pane (a) of Fig. 26.4 corresponds to an essentially unrestrained midpoint, with the expected coalescence
at roughly λ = 43. As ks increases in pane (b); however, the coalescence value drops precipitously. A further increase in ks
leads to a critical mode three/four coalescence (pane (c)) and eventually, for sufficiently high ks , modes one and three of the
double-bay panel lie atop modes one and two of the single-bay panel, as expected (pane (d).
A more comprehensive sweep of ks is shown in Fig. 26.5, with the flutter dynamic pressure shown for each case. This
dynamic pressure was obtained through the use of MATLAB’s fzero root-finding routine to find λflutter for each ks point.
Beginning from the unrestrained λflutter = λ∗ /8 = 43, a smooth drop towards zero flutter dynamic pressure3 occurs for
increasing ks , followed by a region of highly erratic behavior. Eventually, the flutter boundary settles into its expected value
of λflutter = λ∗ = 343.
While perhaps initially surprising, the mechanism for this behavior is straightforward. As the midpoint spring stiffness
increases from zero, mode one of the initially unrestrained panel must increase in frequency, transitioning from a
pinned/pinned “first-bending” shape of length 2L to a more complicated pinned/clamped shape of length L, which is mode
two of the fully restrained double-bay panel. The pinned/pinned second bending shape—mode two of the unrestrained panel
but mode one of the rigidly fixed panel—has a node at the midpoint and is not affected by the increased spring stiffness.
2 Allflutter dynamic pressures remain nondimensionalized to the bay length L for ease of comparison to the fully rigid support
3 It
is not entirely clear whether this corresponds to true “zero” dynamic pressure, or is simply a very low value which caused numerical problems
with fzero. Per the earlier contention that aerodynamic damping is included in the state-space formulation, it seems unlikely that a zero-pressure
boundary should actually occur.
Fig. 26.5 Flutter dynamic pressure as a function of midpoint restraint stiffness ks
Fig. 26.6 The evolution of first two modes and frequencies with vertical restraint stiffness ks . Mode one of the unrestrained panel undergoes
substantial changes in shape and frequency; mode two of the unrestrained panel is unaffected by altered stiffness
Thus, the two frequencies are guaranteed to cross, leading to flutter at or near zero dynamic pressure. The transition process
is illustrated for the first two modes in Fig. 26.6.
This same behavior affects every odd and even mode pair in the structure, with each pair crossing at a larger value of
ks , as demonstrated in Fig. 26.7 for the 20 structural modes retained for this study. Crossings of higher modes explain the
mode three/four coalescence observed in Fig. 26.4 and the erratic behavior of the flutter boundary in Fig. 26.5. Significantly,
there does appear to be a trend which indicates some type of “cutoff stiffness,” above which higher mode crossings no longer
influence the flutter boundary. This is encouraging, since physical multibay systems are of course not infinitely stiff but are
182 J. D. Schoneman
Fig. 26.7 Evolution of the double-bay panel’s first 20 dry natural frequencies as a function of midpoint restraint stiffness ks
not typically prone to unexpected high-frequency instability. Determination of this cutoff stiffness is impaired in this case by
the use of only 20 modes in the model, but is surely of interest for further examination.
The mechanism in play here bears a substantial similarity to that observed in plates with initial curvature [3]. Just as shown
above, the odd natural frequencies of curved plates will cross over the even frequencies at a given level of curvature, leading
to a very low flutter dynamic pressure. The major difference between the two situations is that in the case of the curved panel,
the effects of initial static loading substantially alter the flutter boundary, smoothing out the effects of the frequency crossing.
Under the model studied here, no initial static loading on the panel exists.
26.5 Discussion and Closing Remarks
Examination of a double-bay panel with an intermediate vertical restraint of variable stiffness has demonstrated a catastrophic
drop in flutter dynamic pressure under certain parameters. While further verification via alternate methods is in order, the
underlying physical mechanism causing this behavior seems entirely plausible. Moving forward, a detailed literature survey
is needed to search for any equivalent work. Assuming no equivalent work is found, the next step is to attempt the extension
of an exact analytical solution to the conditions considered here. From a theoretical standpoint, perhaps the most pressing
question concerns the possible existence of a stiffness threshold above which the effects of the finite restraint stiffness are
not observed.
If it is indeed the case that no prior work on this specific line of questioning exists, one would suppose that multibay
designs in common use are indeed sufficiently stiff to avoid unexpected flutter due to frequency crossings. As such, these
results are perhaps of limited use from a design perspective, except as a cautionary note. A more interesting impact could
be on the design of future flutter experiments, in particular the DEP. While the AEDC Tunnel C dynamic pressures are not
sufficient to push even a single-bay version of the DEP to its flutter boundary, proper selection of a crossrib thickness may
allow the observation of flutter during a future experimental campaign.
To support such a possibility, the concepts explored in this work must be extended to also consider the effects of pressure
differentials, thermal buckling, cavity acoustics, and a more geometrically representative three-dimensional panel model.
The general concept of varying the DEP crossrib thickness to tune the frequencies of interacting modes should hold, but
will of course be significantly complicated by these additional effects. A comprehensive understanding of the simpler case
studied here will maximize the changes of realizing the desired effects in practice.
References
1. Dowell, E.: Flutter of multibay panels at high supersonic speeds. AIAA J. 2(10), 1805–1814 (1964)
2. Strike, W.T.: Calibration and Performance of the AEDC/VKF Tunnel C, Mach Number 4, Aerothermal Wind Tunnel. Arnold Engineering
Development Center, Arnold Air Force Station, TN (1982)
3. Dowell, E.: Nonlinear flutter of curved plates. AIAA J. 7(3), 424–431 (1969)
4. Nydick, I., Friedmann, P., Zhong, X.: Hypersonic Panel Flutter Studies on Curved Panels, in 36th Structures, Structural Dynamics, and Materials
Conference, New Orleans (1995)
5. Dowell, E.: Aeroelasticity of Plates and Shells. Springer (1975)
6. S. C. Dixon, G. E. Griffith and H. L. Bohon, Experimental investigation at Mach number 3.0 of the effects of thermal stress and buckling on the
flutter of Four-Bay aluminum alloy panels with Lengh-width ratios of 10. NASA TN D-921 (1961)
7. S. C. Dixon and C. Shore: Effects of differential pressure, thermal stress, and buckling on flutter of flat panels with length-width ratio of 2,
NASA TN D-2047 (1963)
8. Cook, Malthus, Plesha and Witt: Concepts and Applications of Finite Element Analysis, 4th edn. Wiley (2001)
Chapter 27
Nonlinear System Identification of a Jointed Structure Using
Full-Field Data: Part II Analysis
Giancarlo Kosova, Mengshi Jin, Mattia Cenedese, Wei Chen, Aryan Singh, Debasish Jana, Matthew R. W. Brake,
Christoph W. Schwingshackl, Satish Nagarajaiah, Keegan J. Moore, and Jean-Philippe Noël
Abstract Mechanical joints have a significant influence on the dynamic response of assembled structures. Due to friction,
wear, and non-idealized boundary conditions, joints introduce significant nonlinearity into the dynamics of assembled
structures. To better understand and, in the future, tailor the nonlinearities, accurate methods are needed to characterize
the dynamic properties of jointed structures. In this research, the response analysis for a beam with a bolted lap joint is
studied with the help of several available identification techniques. The experimental setup and data capture are described in
Part I of this work, providing high spatial resolution data for a variety of excitation methods. The nonlinear identification of
the data is the focus of this paper, aiming to perform nonlinear modal analysis and to localize the nonlinear characteristics of
the structure with a series of different approaches.
Keywords Jointed structures · Nonlinear dynamics · Nonlinear system identification · Full-field digital image correlation
data · Backbone and damping curve
27.1 Introduction
In order to be able to model the dynamics of jointed structures, a lot of effort has been devoted in the last years for the
development of methods that identify the characteristics of such systems [1]. Due to friction, wear, and non-idealized
boundary conditions, joints introduce significant nonlinearity into the dynamics of assembled structures. Several methods
to identify jointed structures are present in the literature and we refer to [2] for a comparison of some common methods
to analyze free decay measurements. These methods have been mostly applied to single point responses. The focus of
G. Kosova ()
Siemens Industry Software, Leuven, Belgium
Aerospace and Mechanical Engineering Department, University of Liège, Liège, Belgium
M. Jin · W. Chen
School of Aerospace Engineering and Applied Mechanics, Tongji University, Shanghai, China
M. Cenedese
Institute for Mechanical Systems, ETH Zürich, Zürich, Switzerland
A. Singh · K. J. Moore
Department of Mechanical and Materials Engineering, University of Nebraska-Lincoln, Lincoln, NE, USA
D. Jana
Civil and Environmental Engineering, Rice University, Houston, TX, USA
M. R. W. Brake
C. W. Schwingshackl
Department of Mechanical Engineering, Imperial College London, London, UK
S. Nagarajaiah
Civil and Environmental Engineering, Rice University, Houston, TX, USA
J.-P. Noël
Department of Mechanical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands

186 G. Kosova et al.
this research is to push the boundaries of recently developed nonlinear system identification techniques by combining
new methods and experimental techniques. More specifically, we aim to evaluate the level of agreement in the identified
nonlinearities between different methods and to assess how full-field digital image correlation (DIC) [3] data can augment
the identification of nonlinear dynamical behavior, with a focus on localization.
After illustrating the adopted methods and the experimental setting, we first compare some selected methods on the
identification of accelerometer data of free decay vibrations of a jointed structure. Moreover, we show the consistency of
DIC data with accelerometer measurements and we describe the outcome of a novel localization assessment.
27.2 Methods
For the data analysis, we adopt the methods briefly described in the next lines and point out detailed references.
• HT: The analytical signal is the sum of the original signal and its Hilbert transform (HT) [4]. When it is written in polar
coordinates, the instantaneous amplitude and instantaneous frequency can be calculated.
• PFF: Peak Finding and Fitting method (PFF) [5] calculates instantaneous frequency using two successive time instants
at which the signal reaches local maxima and minima. By improving the extraction of instantaneous amplitude, it obtains
the damping ratio with less noise than the HT method.
• SSM: Nonlinear identification is based on spectral submanifolds [6] as they shape the dynamics of damped, multi-
degree of freedom mechanical systems. The method first constructs a data-driven polynomial state space model and
then computes the spectral submanifolds of this model in order to analyze the reduced dynamics.
• DMD: Dynamic Mode Decomposition (DMD) [7] finds a coherent spatio-temporal pattern from the high dimensional
measurement to comprehend the underlying dynamics of the structure. Window-based Hankel DMD [8] captures the
time-varying modal parameters of a mildly nonlinear system, but the assumption being, this type of system behaves like
a linear dynamical system in a small time window.
• OPERS: Orthogonal-projection-based estimation of residual space [9] calculates the part of the response uncorrelated to
the input force that contains information about the nonlinear internal forces. Nonparametric detection and localization are
performed by analyzing this response at different locations of the structure.
27.3 Analysis
Shaker ringdown test was performed on a half Brake-Reuß beam (HBRB) that is modified from a BRB [1] where the
tightening torque is 10 Nm. We exploited the shaker in order to obtain approximate single-mode responses. The experimental
setup is shown in Fig. 27.1 (a) and (b). The acceleration response was measured with the sampling frequency of 6400 Hz (Fig.
27.1 (c)). Meanwhile, high speed cameras were deployed to measure the displacement response. More detailed information
Fig. 27.1 The setup of the HBRB shaker ringdown test, showing from (a) side and (b) top views, and an accelerometer time history measurement
(c) obtained with shaker ringdown testing
27 Nonlinear System Identification of a Jointed Structure Using Full-Field Data: Part II Analysis 187
Fig. 27.2 The comparison between different methods, showing (a) backbone curve and (b) damping ratio curve
Fig. 27.3 The comparison between the results of the accelerometer and DIC points
about the experiment can be found in [10]. The aforementioned methods are applied to the signal measured by the blue
accelerometer from time t = 30.3 s to 40 s.
It is demonstrated in Fig. 27.2 that all of the different methods show great agreement in the identification of both
instantaneous frequency and damping ratio. The instantaneous frequency first decreases and then increases with the increase
of the amplitude, indicating softening and hardening behaviors respectively. However, frequency variations remain within a
0.5 Hz range. Instead, the damping ratio keeps rising with the amplitude increasing, reaching a value which is almost 7 times
higher than the one at the zero amplitude limit. Although the signal was further truncated from 30.4 s to 32.8 s to be analyzed
using the HT method, there are still end effects in the obtained curves, which is one of the drawbacks of this method.
DIC [3] was applied to the images captured using high speed cameras in order to extract displacement response. After
the extraction, the PFF method was used to obtain the instantaneous frequency and damping ratio. The comparison of results
between the DIC and accelerometer data is illustrated in Fig. 27.3. Consistent results between the accelerometer data and the
DIC data are achieved, but there is a lack of accuracy appears at low amplitudes, after two seconds of signal decay.
In Fig. 27.4 it is possible to see that the part of the response uncorrelated with the input force, Xu, strongly highlights
nonlinear distortions, such as harmonics of the fundamental modes and their combinations. It is expected that the amplitude
of the uncorrelated response far from the natural frequencies, in particular at harmonics frequencies, should be higher at
locations close to the nonlinear source, that is the lap joint. Figure 27.4 illustrates that this happens only for some harmonics,
but not for all of them. Thus, only limited information about localization can be obtained for the system under analysis.
27.4 Conclusion
We compare the performance of different nonlinear identification methods on single-mode ringdown data extracted from a
jointed structure. In line with literature results [1], the principal nonlinear behavior of the system is a substantial change
in instantaneous damping ratio along signal decay. Backbone and damping curves extracted with the methods using
accelerometer data are in good agreement. Moreover, the PFF method shows consistent results between the DIC data
188 G. Kosova et al.
Response Autopower Spectrum (solid lines) - Uncorrelated Response Autopower Spectrum (dotted lines)
8
10
X1
X2
76.2 f1 X3
293.4 f2
106 X4
X5
577.7 f3 X6
X7
4
10 Xu 1
80.1 296.3
Xu 2
[(m/s 2)2]
Xu 3
f1*4=304.7 Xu 4
102 586.5 Xu 5
f1*2=152.34 303.7 f3-f1=501.5 Xu 6
152.3 f1*3=228.5 369.9 501.4 2*f2= 586.8 Xu 7
0 217.2
10 227.9 f1+f2=369.6 445.7
521.5 658.8
10-2 f4 (810.9)-2*f1= 658.6

f1*3+f2=521.9
f2-f1= 217.2
2*f1+f2=445.7
-4
10
0 100 200 300 400 500 600 700 800
Frequency [Hz]
Fig. 27.4 Autopower spectrum of the response and its nonlinear part to a shaker pseudorandom excitation
and accelerometer measurements. As a next step of this research, we aim to further investigate DIC data and to highlight
advantages and possible disadvantages of the several nonlinear identification methods in the context of jointed structures.
27.5 Acknowledgements
The authors are grateful to the SIEMENS and South Central Imaging for their sponsorship during the experimental testing.
One of the authors (GK) is supported by European Union’s Marie Skłodowska-Curie grant No 764547.
References
1. Brake, M.R.W. (ed.): The Mechanics of Jointed Structures: Recent Research and Open Challenges for Developing Predictive Models for
Structural Dynamics. Springer, Cham (2017)
2. Jin, M., Brake, M.R.W., Song, H.: Comparison of nonlinear system identification methods for free decay measurements with application to
jointed structures. J. Sound Vib. 453, 268–293 (2019)
3. Chen W., Jin M., Lawal I., et al. Measurement of Slip and Separation in Jointed Structures with Non-flat Interfaces, Mechanical Systems and
Signal Processing, 134 (2019)
4. Feldman, M.: Hilbert transform methods for nonparametric identification of nonlinear time varying vibration systems. Mech. Syst. Signal
Process. 47(1–2), 66–77 (2014)
5. Jin, M., Chen, W., Brake, M.R.W., et al.: Identification of instantaneous frequency and damping from transient decay data. J. Vib. Acoust..
under review
6. Szalai, R., Ehrhardt, D., Haller, G.: Nonlinear model identification and spectral submanifolds for multi-degree-of-freedom mechanical
vibrations. Proc. Roy. Soc. Lon. A Math. Phys. Eng. Sci. 473(2202), (2017)
7. Schmid, P., Sesterhenn, J.: Dynamic mode decomposition of numerical and experimental data. J. Fluid Mech. 656, 5–28 (2010)
8. Arbabi, H., Mezic, I.: Ergodic theory, dynamic mode decomposition, and computation of spectral properties of the Koopman operator. SIAM
J. Appl. Dyn. Syst. 16(4), 2096–2126 (2017)
9. Kolluri, M.M., Allemang, R.J., Phillips, A.W.: Non-parametric detection and localization of structural nonlinearities using orthogonal
projections. Mech. Syst. Signal Process. 123, 455–465 (2019)
10. Singh, A., Chen, W., Jana, D., et al.: Nonlinear Identification of a Jointed Structure Using Full-Field Data: Part I Experimental Investigation,
IMAC XXXVIII a Conference and Exposition on Structural Dynamics, Houston, TX (2020)
Chapter 28
Control Parameters in Non-linear Properties of Linear Guideway
in Lateral Direction
Ting-Yen Wu, Yi-Chun Lo, and Yum Ji Chan
Abstract To build reliable digital twins of machine tools (also known as “virtual machine tools”), the dynamic properties
of nonlinear components in the model should be close to the actual counterpart. The nonlinear dynamics mean that the
frequency response functions (FRFs) derived from tap tests are inaccurate because of difference in excitation levels. In this
study, dynamic behavior of a linear guideway is sought experimentally using harmonic excitation. It is found that the dynamic
properties are affected by (1) excitation level, (2) lubrication, (3) specified preload and (4) static lateral load, and the dynamic
properties cannot be fully described using the Hertzian contact model.
Keywords Non-linear mechanics · Linear guideway · Contact mechanics · Preloaded contact
28.1 Introduction
To reduce the time spent on product development, manufacturing processes are simulated in digital twins of machines such as
machine tools [1]. Such simulation is required to avoid instability in machining, also known as chatter. Occurrence of chatter
depends on the frequency response function (FRF) of the machine tool and material parameters. Therefore, for accurate
prediction, dynamic properties of digital twins of machine tools, also known as “virtual machine tools”, should be close to
the actual counterparts. These properties are usually derived from experimental modal analysis from tap tests with impact
hammers, but such a method cannot obtain nonlinear properties of structures, because excitation levels in actual operation
are periodic. Unfortunately, machine tools contain numerous nonlinear elements such as joints and rolling contacts.
This work is focused on linear guideways as shown in Fig. 28.1. Linear guideways, as a component of feed drive, keep
the motion of moving parts in specific directions with minimal force [2], and this is usually achieved using ball-rail joints in
parallel. The main structure of the linear guideway is divided into three parts: rail, carriage and rolling body (steel balls). In
this study, the control parameters determining the dynamic properties of linear guideways are sought.
In this study, FRFs of selected linear guideways are obtained using harmonic excitation with shakers, unless otherwise
stated. Afterwards, modal and physical properties of the physical system are sought. Natural frequencies and modal damping
ratios are mainly sought using the Circle method, with the Line-fit method adopted as a back-up [4]. The equivalent stiffness
of the rolling contact and mass of the carriage are sought by assuming the system being a single degree-of-freedom system:
the rail is assumed to be rigid and fixed to the ground and the carriage is rigid.
28.2 Modal Property Estimation with Linearized Model
The track, carriage and rolling elements (Fig. 28.1 (right)) are specified to have interference fit, therefore the rail and the
carriage experience internal preload. Such internal preload provides stiffness to the structure and load-carrying ability of the
linear guideway. However, the relationship of interference and stiffness is nonlinear, and external lateral force applied to a
carriage would affect the distribution of contact stress. The nonlinear mechanics between rolling ball with the carriage and
T.-Y. Wu · Y.-C. Lo · Y. J. Chan ()

Department of Mechanical Engineering, National Chung Hsing University, Taichung, Taiwan

190 T.-Y. Wu et al.
Rail
Carriage
Preload due to interference fit

End Cap
End Sealing 45°

Gasket Fp
Ball 45°
Retainer
Fig. 28.1 Schematic of linear guideway (Adapted from [3])
the rail can be modelled using the Hertzian contact [5]. The following points are assumed to simplify the contact mechanics
problem:
1. Friction on contact surfaces and tangential force are neglected
2. The direction of the contact force is perpendicular to the contact surface
3. Deformation is entirely elastic and the contact surface is much smaller than the radius of the object
4. The sum of the forces applying on the objects in contact is equal to the integral of the stress on the contact area
According to Hertzian contact mechanics, the relationship between the contact force F and the positive deformation δ of
the contact point can be expressed in Equation (28.1) [6], while the linearized normal contact stiffness Kn can be estimated
by differentiating Equation (28.1):
3
F = Kh δ 2 (28.1)
dF 3 1 3 2 1
Kn = = Kh δ 2 = Kh3 F 3 (28.2)
dδ 2 2
where Kh is a constant determined by the geometry of the contact surface and the mechanical properties of the material [6].
Based on the equations, a linearized numerical model of linear guideway is built to simulate the effects of preload level
on the dynamic properties. The contact stiffness between carriage and rail is modelled by linearized spring elements, with
the number of spring elements equal to the number of balls. In this work, there are 60 spring elements to represent the 15
rolling elements in each of the 4 grooves.
In this study, the nominal preload levels of the “light-” and “medium-” preloaded linear guideways are 100 N and 400 N,
respectively. Assuming the carriage, steel ball and rail are made of steel, and the radii of the groove and steel balls are
assumed to be 2.8 mm and 2.78 mm, respectively. By the use of Equations (28.1) and (28.2), the normal contact stiffness
levels are Kn = 27.7 N/μm and Kn = 44.0 N/μm for the light- and medium- preloaded linear guideways, respectively. The
first three modes of the linear guideway are shown in Fig. 28.2 and it is found that mode order does not change within the
preload range studied.
28.3 Overview of Experimental Investigation
A total of 6 tests were carried out in this study. In the first part, the extent of linear guideway behavior nonlinearity is
studied using (i) harmonics detection, (ii) FRF dependence on excitation level and (iii) frequency sweep direction. In terms
of physical parameters, effects of (iv) lubrication, (v) internal preload and (vi) static lateral load on dynamic behavior are
studied. The linear guideway studied and the measurement points used are shown in Fig. 28.3. The carriage (165 g) is covered
with a metal cap (425 g) where feasible. Uniaxial accelerometers are set up at Points 2 (opposite to excitation point) and 7
(base of the platform). While response at Point 2 is always captured, other measurement points are included where necessary.
28 Control Parameters in Non-linear Properties of Linear Guideway in Lateral Direction 191
Rolling mode Bending mode Lateral mode

Light preloaded 2990 Hz 8394 Hz 9766 Hz
Medium preloaded 3490 Hz 10059 Hz 11850 Hz
Fig. 28.2 Mode shapes and natural frequencies of linear guideway models
8 4 1
5 2
7
6 3
Fig. 28.3 Top view of experiment setup, with measurement points marked in red dots and triangle
100 100
0 30
0 30
0 30
30 60
Interface
30 60
10–1 30 10–1 60
10–2 10–2
680 700 720 740 760 780 800 680 700 720 740 760 780 800
Frequency (Hz) Frequency (Hz)
Fig. 28.4 Variation of FRFs under random excitation over 60 minutes
In tests involving stepped sine excitation, measurement at each frequency lasts for 30 seconds to reduce the effect of noise,
and FRFs are calculated based on power spectral densities [4].
In initial tests, long-time tests were carried out with random excitation, and measurements were taken in 30-minute
intervals. In some of the tests, FRFs may change in the first 30 minutes but such change is minimal in the subsequent
30 minutes (Fig. 28.4). Therefore, a hardware warm-up of 30 minutes with harmonic excitation of suitable excitation level
is carried out, before actual signals for steady-state dynamic behavior analysis are captured.
192 T.-Y. Wu et al.
28.4 Extent of Nonlinearity in Dynamic Behaviour
28.4.1 Harmonics
It is known that harmonic excitation on a nonlinear system results in response components at frequencies other than the
excitation frequency [7]. Although a higher excitation level can reduce the influence of noise, nonlinear phenomena such as
harmonics become more significant, as shown in Fig. 28.5 and Table 28.1. For example, the relative amplitude of the third
harmonics in the acceleration spectrum is increased from 7.53 × 10−5 to 6.42 × 10−3 , the latter of which correspond to a
relative acceleration amplitude of 8%. Therefore, observable nonlinear behavior exists in the system.
28.4.2 FRF Dependence on Excitation Level
In order to observe the influence of the excitation level on the FRF in this system, the structure is excited using ascending
stepped sine excitation of various levels with Δf = 1 Hz. The repeatability of the test was validated multiple measurements.
Results show that the peak inertance and natural frequency increase and reduces, respectively, with increasing excitation
level (Fig. 28.6). From the natural frequency and damping ratios obtained from either the circle method or the line-fit
method, modal mass, modal stiffness, and equivalent viscous damping coefficient are calculated based on Point-FRF and
single degree-of-freedom assumptions. Results in Table 28.2 show reduction in both damping ratio and the natural frequency
with increasing excitation level. However, the identified modal mass of exceeding 30 kg does not match the carriage mass.
The source of such discrepancy is unknown at this stage.
105
100
Acceleration PSD
10-5
10-10
10-15
0 500 1000 1500 2000 2500
Frequency (Hz)
Fig. 28.5 Acceleration power spectral densities under (thick line) 2-N excitation at 725 Hz and (thin line) 20-N excitation at 718 Hz
Table 28.1 Relative amplitudes of power spectral densities at selected frequencies under different excitation levels
Frequency (Hz) 725 1450 2175
2N Relative PSD Amplitude 1 3.73 × 10−6 7.53 × 10−5
Frequency (Hz) 718 1436 2154
20 N Relative PSD amplitude 1 3.47 × 10−5 6.42 × 10−3
0.16
0.14
0.12
Inertance(g/N)
0.1
0.08
0.06
0.04
0.02
2N 5N 10N 15N 20N
0
710 715 720 725 730 735 740 745
Frequency (Hz)
Fig. 28.6 FRFs under various excitation magnitudes
Table 28.2 Modal parameters of different magnitude of excitation force

Excitation force 2N 5N 10 N 15 N 20 N
ωn (Hz) 725.95 722.43 721.11 718.59 718.47
ξ (%) 2.15 1.71 1.21 1.08 0.81
M (kg) 35.4 28.9 34.8 36.8 46.3
K (N/μm) 735 595 714 749 944
C (ns/m) 5675.2 4310.9 3437.4 3238.2 3085.9
28.4.3 FRF Variance on Sine Sweep Direction
Experiments were carried out with a cycle of increasing and decreasing stepped harmonic excitation of 2 N with Δf = 1 Hz.
The FRFs shown in Fig. 28.7 shows that the FRF is largely independent of sine swept direction.
28.5 Dependence of Dynamic Properties on Physical Parameters
28.5.1 Lubrication
After the experiment above were carried out, the linear guideway is lubricated. The lubricated linear guideway is then excited
with a 2 N-amplitude ascending stepped sine excitation with Δf = 1, 0.5, and 0.2 Hz. It is found from Fig. 28.8(a) that the
amplitude of FRF is gradually increased and the peak of inertance shifts to the right during experiments. Such a tendency is
different from those shown in Fig. 28.4. The variation of peak inertance in Δf = 1 Hz and Δf = 0.5 Hz is 1.77%, whereas
the variation between Δf = 0.5 Hz and Δf = 0.2 Hz is 2.14%. After the 2-N amplitude tests are carried out, another warm-
up period is applied before repeating the test with a 20-N excitation signal. The peak inertance increases slightly during
experiments in the 20-N tests as well (Fig. 28.8), but the variation is smaller than the results under 2-N excitation.
FRFs of different excitation levels with f = 1 Hz are taken to determine their modal parameters. On the one hand, both
the damping ratio and the natural frequency are reduced under a higher excitation level (Table 28.3). On the other hand, both
194 T.-Y. Wu et al.
10-8
3.8
Receptance (m/N)
3.6
3.4
3.2
Upward sweep
Downward sweep
3
734 736 738 740 742 744 746
Frequency (Hz)
Fig. 28.7 FRFs of linear guideway under different sweep direction of stepped harmonic excitation
0.1 0.12
0.11
0.09
0.1
Inertance(g/N)
Inertance(g/N)
0.08 0.09
0.07 0.08
0.07
0.06 f=1Hz f=2Hz
f=0.5Hz 0.06 f=1Hz
f=0.2Hz f=0.5Hz
0.05 0.05
725 730 735 740 745 710 715 720 725 730 735 740
(a) (b)
Fig. 28.8 Inertance FRFs of lubricated linear guideways with various frequency steps and excitation levels: (a) 2 N, (b) 20 N
Table 28.3 Modal parameters of Excitation level 2N 20 N

lubricated linear guideway
ωn (Hz) 737.60 728.40
ξ (%) 2.92 1.53
M (kg) 23.9 30.0
K (N/μm) 513 627
Light preloaded
100 Medium preloaded
Inertance(g/N )
10-1
10-2
400 500 600 700 800 900 1000 1100 1200 1300
Frequency (Hz)
Fig. 28.9 FRFs of different preloaded linear guideways in direction of excitation
Table 28.4 Mode number and Preload level Natural frequencies (Hz)
natural frequency of different
preloaded linear guideways Light 485.56 640.86 969.21 1197.41
Medium 483.80 649.37 1062.82
Table 28.5 Natural frequencies Stinger material/shape Natural frequencies (Hz)

between 400 Hz and 1300 Hz
with different stinger materials Steel, necked 473.35 650.85 1082.44
Aluminium, uniform bar 715.64
the identified stiffness and mass are increased at the same time. Furthermore, comparing the results in Tables 28.3 with those
in Table 28.2, lubrication on the linear guideway sample reduces the natural frequency and damping ratio.
28.5.2 Internal Preload Level
In the next experiment, lightly-preloaded and medium-preloaded linear guideways are tested to study the dependence of
dynamic properties on preload due to geometric interference. To detect rolling and bending, six measurement points are
arranged on a carriage, with three measurement points are measured simultaneously at one trial. A swept sine signal with
frequency increasing from 400 Hz to 1300 Hz is applied to the carriages with different preload levels.
While the natural frequencies at about 485 Hz and 650 Hz do not change with preload level (Fig. 28.9 and Table 28.4),
the inertance in the medium-preloaded linear guideway is lower, indicating higher stiffness. Also, the damping ratio in the
medium-preloaded linear guideway is lower. These findings are consistent with those reported in existing literature [8]. It is
suspect that the low-frequency natural frequencies appear due to the dynamics of the stinger. Therefore, the test on medium-
preloaded linear guideways was repeated with swept sine excitation from 400 Hz to 1300 Hz, but an aluminium stinger was
used instead of the steel counterpart. The FRFs are clearly distinguishable above 450 Hz (Fig. 28.10) thus the experiment
setup has to be further refined [9] (Table 28.5).
28.5.3 Static Lateral Load
Static lateral load (Fig. 28.1, right) can arise in many scenarios. For example, if the two rails under a moving table are not
parallel [10], static lateral load may arise if the moving table travels along the rail. Assuming rigid carriages and moving
table, such load will increase the normal force of the steel balls on one side and such normal force decreases on the opposite
196 T.-Y. Wu et al.
100
Steel stinger
Aluminum stinger
Inertance(g/N)
10-1
10-2
400 500 600 700 800 900 1000 1100 1200 1300
Frequency (Hz)
Fig. 28.10 FRFs of medium-preloaded linear guideway with different stingers in direction of excitation
Table 28.6 Modal parameters of Load (N) −10 0 10

linear guideway under selected
lateral load ωn (Hz) 722.34 722.23 722.45
ξ (%) 1.72 1.77 1.62
M (kg) 11.4 11.1 11.2
K (N/μm) 235 228 231
0.32 0.35
-20N
0.3 -15N
-10N
-5N 0.3
0.28 0N
Inertance (g/N)
5N
Inertance (g/N)
0.26 10N
15N 0.25
0.24 20N
0.22 0.2
0.2
0.15
0.18
Peak value at 712 Hz
0.16
0.1
710 715 720 725 730 -20 -10 0 10 20
Frequency (Hz) Preload (N)
(a) (b)
Fig. 28.11 (a) FRF of linear guideway under various static loads (b) Peak inertance (Solid line) and inertance at 712 Hz (Dashed line) under
various static loads
side. In the experiment, the lateral load level varied from −20 N to 20 N in 5-N intervals. Stepped sine excitation with
10-N amplitude is used with ascending frequencies Δf = 1 Hz. To exert static load, the shaker is fixed to the ground and a
charge-type force transducer replaces the IEPE-type counterpart used in other experiments.
The inertance amplitude away from natural frequency (say, at 712 Hz) shows that lateral force-deformation relationship
is similar to cubic function, with variation of some 25.7% (Table 28.6). The asymmetry in the preload-maximum amplitude
trend in Fig. 28.11 (b) is suspected to be cause of residual lateral load. Static lateral load affects the dynamic properties of
the linear guideway, and the amplitude of the FRF is lowest with the actual preload of zero.
28.6 Conclusion
Linear guideway is found to behave nonlinearly, thus steady-state excitation at various amplitudes are used to determine their
dynamic parameters. Experiments show that the dynamic properties are affected by (1) excitation level, (2) lubrication, (3)
specified preload and (4) Static lateral load. Both the natural frequency and damping ratio reduce upon increase of excitation
level. Secondly, natural frequency and damping ratio of linear guideway are reduced after lubrication. Thirdly, the lateral
force-deformation curve is similar to cubic function. These results suggest that the Hertz contact theory alone does not
describe linear guideways accurately. In terms of experimental setup, support condition and stinger dynamics are found to
affect the results. Last but not least, it is recommended to carry out hardware warm-up before obtaining steady-state results.
Acknowledgements The research is supported by the Ministry of Science of Technology, Taiwan with Grant number 108-2221-E-005 -053.
References
1. Altintas, Y., Brecher, C., Weck, M., Witt, S.: Virtual machine tool. CIRP Ann. Manuf. Technol. 54(2), 115–138 (2005)
2. Altintas, Y., Verl, A., Brecher, C., Uriarte, L., Pritschow, G.: Machine tool feed drives. CIRP Ann. Manuf. Technol. 60(2), 779–796 (2011)
3. Sun, W., Kong, X., Wang, B., Li, X.: Statics modeling and analysis of linear rolling guideway considering rolling balls contact. J. Mech. Eng.
Sci. 229(1), 168–179 (2015)
4. Ewins, D.J.: Modal Testing: Theory, Practice and Application, 2nd Edition. Research Studies Press Ltd (2000)
5. Wu, J.S., Chang, J., Hung, J.: The effect of contact Interface on dynamic characteristics of composite structures. Math. Comput. Simul. 74,
454–467 (2007)
6. Fischer-Cripps, A.C.: Introduction to Contact Mechanics, 2nd edn. LLC, Springer Science+Business Media (2007)
7. Worden, K., Tomlinson, G.R.: Nonlinearity in Structural Dynamics Detection, Identification and Modelling, Institute of Physics Publishing Ltd
(2001)
8. Brecher, C., Fey, M., Ba, S.: Identification method for damping parameters of roller linear guides. German Academic Society for Production
Engineering. 505–512 (2012)
9. Ashory, M.R.: High quality modal testing methods. PhD thesis, Imperial College of Science, Technology and Medicine (1999)
10. Liao, M.C.: “Sensor-Embedded Linear Ball Bearing for Linear Guide Way Straightness and Parallelism Monitoring,” Master Thesis
(in Chinese). National Chung Cheng University (2018)
Chapter 29
Inferring Unstable Equilibrium Configurations from Observed
Dynamics
Yawen Xu, Lawrence N. Virgin, and Richard Wiebe
Abstract Unstable equilibria play an important organizing role in nonlinear dynamic systems in a global sense. However,
it is difficult to measure them directly in a physical experiment. In this study, a digital image correlation (DIC) system is
used to capture the transient behavior of a post-buckled beam in which trajectories are generated by repeated impacts. The
dynamic data collected by the DIC system, with relatively high temporal and spatial resolution, is used estimate equilibrium
configurations of the post-buckled beam, including a detection of the presence of unstable equilibria. The results show good
agreement with the equilibrium configurations obtained from two-mode models for the beam.
Keywords Nonlinear dynamics · System identification · Saddle point
29.1 Introduction
Conservative dynamical systems are typically associated with an underlying potential energy, and one can easily imagine in
a nonlinear context a trajectory meandering through phase space, encountering and passing-by multiple stable and unstable
equilibria as it goes. Although primary interest has historically been focused on physically observable stable equilibria; in
nonlinear systems, the unstable equilibria have an important global influence especially in terms of stability in-the-large [1].
Since stable equilibria are typically separated by the unstable equilibria, the unstable states effectively act as thresholds of
the basin of attractions for stable equilibria. However, unlike stable states, the unstable equilibria are challenging to locate
from experimental data directly.
In a previous study, the authors proposed an approach to locate the saddle points in nonlinear dynamic system from
experimental dynamic trajectories [2]. The approach was shown to be effective in locating the saddle points in a system.
In this study, the structural behavior of slender, a mechanically buckled beam is used as test object. A beam under axial
loading beyond the critical load typically exhibits multiple stable equilibrium configurations: it might buckle into different
configurations. Under these post-buckled conditions if the system is subject to lateral forcing, the beam might snap-through
(saddle-node bifurcation) onto the remote complementary stable configuration. This is also true in a dynamical context in
which energy is imparted to the system via short duration strikes from an impact hammer. Whether the initial conditions
are associated with initial position (away from equilibrium) or from initial velocities, the ensuing trajectories explore the
phase space, with traces of unstable equilibria showing up through close encounters, before energy dissipation causes the
trajectories to end up on one of the available stable equilibria. Our goal is to interrogate the transient dynamic data to infer
the presence of an unstable equilibrium configuration (typically saddle points) in the nonlinear system.
Y. Xu () · L. N. Virgin
Department of Mechanical Engineering and Materials Science, Duke University, Durham, NC, USA
R. Wiebe
Department of Civil Engineering and Environmental Engineering, University of Washington, Seattle, WA, USA

200 Y. Xu et al.
29.2 Experiment Setup
The main components of the experiment set-up for the is presented in Fig. 29.1(a). The boundary conditions for the beam
is clamped-clamped. The beam deflections (in time) are captured by a digital image correlation (DIC) system, i.e. a system
that uses two cameras in stereo to capture three-dimensional deformations. Then, the beam is repeatedly disturbed by hand
using an instrumented impact hammer. The disturbances are recorded by the DIC system at a sampling rate of 2000 Hz. The
storage limits the run-time of each experimental test run to about 100 seconds.
A typical time series of the perturbation of the beam is presented in Fig. 29.1(b). The DIC system acquired full-field
measurement, although only a small subset of the spatial data is needed for these kinds of low-order (2 DOF) systems: Fig.
29.1(b) presents the deviation of the center point of the beam as an example. The snap-throughs caused by the hammer can
be clearly observed in the plot, provided the impact is ‘large-enough’.
29.3 Sample Results
Using the linear regression method in Ref. [2] on the perturbation data collected, both stable equilibria for the beam and
one saddle equilibria can be located in the system. Along with the equilibrium configurations, the local dynamics are also
estimated. Figure 29.2 compares the calculated equilibria from a 2 DOF model of the beam with the equilibrium locations
acquired from the linear regression method. The plot suggests that the results from method show good agreement with the 2
DOF model. An additional ‘mirror-image’ asymmetric saddle is also present but not shown here.
(a) (b)
2
1
y c (mm)
0
-1
-2
0 5 10 15 20 25
t (s)
Fig. 29.1 (a) Photograph of mechanically post-buckled beam showing the speckle pattern on the beam that was used by the DIC system.; (b) A
typical time series indicating perturbations of the beam resulting from hammer hits
1.5
2 DOF model
Experiment
1
0.5
y (mm)
0
-0.5
-1
-1.5
0 50 100 150 200 250
x (mm)
Fig. 29.2 Comparison of the equilibrium configurations from the 2 DOF model with the equilibrium configurations from the regression method
29 Inferring Unstable Equilibrium Configurations from Observed Dynamics 201
29.4 Conclusion
This study presents an experimental setup to measure the disturbances of post-buckled beam from impact. With appropriate
force and contact angle, the beam can exhibit snap-through behavior. These perturbation trajectories can be used to determine
the stable and unstable equilibrium configurations. The calculated equilibrium configurations based on the experimental data
shows good agreement with the result from a 2 DOF model of the beam.
References
1. Virgin, L.N.: Vibration of Axially Loaded Structures, vol. 2007. Cambridge University Press (2007)
2. Xu, Y., Virgin, L.N., Ross, S.D.: On Experimentally Locating saddle-points on a potential energy surface from observed dynamics. Mech. Syst.
Signal Process. (2019)
Chapter 30
Bolt Preload Loss Due to Modal Excitation of a C-Beam
Structure
Max Miller, Chris Johnson, Noah Sonne, John Mersch, Robert J. Kuether, Jeff Smith, Jonel Ortiz,
Gustavo Castelluccio, and Keegan J. Moore
Abstract Bolted joints often risk failure due to the loss of fastener preload when subjected to dynamic, multiaxial loads.
This process is a complex problem that depends on multiple attributes such as loading direction, rate, contact within the
threads and the interface, material properties, and many others. Current literature suggests that oscillatory shearing loads
appear to be most detrimental to the loss of preload in threaded fasteners. To investigate the effect of less idealized loading
conditions, an experimental setup employing a bolted c-beam structure is used to study loss of preload from various initial
preloads during harmonic excitation near specific resonant frequencies of the structure. The preload force is measured using
bolts equipped with internal strain gauges and the structure is excited at specific modes via sine dwell excitation with an
electrodynamic shaker. The experiments were designed to measure loss of preload as a function of excitation duration and
strength. A finite element model incorporating a fully-threaded joint is developed in parallel to investigate the effectiveness
of each at measuring and predicting bolt loosening.
Keywords Bolt loosening · Preload loss · Modal analysis · Contact mechanics · Shaker excitation
30.1 Introduction
Bolted joints are commonly used in many applications to join subcomponents in a mechanical assembly. This fastening
technology is easy to assemble and replace parts, and, when used correctly, can properly carry loads across interfaces without
permanently altering the clamped structures. During their lifetime, a typical joint experiences different types of loads over
many cycles depending on the operating environment. One type of common loading condition to bolted joint assemblies is
oscillatory dynamic forces due to vibrations in the main structure. Such dynamic excitation may instigate preload loss by
several possible mechanisms [1] such as material wear in the interface, plastic deformation in the fasteners or structures, or in
extreme cases, complete failure of the fastener under large strains [2]. The preload loss of interest to this research is the case
where loosening in the fastener occurs via backing off of the nut on the threaded bolt, resulting in a gradual or sometimes
sudden loss of preload that clamps the two or more members together. This failure mechanism occurs without permanent
damage to the structure.
M. Miller ()
Graduate Program in Architectural Acoustics, Rensselaer Polytechnic Institute – Troy, New York, NY, USA
C. Johnson
Department of Mechanical Engineering, University of Illinois – Urbana, Champaign, IL, USA
N. Sonne
Department of Aerospace Engineering, University of Colorado – Boulder, Boulder, CO, USA
J. Mersch · R. J. Kuether · J. Smith · J. Ortiz
G. Castelluccio
School of Aerospace, Transport and Manufacturing, Cranfield University – Cranfield, Bedfordshire, UK
K. J. Moore
Department of Mechanical and Materials Engineering, University of Nebraska-Lincoln, Lincoln, NE, USA

204 M. Miller et al.
The preload loss due to nut loosening has often been characterized to occur in two sequential stages [3]. The first stage,
termed Stage I loosening, corresponds to loosening caused by local plastic deformation and wear in the load bearing threads,
which carry the majority of the load during small strains. In this stage, the plastic deformation of threads enlarges the existing
gaps between them, reducing the stiffness and load in the joint. This allows slip to occur at the head of the fastener and the
contacting threads, further reducing the preload due to the nut backing off the threads. Junker first suggested that this preload
loss occurred due to gross slip at these interfaces [4], but there is no rotation between the bolt and the nut during this stage.
Jiang et al. discovered that preload loss in this stage follows a sudden drop in force, caused by cyclic plastic deformation in
the thread contacts [5]. Further studies by Liu et al. found that fretting wear in the threads causes the gentle decay in preload
following the initial drop [6]. Recent work by Zhang further investigated the wear on the first engaged thread, showing that
a combination of abrasive ploughing (from preloading) and spalling wear (from transverse cyclic loading) had occurred [7].
Following Stage I, Stage II loosening is defined as gross rotation between the nut and the fastener. Zadoks found that in
a transversely loaded joint, slip at both the threads and the bolt head were necessary to cause rotational nut loosening [8].
Later, it was found by Pai and Hess that only partial slip at these interfaces was necessary to induce rotation [9]. Translational
motions were found to be the most severe cause of Stage II self-loosening. Zhang used a finite element model of a bolted
joint under transverse loading to show that partial slip occurs due to a cyclic bending moment, causing stick-slip between the
mating threads [3].
Junker had previously shown that the loss of preload is most likely to occur as the result of shear loading on the joint [4].
This led to other studies which were interested in the loosening of bolted joints as a result of idealized loading conditions,
such as pure transverse or axial loading studied in [3, 10]. In many real-world applications, bolted joints experience more
complicated loading schemes due to a combination of these idealized loading cases. Moreover, since bolted joints exhibit
strong nonlinearity, linear models that obey superposition cannot be used to combine the results observed from these idealized
loading cases. For example, vibration, especially at frequencies near a resonance, can cause a detrimental loss of preload in
a structure [11].
This research presents a study of self-loosening in bolted joints of a c-beam structure undergoing harmonic, steady-state
excitation near different resonant frequencies of the assembly. The work shown here was conducted over the course of seven
weeks through Sandia National Laboratories’ Nonlinear Mechanics and Dynamics (NOMAD) summer research institute.
The objective of the project was to investigate loosening in preloaded bolts both experimentally and numerically with
computational mechanics finite element models. An experiment was designed to use electrodynamic shakers to harmonically
excite the beam assembly near resonance, and the axial preload force in the joint is measured using bolts equipped with strain
gages. Analogous to the experiments, a detailed finite element model of the beam assembly was created, including the small
length-scale details of the threads between the bolt and the nut. There has been little research presented in the literature
exploring the connection between modal excitation and the occurrence of bolt loosening on generic structures. A 1987
NASA report [12] made use of a cantilevered and bolted assembly of two plates as a means of achieving shear loading with
sine dwell, sine sweep and random excitation. Imagine a hole drilled through the two plates. As bending modes are excited
the holes become unaligned due to a relative difference in bending radii and a shear forces befall an installed bolt. The
work concluded that displacement amplitudes at resonance were the largest predictors of fastener loosening, however there
was little emphasis on initial fastener preload as the report did not document this value. With so much intervening research
implicating preload as the foremost deterrent of bolt loosening, expanding upon this work is warranted [10].
The paper is organized as follows: Section 30.2 presents the experimental setup with the electrodynamic shaker and strain
gage equipped bolts. Section 30.3 presents and discusses some of the experimental results obtained on the hardware, as
well as a discussion on both experimental challenges and implications of the results. Section 30.4 provides an overview
of computational methods followed by the presentation of computational findings in Sect. 30.5. General conclusions and
suggestions for future work are offered in Sect. 30.6.
30.2 Experimental Setup
The hardware used to explore the self-loosening phenomena under harmonic excitation near resonance is the so-called
c-beam assembly comprised of two identical c-beams bolted together; see Fig. 30.1 for the mechanical drawing with
dimensions. Each c-beam was made from 4340 steel alloy and was labeled either B9A or B9B to differentiate their position
in the setup. Based on previous research [13], the individual beams were tested to measure the mode shapes and natural
frequencies with free-free boundary conditions. Based on these previous results, the material properties for B9A and B9B
were calibrated to match the first six natural frequencies of the individual piece-parts. The c-beam assembly has two well
defined bolted interface on each end with raised pads to isolate the jointed interface in the structure.
30 Bolt Preload Loss Due to Modal Excitation of a C-Beam Structure 205
SECTION A-A
0.25
1.25 1.00, typ A 0.38
A
1.00
2.00 16.00
20.00
Fig. 30.1 Assembly of bolted, symmetric c-beams (dimensions in inches)
Fig. 30.2 The experimental setup consisted of a shaker connected to an amplifier (not shown) which was fed by the Siemens data acquisition
system. The stinger from the shaker held the force transducer on the end which was chemically adhered to the c-beam structure
A photograph of the experimental setup is shown in Fig. 30.2. Both the bolts and nuts used join the individual beams are
SAE Grade 9 zinc-plated steel. The bolts have internal strain gauges installed within the shank, produced by STRAINSERT
[14], allowing for real-time preload measurements. The schematic in Fig. 30.3 shows the operation of the force measuring
bolt. The c-beam structure was preloaded to an initial value, measured via the strain gage, and then freely suspended from a
semi-rigid test frame using soft bungee cords. Four triaxial accelerometers (Endevco Corp “Isotron” model #65-10) were
placed on the c-beam structure in various locations to measure the vibration response during excitation. A drive-point
accelerometer was placed at the location opposite of the input force gage to measure the induced acceleration at the drive-
point. A modal shaker (model number: MT-161) and an amplifier (model number: PA-138-1) were controlled using a Siemens
LMS SCADAS 24-bit data acquisition system to excite the c-beam structure [15]. A PCB Piezotronics force transducer
was bonded to the c-beam assembly in line with a stinger coupling the shaker and structure [16]. The shaker is driven at
frequencies corresponding with the first two modes of the c-beam structure. The in-phase and out-of-phase first bending
modes studied in this research are shown in Table 30.1. Bolt preload forces (provided via the strain bolt calibration) were
applied equally on both bolted connections at loads between 70 – 1000 lbf (311 – 445 N). Simcenter hardware and software
(courtesy of Siemens LLC) were used to collect input force, accelerometer, and strain bolt readings [15]. The raw data was
smoothed by averaging samples collected over one period corresponding with the modal frequency.
SIGNAL
CONDITIONER
INTERNAL FULL
BRIDGE STRAIN
CIRCUIT
Fig. 30.3 Schematic of the strain bolts used to determine preload [14]. The internal strain gauge calculates the axial strain in the bolt which can
then be used to determine the axial force, or preload, observed in the bolt
Table 30.1 Modal frequencies as a function of bolt preload

Mode 1
Preload [lbf ] Frequency [Hz]
70 275
160 281
500 287
700 238
1000 288
Mode 2
Preload [lbf ] Frequency [Hz]
70 325
160 337
500 350
700 350
1000 350
30.3 Experimental Results and Discussion
An experimental modal analysis was performed to identify the natural frequencies of the bolted assembly prior to performing
the shaker-excited loosening experiments. A roving impact hammer test was performed on the preloaded structure at 66
various points on the beams to spatially resolve the mode shapes. The roving hammer impact force was kept to a low level
between 8-15 lbf (36 – 67 N) to avoid contaminating the measured forced response functions (FRFs) with nonlinearity from
the joint. The results for the first two elastic modes are shown in Table 30.1. The data suggests that as the preload forces
increase, the joints stiffen and the modal frequencies shift to higher values until plateauing to a constant value. These two
modes were targeted for the bolt loosening experiments since the mode shapes loaded the joints in different ways. The first
out-of-phase bending mode applies a bending moment on the bolt head and nut during its deflection, while the first in-phase
bending mode applies a shear load in the joint.
After characterizing the vibration modes with the roving hammer impact tests, several sine dwell experiments were
performed on the c-beam assembly up to a maximum of 10 lbf (44.5 N) input force, 30 minutes in duration, and at the
preloads and modes listed in Table 30.1. These experiments were performed to scope the test parameters required to observe
loosening in the test setup. A majority of these experiments revealed preload losses of less than 5% of initial preload following
five minutes of excitation. It was discovered that only mode 1 produced complete loosening (visual rotation of the nut off the
beam) if the preload level were low enough, e.g. finger tightened. Consider the preload values listed in Table 30.1 in relation
to the maximum load rating of the strain bolts of 5000 lbf (22.2 kN). Fernando discusses, experience guided, installation
Fig. 30.4 Experimental results after applying the sliding average. The trend starts as an exponential decline and eventually starts to linearize
preload values of at least 65% of bolt yield load for fasteners in normal vibratory environments [10]. The scoping tests
revealed that the available 25 lbf (111 N) modal shaker to be insufficient for eliciting bolt loosening at more than 1% of bolt
yield.
After performing the scoping tests, the structure was harmonically excited at the frequency of the in-phase bending mode
for five minutes with the shaker attached at the center of the beam, as shown in Fig. 30.2. The root mean square (RMS)
of the input force supplied to the structure was 4.2 lbf (19 N). Figure 30.4 shows a plot of the bolt preload forces in each
as a function of time for the five-minute experiment; each bolt was torqued to approximately 5 in-lbf (0.6 N.m) resulting
in a nearly 176 lbf (782 N) preload force. It should be reiterated that these preload levels are small (i.e. loosely tightened)
since initial scoping experiments revealed that higher levels of preload did not produce any bolt loosening measurements.
The raw time histories of the bolt strain gages were oscillatory, so a moving average was used to smooth the raw data, such
that the lines in Fig. 30.4 represent the average of the bolt preload over time. This method averages a complete period of
raw data centered on a sample, then was slid over by one sample, and repeated. The preload measurements reveal that over
a five-minute period, the preload drops exponentially and then begins to decay in a linear fashion. Overall, the bolt preload
dropped by approximately 5 to 6 lbf (22 to 27 N), or about 3% of the initial preload. Previous research in the literature found
that the loss of preload begins with an exponential decay and moves into a period of linear decline before reaching Stage II
loosening [9]. The trend observed in Fig. 30.4 resembles the Stage I trend observed by Zhang [7].
In an attempt to observe Stage II self-loosening in the joints, longer duration tests were conducted using mode 1 excitation
at 5 lbf RMS (22 N). Figure 30.5 shows the results of both fifteen-minute and thirty-minute tests. These preloads were
normalized by their maximum value (170 lbf (756 N) nominal) to better show the trend observed in each test. The preload
shows a similar trend as the shorter test as the preload initially drops exponentially then decays linearly. The difference in
severity of the decreasing trends in Fig. 30.5 requires some explanation. With only small losses in preload observed for the
shorter run (the preload was still in the desired nominal range) the longer duration test followed without adjusting the preload
of the bolts. The bulk of Stage I loosening had occurred during the first 5 minutes of the test. Essentially the longer duration
test picked up where the shorter test left off; the linear tail of the Stage I loosening trend. This observation agrees with the
division of bolt loosening into two primary stages. During Stage I, plastic deformation of the threads results in a sudden loss
of preload until plateauing as the imperfectly matched threads “settle in” to place. The plateauing is the strongest indicator
that Stage I loosening was observed. The bulk of collected data follows the Stage I loosening trend observed by Pai et al. [9],
however, further experimental measurements are needed to validate this observation.
Fig. 30.5 Normalized preload results from longer duration tests
The experimental setup observed small losses in preload over long periods, with lightly torqued fasteners. This suggests
that a larger shaker was required to achieve Stage II loosening at realistic preloads since these experiments were near the
limit of the shaker’s output force rating. A brief derivation is presented here to explain some of the challenges and limitations
with resonance testing with an electrodynamic shaker. Consider the impedance equation
F k
Z mech = = iωm + c + . (30.1)
v iω
Here the complex mechanical impedance, Z mech , of a damped and harmonically driven parallel mechanical oscillator is
expressed as the ratio of the driving force, F, and the resulting velocity, v. The angular frequency is represented by ω, the
mass by m, the damping coefficient by c, the spring constant by k, and the square root of negative one by i.
At resonance the angular frequency is

k
ωr = . (30.2)
m
Substitution of ω = ωr into Eq. 30.1 results in the cancellation of the mass and stiffness terms and minimization of the
impedance,
Z mech,res = c. (30.3)
The power supplied to the structure at resonance is purely resistive, serving only to overcome mechanical damping losses.
To maintain the input force at a constant value here requires relatively large shaker output velocities. Large test article input
velocities combined with a low mechanical impedance translates to increased power consumption. The amplifier attempts
to supply the power by increasing the voltage in current regulation mode or the current in voltage regulation mode. The
electromechanical coupling within a modal shaker includes Lenz’s law linking velocity with voltage, while Ampère’s
law couples force and current. Elevated shaker output velocities correspond with equally large armature velocities, and a
proportionally large back electromotive force (EMF) opposing supplied amplifier voltage. The amplifier must overcome
the back EMF and meet the power requirement of the load. With a sufficiently powerful amplifier in current regulation
mode, a voltage increase could potentially avert shaker “force drop off” [17] at resonance. In voltage regulation mode a
current increase might do the same. The available equipment (25 lbf (111 N) modal shaker) and the lightly damped c-
beam structure invariably resulted in current limiting the shaker’s operation well before the shaker’s output force rating was
reached. Operating in voltage regulation mode, at the current limit of the shaker, was also unsuccessful concerning reaching
the specified output force of the shaker. Amplifier performance is suspected in this case but remains an issue for future
consideration. In short, shaking the structure at resonance is more demanding of the entire excitation chain when the driven
article’s resonance behavior is significant relative to the shaker’s operating characteristics. The structure-shaker interaction
was foreseen by Kerley [12], who recommended sizing the shaker at five times the mass of the device under test.
Table 30.2 Material parameters used for each component in the FEM
Component Young’s Modulus [ksi] Density [104 lbf-s2 -in−4 ] Poisson Ratio Hardening Modulus [ksi] Yield Strength [ksi]
Beam B9A 30,850 7.30 0.283 1.26 275
Beam B9B 30,470 7.40 0.283 1.26 275
Bolts 30,450 7.35 0.283 1.26 150
Nuts 30,450 7.35 0.283 1.26 150
The complex electromechanical coupling also complicates consideration of shaker placement. During the summer project,
the team utilized trial and error to evaluate shaker positioning, quantified by input power measurement. Incidentally, the
preceding impedance at resonance discussion was demonstrated by the active input power, the product of the RMS input
force and velocity, driving the structure never exceeding 5 W. The highest input power was realized by shaking the structure
at one quarter of the length of the assembly. Further modeling of the entire electromechanical system will allow a more
precise determination of the optimum excitation point and to better choose the size of the shaker needed to reach the desired
force levels.
30.4 Computational Analysis
A three-dimensional finite element model of the bolted c-beam structure was created using Cubit 15.4 [18]. The model
was constructed of 8-node first-order hexahedral elements. It consists of two c-beam halves (1250 hex elements), two nuts
(26,880 hex elements), and two fasteners (356,430 hex elements). The fastener and nut geometry were created by sweeping
a circular cross-section in a helical fashion similar to previous work by Grimmer et al. [19]. This thread shape allowed the
usage of hex elements to model thread-on-thread contact. The finite element software, SIERRA – Solid Mechanics [20], was
used to apply preloads to the fasteners (step 1) and to apply a harmonic force input (step 2). Preloads were applied using a
built-in preload subroutine that allowed a specified preload to be obtained for each joint. The subroutine applied incremental
rotations to the bolts until the desired load was reached. An elastic-plastic material model was used for all the components
(see Table 30.2). Parameters for the c-beams were borrowed from previous work using the experimental specimens [13].
The fasteners and nuts used properties from the strain gauge bolt specifications [14]. Additionally, a stick-slip friction model
was implemented using a static coefficient of friction of 0.6 and a kinetic coefficient friction of 0.2, taken from literature on
steel-on-steel contact [21].
The harmonic force input followed the functional form f (t) = A sin ωt, where A is the amplitude, ω is the angular
frequency, and t is time. The amplitude A varied from 30-2400 lbf (134 N to 10.7 kN), and ω was varied depending on the
vibration mode of interest and the preload (higher preloads had higher values of resonant frequency ω). Thus, the linear
frequencies predicted from the preloaded model ranged from 241-288 Hz for mode 1 and 321-350 Hz for mode 2. The
harmonic loading scheme was applied for up to 80 milliseconds to the c-beam structure, corresponding to approximately
22 cycles of sine loading. This force was applied on a node at the center of beam B9B, similar to the location tested during
the experiments. No other loadings or boundary conditions were specified, thus the beam assembly was in a free-free state.
Explicit analysis was used for both the preloading and the vibrational loading steps due to the complex contact conditions
within the threaded connections. This limited the maximum number of cycles that could be simulated during vibration
assuming a maximum of two day run time for each load case using parallel computing with distributed memory.
30.5 Computational Results and Discussion
The simulations and the results were divided into two steps. The first step was to apply preload to the fasteners by rotating
the nut onto the threaded bolt. The second step of each simulation applied an external harmonic force near resonance of the
preloaded structure in the first step. The contour plots in Fig. 30.6 show the von Mises stress for the first joint (labeled bolt
1 throughout) for one of the preload forces to a level of 150 lbf (670 N). In this case, the shank of the bolt appears to carry
a uniform load as evidenced by the constant contour of stresses. Gradients in the stress state appear near the bolt head and
threads due to geometry of the modeled joint. The maximum stress in the static preload analysis is in the first contacting
thread in the fastener/nut interaction. At a preload of 150 lbf (670 N), there was no plastic deformation observed in any
Fig. 30.6 Von Mises stress in joint 1 of the c-beam structure after preloading to 150 lbf
location of the finite element model. Other preload cases were simulated, and the results showed the plastic deformation
occurred at a preload of 2800 lbf (12.5 kN), which occurred in the first contacting thread of the fastener.
Prior to applying resonant loads with the externally applied force in the second step of the simulation, a simulation without
harmonic loading was ran for the preloaded structure. These simulations are referred as unloaded simulations. The unloaded
simulation for a 150 lbf (667 N) preload is shown in Fig. 30.7, where the preload was monitored in each bolt by summing
the normal forces acting on the bottom of each bolt head. In the unloaded simulations, preload loss was observed in all
simulated preload levels. The decay in preload was linear for all cases, such as the decay shown in Fig. 30.7, with higher
preloads having larger magnitude slopes. Rotation was found to follow incremental slip for preloads below 500 lbf (2.2 kN),
and cyclic rocking rotation for preloads 500 lbf (2.2 kN) and over. Decay in the preload was likely due to residual stress
waves resulting from the preloading step, allowing self-loosening in the fasteners. Movement in the bolt/nut, especially at
higher preload levels, allowed the joint to essentially ratchet itself off through incremental slip. This result highlights some
of the challenges with using explicit methods for nonlinear contact mechanics problems.
The step two simulation results in Fig. 30.8 show the bolt preload and nut rotation for bolts preloaded to 150 lbf (667 N)
with a 25 lbf (111 N) harmonic force applied at the shaker input location. This simulation covered roughly 20 cycles of
sine loading at 275 Hz for mode 1 excitation and 325 Hz for mode 2. Monitoring the preload levels, the harmonic forcing
appeared to have caused loosening in the joints over the course of roughly 70 ms. For the mode 1 simulation (top row), the
preload in the bolts is spikes up throughout the loading scenario due to the bending of the c-beams during mode 1 excitation,
which press up against the bottom of the bolt head. This trend is observed for mode 2 as well but to a lesser degree. Thus,
the ‘true’ preload in the joint is closer to the minimum value of each oscillation, where the beam is nominally flat with the
bolt head surface.
In the mode 1 simulation, the preload drops sharply at around 50 milliseconds. In the mode 2 simulation, the preload
loss is linear, which could be largely due to the loosening behavior seen in the ‘unloaded’ simulation. The maximum force
amplitude is much smaller compared to the mode 1 simulation. Preload also drops less, reaching 110 lbf (489 N) at the end
of the simulation. Rotation of each bolt was calculated using node positions of a node at the center of the bolt head and
a node at a corner of the head. The rotation follows a cyclic pattern, with decreases in rotation due to the displacement in
the structure from the harmonic loading. True rotation in the joint follows a stick-slip pattern, with only positive increases
in the rotation. The maximum rotation in the bolt increases with each cycle, indicating that the joint is loosening through
ratcheting. At 50 ms, the periodic behavior is briefly suppressed, suggesting the possibility of complete slip between the bolt
and the nut for this brief period. Rotation in the nut (not shown) was roughly 1.6-1.9 times the value of the bolt rotation for
all simulations.
The results in Fig. 30.8 are difficult to interpret knowing that the bolt loses preload under a ‘no load’ condition. Subtracting
off the data from the unloaded simulation loosening provides a better idea of what degree of the loosening is due to the
Fig. 30.7 Preload in (blue) bolt 1 and (orange) bolt 2 during an unloaded simulation (no applied boundary conditions/loading) on a 150 lbf
preloaded structure
Fig. 30.8 Preload and relative rotation in each joint during mode 1 (top row) and mode 2 (bottom row) excitation at 25 lbf (111 N) for a structure
preloaded to 150 lbf (667 N)
harmonic force input. While it is not quantifiably correct to apply superposition to bolt self-loosening, it can provide some
insight into the degree of loosening caused by the resonance. The results with the subtraction of unloaded response is shown
Fig. 30.9 Preload and relative rotation in each joint during mode 1 (top row) and mode 2 (bottom row) excitation at 25 lbf (111 N) for a structure
preloaded to 150 lbf (667 N) after subtracting off the unloaded empty simulation data
Fig. 30.10 Preload (left) and rotation (right) in bolt 1 for a mode 1 excitation at 288 Hz, 100 lbf input force for a structure preloaded to 2400 lbf
[10.7 kN]
in Fig. 30.9. In the mode 1 simulation, the sudden decrease in preload becomes clearer at the 40 to 50 ms mark, along with a
break in the periodic rotation behavior, resulting in a final preload of roughly 50 lbf. (222 N). The mode 2 simulation shows
less preload loss, with a final preload level of 140 lbf (622 N) or about 6.7% of the initial preload. Higher preloads up to 2400
lbf (10.7 kN) were simulated under harmonic loading conditions, but the preload loss was difficult to distinguish between
losses due to the unloaded simulation loosening versus loss due to the forcing input. The plot in Fig. 30.10 shows that at a
higher shaker force input of 100 lbf (445 N), loosening could be observed in 2400 lbf (10.7 kN) preloaded joints under mode
1 excitation, where about 200 lbf (889 N) of preload, or 8.3%, was lost due to a rotation in the bolts/nuts of roughly 0.2-0.6
degrees. In this figure, the unloaded simulation data has been subtracted from the raw preload and rotation data.
Thus, the model demonstrates the ability to simulate bolt self-loosening under resonance using a fully-threaded, high-
fidelity FEM. During each harmonic loading simulation, the contacting threads in each joint experience oscillating contact
Fig. 30.11 Snapshot of the contact stress in joint 1 during mode 1 excitation at 25 lbf (111 N)for a structure preloaded to 150 lbf (667 N)
stress. A contour plot of the stress state during oscillation is shown in Fig. 30.11 shows how the contact in the threads is
intermittent and not uniform around the circumference of the joint. The oscillation in the thread contact is not consistent
across the whole joint, causing some fraction of the threads to be in contact while others are not. This intermittent contact
allows the bolt and the nut to rotate with respect to each other through partial slip. This is similar to behavior investigated
previously by Pai and Hess [9].
30.6 Conclusions
A direct comparison of the experimental and computational results is not appropriate given the different stages of self-
loosening observed in the studies. Experimentally, Stage I preload losses were observed with a suspected cause of plastic
deformation of bolt thread tips. The classification of Stage I follows from preload loss decay trends, specifically initial
exponential drops followed by linear decreases plateauing with time. The initial exponential drop is not present when repeat
modally excited loosening tests are performed without intermediately torqueing fasteners. This observation suggests the
initial exponential preload drop follows from conditions set by fastener installation. For vibratory environments, torqueing
fasteners slightly beyond the desired preload may compensate the predictable effect. Relative fastener rotation was not
observed during these studies due to the limited structure input power. Mode 1 excitation was used for the majority of tests
to maximize c-beam input power.
Computationally, self-loosening followed that of Stage II behavior with gross slip between the bolt and nut. The Stage II
classification stems from the demonstration of relative fastener rotation using nodal displacements, as well as large preload
losses observed in many tests. A clear benefit of an accurate computational model is the ability to specify a desired input
force. Stage I loosening in the computational model may be realized by artificially lowering the elastic moduli of the threads
or increasing the fidelity of the thread mesh. Doing so may lower the preload value at which plastic deformation is observed
in the modeled threads, more closely matching experimental observations. Modeling the threads as a separate body bonded
to the bolt shank may help preserve other mechanics of the joint. Such an approach may help more realistically model
complexities in both thread geometry imperfections and material properties. Sufficient plastic deformation in the thread tips
may help better match the experimental Stage I results. To this end, imperfectly matched internal and external threads, more
in line with reality, may see preload losses as the threads settle in under dynamic excitation. Fretting wear, also implicated
in Stage I loosening, is not currently computationally feasible concerning finite element analysis. Here thread mesh fidelity
may help better resolve contact surfaces.
Acknowledgements This research was conducted at the 2019 Nonlinear Mechanics and Dynamics (NOMAD) Research Institute supported by
Sandia National Laboratories and hosted by the University of New Mexico. Sandia National Laboratories is a multi-mission laboratory managed
and operated by National Technology and Engineering Solutions of Sandia, LLC., a wholly owned subsidiary of Honeywell International, Inc.,
for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA-0003525. The authors would also like to
thank Bill Flynn from Siemens Industry Software NV for supplying the data acquisition systems used to collect the experimental measurements
presented throughout this work. SAND2019-12525 C.
References
1. Bickford, J.H.: An Introduction to the Design and Behavior of Bolted Joints, 3rd edn. Taylor & Francis Group, New York (1995)
2. Mersch, J.P., Smith, J.A., Johnson, E.P.: A Case Study for the Low Fidelity Modeling of Threaded Fasteners Subject to Tensile Loadings at
Low and High Strain Rates. ASME Pressure Vessels and Piping Conference, PVP2017-65518, ASME, Waikoloa, HI (2017)
3. Zhang, M., Jiang, Y., Lee, C.-H.: Finite element modeling of self-loosening of bolted joints. J. Mech. Des. 129(2), 218 (2007)
4. Junker, G.H.: New criteria for self-loosening faster under vibration. 78, 314–335 (1969)
5. Jiang, Y., Zhang, M., Lee, C.-H.: A study of early stage self-loosening of bolted joints. J. Mech. Des. 125(3), 518 (2003)
6. Liu, J., et al.: Experimental and numerical studies of bolted joints subjected to axial excitation. Wear. 346–347, 66–77 (2016)
7. Zhang, M., Lu, L., Wang, W., Zeng, D.: The roles of thread wear on self-loosening behavior of bolted joints under transverse cyclic loading.
Wear. 394–395, 30–39 (2018)
8. Zadoks, R.I., Yu, X.: An investigation of the self-loosening behavior of bolts under transverse vibration. 208, 189–209 (1997)
9. Pai, N.G., Hess, D.P.: Three-dimensional finite element analysis of threaded fastener loosening due to dynamic shear load. Eng. Fail. Anal.
9(4), 383–402 (2002)
10. Fernando, S.: Mechanisms and prevention of vibration loosening in bolted joints. Aust. J. Mech. Eng. 2(2), 73–92 (2017)
11. DellaCorte, C., Howard, S.A., Hess, D.P.: “Preload Loss in a Spacecraft Fastener Via Vibration-Induced Unwinding,” NASA/TP—2018-
219787, NASA Technical Reports Server (NTRS). U.S. National Aeronautics and Space Administration, Washington, D.C. (2018)
12. Kerley, J.J.: “An Application of Retroduction to Analyzing and Testing the Backing off of Nuts and Bolts during Dynamic Loading,” NASA
Technical Memorandum 4001, NASA Technical Reports Server (NTRS). U.S. National Aeronautics and Space Administration, Washington,
D.C. (1987)
13. Fronk, M., Guerra, G., Southwick, M., Kuether, R.J., Brink, A.R., Tiso, P., Quinn, D.D.: Predictive modeling of bolted assemblies with surface
irregularities. In: 37th International Modal Analysis Conference (IMAC XXXVII), Orlando, Florida (2019)
14. Strainsert. Standard Internally Gaged Hex Head Cap Screws, SXS Series (2019)
15. Simcenter SCADAS. https://www.plm.automation.siemens.com/global/en/products/simcenter/simcenter-physical-testing.html. Siemens AG
(2019). Last accessed, Sept 2019
16. Force Sensors. https://www.pcb.com/sensors-for-test-measurement/force-sensors. PCB Piezotronics (2019). Last accessed, Sept 2019
17. Varoto, P.S., Rodrigues de Oliveira, L.P.: On the force drop off phenomenon in shaker testing in experimental modal analysis. Shock and Vib.
9(4), 165–175 (2002)
18. Geometry and Mesh Generation Toolkit. Ver 15.4 2019. https://cubit.sandia.gov/. Last accessed, Sept 2019
19. Grimmer, P.W., Mersch, J.P., Smith, J.A., Veytskin, Y.B., Susan, D.F.: Modeling empirical size relationships on load-displacement behavior
and failure in threaded fasteners. AIAA Scitech 2019 Forum. 1–10 (2019)
20. Sierra Solid Mechanics Team, “Sierra/Solid Mechanics 4.52 User’s Guide,” Sandia National Laboratories, Albuquerque, NM, 2018.
SAND2019-2715
21. CRC Handbook of Physical Quantities. Boca Raton, FL: CRC Press, 1997: 145-156
Chapter 31
Nonlinear Dynamic Analysis of Bolted Joints: Detailed
and Equivalent Modelling
N. Jamia, H. Jalali, J. Taghipour, M. I. Friswell, and H. H. Khodaparast
Abstract A standard finite element analysis of individual components in aero engine and other systems shows a high
accuracy compared to experimental measurements of the system response. However when it comes to assemblies, the
conventional linear approaches fail to deliver good accuracy. This is due to the uncertain physical phenomena in the contact
interface of the joints. A nonlinear contact problem is introduced by the joint and influences the overall dynamic behavior
of the engine assembly. Therefore, the linear dynamic models must be coupled with nonlinear analysis of the assembly to
investigate the accurate dynamics of the nonlinear system. Flanges are widely used joints that represent the main source
of nonlinearities in assemblies. In this study, a finite element simulation of two bolted flanges is considered to identify the
nonlinear behavior of the bolted flange joint caused by the presence of friction in contact interfaces. A detailed model of a
bolted joint was built in ANSYS in order to evaluate the energy dissipated in a bolted joint and therefore provide accurate
modeling of joint interfaces. Due to the high cost of the detailed model, an equivalent model was derived and predictions
from this model are compared to the detailed model results in order to provide a robust model for designing bolted joints.
Keywords Bolted flange · Nonlinear analysis · Detailed model · Equivalent model
31.1 Introduction
Bolted joints are widely used due to their simplicity in assembling mechanical structures. However, despite their simplicity,
their inherent dynamics is too complex to be easily analyzed. Bolted lap and flange joints show specific nonlinear behaviour
particularly at higher excitation amplitudes resulting from nonlinear slip and slap mechanisms in their contact interface.
These mechanisms manifest usually themselves as nonlinear stiffness and damping effects and the transfer of mechanical
energy from lower to higher frequencies. Due to the micro-slip mechanism between the mating surfaces, a hysteresis
phenomenon is present in bolted joint systems when a cyclic load is applied. The area inside a hysteresis loop represents the
amount of dissipated energy in the contact interface in one cycle [1]. Many efforts have been made in the past to model the
nonlinearities associated when hysteresis phenomena. It should be mentioned that any useful joint model must be capable
of reproducing the properties of the contact interfaces in terms of energy dissipation and nonlinear stiffness. Moreover, the
model parameters must be deducible from joint level experimental or very fine-scale finite element results and integration of
the joint model into structural-level models must be also practical [2].
The first attempt to model hysteresis behaviour was undertaken by Timoshenko [3]. Later, Iwan [4, 5] extended the
approach suggested by Timoshenko which assumes that a general hysteretic system consists of a large number of elastoplastic
elements having different yield levels. Gaul et al. [6] experimentally studied the nonlinear damping and response function
characteristics of bolted joints. Lenz and Gaul [7] and Gaul and Lenz [8] studied the behaviour of a bolted joint when it is
subjected to longitudinal and torsional forces. They used the Valanis model, well known from plasticity [9], as a reduced
order lumped model and showed that this model is capable of describing the nonlinear dynamic behaviour of a bolted joint
contact interface. Segalman [10] used the parallel series Iwan model in series with a soft element to represent respectively
the contact patch and the rest of the joint and showed that this model is a good candidate for capturing the nonlinear physics
N. Jamia () · J. Taghipour · M. I. Friswell · H. H. Khodaparast

College of Engineering, Swansea University, Swansea, UK
H. Jalali
Department of Mechanical Engineering, Arak University of Technology, Arak, Iran

216 N. Jamia et al.
involving in contact interfaces. Considering the softening effect at the contact interface of a bolted lap-joint, Ahmadian and
Jalali [11] proposed a lumped joint model consisting of linear and cubic springs and linear damping elements. Yuan et al.
[12] used the adjusted Iwan model to simulate the behaviour of a bolted joint.
In this paper, the microslip behavior at the contact pair interface between two flanges is analysed with two finite element
models of a bolted joint; a detailed 3D FE model and a 1D equivalent model. The modelling details are described and a
comparison of natural frequencies and micro slip behavior is performed between the two models. Then static and dynamic
analyse is were simulated to generate the force-displacement curve for both cases and comparisons between the results of
the two models are performed.
31.2 Detailed Fe Modelling
In this section a detailed three-dimensional finite element model of two bolted joints is developed. The FEM analysis is
carried out using the commercial software ANSYS Workbench.
A three bolted joint consisting of three M10 bolts, three nuts, three washers and two rectangular flanges is considered. The
two flanges are connected through the three M10 bolts as shown in Fig. 31.1. Structural steel with the following parameters
was assigned as the joint material; E = 200GPa, ν = 0.3 and ρ = 7.85 g/cm3 , where E is the Young’s modulus, ν is the
Poisson’s ratio and ρ is the mass density. The different design parameters of the flange are given in Table 31.1.
Three types of contact embedded in ANSYS are used in the model; bonded, frictionless and frictional. Frictionless
behavior allows the parts to slide relative to one another without any resistance. This type of behaviour was assigned to
the contact between the bolt-shank and flanges holes and the bolt-shank to washer interface. The frictional contact provides
shear forces between the parts in contact and was assigned between the contact bolt head to the top surface of the upper
flange, the nut interface to the bottom surface of the lower flange and between the upper flange and the lower flange. Finally
a bonded contact was assigned to the contact nut interface to the bolt. The contact interfaces of the bolted joint are modelled
in ANSYS based on the type of the contact between parts. In the normal direction, ANSYS prevents the contacting bodies
from interpenetration by enforcing contact compatibility using contact algorithms. For the case of nonlinear solid body
contact, the Augmented Lagrange and Pure Penalty formulations are chosen for frictional or frictionless and bonded contacts
respectively. These methods combine robustness and high accuracy of the nonlinear contact solution. These two formulations
are a penalty-based contact formulation. Friction behaviour is considered following Coulomb’s law Ftangential ≤ μ Fnormal
where μ is the coefficient of static friction. Once the tangential force Ftangential exceeds μ Fnormal sliding will occur. If μ is
zero then the contact behaves as fricitonless. In this model, a coefficient of friction μ = 0.2 was assigned to the contacting
area between the two flanges.
Regarding loads and boundary conditions, the lower side of the bottom flange was constrained and a bolt axial load
equals to 5kN was applied to the three bolts. A pretension in the bolts was generated at the mid-plan of the bolt. To specify
the pretension in the bolt, a local coordinate system was defined with the Z-axis along the bolt length as shown in Fig.
31.2(a). To provide high accuracy in the contact analysis, the model is meshed with a fine mesh in the contact area, bolts,
Fig. 31.1 Bolted flange design
Table 31.1 3-D model geometry Description Measurement

parameters
Flange length 100 mm
Flange height 200 mm
Flange thickness 7 mm
Flange width 50 mm
Bolts M10 mm
31 Nonlinear Dynamic Analysis of Bolted Joints: Detailed and Equivalent Modelling 217
(a) (b)
Fig. 31.2 (a) Pretension in the bolts (b) Mesh details of the model
(a) (b)
Fig. 31.3 (a) Contact status (b) Contact pressure at the contact interface
Table 31.2 Simulated natural frequencies (Hz)

– ω1 ω2 ω3 ω4 ω5 ω6
Simulated 34.69 181.18 245.88 419.95 590.04 760.2
nuts and washers. Tetrahedral elements were used in the mesh as shown in Fig. 31.2(b). The assembly has 16,248 elements
and 84,317 nodes.
Due to the presence of the bolt pretension, the simulation is performed over two load steps. During the first load step, the
different assembly interference are calculated to provide the prescribed preload and reaction forces into the bolts and mimic
the assembly of the two flanges. Once the bolt pretension load is achieved, then calculated reaction force agrees with the
applied preload in the bolts. During the second load step, the external load was applied to the bolted joint.
The contact status shown in Fig. 31.3(a) shows the presence of sliding and sticking behaviour in the contacting area
between the two flanges. Sliding is caused by the elements that slide on the surface of the contact; however sticking is caused
by the contact elements that cannot move and therefore penetration occurs. The resulting contact pressure at the contact
interface given in Fig. 31.3(b) shows a nearly uniform distribution.
To determine the natural frequencies and mode shapes of the system, an eigenfrequency study was performed using
ANSYS Workbench. A fixed constraint was applied to the base of the lower flange. The first 6 natural frequencies and their
corresponding mode shapes are given in Table 31.2 and Fig. 31.4, respectively.
218 N. Jamia et al.
Mode 1 Mode 2 Mode 3 Mode 4 Mode 5 Mode 6
Fig. 31.4 The first six modes of the bolted flanges model
(a) (b)
Fig. 31.5 Force-displacement curve obtained from (a) static loading (b) dynamic loading
The interest in this paper is on the bending modes, and their corresponding natural frequencies will be compared to the
equivalent model given in the next section.
A static load was applied to the lower face of the upper flange, close to the contact interface between the two flanges.
This will help to excite the nonlinear behaviour of the joint interface and creates a sliding behavior in the contact interface. A
parametric study was performed where the amplitude of the static load was varied and the relative displacement at a point of
the contact interface between the two flanges was monitored. Fig. 31.5(a) shows the behaviour of the relative displacement at
the contact interface between the two flanges under the action of the static external load. The graph shows a linear behavior
while the force is between -3kN and 3kN and beyond that range, a nonlinear behaviour occurs. This static study gave an
approximation of the value of static force starting from which the structure starts to behave nonlinearly.
A dynamic analysis was performed where a periodic force is applied to the bolted joint. This force is defined as
F(t) = |F| sin (ω t) where | F | = 4 kN and ω = 15 Hz. A simulation was performed for a period of 4 s with a time step
equal to 0.0025 s and the relative displacement of a point from the contact area of the bolted flanges is calculated for each
time step. Fig. 31.5(b) shows the curve of the external load versus the relative displacement at the interface contact between
the two flanges for a single cycle of loading-unloading. It can be observed that the curve shows a hysteresis loop which has
a parallelogram shape where two stages can be identified. The first stage where the variation of the relative displacement is
very slow relatively with the variation of the applied load. This behavior corresponds to sticking. Increasing the applied load
leads to gross slip and a microslip occurs during the transition between the stick and gross slip. This loop is used to quantify
the energy dissipated in the joint contact due to the damping of the joint. The detailed model has a very high cost in terms of
computational time, and therefore the numerical limitations in terms of mesh and time step are the reason behind the noise
in the data shown in Fig. 31.5(b). Therefore, a 1D equivalent model was performed in the next section.
31.3 Equivalent Beam Modelling
An effective approach to construct an equivalent model for the bolted flange structure considered in previous section is
through using beam elements for modelling the beam sections and using appropriate joint models for considering the effects
of slip and slap mechanisms in the contact interface. A schematic of FE model of the structure is shown in Fig. 31.6. DOF
constraints (or MPCs) are used to relate the movement of the end nodes in modelling two perpendicular beam sections.
Beam elements are two-nodded with {u, v, θ z } as the DOFs at each node. Joint elements are used to model the effect of
contact interface dynamics in the FE modelling. Because of the two slap and slip mechanisms which are likely to develop
in normal, i.e. x, and tangential, i.e. y, directions of the contact interface, the following models are proposed to consider the
effects of slip and slap mechanism.
Micro/macro slip can initiate in tangential direction of the contact interface if loading and bolt preload conditions are
enough to initiate them. Prior to micro-slip the behaviour of contact interface is linear under small loading amplitudes. It
is assumed in this paper that macro-slip will not happen in the contact interface and the effort is made to model micro-slip
phenomenon. The Jenkin’s element is used to model the micro/slip in tangential direction of the bolt jointed flange as is
shown in Fig. 31.7.
In Fig. 31.7 kt is the tangential stiffness, μ is the friction coefficient and FN is the pointwise normal force applied to the
contact interface due to bolt clamping force. The force displacement relationship for the Jenkins element can be expressed
as,
Fig. 31.6 FE modelling of bolt jointed flange
Fi
vi +μP
i Sliding
Sticking
Friction
kt 0
Sticking
µFN
Sliding
Fj –μP
vj
j 0
Displacement
(a) (b)
Fig. 31.7 (a) Jenkin’s element used for modelling slip mechanism (b) Corresponding force-displacement diagram
220 N. Jamia et al.
Table 31.3 Comparison of the – ω1 ω2 ω5

simulated and updated bending
natural frequencies (Hz), Simulated 34.69 181.18 590.04
kt = 1341063 N and Updated 34.92 180.97 589.87
m
kn = 62698201 N Error 0.66 −0.11 −0.03
m

1 −1 vi −1
k =F , Fi = −Fj = F (31.1)
−1 1 vj 1
where,

kt , if slider was not sliding in previous iteration
k= (31.2)
0, if slider was sliding in previous iteration

±μFN , if slider is sliding
F = (31.3)
kt u1 , if slider is not sliding but has slid bef ore
u1 = ui − uj − us (31.4)
The slip element shown in Fig. 31.7 is used between each two adjacent i and j nodes in the bolted section of the structure as
it is shown in Fig. 31.6. In using the slip elements it is assumed that all slip elements have the same tangential stiffness kt . Due
to a variable distribution of normal pressure in the contact interface, different FN are considered for different slip elements
used in the contact interface. The distribution of pointwise normal forces FN is similar to the normal pressure distribution
in the contact interface. It is assumed that in the simulated structure in the previous sections, and due to the location of the
applied load, the slap mechanism does not activated. Therefore, the behaviour of the contact interface in normal direction is
considered linear and a linear stiffness kn is used in the joint model to represent the normal stiffness of the contact interface.
The normal and tangential stiffness coefficients can be identified from simulated linear dynamic properties of the structure
such as natural frequencies presented in previous sections. In next section identification of the linear and nonlinear joint
model parameters are considered.
31.4 Joint Model Identification Results
First three bending natural frequencies presented in Table 31.2 were used and the contact stiffness coefficients in tangential
and normal directions were identified. Table 31.3 shows the simulated and updated natural frequencies and their difference.
Results presented in this Table shows that the equivalent beam model is well capable in reproducing the linear behaviour of
the detailed model presented in previous sections:
Next identification of the distribution of the normal forces FN is considered. In Fig. (31.8a) an identified distribution for
normal forces and in Fig. (31.8b) the force-displacement curve are compared for equivalent model and detail modelling.
It is worth noting that since the linear model presented in Table 31.3 is a well accurate model, it is possible to use a proper
identification approach to match the force-displacement diagram from detailed and equivalent models.
31.5 Conclusion
In this paper, a detailed model of bolted flanges was performed to capture the friction behaviour in the contact interface due to
microslip phenomena. A static and dynamic load were applied to the joint and the force-displacement curves and hysteresis
loops were obtained for the contact interface. Since detail modelling is not possible in real engineering applications, an
equivalent model was constructed by using beam element and a proposed joint element. The linear and nonlinear parameters
of the equivalent model were obtained by using the results from detailed modelling approach.
3000 6000
2500 4000
Static force (N)

2000 2000
FN(N)
1500 0
1000
–2000
500
–4000
0
1 2 3 4 5 6 7 8 9 10 11 –6000
nodes in contact interface –0.03 –0.02 –0.01 0 0.01 0.02 0.03 0.04
Relative displacement (mm)
(a)
(b)
Fig. 31.8 (a) an initial identification for distribution of FN (b) Comparison between force-displacement diagrams
Acknowledgements The authors gratefully acknowledge the support of the Engineering and Physical Sciences Research Council through the
award of the Programme Grant “Digital Twins for Improved Dynamic Design”, grant number EP/R006768.
References
1. Segalman, D.J., Starr, M.J.: Relationships among Certain Joint Constitutive Models. Sandia National Laboratories, Albuquerque, NM (2004)
2. Segalman, D.J., Paez, T., Smallwood, D., Sumali, A., Arbina, A.: Status and Integrated Road-Map for Joints Modeling Research. Sandia
National Laboratories, Albuquerque (2003)
3. Timoshenko, S.P.: Strength of Materials. D. Van Nostran company, New York (1930)
4. Iwan, W.D.: A distributed-element model for hysteresis and its steady-state dynamic response. J. Appl. Mech. Transact. ASME. 893–900
(1966)
5. Iwan, W.D.: On a class of models for the yielding behavior of continuous composite systems. J. Appl. Mech. 89, 612–617 (1967)
6. Gaul, L., Nackenhorst, U., Willner, K., Lenz, K.: Nonlinear Vibration Damping of Structures with Bolted Joints. In: Proceedings of the 12th
International Modal Analysis Conference, Hawaii (1994)
7. Lenz, J., Gaul, L.: The Influence of Microslip on the Dynamic Behavior of Bolted Joints. In: 13th International Modal Analysis Conference,
Nashville (1995)
8. Gaul, L., Lenz, J.: Nonlinear dynamics of structures assembled by bolted joints. Acta Mech. 125, 169–181 (1997)
9. Valanis, K.C.: Fundamental consequence of a new intrinsic time measure-plasticity as a limit of the Endochronic theory. Arch. Mech. 32,
171–191 (1980)
10. Segalman D.J., “An Initial Overview of Iwan Modeling for Mechanical Joints”, Technical Report, Sandia National Laboratories, SAND2001-
0811, 2001
11. Ahmadian, H., Jalali, H.: Identification of bolted lap joint parameters in assembled structures. Mech. Syst. Signal Process. 21, 1041–1050
(2007)
12. Yuan, P.P., Ren, W.X., Jian, Z.J.: Dynamic tests and model updating of nonlinear beam structures with bolted joints. Mech. Syst. Signal Process.
126, 193–210 (2019)
Chapter 32
The Relevance of Nonlinear Normal Modes for Randomly
Excited Nonlinear Mechanical Systems
Thomas Breunung and George Haller
Abstract Nonlinear normal modes are of great importance to understand the dynamical behavior of nonlinear mechanical
systems and serve as natural candidates for model order reduction. Therefore, their accurate computation as well as
mathematical foundation has extensively been studied. While numerous results are available for unforced and periodically
forced nonlinear mechanical systems, the case of random external forcing is commonly not considered in the literature. Here,
we clarify the relevance of nonlinear normal modes in the case of small white noise excitation. We demonstrate our results
on explicit mechanical systems.
Keywords Nonlinear normal modes · Reduced order modelling · Stochastic dynamical systems · Spectral
submanifolds · Gaussian white noise
To account for parameter uncertainty, unmodeled degrees of freedom or unknown disturbances in realistic engineering
structures, the use of statistical methods is unavoidable (c.f. Lutes and Sarkani [1]). Commonly Gaussian white noise is
considered to account for such random perturbations. Whether the deterministic concept of a nonlinear normal mode is of
relevance under uncertainties or random external excitation is, however, unclear.
Various definitions of nonlinear normal modes have been proposed in the literature (cf. Kerschen et al. [2] for a review)
and multiple computational algorithms for either the unforced or periodically forced nonlinear mechanical systems have been
developed (c.f. Renson et al. [3]). Recently, Haller and Ponsioen [4] identified the smoothest nonlinear continuation of an
a spectral subspace of the linearization as spectral submanifolds (SSMs). While the importance and relevance of nonlinear
normal modes for e.g. reduced order modelling, has proven valuable in experimental settings, the computational tools assume
a completely deterministic nature of the system. This assumption, however, does not generally hold for realistic structures
which are subject to parameter uncertainties and external (unmodeled) disturbances.
Whether counterparts to the deterministic dynamical features serving as natural candidates for model order reduction,
such as normally hyperbolic invariant manifolds or stable/center manifolds of fixed points, exist for a randomly perturbed
mechanical system is intricate to establish mathematically. Berglund and Gentz [5] assume an idealized slow-fast
decomposition in the deterministic system, while the stable and center manifold established by Arnold [6] generally depend
on the specific realization of the random process. Therefore each realization results in a different reduced order model, which
leads to significant computational costs and limits relevance of the reduced order model significantly.
Independent of the realization is the probability density function, which depends on the position and velocity and gives the
probability that the mechanical system is at the specified location in the state space. In the case of the Gaussian white noise
excitation, the time evolution of the probability density function is governed by the classical Fokker-Planck equation (c.f.
Risken [7]). Explicit solutions of the Fokker-Planck equation are only available in specific cases and approximate numerical
methods discretizing the Fokker-Planck equation or Monte Carlo simulations are computationally expensive (c.f. Soize [8]).
Recently Haller et al. [9] derived tools to identify material barriers, i.e. material surfaces with extremal diffusive transport
of the probability density across. By the definition of these barriers the transport is purely driven the by small stochastic
perturbations of the otherwise deterministic system. E.g. perfect barriers block the transport of probability across at leading
order completely and thereby indicate regions of the phase space that trajectories generally not penetrate. Similarly ridges of
the diffusion barrier strength (DBS) highlight the surfaces with strong diffusive transport, i.e. regions in which trajectories
T. Breunung () · G. Haller

Institute for Mechanical Systems, Department of Mechanical and Process Engineering, Zürich, Switzerland

224 T. Breunung and G. Haller
0.2 0.045
Perfect barriers
Fast stable Eigenspace
Slow manifold Fast Spectral Submanifold
0.15 0.04
0.035
0.1
0.01
0.03
0.05 0.008
modal coordinate p3
0.006
Velocity
0.025
DBS
0.004
0
0.002
0.02
0
-0.05
0.015 -0.002
0.5
-0.004
-0.1 -0.006
0.01
-0.008
0
-0.15 0.005 -0.01
-1 -0.8 -0.6
-0.4 -0.2
0 0.2
-0.2 0 0.4 0.6 -0.5
0.8 1
-0.5 0 0.5 1 1.5 modal coodrinate p1
Postion modal coordinate p2
(a) Duffing oscillator (1) with the non-dimensional (b) Shaw-Pierre example (2) with the non-dimensional
parameters = 3.1, = 1, = 1 and = 0.1. The parameters 1 = 2 = 1, 1 = 2 = 3 = 0.1, 1 =
identified DBS ridge closely aligns with the slow 2 = 3 = 1, = 1 and = 0.1. The identifies perfect
manifold connecting the saddle at (0,0) with the stable barriers align closely with the SSM tangent to the fast
fixed point at (1,0). modal subspace. The SSM is computed by with the
SSMtool [12].
Fig. 32.1 Identifiers from Haller et al. [9] applied to the Duffing (1) and the Shaw-Pierre Example (2). (a) Duffing oscillator (1) with the non-
dimensional parameters c = 3.1, k = 1, κ = 1 and ν = 0.1. The identified DBS ridge closely aligns with the slow manifold connecting the saddle
at (0,0) with the stable fixed point at (1,0). (b) Shaw-Pierre example (2) with the non-dimensional parameters m1 = m2 = 1, c1 = c2 = c3 = 0.1,
k1 = k2 = k3 = 1, κ = 1 and ν = 0.1. The identifies perfect barriers align closely with the SSM tangent to the fast modal subspace. The SSM is
computed by with the SSMtool [12]
cumulate. Both identifiers (perfect barriers and DBS) can be computed based on purely deterministic quantities of the
dynamical system, whereby computationally expensive numerical methods such as Monte-Carlo approximations can be
omitted.
In this talk, we apply the methods pioneered by Haller et al. [9] to mechanical systems excited by Gaussian white noise
with small intensity. After rewriting the Fokker-Planck equation for the mechanical system in a suitable form to apply the
results of Haller et al. [9], we numerically investigate whether the distinguished transport extremizers align with known
deterministic quantities such as the slow and the fast stable manifold of stable fixed points. More specifically we investigate
the modified Shaw-Pierre example [10] in the form of
3
m1 0 c1 + c2 −c2 k1 + k2 −k2 κq1 √ 1
q̈ + q̇ + q+ = ν dw, (32.1)
0 m2 − c2 c2 + c3 − k2 k2 + k3 0 0
where the intensity ν is a small parameter (ν 1) and dw indicates Gaussian white noise. Further, we investigate the classical
Duffing oscillator [11].
√
q̈ + cq̇ − kq + κq 3 = ν dw. (32.2)
For the systems (32.1) and (32.2), we plot the identifiers defined by Haller et al. [9]in Fig. 32.1. In the case of the Duffing
oscillator (32.2) the slow manifold connecting the saddle type fixed point at the origin with the stable node at q = 1 is
successfully identified as a DBS ridge (cf. Fig. 32.1a). In the case of the Shaw-Pierre example (2) the fast stable submanifold
is successfully identified as perfect barrier for the stochastically excited nonlinear mechanical systems (cf. Fig. 32.1b).
References
1. Lutes, L.D., Sarkani, S.: Random Vibrations: Analysis of Structural and Mechanical Systems. Butterworth-Heinemann (2004)
2. Kerschen, G., Peeters, M., Golinval, J.C., Vakakis, A.F.: Nonlinear normal modes, part I: a useful framework for the structural dynamicist.
Mech. Syst. Signal Process. 23(1), 170–194 (2009)
32 The Relevance of Nonlinear Normal Modes for Randomly Excited Nonlinear Mechanical Systems 225
3. Renson, L., Kerschen, G., Cochelin, B.: Numerical computation of nonlinear normal modes in mechanical engineering. J. Sound Vib. 364,
177–206 (2016)
4. Haller, G., Ponsioen, S.: Nonlinear normal modes and spectral submanifolds: existence, uniqueness and use in model reduction. Nonlinear
Dynam. 86(3), 1493–1534 (2016)
5. Berglund, N., Gentz, B.: Geometric singular perturbation theory for stochastic differential equations. J. Different. Eq. 191(1), 1–54 (2003)
6. Arnold, L.: Random Dynamical Systems. Springer, Berlin, Heidelberg (1995)
7. Risken, H.: The Fokker-Planck Equation. Springer, Berlin, Heidelberg (1996)
8. Soize, C.: The Fokker-Planck equation for stochastic dynamical systems and its explicit steady state solutions, vol. 17. World Scientific (1994)
9. Haller, G., Karrasch, D., Kogelbauer, F.: Barriers to the transport of diffusive scalars in compressible flows. SIAM J. Appl. Dynam. Sys. 19(1),
86-123 (2020)
10. Shaw, S.W., Pierre, C.: Normal modes for non-linear vibratory systems. J. Sound Vib. 164(1), 85–124 (1993)
11. Guckenheimer, J., Holmes, P.: Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields Corr, 7th edn. Springer, New
York (2002)
12. Ponsioen, S., Pedergnana, T., Haller, G.: Automated computation of autonomous spectral submanifolds for nonlinear modal analysis. J. Sound
Vib. 420, 269–295 (2018)
Chapter 33
Analysis of an Actuated Frictional Interface for Improved
Dynamic Performance
M. Lasen, Y. Sun, C. W. Schwingshackl, and D. Dini
Abstract Friction in assembled structures is of great interest due to its ability to reduce the vibration amplitude of critical
components. The nonlinear behaviour of a structure depends on a variety of physical parameters. Among these parameters,
the contact pressure distribution and the contact area have shown to be critical for the behaviour of the joint and the responses
of assembled structures. In most application cases the impact of the interface geometry is not considered as a design
parameter, although some attempts have been reported to shape the interface geometry for a specific dynamic response.
Taking this idea of designing an interface geometry for a better dynamic performance a step further, the concept presented
here propose an actively controlled interface geometry and contact pressure distribution, to change the joint behaviour during
a vibration cycle. The concept consists of a device capable of manipulating the shape and pressure of a flexible membrane in
contact with a rigid punch, subjected to a normal load and a tangential excitation, via a row of piezoelectric actuators.
Keywords Frictional damping · Semi active control · Piezoelectric actuator · Pressure distribution · Contact interfaces
33.1 Introduction
Assembled structures contain a multitude of joints that can introduce frictional dynamics during vibration events [1].
Frictional joints in assembled structures have been extensively studied over the last decade as means of robust energy
dissipation to extend the life of critical components.
This energy dissipation depends on the frictional parameters of the contact area, which for a simple elastic-Coulomb
model are: normal load and its distribution, sliding distance, tangential stiffness, kt , and friction coefficient, μ [2]. Two
surfaces in contact, such as bolted flange or lap joints, will show a non-uniform pressure distribution leading to potential
microslip at different locations during a vibration cycle [3]. Before the onset of local slip (microslip), the linear response is
strongly dependant on the contact stiffness, kt [4]. With increasing microslip, the response departs from a linear condition
to a softened response [5]. Since microslip can heavily depend on the pressure distribution at the contact interface [6], some
recent research has focused on trying to optimise the interface for a better dynamic behaviour [7, 8], showing promising
changes to the dynamic performance in terms of better damping, or a more robust response. The permanent modification of
the interface geometry allows an improved global performance, but ones designed cannot respond to changes in the dynamic
behaviour. Very few cases have been presented where an active control of the interface load was used to change the frictional
interface behaviour. In [9], an On-Off control to lower the normal load when the tangential force changes direction, leading
to some extra slip, and consequently more damping. The normal load in a bolted joint is controlled continuously in [10],
controlling both the friction limit and the sliding distance in the hysteresis loop.
This paper proposes a system that utilises the current understanding of the interface geometry dependence of the hysteretic
cycle with the idea of an active variation of normal load distribution, to achieve a change of the energy dissipation of the
system while keeping the total normal load constant.
M. Lasen () · Y. Sun · C. W. Schwingshackl · D. Dini

Department of Mechanical Engineering, Imperial College London, London, UK

228 M. Lasen et al.
Fig. 33.1 Model. (a) Isometric and CCFF Boundary condition in the membrane. (b) XY midplane cut
33.2 A Novel Actuated Frictional Interface
The presented concept is based on the idea to modify the microslip behaviour of an interface by actively controlling its
pressure distribution. In the proof of concept design this will be achieved by the variation of the shape of a membrane
which forms one part of the interface via three piezoelectric actuators. Figure 33.1 shows the arrangement, with a rigid
punch sitting on top of the membrane, and the three actuators underneath it. Depending on which actuator is being activated,
different pressure distributions can be achieved for a constant overall normal load.
To evaluate the ability of the proposed system to influence the shape of a hysteresis loop, it is modelled in
ABAQUS/Standard. The punch is modelled as a solid with a flat interface of 6.6 × 6.5 mm2 and 10 mm radius rounded
edges. The membrane consists of a 200 μm thick steel plate with a Young’s modulus of E = 200 GPa, and a Poisson Ratio
of ν = 0.3.
Each piezo interface is modeled as a rigid sphere with a radius of 5 mm, while the piezo actuator is represented by a linear
spring with the characteristic piezoelectric stiffness (24 N/μm), and the applied drive voltage is represented by a spring base
displacement (6.5 μm, corresponding to 100 V, indicated with red arrows in Fig. 33.1). The Punch-Membrane interface is
modelled as a frictional interface via the penalty method in the tangential direction (μ = 0.67) and a limit elastic slip of 2%
and a hard contact in the normal direction, while the membrane-piezo interface is set to be frictionless. The mesh along the
entire interface is very fine (0.05 μm element length) to ensure accurate results, leading to a total mesh size of 234,580 linear
3D brick elements. The membrane in clamped-clamped-free-free as displayed in Fig. 33.1. The punch is constrained to move
in the vertical and longitudinal directions only and the linear springs and the 5 mm curved tip of the piezoelectric actuators
also move in the vertical direction only.
33.3 Analysis
To compute a full hysteresis cycle for different piezo activation patterns, three analysis steps are required.
In a first step, a Normal Load equal to 100 N is applied to the punch downward, leading to bending of the membrane and
compression of the actuators.
In a next step one or more piezoelectric actuators are activated. For the purpose of showing the principles on which the
technique is based, in this paper two configurations are used, one where the middle actuator is active (‘0-1-0’) and one where
the two lateral actuators are activated (‘1-0-1’). The resulting shapes and contact pressure distributions of the membrane can
be seen in Fig. 33.2. Unsurprisingly the 010 configuration leads to a Hertzian type contribution, while the 101 contribution
shows a slightly skewed double peak distribution due to the bending effect of the membrane.
In the final step a cyclic sliding displacement of the punch, characterized by 8 μm sliding distance amplitude, is prescribed
in the tangential direction. The resulting frictional force is measured via the rigid punch, allowing the assembly of the
hysteresis loop. Time and spatial discretisation are very fine to capture microslip accurately and to ensure convergence. The
time increment of the quasi static analysis is set to a maximum of 0.000625 of the total time needed to complete an entire
vibration cycle, resulting in 1600 increments used to resolve the relative motion of the nodes at the interface.
33 Analysis of an Actuated Frictional Interface for Improved Dynamic Performance 229
Fig. 33.2 Contact Pressure Distribution (above) and 10x vertical displacement in the midplane (below) before the tangential loading for
configurations: (a) 0-1-0 and (b) 1-0-1
Fig. 33.3 (a) Hysteresis Loop and (b) Time history and prescribed displacement in the punch for all configurations
The results in Fig. 33.3 show, that the two patterns indeed lead to slightly different hysteretic behaviour. Pattern 101
appears stiffer during pre-sliding while pattern 010 is softer. This can be attributed to the number of active contact elements,
since a larger number will lead to a stiffer interface. The effect on the hysteresis loops in this case is a 1.6% increase in
dissipated energy for the 101 pattern. Although difficult to see, a very small change to the microslip transition can also be
observed.
The analysis for this simple case shows that re-shaping the interface allows to shape the hysteresis loops; future work will
focus on the optimisation of the configuration and on the identification the best control configuration and strategy to obtain
more energy dissipation with the aim to improve frictional damping performances.
33.4 Conclusion
This investigation explores the effect of the shape manipulation of contact interfaces to induce changes in the contact pressure
distribution of an interface, with the purpose of controlling the total dissipated energy per cycle. It has been shown that the
change from a Hertzian contact to a two points contact induced by the use of piezoelectric actuation leads to a small increase
in dissipated energy, which can be mainly attributed to a change of overall contact stiffness at the interface, which in turn
depends on the increase in the contact area.
Acknowledgements ML acknowledges Imperial College London for its financial support through a President’s Scholarship.
230 M. Lasen et al.
References
1. Gaul, L., Nitsche, R.: The role of friction in mechanical joints. Applied Mechanics Review. 54(2), 93 (2001)
2. Schwingshackl, C., Petrov, E., Ewins, D.: Effects of contact interface parameters on vibration of turbine bladed disks with underplatform
dampers. J. Eng. Gas Turbines Power. 134, (2012)
3. Menq, C., Bielak, J., Griffin, J.: The influence of microslip on vibratory response, part I: a new microslip model. J. Sound Vib. 107(2), 279–229
(1986)
4. Kartal, M., Mulvihill, D., Nowell, D., Hills, D.: Measurements of pressure and area dependent tangential contact stiffness between rough
surfaces using digital image correlation. Tribol. Int. 44, 1188–1198 (2011)
5. Yang, B., et al.: Stick–slip–separation analysis and non-linear stiffness and damping characterization of friction contacts having variable normal
load. J. Sound Vib. 210(4), 461–481 (1998)
6. Pesaresi, L.: Modelling the nonlinear behaviour of an underplatform damper test rig for turbine applications. Mech. Sys. Signal Proces.,
Elsevier. 85, 662–679 (2017)
7. Dosogne, T., et al.: Experimental assessment of the influence of interface geometries on structural dynamic response. Dynam. Coupled Struct.
4, (2017)
8. Gastaldi, C.: The relevance of damper pre-optimization and its effectiveness on the forced response of blades. J. Eng. Gas Turbines Power.
140, (2018)
9. Dupont, P., et al.: Semi-active control of friction dampers. J. Sound Vib. 202(2), 203–218 (1997)
10. Gaul, L., et al.: Active damping of space structures by contact pressure control in joints. J. Struct. Mech. 26(1), 81–100 (1998)
Chapter 34
Vibration Reduction of a Structure by Using Nonlinear Tuned
Vibration Absorbers
Muhammed Emin Dogan and Ender Cigeroglu
Abstract Tuned Vibration Absorbers (TVA) are commonly used in reducing undesirable vibrations of mechanical
structures. However, TVAs work in a very limited frequency range and if the excitation frequency is outside of this range, they
become ineffective. In order to solve this problem, researchers started to consider nonlinear TVAs for vibration attenuation.
In this study, dynamic behavior of a linear Euler-Bernoulli beam coupled with a nonlinear TVA is investigated. The system
is subjected to sinusoidal base excitation. Parameters of the nonlinear TVA is optimized to minimize vibration amplitudes
of the primary system. Assumed modes method is used to model the Euler-Bernoulli beam. Nonlinear differential equations
of motion are converted to a set of nonlinear algebraic equations by using Harmonic Balance Method (HBM). The resulting
set of nonlinear algebraic equations is solved by Newton’s Method with Arc-Length continuation. Nonlinearities used in the
TVA are cubic stiffness, cubic damping and dry friction damping. Hill’s method is used to evaluate stability of the solutions
obtained. Results of the system with optimum nonlinear TVAs are compared with that of optimum linear TVA. Although,
NES show to exhibit good vibration reduction performance – which is in parallel with the results given in literature, due to
instability of the frequency domain solutions, it is observed that, actually, it is not as effective as other nonlinear TVAs.
Keywords Nonlinear tuned vibration absorber · Nonlinear vibrations · Nonlinear energy sink · Harmonic balance
method
34.1 Introduction
TVA is a device to attenuate excessive vibration at resonances. It was invented by Frahm in 1909 [1]. Den Hartog [2]
carried out the first theoretical study in which a TVA with a viscous damper attached on an undamped SDOF system and
subjected to harmonic excitation is considered. It had been noticed that frequency response curves of the main mass pass
through two invariant points for different damping values. Optimization was performed by changing the ratio of the natural
frequency of the TVA to that of the main system until response amplitudes of the invariant points were equal. Respectively,
damping of TVA was altered such that slope of the frequency response curve at the invariant points were zero. TVAs applied
on continuous system are studied in the literature extensively [3–6]. Although many researchers still interested in linear
vibration absorbers, there are a lot of studies on nonlinear vibration absorbers to further improve the effectives of TVAs.
Inaudi in 1995 [7] introduced a TVA with friction dampers (FTVA). Systems with strong nonlinear attachment bring a
new concept, which is Nonlinear Energy Sink (NES) [8, 9]. This concept also covers grounded vibration absorbers i.e. the
nonlinear element is located between vibration absorber and the ground [10]. Steady state dynamics of linear Euler-Bernouli
beam with NES under harmonic forcing is studied in [11]. The nonlinear element in the NES is the cubic stiffness. Authors
stated that NES performs better than linear TVA if the forcing is the same or less than designed force amplitude. Another
study [12] compares the performance of linear TVA and NES. In this study authors stated that properly tuned linear TVA
outperforms NES.
M. E. Dogan
ASELSAN, Ankara, Turkey
E. Cigeroglu ()

232 M. E. Dogan and E. Cigeroglu
In this study Fixed-fixed linear Euler Bernouli beam equipped with nonlinear TVAs are investigated. By the use of multiple
TVAs and nonlinear elements, it is aimed to suppress the vibrations of the beam in a broader frequency range.
34.2 Mathematical Modelling
Fixed-fixed linear Euler Bernouli beam equipped with a nonlinear TVA is shown in Fig. 34.1. Where U(t) is the base
excitation, W(y, t) is the relative displacement of the beam with respect to supports, L is the length of the beam, EI is the
modulus of rigidity, A is cross sectional area, ρ is density. NL is the nonlinear element, kTVA and cTVA are stiffness and
damping elements of the TVA. xTVA is the relative displacement of the TVA with respect to supports.
Displacement of the beam can be assumed as
n
W (y, t) = φi (y)xi (t).# (34.1)
i
Where, φ i (y) are the mode shapes of the linear fixed-fixed Euler-Bernoulli beam and, xi (t) are the unknown modal coefficients
which are functions of time. Potential and kinetic energies can be defined as follows
L 2 2
V = 12 EI 0 ∂ W∂y(y,t)
2 dy + 12 kT V A (xrel )2
L 2 2 , # (34.2)
T = 12 ρA 0 ∂W∂t(y,t) + ∂U∂t(t) dy + 12 mT V A ∂x∂t TVA
+ ∂U∂t(t)
Dissipative function is as
2
∂x rel
D = 12 cT V A ∂t .# (34.3)
Where xrel (t) = W(La , t) − xTVA (t) i.e. relative displacement between TVA and point of the beam where TVA is connected.
Nonlinear force is considered as an external forcing hence it is included in the work done by nonconservative forces. Rayleigh
damping [13] is used to model the inherent damping of the beam. Assumed modes method is used and only the first 5 modes
are taken into consideration. Lagrange’s equation is given as follows

∂t ∂ ẋi − ∂xi + ∂xi + ∂ ẋi = Qi .#
∂ ∂T ∂T ∂V ∂D
(34.4)
Substituting Eqs. (34.1), (34.2) and (34.3) into Eq. (34.4) equation of motion is obtained as follows
Mẍ + Cẋ + Kx + fnon (x) = fext (t) , # (34.5)
where
fext = FM Ü (t)# (34.6)
FM is property of the system related to mass.
Uniform Beam:
t)
Fig. 34.1 Linear System Equipped with Nonlinear TVA

34 Vibration Reduction of a Structure by Using Nonlinear Tuned Vibration Absorbers 233
Harmonic Balance Method (HBM) is used to determine the steady state response of system. In HBM, responses and
nonlinear forces are represented in terms of Fourier series and substituted into the nonlinear differential equations of motion
which results in a set of nonlinear algebraic equations. Considering only the first harmonic, displacement response of the
system can be represented by a single term Fourier series as
x (θ ) = xs sin (θ ) + xc cos (θ ) .# (34.7)
Where θ = ωt. Similar to response, nonlinear internal forcing vector can be represented by Fourier series. Sine and cosine
coefficients can be obtained from the following relations
s =

1 2π
fn,j π 0 fnon,j (θ ) sin (θ ) dθ # (34.8)
c = 1 2π
fn,j π 0 fnon,j (θ ) cos (θ ) dθ
Substituting Eqs. (34.7) and (34.8) into Eq. (34.5) the following nonlinear set of algebraic equations in residual form is
obtained
s s s
−ω2 M + K −H + ωC x f f
r (x, ω) = + nc − exc = 0# (34.9)
H + ωC −ω2 M + K xc fn fcexc
Newton’s Method with Arc-Length Continuation is used to solve Eq. (34.9).

A single step of Newton’s Method with arc-length continuation is given as [14].
∂r ∂r
! ( *
i+1 ∂x ∂ω r (xi , ωi )
qk = qk − i # (34.10)
∂h ∂h
∂x ∂ω x &ω h (xi , ωi )
i i
T T
where q = x ω and h (x, ω) = qk − qk−1 qk − qk−1 − s 2 = 0, s is the arc-length parameter.
Application of HBM results in determination of the periodic solutions of the system; however, it does not give any
information about the stability of them. It should be noted that not all the solutions obtained are stable. In this study, stability
of the solutions obtained is determined by Hill’s Method details of which can be found in [15].
34.3 Results
In this study, TVA with cubic stiffness (NES), cubic damping or dry friction damping is considered. TVAs parameters are
optimized by using genetic algorithm (GA) of MATLAB. Optimum values are used as an initial guess in a gradient based
optimization for further improvement of the results obtained. Two different cost functions are considered in the study. The
first is the maximum response amplitude, since the aim of the vibration suppression is to minimize the maximum vibration
amplitude. The second cost function is the area under the normalized displacement curve. It should be noted that considering
only the resonance frequency may not give the desired suppression characteristic for multiple degree-of-freedom systems
(Fig. 34.2).
When the weight of area cost function is zero, i.e. 100% Max, optimization resulted in three peaks with equal amplitudes
(green line). When the weight of the maximum displacement cost function is zero, i.e. 100% Area, amplitude of the first
resonance peak is larger than the others. This is due to the fact that the area under the frequency response function is
minimized without considering the amplitudes of the resonance peaks. When considering the general physical behavior,
combination of the first and the second cost function is more effective. It is selected as the ultimate cost function in further
analysis (Fig. 34.3).
Single TVA with friction damper has higher amplitude than linear response between normalized frequencies of 0.7 to
0.95. After that value, it has better suppression regime. Configuration of double TVAs with friction dampers has higher
amplitude in a narrow region of normalized frequencies between 0.8 to 0.95. For the rest of the frequency range, it has
slightly better suppression regime. Single TVA with cubic damping, single linear TVA and multi linear TVA have similar
suppression behavior. TVA with cubic stiffness element (so called NES) seems to have a better suppression behavior than
the others; however, it is ineffective due to its stability behavior. There is a frequency internal with no stable solution. Time
domain analysis is performed for the TVA with cubic stiffness. Steady-state displacement and velocity values at the previous
234 M. E. Dogan and E. Cigeroglu
Fig. 34.2 Comparisons of Cost Functions
Fig. 34.3 Vibration Suppression in First Mode

34 Vibration Reduction of a Structure by Using Nonlinear Tuned Vibration Absorbers 235
frequency point are used as the initial condition of the next frequency point and steady-state responses are recorded. In case
steady-state response is not obtained, amplitude in the final one second is recorded. It can be seen according to the time
domain results, performance of TVA with cubic stiffness element (NES) is below all other TVA types, since time domain
solutions are actually much larger than the unstable solution obtained in the solution process.
34.4 Conclusion
In this study, vibration reduction of structures by using linear and nonlinear Tuned Vibration Absorbers is studied. Studies are
carried out on a fixed-fixed Euler Bernoulli Beam equipped with TVAs. The system is subjected to sinusoidal base excitation.
Linear and nonlinear TVA configurations are optimized to reduce vibration response of the primary structure. The nonlinear
elements considered in this study are cubic damping, dry friction damping and cubic stiffness. Besides a single TVA, use of
multiple TVAs is as well considered. From the results obtained TVA with cubic stiffness seems to be most very effective one
in terms of vibration suppression; however; not all the solutions obtained are stable. Further investigation is carried out and
it is observed that time solutions in that range have higher amplitudes and they are not periodic. Therefore, TVA with cubic
stiffness is not as effective as expected. Two TVAs with dry friction damping has better performance for frequencies higher
than the normalized frequency of 1 While maintaining similar performance in other regions compared to TVA with cubic
damping and linear TVA.
References
1. Frahm, H.: Device for damping vibrations of bodies. US Patent No. 989958. (1911)
2. Den Hartog, J.P.: Mechanical Vibration, N.Y 3rd ed. McGraw-Hill, New York (1947)
3. Young, D.: Theory of dynamic vibration absorbers for beams, Proceedings of the First U.S. National Congress of Applied Mechanics. 83(12),
91–96 (1952)
4. Jacquot, R.G.: Optimal dynamic vibration absorbers for general beam systems. J. Sound Vib. 60(4), 535–542 (1978)
5. Warburton, G.B., Ayorinde, E.O.: Optimum absorber parameters for simple systems. Earthq. Eng. Struct. Dyn. 8(3), 197–217 (1980)
6. Özgüven, H.N., Çandır, B.: Suppressing the first and second resonances of beams by dynamic vibration absorbers. J. Sound Vib. 111(3),
377–390 (1985)
7. Inaudi, J.A., Kelly, J.M.: Mass damper using friction-dissipating devices. J. Eng. Mech. 121(1), 142–149 (1995)
8. Gendelman, O.V., Vakakis, A.F., Manevitch, L.I., McCloskey, R.: Energy pumping in nonlinear mechanical oscillators I: dynamics of the
underlying Hamiltonian system. J. Appl. Mech. 68(1), 34–41 (2001)
9. Vakakis, A.F., Gendelman, O.V.: Energy pumping in nonlinear mechanical oscillators II: resonance capture. J. Appl. Mech. 68(1), 42–48
(2001)
10. Jiang, X., McFarland, D.M., Bergman, L.A., Vakakis, A.F.: Steady state passive nonlinear energy pumping in coupled oscillators: theoretical
and experimental results. J.Nonlinear Dynam. 33(1), 87–102 (2003)
11. Parseh, M., Dardel, M., Ghasemi, M.H.: Investigating the robustness of nonlinear energy sink in steady state dynamics of linear beams with
different boundary conditions. Commun. Nonlinear Sci. Numer. Simul. 29(1–3), 50–71 (2015)
12. Gourc, E., Elce, L.D., Kerschen, G., Michon, G., Aridon, G., Hot, A.: Performance comparison between a nonlinear energy sink and a linear
tuned vibration absorber for broadband control. Nonlinear Dynam. 1, 83–95 (2016)
13. Connor, J., Laflamme, S.: Structural Motion Engineering (2014). https://doi.org/10.1007/978-3-319-06281-5
14. Cigeroglu, E., Samandari, H.: Nonlinear free vibration of double walled carbon nanotubes by using describing function method with multiple
trial functions. Phys. E. 46, 160–173 (2012)
15. Detroux, T., Renson, L., Masset, L., Kerschen, G.: The harmonic balance method for bifurcation analysis of large-scale nonlinear mechanical
systems. Comput. Methods Appl. Mech. Eng. 296, 18–38 (2015)
Chapter 35
Simulation-Free Reduction Basis Interpolation to Reduce
Parametrized Dynamic Models of Geometrically Non-linear
Structures
Christian H. Meyer and Daniel J. Rixen
Abstract Virtual design studies for the dynamics of structures that undergo large deformations, such as wind turbine blades
or Micro-Electro-Mechanical Systems (MEMS), can be a tedious task. Such studies are usually done with finite element
simulations. The equations of motion that result from the finite element discretization typically are high-dimensional and
nonlinear. This leads to high computation costs because the high-dimensional nonlinear stiffness term and its Jacobian
must be evaluated at each Newton-Raphson iteration during time integration. Model reduction can overcome this burden
by reducing the high-dimensional model to a smaller problem. This is done in two steps: First, a Galerkin projection on a
reduction basis, and, second, hyperreduction of the geometric nonlinear restoring force term.
The first step, namely finding a proper reduction basis, can be performed by either simulation-based or simulation-free
methods. While simulation-based methods, such as the Proper Orthogonal Decomposition (POD), rely on costly preliminary
simulations of full high-dimensional models, simulation-free methods are much cheaper in computation. For this reason,
simulation-free methods are more desirable for design studies where the amount of the so called ‘offline costs’ for reduction
of the high-dimensional model are of high interest. However, simulation-free reduction bases are dependent on the system’s
properties, and thus depend on design parameters that typically change for each design iteration. This dependence must be
taken into account if the parameter space of interest is large.
This contribution shows how design iterations can be performed without the need for expensive simulations of the high-
dimensional model. We propose to sample the parameter space, compute simulation-free reduction bases at the sample
points and interpolate the bases at new parameter points. As hyperreduction technique, the Energy Conserving Sampling
and Weighting method and the Polynomial expansion are used for hyperreduction of the nonlinear term. In this step, we
also avoid simulations of the high-dimensional nonlinear model. The coefficients of the hyperreduction are updated in each
design iteration for the new reduction bases.
A simple case study of a shape parameterized beam shows the performance of the proposed method. The case study also
accounts for a last challenge that occurs in models that are parametric in shape: The topology of the finite element mesh must
be maintained during the design iterations. We face this challenge by using mesh morphing techniques.
Keywords Nonlinear model reduction · Parametric model reduction · Basis interpolation · Parametric hyperreduction ·
Optimization
35.1 Introduction
A typical task for engineers in development and design is to optimize structures with respect to certain requirements.
This task can be very challenging when it is hard to predict if a certain design will fulfill a requirement. One example is
the requirement for a certain (dynamic) behavior of structures that undergo large deformation. These structures appear in
applications like wind turbine blades [1, 2], Micro-Electro-Mechanical-Systems (MEMS) [3] or compliance mechanisms.
The dynamic behavior of these structures are often analyzed by finite element simulations. They allow design studies at
low-cost compared to experiments. However, these simulations can require large simulation times.
C. H. Meyer () · D. J. Rixen

Chair of Applied Mechanics, Faculty of Mechanical Engineering, Technical University of Munich, Garching, Germany

238 C. H. Meyer and D. J. Rixen
The reason for this issue can be demonstrated with the equations of motion
M ü + f (u, u̇) = F (t) (35.1)
where u ∈ RN are the nodal displacements of the finite element mesh, M ∈ RN ×N is the mass matrix, f (u, u̇) ∈ RN is the
internal restoring force and F ∈ RN are the external forces. First, the number of degrees of freedom N can be very large
for industrial problems. Second, the internal force term f that is nonlinear for systems that undergo large deformations, is
expensive to evaluate. And lastly, design studies require to solve the equations several times due to parameter dependence of
the system quantities. This dependency can be expressed by
M(p) ü + f (p, u, u̇) = F (t) (35.2)
where p describes the vector of design parameters.

It is highly desired to reduce the simulation time for design studies. One approach to achieve this goal is to reduce the
equations of motion by a Galerkin projection
V T M(p)V q̈ + V T f (p, V q, V q̇) = V T F (t) (35.3)
where V ∈ RN ×n is called reduction matrix. The Galerkin projection (35.3) forces the solution to live in a subspace V that
is spanned by the column vectors of V . This reduces the number of unknowns from N to n. The challenge to perform a
successful reduction with this technique is to provide a reduction matrix that is able to span a good subspace for an accurate
approximation while still having a low dimension n.
The Galerkin projection on its own is not sufficient to reduce simulation time. The reason is that the projected nonlinear
internal force term V T f (p, V q, V q̇) still must be evaluated by the assembly of all element forces. Methods that are able to
speed up this evaluation are called hyperreduction. Hyperreduction methods approximate the projected nonlinear force term
by a term
f hr (p, q, q̇) ≈ V T f (p, V q, V q̇) (35.4)
that is computationally much cheaper to evaluate. The reduced order model (ROM) can then be described by
M̄(p) q̈ + f hr (p, q, q̇) = V T F (t) (35.5)
where M̄(p) = V T M(p)V is the reduced mass matrix and f hr describes a hyperreduced projected nonlinear force term
that is fast to evaluate.
To perform faster design studies, the following challenges must be met:
1. For each design value p̂ ∈ Pdesign a reduction basis V (p̂) must be computed that leads to a good accuracy of the reduced
order model.
2. The overall simulation time for performing design studies with reduced order models (35.5) must be significantly smaller
than with full order models (35.2).
This contribution proposes a workflow to perform faster design studies with model reduction by avoiding time consuming
solutions to high dimensional models. The remainder of this paper is outlined as follows: Sect. 35.2 describes a simulation-
free reduction basis for geometric nonlinear structures known from literature: Static Modal Derivatives. Section 35.3 shows
how reduction bases can be interpolated for studies with one parameter. This can be used to gain a reduction basis at
a new design point by interpolating already computed reduction bases at other design points. Section 35.4 recaps two
hyperreduction techniques known from literature: The Polynomial Expansion and the Energy Conserving Sampling and
Weighting method. In contrast to many other contributions, it is also shown how these methods can be used without the need
for a time integration of the high dimensional model. Section 35.5 illustrates the performance of the proposed workflow
at a case study. The case study is a shape parameterized cantilever beam loaded at the tip. The results are summarized in
Sect. 35.6 which also contains a conclusion.
35 Simulation-Free Reduction Basis Interpolation to Reduce Parametrized Dynamic Models of Geometrically Non-linear Structures 239
35.2 Static Modal Derivatives for Simulation-Free Reduction Bases
The success of the Galerkin projection (35.3) is highly dependent on the proper choice of the reduction matrix V . One can
find several methods to compute a reduction basis in literature. These can be classified in two groups: Simulation-based
and simulation-free methods. Simulation-based methods, like the Proper Orthogonal Decomposition of training sets [4], rely
on data of time integration of full order models. In contrast, simulation-free methods avoid these computations. We desire
low computation costs for gaining a reduction basis. Therefore, we will use a simulation-free method in this contribution:
A combination of modes of the linearized system and their static modal derivatives [5]. To begin, the first m modes of the
linearized system are computed by solving the eigenproblem

∂f (u, u̇)
(K − ωi2 M)φ i =0 with K := (35.6)
∂u u=u̇=0
where K is the linearized stiffness matrix. The static modal derivatives φ i,j are defined by
φ i,j = −K −1 · (∇φ j K) · φ i (35.7)
where ∇φ j is the directional derivative with respect to u in the direction of mode φ j . The reduction basis V is then built by
concatenation

V raw = φ 1 , φ 2 , . . . , φ m , φ 1,1 , φ 1,2 , . . . , φ m,m . (35.8)
No selection criteria of the static modal derivatives is used in this contribution. All static derivatives of the modes that are
taken into the reduction matrix are also inserted. Selection criteria to reduce the size of the reduction matrix can be found,
e.g., in [6]. Afterwards, the reduction matrix is deflated by computing a singular value decomposition

V raw = σi ui wTi (35.9)
i=1
and taking the first n left singular values as reduction matrix
V = [u1 , u2 , . . . , un ] . (35.10)
35.3 Geodesic Interpolation of Reduction Bases
The accuracy of the reduced solution to equation (35.3) only depends on the subspace V which is provided by the column
vectors of the reduction matrix V ∈ RN ×n . If the solution u(t) to the high dimensional model can be represented within
the subspace V with good accuracy, the solution to the reduced system will also be accurate. The subspaces which are
provided by the reduction bases computed with equations (35.6, 35.7, 35.8, 35.9, and 35.10) can differ for changes in
the parameters p. Thus, one would need to compute a new reduction basis in each design iteration. This can be a time-
consuming task for large models. One idea to reduce these computation costs is to compute only a few reduction bases at
some samples p s ∈ Psample ⊂ Pdesign and gain the other bases by interpolating between these samples.
The most natural way to perform these interpolations is to interpolate linearly along the shortest path between the samples.
The definition of the shortest path requires the definition of a manifold on which the reduction bases are interpolated. As
we have stated above, the accuracy of a reduced order model only depends on the subspace that is provided by a reduction
basis V ∈ RN ×n . Therefore, it seems natural to use a manifold that describes all n dimensional subspaces that are embedded
in RN . This manifold is known as Grassmann manifold G(n, RN ). An extensive overview about calculations on this manifold,
including the computation of geodesic paths, can be found in [7]. Amsallem has used this approach for a similar application,
namely the interpolation of reduction bases in [8]. The main results from [8] to compute an interpolation on this manifold
for a single parameter are recapped in the following paragraph.
We assume that possible design changes in our model can be described by just one parameter p = [p]. It is also assumed
that reduction bases V (p s ) have already been computed at the sample points
{ps1 , ps2 , . . . , p sNs } = {[ps1 ], [ps2 ], . . . , [psNs ]} =: Psample . (35.11)
The column vectors of two reduction matrices V j := V (psj ) and V k := V (psk ) span the euclidean subspaces Vj and
Vk , respectively. The computation of a reduction matrix that spans a subspace which is linearly interpolated on a geodesic
between Vj and Vk on the Grassmann manifold can be done in two steps: First, a singular value decomposition of
−1
I N − V j V Tk V k V Tj V k = U W T (35.12)
is computed. Then the interpolated reduction basis Ṽ at p is determined by

p − psj −1 p − psj −1
Ṽ (p) = V j W cos tan () + U sin tan () . (35.13)
psk − psj psk − psj
35.4 Simulation-Lean Hyperreduction
The Galerkin projection (35.3) is not sufficient to reduce computation time. As already stated in Sect. 35.1, a hyperre-
duction (35.4) is needed to accelerate the evaluation of the projected nonlinear internal force term. We use two different
hyperreduction techniques: The Polynomial Expansion [9] and the Energy Conserving Sampling and Weighting method [10].
They are summarized in the following paragraphs.
35.4.1 The Polynomial Expansion
Assumes that the nonlinear force term can be written as a polynomial of third order
(1) (2) (3)

(f hr,poly )i := K̄ ij q j + K̄ ij k q j q k + K̄ ij kl q j q k q l ≈ (V T f (V q))i (35.14)
where Einstein notation applies. Equation (35.14) is even exact for models with linear constitutive relations, such as the
St. Venant Kirchhoff material. The coefficients of the tensors
(1) (2) (3)
K̄ ∈ Rn×n , K̄ ∈ Rn×n×n , K̄ ∈ Rn×n×n×n (35.15)
can be identified via various techniques [11]. We use finite differences of the tangential stiffness matrix to identify K (2)
and K (3) .
35.4.2 The Energy Conserving Sampling and Weighting
Method [10] approximates the nonlinear force vector by a reduced assembly. Instead of assembling the forces of all
elements e ∈ E, only a subset Ẽ ⊂ E is evaluated. The hyperreduced global force vector

f hr,ECSW := ξe V T LTe f e (Le V q) ≈ V T LTe f e (Le V q) = (V T f (V q)) (35.16)
e∈Ẽ⊂E e∈E
is assembled by weighting their contribution with positive weights ξe . Computation time is significantly reduced due to the
smaller number of elements where f e needs to be evaluated.
The element set Ẽ and the weights ξe are chosen such that the virtual work of some training displacements uτ = V q τ is
retained up to a certain relative tolerance ε. Thus,
|δWhr (q τ , 0) − δW (q τ , 0)| ≤ ε · δW (q τ , 0)
⇔ δq T f hr (q τ , 0) − δq T V T f (V q τ , 0) ≤ ε · δq T V T f (V q τ , 0)
is satisfied for each variation δq and training displacement q τ ∈ T , where T = {q τ,1 , q τ,2 , . . .} is the set of training
responses. Details about the algorithm to compute these parameters can be found in [10].
Representative training displacements T are crucial to gain an accurate reduced order model (35.5) with the ECSW. One
can use training displacements resulting from time integrations of full order models. But this is disadvantageous for design
studies because it is desired to avoid time integration of full order models. A cheaper approach, called nonlinear stochastic
Krylov training sets, has been proposed by Rutzmoser in [12]. It only relies on nonlinear static solutions of the projected
equations (35.3). This is significantly cheaper than time integration of (35.3). First, a q-dimensional force space

−1 −1 −1 −1
F kry, raw = V T b, (K̄ M̄)V T b, (K̄ M̄)2 V T b, . . . , (K̄ M̄)(q−1) V T b = Kq (K̄ M̄), V T b (35.17)
is generated that is assumed to provide a good subspace for the internal force vectors f . The function Kq defines a q-
dimensional Krylov subspace. The vectors are orthonormalized such that
F Tkry K −1 F kry = I . (35.18)
Afterwards, e stochastic linear combinations
f rand,i = F kry · ni (35.19)
of force vectors from this space are generated. The e amplifiers ni ∈ Rq are generated randomly as described in [12].
Training displacements q τ,i,l are then generated by solving k · e nonlinear static problems
l
V T · f (V q τ,i,l ) = f with l ∈ {1, 2, . . . , k} and i ∈ {1, 2, . . . , e} . (35.20)
k rand,i
35.5 Case Study
The proposed methods are tested at a simple case study. The case study is a parameterized cantilever beam which is
illustrated in Fig. 35.1. The beam has a taper at position xtaper . The goal of the design study is to quantify the effect of
x
the taper’s position p = taper
l on the tip displacement utip due to a harmonic excitation force at the tip. Parameters of interest
are Pdesign = {0.30, 0.31, 0.32, . . . , 0.70}.
The beam is meshed with 2,507 nodes and 1,148 triangular elements of second order. The taper position is changed by
moving the node coordinates of the nodes. This is illustrated in Fig. 35.2. This procedure is necessary to maintain the physical
meaning of each degree of freedom. Otherwise an interpolation between reduction bases computed with equations (35.12
and 35.13) would fail.
xtaper F
thickness b, Young’s modulus E, Poisson’s ratio ν
d
l
Fig. 35.1 Tapered cantilever beam. The design parameter is the position of the taper p = [xtaper / l]. Fixed dimensionless parameters: F
=
) El 2
2 · 10−9 · (sin(0.0503 τ ) + sin(0.3139 τ ) + sin(0.1300 τ )), ν = 0.49, hl = 0.05, dl = 0.03, bl = 0.1, τ = Eρ tl
Fig. 35.2 Morphing of Meshes. Middle: The reference mesh with taper position at p = 0.5. Top and bottom: Morphed meshes. Only the nodal
coordinates are changed. The topology of the mesh is maintained
Fig. 35.3 Tip displacement over time for p = 0.4 (full solution)
The best sampling points for the design study are unknown in advance. We test the performance of the proposed method
at two taper positions (0.40 and 0.60) with different interpolation points. Figures 35.3 and 35.7 show the solutions for the
tip displacement of the full order models. Figures 35.4 and 35.8 show the absolute errors of the tip displacement computed
with reduced order models that are only reduced with a Galerkin projection (no hyperreduction). It is easily seen that a
direct computation of a reduction basis at the desired parameter point gives the best accuracy. One can also conclude that
the accuracy becomes better when interpolation points are used that are closer to the desired parameter point. Despite the
solution of reduced order models with interpolated bases are worse than with bases that are directly computed, the solution
is still good and usable. In contrast, if a reduction basis, that is close to the parameter point, is used directly without an
interpolation, the error becomes too big to have an accurate approximation. The errors of hyperreduced models that are
reduced with the polynomial expansion are depicted in Figs. 35.5 and 35.9. As expected the errors are the same as in the
non-hyperreduced models because we have used a linear constitutive law for the material. The plots in Figs. 35.6 and 35.10
Fig. 35.4 Tip displacement error of reduced models for p = 0.4 (Galerkin projection only, no hyperreduction)
Fig. 35.5 Tip displacement error of reduced models for p = 0.4 (with Polynomial Expansion hyperreduction)
Fig. 35.6 Tip displacement error of reduced models for p = 0.4 (with ECSW hyperreduction)
Fig. 35.7 Tip displacement over time for p = 0.6 (full solution)
show the error of the tip displacement for models that are hyperreduced with the ECSW-method. The training sets for the
ECSW are generated via the NSKTS method. The dimension of the Krylov space for the NSKTS is chosen to q = 4. The
number of stochastic linear combinations is e = 8 and the number of load steps is k = 20. One can see that the errors due to
the hyperreduction with the ECSW increase only slightly (Figs. 35.7, 35.8, 35.9, and 35.10).
Table 35.1 lists CPU times for each step of the case study. All listed values are averages over 10 experiments. The second
column shows the CPU time to compute the reduction basis for the direct and the interpolation method. One can see that the
computation time for the interpolation is much smaller than for the direct method to gain a reduction basis. However, the
offline costs for the hyperreduction is much higher than the offline costs for the basis generation. Thus, the basis interpolation
Fig. 35.8 Tip displacement error of reduced models for p = 0.6 (Galerkin projection only, no hyperreduction)
Fig. 35.9 Tip displacement error of reduced models for p = 0.6 (with Polynomial Expansion hyperreduction)
Fig. 35.10 Tip displacement error of reduced models for p = 0.6 (with ECSW hyperreduction)
Table 35.1 CPU-times in seconds (determined via Python command time.process_time()) for different simulation steps. Offline costs
are needed to reduce a high dimensional model. Online costs are needed to perform a time integration of the reduced order model. All simulations
are done with the code AMfe developed by the Chair of Applied Mechanics, Technical University of Munich
Full t.i.a Offline b.g.b Online Galerkinc Offline Poly3d Online Poly3e Offline ECSWf Online ECSWg
0.4 direct 156.825015 1.992477 112.860184 623.364868 41.923435 567.513856 55.384431
0.6 direct 186.877985 1.863974 137.904871 610.475783 46.795020 568.400610 64.554823
0.4 interp. 156.825015 0.935386 113.222813 601.583104 41.433322 579.790726 56.890483
0.6 interp. 186.877985 0.878111 138.921961 604.121055 50.591013 577.181495 67.968682
a Fulltime integration
b Costs for basis generation (Modes, Static Derivatives and Orthogonalization)
c Costs to run a time integration with a Galerkin projected system
d Costs for determination of coefficients for Polynomial Expansion
e Costs to run a time integration with a system that is hyperreduced with Polynomial Expansion)
f Costs to determine the NSKTS, the element set Ẽ and the weights ξ
e
g Costs to run a time integration with a ECSW-hyperreduced system
only gives a small benefit in this case study. On the other hand, the case study is done with a quite small academic model.
It is expected that the benefit will increase for models with higher dimension N . Nevertheless, it is highly desired to also
reduce the offline costs for the hyperreduction of the models when the parameters are changed.
35.6 Conclusion
Numerical studies to design the dynamics of structures that undergo large displacements can be a tedious task. The simulation
times can be high due to the high dimension of the finite element model and the cost-intensive evaluation of the nonlinear
restoring force term. Model reduction can overcome this burden by reducing the number of degrees of freedom by a linear
combination of basis vectors spanned by a reduction matrix V . This contribution shows how one can compute reduction
bases for design studies efficiently in two steps: First, simulation-free reduction bases are computed at some sampling points.
Second, these bases are interpolated for parameter points that were not sampled. This approach is much cheaper than using
simulation-based methods that rely on cost-intensive time integrations of full order models. It has been shown that the
interpolation is also cheaper than a direct simulation-free computation. The accuracy of the reduced order models that are
reduced with interpolated bases is better than those that are reduced with bases from sampling points. However, though the
basis interpolation is cheap, the total time benefit of the reduction that also includes a hyperreduction is quite low. This is
because the costs to determine the parameters for the hyperreduction are much higher than the costs to compute simulation-
free reduction bases. It is desired to develop methods that reduce the offline costs of the hyperreduction, e.g. also via an
interpolation scheme.
Acknowledgments The presented work is part of the research within the context of the Priority Program 1897 “Calm, Smooth and Smart”,
subproject “Model Order Reduction of Parametric Nonlinear Mechanical Systems for Influencing Vibrations”. The authors are grateful to the DFG
(German Research Foundation) for the financial support.
References
1. Collier, W., Milian Sanz, J.: Comparison of linear and non-linear blade model predictions in Bladed to measurement data from GE 6MW wind
turbine. J. Phys. Conf. Ser. 753(8), 082004 (2016)
2. Rezaei, M., Behzad, M., Haddadpour, H., Moradi, H.: Aeroelastic analysis of a rotating wind turbine blade using a geometrically exact
formulation. Nonlinear Dyn. 89(4), 2367–2392 (2017)
3. Zhao, J., Chen, H.: A study on the coupled dynamic characteristics for a torsional micromirror. Microsyst. Technol. 11(12), 1301–1309 (2005)
4. Willcox, K., Peraire, J.: Balanced model reduction via the proper orthogonal decomposition. AIAA J. 40(11), 2323–2330 (2002)
5. Slaats, P., de Jongh, J., Sauren, A.: Model reduction tools for nonlinear structural dynamics. Comput. Struct. 54(6), 1155–1171 (1995)
6. Tiso, P.: Optimal second order reduction basis selection for nonlinear transient analysis. Conf. Proc. Soc. Exp. Mech. Ser. 3, 27–39 (2011)
7. Edelman, A., Arias, T., Smith, S.: The geometry of algorithms with orthogonality constraints. SIAM J. Matrix Anal. Appl. 20(2), 303–353
(1998)
8. Amsallem, D., Farhat, C.: Interpolation method for adapting reduced-order models and application to aeroelasticity. AIAA J. 46(7), 1803–1813
(2008)
9. Perez, R., Wang, X., Mignolet, M.: Nonintrusive structural dynamic reduced order modeling for large deformations: enhancements for complex
structures. J. Comput. Nonlinear Dyn. 9(3), 031008-1–031008-12 (2014). https://doi.org/10.1115/1.4026155
10. Farhat, C., Avery, P., Chapman, T., Cortial, J.: Dimensional reduction of nonlinear finite element dynamic models with finite rotations and
energy-based mesh sampling and weighting for computational efficiency. Int. J. Numer. Methods Eng. 98(9), 625–662 (2014)
11. Rutzmoser, J.: Model order reduction for nonlinear structural dynamics. Dissertation, Technische Universität München, München (2018)
12. Rutzmoser, J., Rixen, D.: A lean and efficient snapshot generation technique for the hyper-reduction of nonlinear structural dynamics. Comput.
Methods Appl. Mech. Eng. 325, 330–349 (2017)
Chapter 36
Experimental Spectral Submanifold Reduced Order Models
from Machine Learning
Mattia Cenedese and George Haller
Abstract Nonlinear system identification is a challenging problem in experimental modal analysis. It is currently tackled
using a toolbox approach, where different techniques are employed depending on the structural system under investigation,
the identification goals and the type of excitation used. In this contribution, we exploit analytic reduction to spectral
submanifolds combined with machine learning techniques in order to obtain the nonlinear coefficients up to cubic order
of a single-degree-of-freedom reduced order model. The system measurements aimed at model fitting can be performed
using any type of excitation techniques, ranging from free-decay to sine-sweeps or random shaker testing. We illustrate the
accuracy of our method using both simulated and real experimental data.
Keywords Nonlinear system identification · Reduce order model · Spectral submanifolds · Machine learning ·
Nonlinear vibrations
36.1 Introduction
Over the last two decades, there has been a significant investment in developing nonlinear system identification techniques for
multi-degree-of-freedom mechanical systems. When applied to single-mode responses, the primary goal of these techniques
is the reconstruction of the curves representing amplitude dependent properties of frequency and damping. The zero crossing
method, the Hilbert method, the short time Fourier transform, and the wavelet-based identification methods [1, 2, 3] are
popular non-parametric identification methods that typically rely on ringdown trajectories. Instead, force appropriation
[4] and experimental continuation [5] exploit periodic excitation for extracting system information. Parametric methods
often rely on the construction of a proper, potentially large state space model [6]. Each method may be more or less
suitable depending on the specific system to analyze, the available measurements, the possible excitations and the scope
of the identification [2]. Smaller effort has however been devoted to the identification of a simple reduced order model that
efficiently describes systems dynamics starting from available experimental data and prior information.
Spectral submanifolds (SSMs) [7] configure as a theoretical framework for addressing this issue. SSMs represent the
smoothest nonlinear continuation of linear mode shapes and they serve as a basis for exact, local model reduction of multi-
degree-of-freedom mechanical systems subject to arbitrary damping and to small forcing. A 2n-dimensional SSM can be
constructed by computing its parametrization and its reduced dynamics, either from a model [8] or from data [9]. Depending
on the coordinate representation, these maps can be more or less complicated. The approach that better unfold the reduced
dynamics is the polynomial normal form parametrization style [10]. For example, only four parameters govern the reduced
dynamics on a 2-dimensional or single-mode SSM up to quartic order, namely the linearized natural frequency, the linearized
damping ratio and a couple of cubic order parameters representing variations with amplitude in instantaneous frequency and
dissipation. However, the drawback of the normal form parametrization style is that the parametrization can, in principle, be
complex even for simple and localized system nonlinearities.
In this contribution, we propose a data-driven model identification technique for a 2-dimensional SSM that configures as
complementary to the identification approaches mentioned previously. It can be applied to either synthetic of experimental
data, to forced or decaying oscillation measurements. By assuming parametrization and reduced dynamics of polynomial
type, our goal is the identification of the systems coefficients up to cubic order.
M. Cenedese () · G. Haller

Institute for Mechanical Systems, ETH Zürich, Zürich, Switzerland

250 M. Cenedese and G. Haller
1
a) b)
0.5
q1 [m]
0
-0.5
-1
0 50 100 150 200 250 300
t [s]
Fig. 36.1 (a) Representation of the mechanical system in Eq. (36.1) and (b) simulated trajectory with initial conditions on the slow SSM
We adopt machine learning procedures in order to extract these nonlinear maps from data and perform uncertainty
quantification, particularly on the parameters of the reduced dynamics.
36.2 Method Overview with a Numerical Example
We illustrate our method on a specific example. The mechanical system in Fig. 36.1a with m = k = 1 and c = 0.006, its
equations of motion read

q̈1 + 2q1 − q2 + 0.006 (2q̇1 − q̇2 ) − 0.1q1 2 + 0.33q̇12 + 0.5q1 3 + 0.3q12 q̇1 + 0.05q̇13 = 0
(36.1)
q̈2 + 2q2 − q1 + 0.006 (2q̇2 − q̇1 ) = 0.
√
The linearized natural frequencies about the origin are 1 and 3, with damping ratios 0.003 and 0.009 respectively.
According to [7], no resonance condition is satisfied, so there exist an 2-dimensional SSM attached to each of the two linear
modes. Focusing on the slowest one, theoretical results show that the dynamics at quartic order can be reduced to

ρ̇ = −ρ 0.003 + αρ 2 = −ρζ (ρ) ω (ρ)
(36.2)
θ̇ = 1 + βρ 2 = ω (ρ) ,
where ρ denotes the parametrization amplitude, θ the parametrization angle, ω(ρ) the instantaneous frequency and ζ (ρ) the
instantaneous damping ratio. The dynamic model in Eq. (36.2) is related to physical coordinates through an O(4)-polynomial
mapping W (z, z) where z = ρeiθ . For simplicity, we assume to know the linear dynamical parameters exactly and we aim to
estimate α, β and the coefficients of W given the trajectory in Fig. 36.1b. The latter is a simulated decaying oscillation having
initial condition on the SSM of interest and, relying on nonparametric information, we impose prior probability distributions
for every parameter to estimate. For this example, a standard maximum a priori estimate performs effectively and we achieve
the results in Fig. 36.2. These plots show the instantaneous frequency and instantaneous damping ratio curves obtained with
the best-fit parameters in red, while blue lines are computed using the automated model-based routine in [8]. The grey areas
correspond to the loci of curves for parameters values whose reduced order model has 5% or less reconstruction error on the
training trajectory.
36 Experimental Spectral Submanifold Reduced Order Models from Machine Learning 251
1 1
0.8 0.8
max|q1 | [m]
max|q1 | [m]
0.6 0.6
0.4 0.4
0.2 5 % Error Bound 0.2

Model-based SSM
Data-driven SSM
0 0
0.99 1 1.01 1.02 1.03 1.04 1.05 1.06 1.07 0 0.005 0.01 0.015 0.02 0.025
a) [rad/s] b) [-]
Fig. 36.2 Instantaneous frequency (a) and instantaneous damping ratio (b) curves for the slowest SSM of the mechanical system in Eq. (36.1).
Blue lines refers to SSM model based computation as in [8] while red ones correspond to the best fit of our method. Grey areas depict parameter
bounds for which the reconstruction error is 5% or less
References
1. Noël, J.P., Kerschen, G.: Nonlinear system identification in structural dynamics: 10 more years of progress. Mech. Syst. Signal Process. 83,
2–35 (2017)
2. Kerschen, G., Worden, K., Vakakis, A.F., Golinval, J.: Past, present, and future of nonlinear system identification in structural dynamics. Mech.
3. Moore, K.J., Kurt, M., Eriten, M., McFarland, D.M., Bergman, L.A., Vakakis, A.F.: Wavelet-bounded empirical mode decomposition for
measured time series analysis. Mech. Syst. Signal Process. 99, 14–29 (2018)
4. Peeters, M., Kerschen, G., Golinval, J.C.: Dynamic testing of nonlinear vibrating structures using nonlinear normal modes. J. Sound Vib. 330,
486–509 (2011)
5. Renson, L., Gonzalez-Bulega, A., Barton, D.A.W., Neild, S.A.: Robust identification of backbone curves using control-based continuation. J.
Sound Vib. 367, 145–158 (2016)
6. Noël, J.P., Schoukens, J.: Grey-box state-space identification of nonlinear mechanical vibrations. Int. J. Control. 91(5), 1118–1139 (2018)
7. Haller, G., Ponsioen, S.: Nonlinear normal modes and spectral submanifolds: existence, uniqueness and use in model reduction. Nonlin.
Dynam. 86(3), 1493–1534 (2016)
8. Ponsioen, S., Pedergnana, T., Haller, G.: Automated computation of autonomous spectral submanifolds for nonlinear modal analysis. J. Sound
Vib. 420, 269–295 (2018)
9. Szalai, R., Ehrhardt, D., Haller, G.: Nonlinear model identification and spectral submanifolds for multi-degree-of-freedom mechanical
vibrations. Proc. R. Soc. London A Math. Phys. Eng. Sci. 473(2202), (2017)
10. Breunung, T., Haller, G.: Explicit backbone curves from spectral submanifolds of forced-damped nonlinear mechanical systems. Proc. R. Soc.
London A Math. Phys. Eng. Sci. 474(2213), (2018)
Chapter 37
Development of an Experimental Rig for Emulating Undulatory
Locomotion
S. N. H. Syuhri, A. McCartney, and A. Cammarano
Abstract Some creatures, such as eels, snakes and slender fish use body undulations to move the fluid around them and
create propulsion; similarly, travelling waves in structures can be used as an alternative propulsion system. Based on this
bio-inspired mechanism, the application of travelling waves is widespread in engineering, e.g. motors, pump systems and
transport devices. However, generating high amplitude travelling waves is not as easy as generating standing waves, since
travelling waves are generally observed away from resonance. Also, the reflection of the waves at the boundaries can interact
destructively further reducing their amplitude. This paper investigates the characteristics of the experimental rig built for
emulating the propulsion mechanism used in undulatory locomotion, introducing the conceptual design, and investigating
the features that promote travelling waves. Finally, an analysis of the influence of individual components in the design is
carried out.
Keywords Stationary swimmer · Beam characteristics · Experimental rig · Electromagnetic · Undulatory locomotion
37.1 Introduction
Research into the locomotion of animals has been carried out for centuries with several studies taking place in the attempt
of understanding the styles utilised by different species. Aquatic animals have been of particular interest as they are able
to propel through fluids very efficiently and with the use of minimal energy [1]. By swimming in groups, for instance, fish
can recover energy from vortices induced by environment to reduce the cost of locomotion [2]. Evolution also resulted in
creature adopting different swimming styles such as anguilliform, subcarangiform, carangiform and thunniform [3]. The
style mentioned here all produce forward motion via undulations from the animal’s tail, with the difference being in how
much of the body is used for propulsion. The anguilliform swimming style, which is also adopted by eels and snakes, uses
almost the entire body whereas fish and whales only use their tail and a limited part of their body. Eels and snakes produce
the undulations via neurons firing in series down their spines triggering the muscles to create the deformations driving them
forward [4].
The study of undulating swimmers began with the development of theoretical works based on the natural observation such
as locomotion of the eels pioneered by Gray [5]. The paper describes the different movements of fish and compares them to
mechanical forms of propulsion (jet, paddle, screw, etc.). The investigation then focuses on the movement displayed by eels.
Eels’ bodies are comprised of a moving head with a flexible, elastic tail which responds to undulations created by the head.
This creates sinusoidal wave patterns along the length of the eel’s tail. Gray’s work also includes photographic evidence of
the eels swimming style to analyse the effects of the swimming pattern on the resistance created by the fluid.
The research was then furthered by Taylor [6] who expanded the analysis to include sea snakes and marine worms; this
also included geometrical models different than those used in previous works. In Taylor’s study the simplifications used for
the swimmer shape do not permit to analyse the increasing amplitude of body displacement exhibited by some creatures
while swimming. This is addressed in Gray and Hancock’s work [7] which presents a more realistic model of a swimmer
that contemplate larger amplitudes, similar to those observed in nature.
S. N. H. Syuhri () · A. McCartney · A. Cammarano

School of Engineering, University of Glasgow, Glasgow, UK

254 S. N. H. Syuhri et al.
Notable are also the works by Lighthill [8, 9], that investigated the movements of slender fish and how these can produce
an efficient propulsion method when compared to the relative muscular output of sea creatures. In this paper, the author
investigates the boundary flow around a swimmer and concludes that the boundary layer closely surrounding the swimmer
is mostly laminar and as such, the drag experienced is lower than expected, explaining the high efficiency of the process.
The influence of the fluid characteristics, including the effect of the Reynolds number, used to describe the nature of fluid
regime, is detailed in Wu [10]. This work focused on the effect that a large Reynolds number would have on the swimmer
and derived new equations to determine the swimmer’s propulsion and dynamics. Further studies on this topic can be found
in a number of works investigating the effect of Reynold’s number on swimming [11–13].
Shortly after theoretical analysis of the swimming techniques, the first attempts to emulate the motion through engineered
mechanism were reported. The initial studies focussed on the anguilliform swimming style, the easiest swimming style to
recreate. In this case, the motion of the entire body is generated through motion of the head, which induces an oscillatory
motion through the tail – or in the case, the rest of the body. This motion can be recreated in a structure made of a lumped
mass attached to a beam where the muss is displaced laterally [14]. The head can be assumed to be the only moving part,
with the beam reacting to its movement: this translates the lateral movement of the head into forward motion [15]. This
model, however, will allow displacement at one point (in this case the head of the swimmer), while an animal can apply
lateral displacement needed to gain forward motion at any point along its body [16].
An experimental reproduction of this motion was created in [17], where a buoyant swimmer moving forward on the water
surface was observed. The swimmer consisted of an elastic tail with a magnet embedded into its head. The setup presents two
Helmholtz coils at either side of the tank, which induce an electromagnetic field that is used to move the head of the swimmer.
The experiment proved that the swimmer can achieve forward motion, but also highlighted the difficulty in controlling its
position. Therefore, it is desirable to investigate further the interaction with the electromagnetic field, so that a better control
can be attained through by manipulation of the magnetic fields. The aim of this study is to create a stationary model that can
be used to study the dynamic behaviour of the swimmer and its interaction with an external electromagnetic field. For this
purpose a new test rig has been designed that allows for the monitoring of the magnetic field whilst monitoring the behaviour
of the swimmer.
37.2 Structural Design and Operation
In order to excite the head of the swimmer, an electromagnet is used to generate a magnetic fields around a permanent magnet
placed in the head of the swimmer. The electromagnet itself is made of a coil wound around a region of space in which a
magnetic field is required; a controlled current is induced in the coil. Unlike a permanent magnet, the poles and strength of
an electromagnet can be reversed and varied by reversing and altering the current. By giving a periodic oscillation to the
current, the electromagnet can either attract or repel the permanent magnet with periodic forces. This system is used as a
forcing mechanism for the purpose of this work.
The coils are positioned parallel to each other, and are designed to produce a stable magnetic field [18]. The coil distance
allows for the swimmer to be mounted between the coils, springs and supports are used to keep the swimmer in the right
position. The minimum elongation of the springs is 12 mm the head of the swimmer is 20 mm. These geometrical constraints
dictate that the minimum distance between the coils is about 40 mm, but to allow for changes to the design this distance
was made to be 60 mm. In order to have a homogeneous field in the region of interests it has been calculated that the coils
required a minimum radius of 60 mm.
The spool has been designed to hold approximately 1000 turns of wire. This design choice has determined the thickness
and width of the spool as well as its outside diameter. Estimations of these design parameters have been made assuming an
even distribution of the wire around the spool. With 1000 windings the optimum distribution is roughly 30 windings across
by 30 layers of winding. A 0.5 mm copper wire would require the spool to house a coil with a section of 15 mm x 15 mm.
This assumes that the wire is wrapped in the most space-efficient way, which is unlikely to be achieved. Therefore, safety
factors have been applied to determine the width and the height of the spool, which have been dimensioned to be 20 mm and
30 mm respectively.
The magnetic head is attached to the springs and to the beam. This was achieved by using a sandwich configuration where
the magnet and beam are embedded between two plates, hold together with four bolts in the corners. These bolts are also
used to secure the head to the springs. Hence the frames purpose is not only to hold the spools in place, but also to hold the
springs and the swimmer in place.
The exploded view of the full assembly is shown in Fig. 37.1. The final design sees the Helmholtz coils being mounted
to the inside of the uprights of the frame. The 3D printed coils feature a cross that attaches it to the uprights while also
37 Development of an Experimental Rig for Emulating Undulatory Locomotion 255
Fig. 37.1 Exploded view of the full assembly
Fig. 37.2 Cross-section showing spring mounting points
creating a mounting point for the springs to attach to. The stainless-steel springs are expected to have a negligible effect on
the magnetic field, which in good approximation can be considered uniform.
The swimmer itself will consist of a thin beam of metal. The design will consist of two 3D printed sheets which will
sandwich the magnet and beam between them. The two sides will then be held together with screws in each corner which
will also act as the mounting point for the springs. The assembled swimmer and its mounting points are depicted in Fig. 37.2.
37.3 Fabrication and Measurement
37.3.1 Coils Characteristics
The parts involving the spool and frame were made using a Makerbot Replicator + FDM 3D printer with a polylatic acid
(PLA) filament with 10–20% infill. Approximately 800 turns are wrapped around the spools to make the coil. The wire
diameter is of 0.51 mm, the total resistance 32 and 25 for each coil respectively. The discrepancy in the resistance
values is justified by inhomogeneity in the copper and imperfections in the winding process. To avoid any problem from
different phase shift arising from the difference in resistance, these coils are connected in series: the total resistance is
Fig. 37.3 Frequency response of the amplifier in different gain levels
Fig. 37.4 Test setup for characterising magnetic fields
56.5 . Both the signal generator and signal analyser are provided by the SignalCalc Ace powered by the Quattro hardware
platform which has the maximum peak-to-peak voltage (Vpp) of 10 V.
The output signal from Quattro is then amplified by a 48 W power amplifier provided by LDS PA25E. Figure 37.3 shows
the magnitude of the transfer function with respect to frequency of the power amplifier in two different gain levels. The
magnitude of the transfer function over frequency was obtained by connecting the input and output signals of the amplifier
to Quattro for computing the response using the fast Fourier transform (FFT) spectrum analysis. It is seen from the figure
that, for all purposes, the transfer function can be considered constant after 10 Hz.
To obtain the characteristics of the magnetic field, a smaller coil was built and placed between the two coils as shown in
Fig. 37.4. The system obtained is analogous to a single-phase air core transformer with the big and small coils as the primary
and secondary windings, respectively. A random signal of 0.707 V was generated from Quattro with the gain amplifier of 3.
In the first test, the small coil was made of 1000 turns with the wire diameter of 0.2 mm, resulting in the resistance of 45.2 .
The second test was conducted with a slightly bigger coil to reduce the resistance. In the second case the coil was made of
600 turns with wire thickness of 0.41 mm, leading to a resistance of 3 . The frequency response of the first and the second
tests is plotted in Fig. 37.5 indicated with the red and blue line colours, respectively. Here the magnitude refers to the ratio
of the output voltage of the small coil to the input voltage generated by Quattro. As shown, the transfer function is frequency
dependent and converge towards a constant value for high frequencies.
Fig. 37.5 Magnitude of the transfer function of the coil receiver relative to the signal generator with respect to frequency
Fig. 37.6 Test set up for beam characterization
37.3.2 Natural Frequencies of the Beam
To assess the natural frequencies, the beam is mounted on a shaker specifically for vibration and mechanical shock testing
provided by LDS V406. To simulate the boundary conditions experienced in the test rig, the beam was clamped at one side,
leaving the other side free to vibrate. The general layout of the testing method is shown in Fig. 37.6. An accelerometer is
mounted on the base of the beam to measure the acceleration level generated by the shaker. Meanwhile, a laser vibrometer is
used to measure the velocity at the tip of the beam. The output voltage from these measurements are amplified and connected
to Quattro to compute the frequency response function (FRF). The shaker is driven by a random input signal. The signal is
generated via a signal generator and amplified by a power amplifier to provide the power required by the shaker.
The beam, made of a metal material with the dimension of 149.65 × 15.2 × 0.4 mm, was excited with an arbitrary random
signal of 0.707 V. The obtained FRF acquired from the signal analyser is shown in Fig. 37.7. The peaks in the FRF refer to
the natural frequencies of the beam indicated with markers ‘x’, ‘o’ and ‘*’ for the first, second and third natural frequencies,
respectively.
Fig. 37.7 FRF measured at the tip of the beam relative to base excitation
Fig. 37.8 Assembled swimmer rig
37.3.3 Swimmer Characteristics
The same PLA was used for the swimmer’s head as it satisfied the required stress resistance with 10% infill. To ensure that
there is enough movement in the head of the swimmer, an N52 neodymium magnets is used. The N52 grade is the most
powerful readily available magnets with the dimension of 40 mm × 20 mm × 5 mm, resulting in a pull strength of 15.1 kg.
The spring has a spring rate of 210 N/m with an initial tension of 0.66 N. The assembled swimmer and its rig are shown in
Fig. 37.8.
The same test procedure that was used to characterise the beam was used to obtain the natural frequency of the swimmer.
However, the output voltage from the accelerometer is replaced with the output voltage from Quattro to avoid the use of
accelerometer under strong magnetic fields. The signal is then compared to the velocity at the tip of the beam from the laser
vibrometer. The FRF of the measured signal is shown in Fig. 37.9. The markers with the green and blue colours indicate the
natural frequencies of springs and the swimmer, respectively. It is seen that the first-two natural frequencies are displaced as
the presence of the springs as well as the swimmer’s head.
Another peak that appears on the FRF is the torsional bending of the beam indicated by the black marker colour. The
torsional bending was evaluated by measuring the time series of velocity in different locations along the beam width. Figure
37.10a shows the velocity response at the tip of the beam generated at the natural frequency of the torsional bending of
25.75 Hz. For comparison, the same measurement was taken at the fourth peak, 36.72 Hz, as shown in Fig. 37.10b. It is seen
that there is different phase and amplitude in the third peak along the width of the beam. On the contrary, the output has
Fig. 37.9 FRF measured at the tip of the beam relative to the signal generator
Fig. 37.10 Time histories showing torsional vibration on the beam
Fig. 37.11 Time histories of the swimmer at different frequency; blue: laser vibrometer, red: signal generator
only a small discrepancy in the amplitude but the same phase in correspondence of the fourth peak. This can be caused by
either manufacturing imprecision of the head and bad positioning of the swimmer or non-uniformity of the magnetic fields.
Therefore, this discrepancy becomes obvious at the natural frequency for the torsional vibration as shown in Fig. 37.10a.
At the natural frequency, an adjustment in the signal amplitude is needed to avoid that the swimmer hits the boundary: this
can introduce distortion in the measurements and subsequently subharmonics in the response of the beam. This phenomenon
is depicted in Fig. 37.11a, b generated, for instance, at the second peak of the FRF, 19.53 Hz, with the amplitude of 0.3 V
and 0.4 V, respectively. It is seen that there is an obvious difference of the peaks between the two which resulted from the
impact of the swimmer to the boundary at 0.4 V.
At frequencies far away from resonant values, the response becomes invisible as it requires more power to increase the
strength of the magnetic fields. Figure 37.11c, d show the sensitivity of the measurement with respect to time, generated at
10 Hz and 80 Hz, respectively, with the input voltage amplitude of 1 V and maximum gain amplification level. The swimmer
offers a stronger response at 10 Hz as it is close to the first resonant value. However, the response of the beam is almost
negligible at 80 Hz where the closest peak is approximately 10 Hz away.
37.3.4 Swimmer with Head Support
In the experiments presented in the previous sections it has become evident that the torsional motion can have a strong
impact in the measurements. To avoid the structural torsion, an acrylic plate is used to support the swimmer’s head. This
design will avoid rotation of the head around the axis of the beam. The bottom of the head is also covered with the same
material to create smooth contact between the surfaces. The purpose of this support is also to limit the vertical movement of
the swimmer. This also reduce the effects of imbalance of the centre of mass and/or improper positioning of spring. Figure
37.12 shows the assembled model of the head support integrated in the rig. For evaluating the FRF, a maximum 10 Vpp was
generated from the signal generator with the maximum gain level for the amplifier.
Figure 37.13 shows the FRF of the swimmer model with the head support. The first-two natural frequencies obtained in
Fig. 37.13 are similar to that of Fig. 37.7. The discrepancy in the frequency values is caused by the different length of the
clamped section of the beam, where the swimmer requires only 5 mm of the beam length to fully engage with the head’s slot.
Based on this occurrence it is seen that there is not enough force generated by magnetic fields to counter the static friction
force of the surface, leading to stick condition of the head of the swimmer.
The stick state in the swimmer results in shaking a whole rig (including the holder and the coils). The velocity time
histories captured at the tip of the beam are presented in Fig. 37.14. Figure 37.14a, c, d reveal that multiple frequencies
were involved in the measured responses. On the contrary, a clear sinusoidal signal was obtained when excited at the natural
frequency of the beam as shown in Fig. 37.14b. These are an indication that the magnetic fields induce vibration through the
whole frame, as clearly observed in the experimental phase. Clearly, in this case the dynamics of the beam becomes more
complex and the frequency content of the displacement is far richer than expected.
Fig. 37.12 Swimmer with an

additional head support
Fig. 37.13 FRF at stick condition
Fig. 37.14 Time histories at stick conditions generated with forcing frequency of (a) 7.5 Hz, (b) 13 Hz, (c) 19.5 Hz and (d) 25.75 Hz
To achieve the slip condition, an arbitrary initial force was given to the swimmer to overcome the static friction force
between two surfaces. Since the movement of the swimmer was not visible around the natural frequency using random
signal, the input was changed with the sinusoidal wave. This, however, would give a strong peak in the forcing frequency
in the FRF of the signal analyser, attenuating the influence of the modes. Therefore, a noise signal was superimposed to the
sine wave to excite all the modes in the frequency range of interest and therefore generate a frequency response function.
This method was first validated in the proper shaker that was used for evaluating the natural frequency of the beam in Sect.
37.3.2. From the measurement, both the sinusoidal input with noise and random input, not shown here for brevity, give the
same resonant frequencies with good repeatability. The discrepancy, however, is occurred in the magnitude of the FRF in
which this is not of interest in this study.
For the swimmer with the head support, the FRF from signal analyser is shown in Fig. 37.15a, c, e, g. From the figures it
is seen that the responses contain noisy signals with inconsistent pattern of the peaks. The measurements were repeated by
adding lubricants as well as adjusting the percentage of the noise input levels with the purpose to remove the noisy signal
and unveil the modes, but either due to an unsuitable method in analysing FRF using sinusoidal input signals or the friction
between two acrylic surfaces creating unwanted noise resulted in indeterminate FRF patterns.
Further investigations were carried out by probing output signals from the laser vibrometer and signal generator with
respect to time using an oscilloscope. The responses are then depicted in Fig. 37.15b, d, f, h. The notches and sharp peaks
in the time series from the figures demonstrate the presence of stick-slip condition in the swimmer dynamics. “Stick” occurs
when the swimmer changes the direction of the displacement, resulting in the sharp peak of velocity responses. These time
series plots reveal that the damping caused by the friction of two surfaces is high enough to make the swimmer response not
strongly correlated to the input signal. As the frequency increased, the swimmer exhibits very small displacements and the
Fig. 37.15 FRF and time histories at slip condition generated with (a, b) 30 Hz, (c, d) 40 Hz, (e, f) 50 Hz and (g, h) 60 Hz
influence of the spring become negligible. Therefore the well identifiable peaks in Fig. 37.15e, g become more similar to
those observed in Figs. 37.7 and 37.13 where the beam can be considered clamped at one end.
37.4 Conclusion
In this paper, a conceptual design of the experimental rig for emulating undulatory motion has been presented and fabricated.
An investigation of the characteristics of the components of the rig has been undertaken, and the assembled models have been
examined. The designed swimmer presents three natural frequencies under 150 Hz. Two of these frequencies are influenced
by springs used in the assembly, while the other frequency is from the torsional vibration of the beam. The presence of
springs also shifted the first and second resonant frequencies of the beam upwards to 19.53 Hz and 98.24 Hz, respectively.
The torsional vibration occurring in the swimmer is quite disruptive for the experiments: for this reason additional
improvements to the test rig were necessary. A plate to limit the motion of the swimmer was introduced. However, the
friction forced induced by the contact of the head of the swimmer with the plate showed that the undulatory motion of the
swimmer is heavily affected and that further development of the design is required.
It is recommended that the head and support of the swimmer are re-designed so that the influence of the stick slip
phenomenon is reduced, possibly eliminated. This is essential for future studies where the swimmer will be immersed in
liquid environment.
References
1. Cohen N, Boyle JH (2009) Undulatory locomotion. arXiv e-prints

2. Marras, S., Killen, S.S., Lindström, J., McKenzie, D.J., Steffensen, J.F., Domenici, P.: Fish swimming in schools save energy regardless of
their spatial position. Behav. Ecol. Sociobiol. 69(2), 219–226 (2014). https://doi.org/10.1007/s00265-014-1834-4
3. Tytell, E.D., Borazjani, I., Sotiropoulos, F., Baker, T.V., Anderson, E.J., Lauder, G.V.: Disentangling the functional roles of morphology and
motion in the swimming of fish. Integr. Comp. Biol. 50(6), 1140–1154 (2010). https://doi.org/10.1093/icb/icq057
4. Muller, U.K., van Leeuwen, J.L.: Undulatory fish swimming: from muscles to flow. Fish Fish. 7(2), 84–103 (2006). https://doi.org/10.1111/
j.1467-2979.2006.00210.x
5. Gray, J.: Studies in animal locomotion. I The movement of fish with special reference to the eel. J. Exp. Biol. 10(1), 88–104 (1933)
6. Taylor, G.I.: Analysis of the swimming of long and narrow animals. Proc. R. Soc. London Ser. A Math. Phys. Sci. 214(1117), 158–183 (1952).
https://doi.org/10.1098/rspa.1952.0159
7. Gray, J., Hancock, G.J.: The propulsion of sea-urchin spermatozoa. J. Exp. Biol. 32(4), 802–814 (1955)
8. Lighthill, M.J.: Note on the swimming of slender fish. J. Fluid Mech. 9(2), 305–317 (1960). https://doi.org/10.1017/s0022112060001110
9. Lighthill, M.J.: Large-amplitude elongated-body theory of fish locomotion. Proc. R. Soc. London Ser. B Biol. Sci. 179(1055), 125–138 (1971).
https://doi.org/10.1098/rspb.1971.0085
10. Wu, T.Y.-T.: Hydromechanics of swimming propulsion. Part 1. Swimming of a two-dimensional flexible plate at variable forward speeds in an
inviscid fluid. J. Fluid Mech. 46(2), 337–355 (1970). https://doi.org/10.1017/s0022112071000570
11. Eloy, C., Schouveiler, L.: Optimisation of two-dimensional undulatory swimming at high Reynolds number. Int. J. Non-Linear Mech. 46(4),
568–576 (2011). https://doi.org/10.1016/j.ijnonlinmec.2010.12.007
12. Tyson, R., Jordan, C.E., Hebert, J.: Modelling anguilliform swimming at intermediate Reynolds number: a review and a novel extension
of immersed boundary method applications. Comput. Methods Appl. Mech. Eng. 197(25–28), 2105–2118 (2008). https://doi.org/10.1016/
j.cma.2007.07.009
13. Wiggins, C.H., Goldstein, R.E.: Flexive and propulsive dynamics of Elastica at low Reynolds number. Phys. Rev. Lett. 80(17), 3879–3882
(1998). https://doi.org/10.1103/PhysRevLett.80.3879
14. Triantafyllou, M.S., Triantafyllou, G.S., Yue, D.K.P.: Hydrodynamics of fishlike swimming. Annu. Rev. Fluid Mech. 32(1), 33–53 (2000).
https://doi.org/10.1146/annurev.fluid.32.1.33
15. Cox, R.G.: The motion of long slender bodies in a viscous fluid part 1. General theory. J. Fluid Mech. 44(04), (1970). https://doi.org/10.1017/
s002211207000215x
16. Lavie, A.M.: Analysis of the swimming of elastic slender bodies excited by an external force. J. Fluid Mech. 53(4), 701–714 (1972). https://
doi.org/10.1017/s0022112072000436
17. Ramananarivo, S., Godoy-Diana, R., Thiria, B.: Passive elastic mechanism to mimic fish-muscle action in anguilliform swimming. J. R. Soc.
Interface. 10(88), (2013). https://doi.org/10.1098/rsif.2013.0667
18. Jackson, J.D.: Classical Electrodynamics, Third edn. Wiley, Hoboken, New Jersey (1999)
Chapter 38
Identification and Modeling of a Variable Amplitude Fatigue
Experiment Apparatus with Damaged Beam Specimen
Abstract The useful remaining life of engineering structures under variable amplitude (VA) fatigue loading remains a major
unresolved engineering problem. The existing proposed life prediction models are usually based on empirical approximation
from experimental results (Fatemi, Yang Int J Fatigue 20(1):9–34, 1998, Santecchia et al. Adv Mater Sci Eng 2016:1–26,
2016). The variable fatigue experiment apparatus in this extended abstract was designed for simulating structural fatigue with
a high testing frequency, variable R-ratio as well as modifiable experimental layout (Falco et al. J Vib Acoust 136(4):041001,
2014). In previous studies, the inherent nonlinearity of the testing rig was detected, the obtained parameters allow one
to properly use this testing rig within its linear region. As damage accumulates, however, the corresponding dynamic
characteristics of the specimen alter accordingly. Therefore, proper modeling considering the interaction between the inherent
nonlinearity and the damage induced nonlinearity for both (1) opening crack and (2) breathing crack is necessary for future
fatigue life estimation under complex fatigue loading. Here, nonlinear system identification of the lately modified variable
amplitude fatigue experiment apparatus is presented based on a combination of first-principles and data-driven modeling
techniques. Eventually, structure-damage interaction dynamics will be described to model the underlying fatigue evolution
and structural dynamics interactions.
Keywords Nonlinear system identification · Modal parameter estimation · Damage identification · System-damage
interaction · Hysteretic stiffness · Stick-slip dynamics
38.1 Calibration of Accelerometers
The investigation of engineering fatigue mainly relies on experimental observations. And the models for fatigue life
prediction under variable amplitude loading are, most of the time, empirical [1–3]. Here, we investigate the modeling of
nonlinear dynamic characteristics of a damaged beam through experimental modal analysis. First, the nonlinear system
identification is conducted on the fatigue experiment apparatus with a healthy beam.
Before system identification, calibration of transducers was performed in order to make the resulting measurements
comparable in their corresponding units [4]. In the dynamic characterization, the main sensors being used are piezoelectric
accelerometers and Eddy current displacement sensor. Two different series of accelerometers are being used, namely, PCB
333B42 and PCB U352C66 ICP accelerometers. Each series has three accelerometers, among which the measurements
should be comparable such that their dynamic responses to the input are similar (Table 38.1).
Table 38.1 Table of relative calibration information

DAQ channel no. 0 1 2 3 4 5
Sensor channel no. 2 1 3 4 5 7
Sensor series 333B42 U352C66
Sensor no. 12002 18444 18442 14469 20222 20224
Calibration ratio (RMS) 1.000 1.110 1.136 0.215 0.189 0.189
H. Li () · D. Chelidze
Nonlinear Dynamics Laboratory, Department of Mechanical, Industrial and Systems Engineering, University of
Rhode Island, Kingston, RI, USA

266 H. Li and D. Chelidze
Fig. 38.1 Location of the accelerometers used during the calibration

Magnitude (V)
0.04 Ch0 Ch3

0.02 Ch1 Ch4
Ch2 Ch5
0
-0.02
0 0.1 0.2 0.3 0.4 0.5
Time (second)
Magnitude (V)
2 Ch0 Ch3
1 Ch1 Ch4
0 Ch2 Ch5
-1
0 0.1 0.2 0.3 0.4 0.5

Time (second)
Magnitude (V)
0.1 Ch0 Ch3

Ch1 Ch4
0 Ch2 Ch5
-0.1
0 0.1 0.2 0.3 0.4 0.5
Time (second)
Magnitude (V)
2 Ch0 Ch3
Ch1 Ch4
0 Ch2 Ch5
-2
0 0.1 0.2 0.3 0.4 0.5
Time (second)
Fig. 38.2 Plot of load time history of the acquired signals without filtering. Upper two: raw and normalized load time history during noise floor
estimation; lower two: raw and normalized load time history when system under a 23 Hz, 200 mVpp sinusoidal excitation
38 Identification and Modeling of a Variable Amplitude Fatigue Experiment Apparatus with Damaged Beam Specimen 267
0 -20
Ch0 Ch2 Ch4 Ch0 Ch2 Ch4
-10 Ch1 Ch3 Ch5 -25 Ch1 Ch3 Ch5
Magnitude (dB)
Magnitude (dB)
-20 -30
-30 -35
-40 -40
-50 -45
-60 -50
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
Fig. 38.3 Welch’s periodogram of the acquired data for noise floor estimation of accelerometers
The six accelerometers are densely glued to one of the beam support to ensure no bias was introduced by the vibration
mode of the beam support, see Fig. 38.1. The calibration was conducted under an input of sinusoidal signal with a frequency
of 23 Hz and a magnitude of 200 mVpp. And the signals were amplified with a gain of 1 through PCB signal conditioner
498A03 and were not filtered. For the data acquisition, the out put signals were recorded by 6 channels on a National
Instrument data acquisition card, with channel names a0 – a5 (CH1, . . . , CH5, and CH7 from Fig. 38.1). All channels
were sampled with a sampling frequency of 1000 Hz. As for the frequency component of the acquired signals, a Welch’s
power spectral density estimation (periodogram) was obtained by adding a Hanning window (1024 samples long) with
50% overlapping between windowed signal segments. The raw and the RMS normalized load time history data in noise
floor estimation were presented in Fig. 38.2. The noise floor estimation can be observed from Fig. 38.3. Since the fatigue
experiments will be executed under 100 Hz, the noise estimate is around −24 dB. The comparison between the two series
of accelerometers indicates that the 333B46 series has batter signal to noise ratio (SNR) under 200 mVpp, 23 Hz sinusoidal
excitation, comparing to its counterparts. Aside from the 23 Hz resonance frequency, line noise related peaks can be found at
60, 120 Hz in the 333B42 spectral density plot. Further investigation should be done in investigating the source of the 300 Hz
peak, but for the sake of future experiments, this can be mitigated by the application of an antialiasing filter.
38.2 Detection of Nonlinearity
38.2.1 Frequency Response Function Overlay
Frequency response function is widely used in experimental modal analysis methodology which can also be utilized to detect
nonlienarity of the system under investigation. Frequency response function overlay (FRFO) was used here to detect mainly
the nonlinear stiffness of the system under investigation.
After calibrating the accelerometers, the system identification was established using sine swept excitation with up and
down sweeping in order to detect nonlinearity in the system, if there is any. Frequency response functions can be obtained
with various excitation level, this helps identify the certain kind of nonlinearity in the system caused by nonlinear stiffness.
Therefore, three sets of swept-sine loading experiments were designed at three excitation amplitude (they are 100, 300, and
500 mVpp, respectively). The preliminary FRFO results are illustrated in Fig. 38.4 from which it indicates that there exists
softening spring nonlinearity since the resonance frequency depends on the forcing magnitude in a negatively correlated
way. In another word, the resonance frequency of the system decreases with increasing forcing magnitude. Although the
experiments were designed to detect the nonlinearity by up-sweeping and down-sweeping in frequency of sinusoidal loading,
the results indicate relatively high damping in the system. Even though the acquired FRFs are influenced by damping, the
results still support the observation of softening spring nonlinearity. The existence of a higher peak at a lower amplitude of
excitation can be explained from the right sub-figure in Fig. 38.4 which shows the FRF in the vertical direction of the beam
specimen. The higher peak is induced by the vertical resonance frequency at 80 Hz. Therefore, the magnitude of the FRFO
increases with increasing input amplitude.
10 -15
-20
5
-25
Magnitude(dB)
Magnitude(dB)
0 -30
-35
-5
Low amplitude Low amplitude
Mid amplitue -40 Mid amplitue
High amplitude High amplitude
-10 -45
20 40 60 80 100 20 40 60 80 100
Frequency(Hz) Frequency(Hz)
Fig. 38.4 Frequency response function overlay results indicate softening spring nonlinearity in the system under investigation
40 40
Scatter points First loading scatter
Cubic curve fit Unloading and reloading
20 Pressure (psi) 20
Pressure (psi)
0 0
-20 -20
-40 -40
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5
Voltage (displacement sensor) Voltage (displacement sensor)
Fig. 38.5 Left: raw data of the deflection-pressure
38.2.2 Nonlinear Stiffness Identification Under Static Loading
In previous laboratory location, nonlinear stiffness was detected and simplified to a cubic nonlinear stiffness. However,
hysteresis behavior was observed and was not included in the modeling. More importantly, as damage accumulates, the
stiffness characteristics can be much more complex, hence, a more judicious examination has to be considered. Here, aside
from the cubic softening nonlinear stiffness, the reason behind the hysteresis was investigated by loading and unloading
the system statically through the pneumatic cylinders. The deflection-pressure curves were obtained by measuring the
deflection using a Lion precision eddy current displacement sensor, and the corresponding pressure was measured by two
SSI technology P51 series pressure sensors. Two sets of experiments were designed and executed. The pressure in the two
cylinders were set to equal values before initiation of the experiments. A static load will then be applied to the beam from one
side by increasing the pressure in the cylinder on the same side. The first experiment was a pure loading experiments during
which the pressure setup will increase monotonically until 36psi while the second experiment was designed to explore the
hysteresis behavior in the force-displacement relation. This result indicates softening spring nonlinear stiffness (which can
be described by a cubic function) as well as hysteretic behavior, see the right sub-figure in Fig. 38.5. For now, it is hard to
designate a good model for this behavior since dynamic behavior that relates force-velocity will reveal the hysteretic behavior
in detail [5, 6].
38 Identification and Modeling of a Variable Amplitude Fatigue Experiment Apparatus with Damaged Beam Specimen 269
38.3 Future Work
First, dynamic characterization of nonlinear stiffness will be conducted using restoring force reconstruction to further
understand the dynamics of the system under investigation. Later on, damage identification with the comparison of a long
notched beam and beams with equal length macro crack will be performed to provide empirical data for future parameter
estimation and model updating. The last step is to establish a system model with time varying system parameters that takes
the damage propagation into consideration. This will be further served to produce accurate results in real time fatigue life
prediction modeling.
Acknowledgments This work is supported by the National Science Foundation Grant No. 1561960.
References
1. Fatemi, A., Yang, L.: Cumulative fatigue damage and life prediction theories: a survey of the state of the art for homogeneous materials. Int. J.
Fatigue 20(1), 9–34 (1998)
2. Santecchia, E., Hamouda, A., Musharavati, F., Zalnezhad, E., Cabibbo, M., El Mehtedi, M., Spigarelli, S.: A review on fatigue life prediction
methods for metals. Adv. Mater. Sci. Eng. 2016, 1–26 (2016)
3. Falco, M., Liu, M., Hai Nguyen, S., Chelidze, D.: Nonlinear system identification and modeling of a new fatigue testing rig based on inertial
forces. J. Vib. Acoust. 136(4), 041001 (2014)
4. Brincker, R., Ventura, C.: Introduction to Operational Modal Analysis. Wiley, Chichester (2015)
5. Liu, Y., Li, J., Zhang, Z., Hu, X., Zhang, W.: Experimental comparison of five friction models on the same test-bed of the micro stick-slip motion
system. Mech. Sci. 6(1), 15–28 (2015)
6. Klepka, A., Dziedziech, K., Spytek, J., Mrówka, J., Górski, J.: Experimental investigation of hysteretic stiffness related effects in contact-type
nonlinearity. Nonlinear Dyn. 95(2), 1513–1528 (2019)
Chapter 39
Higher-Order Decompositions for Modal Identification and
Model Order Reduction
David Chelidze
Abstract Output only modal analysis is an essential tool for monitoring operations of complex large structures like offshore
platforms or studying complex flow dynamics. Here we consider a higher-order singular value and non-Hermitian matrix
decompositions and describe how they can be used in linear modal analysis to enhance the currently available output only
modal analysis methods such as dynamic mode decomposition or eigenvalue realization algorithm. In addition, we show how
these methodologies can be used for empirical nonlinear modal identification to obtain the slow flow dynamics of nonlinear
dynamical systems. Finally, we show how this information can be used to obtain high-fidelity robust reduced-order models
of nonlinear systems.
Keywords Output only modal analysis · Higher-order singular value decomposition · Higher-order non-Hermitian matrix
decomposition · Reduced-order models
39.1 Introduction
Many advanced engineered systems are designed to operate in nonlinear systems’ domain. This expansion from the linear
system’s domain is caused by the need for more efficient and lighter structures to meet the demands of the modern economic
landscape. When operating in the linear domain, dynamical systems spatiotemporal response is usually analyzed using
separation of variables—i.e., separation of temporal and spatial domains—and the application of the superposition principle.
Linear model and data reduction techniques mainly use Galerkin projection onto a set of spatial basis functions described
by the linear canonical transformations, which are actions in a time-frequency plane (e.g., Fourier, Laplace, Bergman,
Fresnel Transforms). When dealing with transient and intermittent cases, the time and frequency domains are hard to
decouple. However, superposition is still applicable to a set functions or modes localized in both time and frequency,
which are determined, for example, using Wavelet or Chirplet Transforms. In addition to analytical decomposition methods,
statistical/probabilistic characteristics of linear systems can be used to generate empirical decompositions (e.g., Principal
Component Analysis, Independent Component Analysis, Proper Orthogonal Decomposition). More recently, we have also
considered empirical decompositions that can account for both statistical (spatial domain) and temporal (time-frequency
domain) properties of the system such as Smooth Orthogonal Decomposition, Empirical Mode Decomposition, and Dynamic
Mode Decomposition.
Most of the decompositions described above are implicitly using the superposition principle to describe a system in terms
of appropriate modes. In the case of nonlinear systems, when superposition is not applicable, linear decompositions cannot
be expected to reflect or capture the separate and individual modal dynamics in a single linear mode. However, they can still
be used to capture (or embed) the active nonlinear modes in a linear subspace of the original system’s phases space, if these
active modes are spatially bounded. In many practical applications this is reduced to looking for most energetic “modes”
or a linear subspace that maximally captures the system’s energy (e.g., Proper Orthogonal Decomposition or Principal
Component Analysis); to identify statistically independent subspaces (e.g., Independent Component Analysis); capturing
best linear approximation of the nonlinear mode (e.g., Dynamic Mode Decomposition); or identifying subspaces containing
specific temporal of frequency characteristics (e.g., Empirical Mode Decomposition, Krylov Subspace Analysis, and Smooth
Orthogonal Decomposition).
D. Chelidze ()
Nonlinear Dynamics Laboratory, Department of Mechanical, Industrial and Systems Engineering, University of Rhode Island, Kingston, RI, USA

272 D. Chelidze
39.2 Background
In output-only experimental modal analysis, we usually work with a sampled field matrix X ∈ Cm×n , where Xij = f (xj , ti )
for (i = 1, . . . , m and j = 1, . . . , n) is the measurement of the field variable at spatial point xj and temporal time ti . We are
interested in variations about the mean-field and, therefore, we assume that the columns of the matrix X are zero mean or we
subtract the mean from each column. Thus, we are seeking decompositions in the form

n

X = U VH = σk uk vkH , such that uk vkH = 1 , (39.1)
k=1
where superscript (·)H indicates a Hermitian or conjugate transpose of (·). In addition, for m ≥ n and if X has full column
rank, U = [u1 · · · un ] ∈ Cm×n is composed of unit vectors ( ui = 1), V = [v1 · · · vn ] ∈ Cn×n is composed of unit vectors
( vi = 1), and = diag(σi ) ∈ Cn×n . Alternatively, for m ≤ n and if X has full row rank, U = [u1 · · · um ] ∈ Cm×m
and V = [v1 · · · vn ] ∈ Cn×m are composed of unit vectors, and = diag(σi ) ∈ Cm×m . If matrix X is rank deficient or
rank(X) < min(m, n), we can still find its decomposition in r-dimensional subspace, where
r = min (m, n, rank(X)) , (39.2)
where U ∈ Cm×r , V ∈ Cn×r , and ∈ Cr×r .

For a discrete set of scalar field matrices {Xi }N
i=1 , decompositions are usually the solution to some form of the
(generalized) eigenvalue problem involving cross-covariance matrices
Kij XiH Xj or Pij Xi XjH , (39.3)
where (·)H indicates Hermitian or conjugate transpose of (·), Xi ∈ Cmi ×n for Kij , and Xi ∈ Cm×ni for Pij . If we only
consider Hermitian autocovariance matrices (i.e., i = j ), then decompositions are based on diagonalizing either a singular
autocovariance matrix (e.g., singular value decomposition), a simultaneous decomposition of two autocovariance matrices
(e.g., generalized singular value decomposition), or simultaneous diagonalization of a set of autocovariance matrices (e.g.,
higher-order generalized singular value decomposition).
39.3 Higher-Order Generalized Singular Value Decomposition
For a discrete set of scalar field matrices {Xi }N

i=1 , higher-order generalized singular value decomposition (HO-GSVD)
Xi = Ui iV
H
, (39.4)
N
where Xi ∈ Cmi ×n i=1 (mi > n), Ui ∈ Cmi ×n are unitary, i ∈ Rn×n are diagonal, and V ∈ Cn×n is a full rank matrix of
unit vectors. Therefore, we can express our Hermitian autocovariance matrices as
Kii = V H
i Ui Ui iV
H
=V 2 H
iV (39.5)
Thus,
−1 −2 −1 −1 −2
Kii Kjj =V 2
i j V and Kjj Kii = V −H j
2 H
iV , (39.6)
which allows us to formulate the corresponding generalized eigenvalue problems in the matrix form as
−1
Kjj V = Kii−1 V 2ij and Kii W = Kjj W 2ij , (39.7)
where
39 Higher-Order Decompositions for Modal Identification and Model Order Reduction 273
−2
W = V −H and 2ij = 2
i j = diag(λ2ij,k ) . (39.8)
Therefore, we can consider the following Hermitian eigenvalue problem

−1
Kii Kjj vk = λ2ij,k vk , (39.9)
or the following Hermitian generalized eigenvalue problem in terms of adjoint basis to vi
Kii wk = λ2ij,k Kjj wk , (39.10)
or in a matrix form
Kii W = Kjj W 2ij . (39.11)
Once we know W , we can get modes using V = W −H . Since Eq. (39.9) should be true for all i and j , we sum Eq. (39.9)
over all i and j = i to get
Sv = λv (39.12)
where

N N
−1
1 −1
S= Kii Kjj + Kjj Kii−1 ,
N(N − 1)
i=1 j =i+1
(39.13)

N N
−1
1
λ= λ2ij + λ2j i .
N(N − 1)
i=1 j =i+1
Then, after making all the columns vi in V unitary, we can determine the other components by defining:

Mi mi,1 , · · · , mi,n = Xi V −H , (39.14)
and by letting

i = diag mi,k ⇒ Ui = M i −1
i . (39.15)
39.4 Higher-Order Non-hermitian Decomposition

N
For a set of matrices Xi ∈ Cmi ×n i=1 (where m1 = max(mi ) with all mi > n) higher-order non-Hermitian decomposition
(HO-NHD) decomposes them into
Xi = !i "H , (39.16)
where " = [φ1 , φ2 , · · · , φn ] ∈ Cn×n are right basis vectors or mode shapes, !i = [ξ1i , ξ2i , · · · , ξni ] ∈ Cmi ×n the
corresponding time coordinates for each data matrix. For HO-NHD we consider non-Hermitian cross-covariance matrices
for i = k
Kik = XiH Xk . (39.17)
However, if mk = mi , we cannot do this calculation directly. Therefore, we let mik = min (mi , mk ) and truncate each
data matrix in the pair to the same number of mik rows to get X̂i ∈ Cmik ×n and X̂k ∈ Cmik ×n and obtain
274 D. Chelidze
ˆH
Kik = X̂iH X̂k = "! ˆ H
i !k " . (39.18)
The basic assumption HO-NHD is stated as
î =!
! ˆ i , (39.19)
where i ∈ Cn×n is a diagonal matrix of some characteristic values and ! represents some reference time coordinates. Now,
assuming each of the cross-covariances has full rank, we derive the following expression
−1
Kik Kj−1 ˆ H ˆ H "!
k = "! i ! k "
ˆH ˆ H
j !k "
−1
−1 −1 −1
=" i ! ˆ
ˆ H! k ! ˆ
ˆ H! j "
(39.20)
k
−1 −1
=" i j " .
Therefore
Kik Kj−1
k = "ij "
−1
, (39.21)
where
−1
ij = i j . (39.22)
This Eq. (39.21) can also be written as the equivalent non-Hermitian eigenvalue problem
Kik Kj−1
k φ = λij φ , (39.23)
or the corresponding non-Hermitian generalized eigenvalue problem in terms of the adjoint eigenvectors
Kki ψ = λij Kkj ψ , (39.24)
which gives the corresponding adjoint eigenvalue problem:

−1
Kkj Kki ψ = λij ψ . (39.25)
Example 1 (State-Variable Decomposition for Modal Identification)

State-Variable Decomposition (S-VD) was proposed to identify complex vibration modes, and the associated damping ratios
and frequencies, of the state-variable model of a multi-degree-of-freedom linear system. In this approach, we obtain the
S-VD by solving the following asymmetric generalized eigenvalue problem
KXẊ ψ̃j = λ̃j KXX ψ̃j (39.26)
or its adjoint problem
KẊX φ̃j = λ̃j KXX φ̃j , (39.27)
where KXX is an autocovariance matrix of a free-vibration time series of state-variables in X, and KẊX = KXT Ẋ is
the corresponding cross-covariance matrix between the time derivative Ẋ and X. Then, the complex
vibration modes are
approximated by the columns of " ˜ and the corresponding eigenvalues λ̃ approximate −ζj ωj ± iωj 1 − ζj , where ζj and
ωj are the damping ratios and the natural frequencies of the j -the mode.
Example 2 (Dynamic Mode Decomposition)
Dynamic mode decomposition (DMD) works with snapshots of the fluid-flow vector field X1 = [x1 , x2 , · · · , xm ] ∈ Cn×m
and its image one-time-step later X2 = [x2 , x3 , · · · , xm+1 ] ∈ Cn×m to identify a best linear mapping between them
X2 = AX1 , (39.28)
where spatial information is in columns and temporal information is in rows. DMD aims to identify the proxy to a generally
nonlinear dynamical system
ẋ = f (x, t) (39.29)
by finding the approximate locally linear dynamical system
ẋ = Gx (39.30)
similar to the one for the damped linear vibrations. Then the solution to this approximate equation can be written as
xi = x(ti ) = "eti ξ0 , (39.31)
where " ∈ Cn×r and = diag(λi ) ∈ Cr×r are the r dominant eigenvectors and eigenvalues of the matrix G and ξ0 contains
the initial amplitudes of each mode or "ξ0 = x1 . Therefore, we can write that
xi+1 = "eti+1 ξ0 = "e(ti + t)

ξ0 = "e t eti ξ0 = "eti ξ0 , (39.32)
where = e t . Thus, we can also write
xi+1 = ""−1 "eti ξ0 = ""−1 xi = Axi , (39.33)
where
A = ""−1 (39.34)
provides the eigendecomposition to matrix A with eigenvectors φi and the corresponding eigenvalies ωi = eλi t. Therefore,
in DMD, the basic assumption is that
X1 = "! (39.35)
Then considering Eq. (39.32) we can also write
X2 = "! (39.36)
Plugging this back into Eq. (39.28) gives
"! = A"! , (39.37)
which results into
A" = " , (39.38)
which is the eigen-decomposition of A steming from the corresponding eigenvalue problem
Aφi = ωi φi . (39.39)
Therefore, to find the dynamical modes, we need to find the eigenvalues of A.

In DMD scenarios, it is common to have n m and usual pseudo-inverse based least-squares solution to finding the
matrix A
A = X2 X1† , (39.40)
276 D. Chelidze
where X1† = X1H (X1 X1H )−1 , is not possible since the rank of the matrix X1 is at most m. Therefore, we have to look for a
regularized solution to find the eigenvectors of the A matrix. The standard approach is to use the r-rank TSVD of the matrix
X1
X1 ≈ Ũ ˜ Ṽ H ⇒ A ≈ X2 Ṽ ˜ −1 Ũ H , (39.41)
where r ≤ min (rank(X1 ), m), to get an (r × r)-dimensional similar matrix to A
Ã = Ũ H AŨ = Ũ H X2 Ṽ ˜ −1 . (39.42)
Then using the eigendecomposition of Ã,
ÃW = W , (39.43)
one can reconstruct the needed DMD modes as
" = X2 Ṽ ˜ −1 W −1 . (39.44)
We advocate the alternative approach based on GSVD of the matrix pair X1 and X2 . Using Eq. (39.31) we can write
X1 = "!H
1 , (39.45)
where the matrix !1 contains time coordinates

t
1 = e
!H 1 ξ , e t2 ξ , · · · , e tm ξ (39.46)
0 0 0 .
We can also write that
X2 = "!H
2 , (39.47)
where
t
2 = e
!H 2 ξ , e t3 ξ , · · · , e tm+1 ξ
0 0 0

= e(t1 + t) ξ0 , e(t2 + t) ξ0 , · · · , e(tm + t) ξ0 (39.48)

= e t et1 ξ0 , e t ξ0 , · · · , etm ξ0 = e t !H 1 .
Therefore, we get that
X2 = "e t !H
1 . (39.49)
Now, we can use the r-rank trancated GSVD of the matrix pair X1 and X2
X1 ≈ Ũ1 ˜ 1 Ṽ H and X2 ≈ Ũ2 ˜ 2 Ṽ H (39.50)
to approximate A as
A ≈ Ũ2 ˜ 2 ˜ 1−1 Ũ1H (39.51)
and use it to obtain an (r × r)-dimensional matrix similar to A and the corresponding eigendecomposition
Ã1 = Ũ1H AŨ1 = Ũ1H U2 ˜ 2 ˜ 1−1 = W1 W1−1 . (39.52)
Then
AŨ1 W1 = Ũ1 W1 (39.53)
and, thus, the corresponding DMD modes are
" ≈ Ũ1 W1 . (39.54)
We can also get another (r × r)-dimensional matrix similar to A and the corresponding eigendecomposition
Ã2 = Ũ2H AŨ2 = ˜ 2 ˜ 1−1 Ũ1H Ũ2 = W2 W2−1 . (39.55)
Then
AŨ2 W2 = Ũ2 W2 (39.56)
and the corresponding DMD modes are
" ≈ Ũ2 W2 . (39.57)
Finally, we can take the arithmetic mean to get a better estimates of the DMD modes
1
"= Ũ1 W1 + Ũ2 W2 . (39.58)
2
39.5 Results
39.5.1 Free Linear Response of the Damped Cantilever Beam
As a first example, we consider a damped free vibration of a cantilever beam. In particular, we will use 10 mode
approximation to the transverse displacement of a cantilever beam

10
w(x, t) ≈ ξi (t)φi (x) (39.59)
i=1
where the vibration modes are given by
sin βi + sinh βi
φi (x) = sin βi x − sinh βi x − (cos βi x − coshβi x) , (39.60)
cos βi − coshβi
where, assuming x ∈ [0, 1], first ten β = [1.8751, 4.69409, 7.85476, 10.9955, 14.1372, 17.2788, 20.4204, 23.5619, 26.7035,
29.8451] are obtained from the transcendental characteristic equation for the corresponding eigenvalue problem. We
randomly assigned the damping ratios to each mode as ζ = [0.0395, 0.1710, 0.1018, 0.0781, 0.1014, 0.0739, 0.0763,
0.1247, 0.1223, 0.1225]. We sampled the response at 10 spatial points evenly distributed along the length of the beam with
the corresponding modal coordinates
)
ξi = ai e−ζi ωi t cos ωd,i t with ωi = βi2 and ωd,i = ωi 1 − ζi2 , (39.61)
where coefficients ai are obtained by assigning zero initial velocity ẇ(0) = 0 with w(0) = [0.05, 0.05, 0.05, 0.05, 0.1,
0.12, 0.25, 0.5, 1, 2] and time was varied from 0 to 20 s using 0.001 s sampling period.
The results of the modal identification are shown in Fig. 39.1. While other methods are capable to identify these modes,
HO-NHD with just one transient free response data can identify not only the mode shapes and modal coordinates but also
the corresponding frequencies and damping ratios.
278 D. Chelidze
2 2 0.02
0.05 1000
1
2
1 0 0
0
0 -2 -0.02 -0.05 800
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 -0.02 0 0.02 0 0.06
2 2
600
1 0.05 0.1
3
4
0 0 400
0 0
-0.05
-1 -2 -0.1 -0.1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 -0.1 0 0.1 0 200 actual
estimated
2 2 0.2 0
1 0.1 0 500 1000
5
6
0 0 0 0
-0.1
-1
-2 -0.2 -0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 -0.2 0 -0.2 0 0.2 0.2
2 2 0.2
1 0.1
0
7
0 0 0.15
0 -0.1 -0.2
-1 -0.2 -0.4
-2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 -0.2 0 -0.2 0 0.2
0.1
2 2
0.2 0.2
0 0 actual
10
9
0 0
-0.2 -0.2 estimated
0.05
-0.4 -0.4
-2 -2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0 0.05 0.1 0.15 0.2
Fig. 39.1 Identified modes (circles) compared to the real modes (lines) (left plot), identified time coordinates (center plots), and identified modal
parameters vs real modal parameters (right plots) for the damped cantilevered beam
1 1
1 1 10 0 20
0
0.5 0.5 -5
0.5 0.5 0 -10 10
0 0 -10
-10
0 0 -20 0 -15
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 0 10 20 30 0 10 20 0 10 20 0 5 10
1 1 1 1 15 0
2 6
0.5 0.5 0.5 10 -2
0.5 0
4
0 0
5 -4
2
0 0 0
-2 -6
0
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 0 5 10 0 2 4 0 2 4 0 2 4
1 1 1 1 3 0 0.01 0.01
0.5 2 -1
0.5 0.5 0.5 0 0
1 -2
0 0 0 0 0 -3 -0.01 -0.01
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 -1 0 1 2 -1 0 1 2 -0.01 0 0.01 -0.01 0 0.01
1 1 1 1 0.02
0.01 0.02
0.01
0.5 0.5 0.5 0.5 0
0 0
0
-0.01 -0.02
0 0 0 0 -0.01
-0.02
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 0 0 -0.02 0 0.02 -0.02 0 0.02
1 1 1 1 0.05 100
0.05 100
0.5 0.5 0 0 0 0
0 0
-100
0 0 -1 -1 -0.05
-0.05 -100
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 -0.05 0 0.05 -0.05 0 0.05 -100 0 100 -100 0 100
Fig. 39.2 Identified modes (left plot) and the corresponding time coordinates (left plots) from the free response of the double-well damped
cantilevered beam
39.5.2 Free Nonlinear Response of the Damped Cantilever Beam
Now, we consider the same cantilever beam, but we attach a nonlinear spring at the free end. Here simulation is conduced
using integration of ordinary differential equations corresponding the described system. Time was varied from 0 to 10 s using
0.001 s sampling period. The results of the modal identification are shown in Fig. 39.2.
Chapter 40
Utilizing Modal Testing for Monitoring the Structural Health
of Wind Tunnel Facility Hardware
Kenneth Pederson and Vicente Suarez
Abstract The 10- by 10-Foot Abe Silverstein Supersonic Wind Tunnel (10 × 10) is the largest and fastest wind tunnel
facility at NASA’s Glenn Research Center(GRC) and is specifically designed to test supersonic propulsion components from
inlets and nozzles to full-scale jet and rocket engines (10 × 10 Abe Silverstein Supersonic Wind Tunnel, https://www1.grc.
nasa.gov/facilities/10x10/). Recently, a critical part of the wind tunnel failed and required a redesign before reintegrating
into the facility. The design requirements of this new component required that clearances between large metallic components
exist, which have the potential for undesirable nonlinear dynamics to occur, in particular rattling. Rattling is feared to occur
when the wind tunnel is being operated in certain flow regimes that induce cyclic aero loads on the new component near
its natural frequencies. This paper describes the approach taken to better understand and resolve this vibration problem
using modal testing. A modal test was developed and executed by GRC’s Structural Dynamics Lab to quantify the modal
parameters of the structure, namely which specific excitation frequencies caused the structure to rattle. These results were
shared with facility operators as frequency ranges that should be avoided to ensure maximum lifespan of the new structure.
Additional means of structural health monitoring (SHM) as well as Vortex shedding are briefly discussed in this paper.
Keywords Non-linear dynamics · Rattle · Structural health monitoring · Modal test · Aerospace
40.1 Introduction
Modal testing and analysis has been and continues to be used extensively in practice as a means to check the structural health
of a system. Sources such as Hsieh (2006) [1] and Kim (1993) [2] provide excellent literature reviews as well as get into the
history and methodology of structural health monitoring utilizing modal data.
The motivation for this project came when the 10- by 10-Foot Abe Silverstein Supersonic Wind Tunnel (10 × 10) at
NASA Glenn Research Center began redesigning a cooling unit with the hopes of increasing its lifespan. The cooler is used
to reduce the temperature of post combustion gases coming from various engines running in the wind tunnel. It is comprised
of an array of thin walled tubes supported at various axial locations and can be seen in better detail below in Fig. 40.1. In the
previous designs, these tube supports were press fit between the tube and a bore through a piece of sheet metal. However,
the new design would have a small clearance between the tubes and the bores to allow thermal growth of the tube to not be
impeded by the supports or vice versa. A simplified schematic of one of these tube and bore supports can be seen below in
Fig. 40.2. Although these clearances alleviated a potential thermal expansion issue, it introduced a new failure mode. The
clearances between the tubes and their bores now allowed for rattling to occur within the support which could lead to cracks
within the support structure, damage to individual tubes, and contribute to accelerated fatigue of critical components.
This paper serves predominately to describe a modal test done on this new cooler design to characterize the structure and
predict what types of excitations induced rattling between the tubes and their bores. The analysis portion of this paper
describes how select structural health monitoring techniques such as harmonic distortion are used in conjunction with
Campbell diagrams to identify nonlinear behavior in the test data. Lastly, the conclusions section provides a summary of
the guidance given to the wind tunnel operators in light of the results from the analysis section and includes some future
work topics to increase the effectiveness of this study.
K. Pederson () · V. Suarez

NASA Glenn Research Center, Cleveland, OH, USA

280 K. Pederson and V. Suarez
Fig. 40.1 Test article setup
Fig. 40.2 Tube in bore schematic
40.1.1 Failure Mode
Rattling is a well-documented phenomena in modal and vibration testing and analysis. The non-linear nature of rattle and
how it can be applied to structural health monitoring is discussed in detail in Worden (2007) [3]. This article describes non-
stationary and nonlinear behavior such as rattling as a failure mode in itself. This cooler is still operable when the tubes
rattle, but the concern is this rattle will lead to a severe failure as described above. Nonlinear system identification methods
are described in detail such as Harmonic distortion where a structure is evaluated by applying a constant frequency sine
wave and the response is measured for higher harmonics. This methodology is applied in the analysis portion of this paper.
40 Utilizing Modal Testing for Monitoring the Structural Health of Wind Tunnel Facility Hardware 281
Worden mentions that the primary usage of structural health monitoring in practice is with rotating machinery so many of
the techniques that are describe require slight modification to be applied to modal testing and aero-structures in general.
During operation, fluid flow over these tubes will cause vortices to be shed around each tube in a phenomena known as
vortex shedding. Vortex shedding is a well-studied phenomena and background can be found in a multitude of fluid mechanics
textbooks. In 1984 P.W. Bearman wrote a literature review called “Vortex Shedding From Oscillating Bluff Bodies” [4] which
is an excellent overview of many studies on vortex induced oscillations similar to the structure being tested in this paper.
Bearman describes how a flexible body enduring vortex shedding will begin to oscillate due to the fluctuating pressure fields
and can have an amplitude as large as 1.5× to 2× the body’s diameter. Bearman also mentions the oscillations induced by
vortex shedding routinely are more severe than those induced by turbulent boundary layers. For a further and more complete
discussion on vortex-induced oscillations, a 1977 book by Blevins titled “Flow-Induced Vibration” is a good reference.
40.2 Modal Test
A modal characterization of the structure was proposed to identify what frequencies would induce rattling. These frequencies
in which major tube rattling was induced would better inform the fluid dynamics team and allow them the ability to identify
what tunnel parameters could potentially induce rattling so they could be avoided.
The cooler being tested is shown in Fig. 40.1.
The test article is comprised of 168 tubes running from the “front side” to the “back side” and a rack that supports the
beams at various axial locations. The rack consists of an outer box with intermediary “plates” that are normal to the tubes.
The plates are metal sheets with bored out clearance holes for each tube and a three of the seven plates are labeled in the
photo as “Plate 1, Plate 4, and Plate 7”. “Stations” are used to describe the region between individual plates. The outer sides
of the cooler are defined in the above picture as “FRONT”, “BACK”, “LEFT”, and “RIGHT”. The basic operation of the
cooler involves cold water flowing through the tubes along the Z-direction as a means to cool down hot post-combustion
gases flowing through the structure in the Y-direction.
A challenge in this test, as in most modal tests, is that replicating the “in operation” boundary conditions of the test article
would be nearly impossible to replicate in this modal test. In this test, mimicking the boundary conditions was deemed
unfeasible To remedy this, it was important that the boundary conditions were simple and well understood to aid in future
analytical work. To accomplish this, air bags were placed at each of the four corners of the structure. These isolate the
structure from the ground induced vibration as well as ensure the rigid body modes are low enough in frequency as to not
couple with the flexible body modes that were relevant to the test.
Figure 40.2 below shows a breakout schematic of one of the tube in bore supports. This is one example of the “tube
and bore” support of which there are 1176 of throughout the cooler split evenly between the 7 plates. The clearance gap is
exaggerated and the tube, bore, and plate thickness dimensions are not to scale.
The test included measuring the force input to the structure through load cells in line with the modal hammers and modal
shakers while measuring the response with 35 accelerometers and two microphones. Response Accelerometers were placed
around the structure using engineering judgement to capture the global modes of the rack. In addition, select supports were
instrumented to capture local rattling of the tube within the bore. An example of this type of accelerometer placement is
shown below in Fig. 40.3.
Two modal shakers were used to excite the structure and can be seen below in Fig. 40.4. The shakers were placed on
pedestals to position them as close as possible to minimize the length of the shaker stinger. The stingers were attached to
custom aluminum flexures that reduce any potential moments due to shaker misalignment to be input into the force sensors.
A multitude of excitation techniques were used to characterize the structure including impact with modal hammers and
random, burst random, and sine sweeps using the modal shakers. The impact and random data was found to be effective at
extracting global modes of the racks. The coherence of the random data was very poor due to the rattling and for this reason
was not effective at identifying rattling. As will be discussed in more detail in the analysis portion the sine sweeps response
data exhibited the rattling clearly.
40.3 Analysis and Results
As vortices are shed around the tube, the fluctuating pressure fields will cause the tubes to oscillate at a specific frequency.
This frequency can be estimated from the fluid properties and flow parameters using the Strouhal number. The Strouhal
Fig. 40.3 Select supports instrumented both on and adjacent to a single tube
Fig. 40.4 Modal shakers positioned on the test article
number is proportional to the vortex shedding frequency and the characteristic length of the body, and is inversely
proportional to the fluid velocity. In this case, the characteristic length of the body is the diameter of the tube. The Strouhal
number is a function of the Reynolds number; however, can be assumed to be about .22 for many flow regimes. Calculating
vortex shedding frequencies from Strouhal numbers has been thoroughly studied for a vast array of flow regimes and is not
trivial for as complex as flow regimes as this cooler will see. For this reason, a more detailed discussion is not included in
this paper. An expansive literature review of these methods can be found in an article written in 1990 by H Sakamoto titled
“A Study on Vortex Shedding From Spheres in a Uniform Flow” [5]. The problem presented by this cooler is significantly
more complex and challenging that a simple cylinder in a fluid field. Compounding the complexity of the free-stream flow
parameters of this cooler is the fact that many of the tubes will not see the uniform free-stream flow. Most closed form
calculations assume a steady uniform flow field impacting a cylinder, however due to the closely packed nature of these
tubes there will be flow interactions between the tubes increasing the complexity of the vortices. In addition, the tubes that
Fig. 40.5 Campbell diagram of accelerometer on tube
are not on the intake face of the cooler will see a non-uniform flow distribution comprised of the vortices being shed by the
tubes in front of them. This paper does not get into how these more complex fluid regimes will affect the problem.
In a very simplistic view of this problem the wind tunnel operators can use the vortex shedding analysis in conjunction
with the modal test results and identify which flow regimes will cause rattling to occur.
As the tubes begin to oscillate, they will begin to rock within the bore. The amplitude of the rocking will be a function
of two main parameters, excitation frequency and amplitude. As the vortex shedding frequency approaches the structures
natural frequencies this rocking will couple with the tube’s natural modes and be amplified by the resonance. The excitation
amplitude will also directly increase the amplitude of the rocking. This phenomena behaves similar to an airplane wing
undergoing flutter where a bending mode frequency lies near the frequency of vortices being shed around it. When the
amplitude of the rocking increases past a tipping point the tube actually becomes detached from the bore and it will begin to
rattle.
The most promising data came from the sine sweeps. These sine sweep rates were set slow enough to fully develop
modes and acquire sufficient data in relevant frequency bands to accurately extract modal parameters. The figures below
show results from sine sweep tests from 1500 Hz to 50 Hz at 1 octave per minute. Campbell diagrams were chosen to plot
Fig. 40.6 Campbell diagram of accelerometer on support
the data below due to their keen ability to identify nonlinear system dynamics. The first plot is of an accelerometer on an
actual tube and the second plot is an accelerometer on the support immediately adjacent to the tube as shown above in Fig.
40.3. These accelerometers were chosen due to their proximity to the rattling events. Harmonic distortion is easily identified
in these diagrams as lines of high g-levels parallel to the input excitation shifted an exact frequency multiple of the excitation
frequency to the right or left (Fig. 40.5).
The vertical axis is time in the sweep, the horizontal axis is response frequency, and the color code is normalized g levels
of response. There are two distinct features on these plots, the diagonal line of high amplitude from top left to bottom right
and the horizontal lines of varying thickness. The diagonal line of high amplitude is due to the frequency of the shaker and
the lines parallel to it are its accompanying sub and super harmonics. At the beginning of the sweep when time is zero the
shaker excitation line is at a frequency value equal to 1500 Hz and at the end of the sweep, about 300 seconds, the line is at
a frequency value of 50 Hz. The horizontal lines with high amplitude indicate a broadband response at a specific excitation
frequency and are an indication that rattling has occurred. The larger the amplitude of these broadband responses are the
more severe the rattling and higher potential for damage. The intersection of the horizontal and diagonal lines represent the
frequencies in which rattling is induced. Figure 40.6 is reshown below in Fig. 40.7 with these intersections identified. The
Fig. 40.7 Rattling frequencies
accompanying frequencies of each intersection are listed in Table 40.1 and are frequencies to be avoided when operating the
wind tunnel.
40.4 Conclusion
Accelerometers on the tubes as well as on the supports directly adjacent to the tubes exhibited instantaneous broadband
response over about 150 Hz – 3 kHz at single shaker excitation frequencies. A linear structure would only exhibit a single
frequency response at the same frequency of the excitation. It is possible for structures to exhibit non-linear phenomena such
Table 40.1 Rattling frequencies Label Frequencies

as noted in Fig. 40.7
F1 1217.465206
F2 949.728238
F3 759.5086677
F4 670.7794989
F5 473.7961395
F6 328.0537758
F7 259.7556309
F8 174.5776203
as super and sub harmonics of the input excitation without directly causing rattle. This broadband response, in addition to
test engineers being able to audibly pick up high pitched rattling during the sine sweeps, confirms that rattling will occur in
this cooler design at specific excitation frequencies. The frequencies listed in Table 40.1 should be used in conjunction with
the vortex shedding analysis suggested in the analysis portion of this paper to advise wind tunnel operators on what flow
regimes should be avoided during tunnel operation.
This work also demonstrated a potential means of in operation structural health monitoring for this cooler system. The
acceleration response of the structure proved to be an excellent indicator of rattling and could be transitioned from a modal
test environment to a wind tunnel environment. This would require ruggedized accelerometers capable of resisting high
temperatures and drag loads. Future work could be performed in developing a real time data acquisition approach to acquire
vibration data from the cooler during operation. Microphones were used to collect response data during the modal to allow
wind tunnel operators to install microphones during operation and have pre-test baseline data to compare to. This would
allow facility operators to perform in-situ modal parameter extractions and identify if rattling is occurring during tunnel
operation. It would also allow for significant changes in the natural frequencies of the cooler to be detected over time, which
could infer structural degradation.
References
1. 10×10 Abe Silverstein Supersonic Wind Tunnel, https://www1.grc.nasa.gov/facilities/10x10/

2. Hsieh, K., Halling, M., & Barr, P. (2006). Overview of vibrational structural health monitoring with representative case studies . J. Bridg. Eng.,
13(3), 304–304. doi: https://doi.org/10.1061/(asce)1084-0702(2008)13:3(304.2)
3. Kim, H., & Bartkowicz, T. (1993). Damage Detection and Health Monitoring of Large Space Structures. In AIAA
4. Worden, K., Farrar, C., Haywood, J., Todd, M.: A review of nonlinear dynamics applications to structural health monitoring. Struct. Control.
Health Monit. 15(4), 540–567 (2008). https://doi.org/10.1002/stc.215
5. Bearman, P.: Vortex shedding from oscillating bluff bodies. Annu. Rev. Fluid Mech. 16(1), 195–222 (1984). https://doi.org/10.1146/
annurev.fluid.16.1.195
Chapter 41
An Efficient Coupled Modal Quasi-static Approach for
Characterizing Non-linear Modal Properties of Prestressed
Structures
Nidish Narayanaa Balaji and Matthew R. W. Brake
Abstract Linear modal analysis is a powerful tool in studying linear dynamical systems with several Degrees-of-Freedom
(DoFs). There has been an increasing interest in how this can be extended to large (in the sense of number of DoFs) non-linear
dynamical systems. The current study proposes an extension to the stationarity of Rayleigh quotients, a classical technique
for linear modal analysis, and demonstrates its applicability to conservative and non-conservative non-linear systems. Apart
from offering a theoretical motivation of modal analysis in non-linear dynamics, the approach also circumvents several
limitations in previous quasi-static non-linear modal analysis methods. The method is demonstrated on a simplified model of
a bolted-joint which includes unilateral springs and elastic dry friction elements describing the non-linearities. The results are
compared with the Extended Periodic Motion Concept (EPMC), a frequency domain approach based on periodic solutions.
Keywords Nonlinear modal analysis · Hysteretic systems · Quasi-static modeling · Rayleigh quotient · Courant-Fischer
theorem
41.1 Introduction
The current paper presents a Nonlinear Modal Analysis (NMA) technique that is based on a nonlinear extension of the
concept of Rayleigh Quotients in linear dynamics [6]. In matrix analysis, which forms the basis of linear modal analysis, the
Courant-Fischer theorem (CF theorem) is a very important result that allows for the reformulation of the linear eigenvalue
problem as a constrained optimization problem with the Rayleigh quotients as the objective function. The results from the
theorem (presented here for the case of distinct eigenvalues) are two-fold:
(a) Each eigenpair formed by an eigenvalue and its eigenvector are local minimizers of the Rayleigh quotient when some
norm of the eigenvector is constrained to be a constant; and
(b) An eigenpair formed by an eigenvalue and its eigenvector is the global minimizer of the Rayleigh quotient when some
norm of the eigenvector is constrained to be a constant, and the eigenvector is constrained to be orthogonal to all the
eigenvectors with eigenvalues below its own eigenvalue.
Since the Rayleigh quotient is quadratic in the eigenvector, the minima in the above are well-defined and it is relatively easy
to obtain them numerically. In the context of linear dynamics, this forms one definition for the modes of undamped multi-
Degree-of-Freedom (DoF) second order dynamical systems. However, there have been recent studies [5] that this property
holds, at least in a weak sense, for damped systems. Although the results are not directly applicable for all problems, such
works have motivated the extension that is explored here.
The developed formulation, described in more detail in Balaji and Brake [1], is shown to be applicable for a broad range
of problems and a specific example with the model of a pre-stressed bolted joint is presented for illustration. In comparing
the technique with existing frequency-domain [3] and quasi-static [4] methodologies, a relative qualitative assessment of the
method is derived.
N. N. Balaji () · M. R. W. Brake


288 N. N. Balaji and M. R. W. Brake
41.2 Theoretical Formulation
The formulation is based upon two main observations on the Rayleigh quotients:
(a) the directional derivatives of the Rayleigh quotients along virtual displacements yield corresponding virtual work
contributions. Mathematically, denoting the directional derivative by Dw (with w being some displacement vector),
+ +
and the Rayleigh quotient (RK (w )), for some stiffness matrix K, by 12 w T Kw , the directional derivative is expressed as
+ + +
1 T
Dw RK (u) = Dw u Ku = w T Ku. (41.1)
+ + + 2+ + + +
(b) the first order optimality conditions of the extremization problem, in Rn space,
1 T
min u Ku
u∈Rn + +
+ ,2 -. /
RK (u)
+
1 T 1
subject to, u Mu = q 2 (41.2)
+ +
,2 -. / 2
RM (u)
+
is given, albeit in an alternate form, using a Lagrange multiplier λ as the set of equations,
w T (K − λM) u = 0 ∀ w ∈ Rn
+ + +
uT Mu − q 2 = 0. (41.3)
+ +
Note that RM (u) is a norm in Rn since M is taken to be positive definite. Here, if w may be interpreted as virtual
+ +
displacements, this is exactly the linear eigenvalue problem for the dynamical system
Mü + Ku = 0. (41.4)
+ + +
For non-linear problems, although neither Rayleigh quotients nor energy scalars may be defined, the “weak form” always
exists and may be interpreted as the first order optimality conditions for some optimization problem. For a dynamical system
of the form,
Mü + Ku + fnl (u) = fstatic + fdyn (t), (41.5)

+ + + + + +
with static solution us satisfying

+
Kus + fnl (us ) = fstatic , (41.6)

+ + + +
the extension of the weak form optimality conditions, using just perturbations from the static solution for the norm condition
corresponds to the non-linear static problem given by
Ku + fnl (u) − λM(u − us ) − fstatic = 0

+ + + + + +
(u − us ) M(u − us ) − q 2 = 0.
T
(41.7)
+ + + +
Here, q may be taken to be “modal amplitudes”, and this system may be solved for different modal amplitudes to obtain
a quasi-static procedure for the modal characterization of the system (Eq. (41.5)). This procedure, referred to as Rayleigh
41 An Efficient Coupled Modal Quasi-static Approach for Characterizing Non-linear Modal Properties of Prestressed Structures 289
Quotient-based Nonlinear Modal Analysis (RQNMA) henceforth, yields non-linear amplitude-dependent modes and mode-
shapes for the underlying system.
41.3 Application Example: Bolted Joint
The current section presents an application of the developed method for an idealized model of the Brake-Reuß Beam (BRB),
a three-bolted lap-joint benchmark [2]. The model is developed using Euler-Bernoulli beam finite elements and a liner
idealization of the bolts in the interface. Unilateral springs and planar elastic dry-friction elements are employed to model in
the contact interface. More details may be found in [1] (Fig. 41.1).
Non-linear modal analysis is conducted using RQNMA and the Extended Periodic Motion Concept [3] (EPMC, which is
a frequency-domain NMA procedure).
Figure 41.2 presents the results of the non-linear modal analysis. As can be observed, the developed approach follows
the peaks on the frequency-response plot closely (the non-linearity is rather exaggerated since bolt-pinning effects are not
considered here). Further, the approach also allows one to obtain amplitude-dependent mode-shapes explicitly, which is
something that is not possible with other quasi-static approaches (see QSMA [4], for instance) consistently.
41.4 Conclusion
The paper presents the formulation and application of a novel non-linear modal analysis procedure based on a non-linear
extension of key results from linear modal analysis theory. The approach is demonstrated to be applicable to prestressed
structures by validating it with non-linear forced responses as well as a frequency-domain NMA technique. Furthermore, the
fact that the method is directly applicable to hysteretic structures implies that the potential for its applicability extends much
beyond frictional problems, to general plastic structures too.
Fig. 41.1 Bolted joint idealization. (Figure from [1])

290 N. N. Balaji and M. R. W. Brake
10 -3
10 -4
10 -5
10 -6
50 100 150 200 250

(b)
(a)
Fig. 41.2 Nonlinear Modal Analysis results for mode 2 of the BRB idealization. (Figures from [1]): (a) the frequency response of the system with
the RQNM and EPMC backbones; and (b) comparison of low- and high-amplitude mode-shapes
References
1. Balaji, N.N., Brake, M.R.W.: A quasi-static non-linear modal analysis procedure extending Rayleigh quotient stationarity for non-conservative
dynamical systems. Comput. Struct. 230, 106184 (2020)
2. Brake, M.R.W. (ed.): The Mechanics of Jointed Structures. Springer, Cham (2017)
3. Krack, M.: Nonlinear modal analysis of nonconservative systems: extension of the periodic motion concept. Comput. Struct. 154, 59–71 (2015).
ISSN 00457949. https://doi.org/10.1016/j.compstruc.2015.03.008. http://linkinghub.elsevier.com/retrieve/pii/S0045794915000978
Process. 118, 133–157 (2019). ISSN 08883270. https://doi.org/10.1016/j.ymssp.2018.08.034
5. Phani, A.S., Adhikari, S.: Rayleigh quotient and dissipative systems. J. Appl. Mech. 75(6), 061005–061005–6 (2008). ISSN 0021-8936. https://
doi.org/10.1115/1.2910898
6. Rao, S.S.: Mechanical vibrations in SI Units. Pearson Education Limited (2017). ISBN 978-1-292-17861-5. https://books.google.com/books?
id=REQ4DwAAQBAJ
Chapter 42
An Assessment of the Applicability and Epistemic Uncertainties
Inherent to Different Classes of Friction Models for Modeling
Bolted Interfaces
Justin H. Porter, Clayton R. Little, Nidish Narayanaa Balaji, and Matthew R. W. Brake
Abstract Modeling the contact in bolted structures is a persisting challenge in the community. One of the greatest obstacles
in developing predictive models is a lack of understanding of the relative epistemic uncertainties (model form errors) inherent
to different choices of contact constitutive modeling approaches. The contact constitutive models affect the stiffness and
hysteresis of the contact resulting in different frequency and damping properties. Multi-Objective Optimization (MOO)
is applied to find solutions that minimize the deviation in both frequency and damping from experimentally measured
properties. Then the concept of non-domination of design parameter sets is used from MOO approaches to develop a
quantitative understanding of the epistemic uncertainties, i.e., the inexactness of each approach to represent experimentally
measured properties. The current study investigates the uncertainty associated with conventionally popular simplified
approaches such as the use of phenomenological models (e.g., Iwan or Bouc-Wen models) and constitutive models (e.g.,
Jenkins models) within a whole-jointed framework.
Keywords Parametric joint identification · Hysteretic modeling · Multi-objective optimization · Non-linear modal
analysis
42.1 Introduction
Parametric identification of bolted joints is an active area of research in the development of computational descriptions of
bolted joints. However, this has continued to be a challenging area since the model form to be assumed is generally unclear.
Although the use of hysteretic model formulations is agreed upon since joints have been observed to show dissipative
behavior [5], there is little consensus about the context-specific appropriateness of each model.
The current paper seeks to build intuition into this aspect of hysteretic models in the context of bolted joint identification.
Matching amplitude-dependent modal characteristics in terms of modal frequencies and damping factors forms the objective
for the identification procedures carried out in the work. The epistemic uncertainty, or model form error, has previously been
observed to be expressed in terms of the inability of the identification to determine a single model that predicts trends in both
the stiffness (natural frequency) as well as the dissipation (damping factor) for a range of response levels [1].
The above framework can be mathematically expressed as a bi-objective optimization problem, with the two objectives
being the frequency and damping factor. In the Multi-Objective Optimization (MOO) discipline, it is common to look at
the parameter-space and objective-space separately. Each design, taken to be a set of parameters represented by a point
in the parameter-space, can be taken to map to a corresponding point on the objective-space through the non-linear modal
simulations. A set of points (images) on the objective space is said to be non-dominated or Pareto-optimal if, no point in the
set dominates any other (in the set) in terms of all the objectives. The corresponding pre-image of this set is known as the
non-dominated designs. The current paper explores the possibility of relating the location of the non-dominated set on the
objective space to epistemic uncertainty inherent to a hysteretic model.
J. H. Porter · C. R. Little · N. N. Balaji () · M. R. W. Brake

e-mail: [email protected]; [email protected]; [email protected]; [email protected]

292 J. H. Porter et al.
42.2 Modeling and Approach
Following recent work such as Ref. [2], the non-linear modal analysis of pre-stressed systems has become significantly
efficient computationally, to the point of enabling parametric identification for fairly large assemblies in a physically
consistent and accurate manner. Using impact response techniques (see [4] for the method followed here and some
comparisons), provides experimental backbones that may directly be correlated with the predictions of the non-linear modal
analysis. As already mentioned, an objective minimization problem will be set up as follows:
⎛ ⎞
⎛ ⎞
⎜f (X) ⎟ Δω(qi ; X) + 2
⎜ ω + ⎟ ⎜ ⎟
⎜ ⎟ ⎜ i
ω(qi ) ⎟
min ⎜
⎜
⎟ ⎜
⎟=⎜
⎟
⎟. (42.1)
+ ⎜ ⎟ ⎜ 2⎟
⎜ ⎟ +
+ ⎠ ⎝ Δ ln ζ (qi ; X) ⎠
X∈Ω
⎝ fζ (X)
i
ln ζ (qi )
Here, ω, ζ denote the natural frequency and damping factor, qi (i = 1, 2, . . . ) denotes a set of amplitude levels, and X ,
+
denote the set of parameters and the parameter space respectively.
There are several techniques in the literature to solve these problems [3], such as weighted minimization, boundary
intersection, classes of evolutionary algorithms, etc. The spherical boundary intersection method is employed for the current
study. This method is a deterministic gradient based approach that transforms the problem into a family of equivalent single
objective constrained optimization problems (with additional parameterizations) through boundary intersection constraints
on the objective space. This allows for one to traverse through the whole non-dominated set (referred to as boundary here)
by solving all the problems in this family. Figure 42.1 presents an overview of the methodology for each friction model.
42.3 Results
Figure 42.2 shows the whole joint modeling approach employed (in Fig. 42.2a) and the non-dominated sets in the objective
space (in Fig. 42.2b) for different friction models employed.
The four- and five-parameter Iwan models and the general Bouc-Wen model produce the best fits over different regions
suggesting that they have the lowest levels of epistemic uncertainty. The four- and five-parameter Iwan models produce nearly
identical boundaries suggesting that the additional parameter is not needed to fit this experimental data. For the Bouc-Wen
models, the general model performs clearly better than the two specific cases of a fixed exponent n. However, the two cases
of n = 1 and n = 2 are more convenient because the model can be analytically integrated to get a direct force displacement
relationship. Over most of the region, the Jenkins model produces the poorest fits expect for when compared to a few points
of the Bouc-Wen n = 2 boundary. This suggests that the Jenkins model has the highest level of epistemic uncertainty. This
is expected because the Jenkins model is generally for an element to element type of interface rather than for patches.
In general, more parameters usually results in a better fit to the data because the model had more flexibility. The one
exception to this is in the region where the Bouc-Wen general and n = 1 models both perform better than the two Iwan
models. In addition, sometimes the additional flexibility is not useful such as the two Iwan models producing nearly identical
boundaries. There is also a region where the general Bouc-Wen model converges to values of n near 1 resulting in a portion
of the boundary being close to the boundary for the special case of n = 1.
The boundary intersection method for the Jenkins simulations converges to several points that are dominated by points on
the boundary. These solutions could be local minimizers or represent a non-convex Pareto front.
All of the models are applied independently in the x and y directions with a linear spring in the normal direction. Viscous
damping equal to the damping at the lowest amplitude of the test data is assumed for all simulations. This assumption of low
amplitude damping helps the Jenkins model since the model has no dissipation until macroslip occurs. However this may be
disadvantageous for other hysteric models, some of which have finite dissipation even for small amplitudes.
42 An Assessment of the Applicability and Epistemic Uncertainties Inherent. . . 293
Fig. 42.1 Overview of assessment methodology for each friction model
RP
RP
RP
K
RP
RP
RP K
RP
RP K
RP
RP
(a) (b)
Fig. 42.2 Results for whole-jointed model: (a) whole-joint interface representation [1]; and (b) the non-dominated fronts
References
1. Balaji, N.N., Brake, M.R.W.: The surrogate system hypothesis for joint mechanics. Mech. Syst. Signal Process. 126, 42–64 (2019)
dynamical systems. Comput. Struct. 230, 106–184 (2020)
294 J. H. Porter et al.
3. Deb, K.: Multi-objective Optimization Using Evolutionary Algorithms, vol. 16. John Wiley & Sons, Chichester (2001)
4. Jin, M., Brake, M.R.W., Song, H.: Comparison of nonlinear system identification methods for free decay measurements with application to
jointed structures. J. Sound Vib. 453, 268–293 (2019)
5. Mathis, A.T., Balaji, N.N., Kuether, R.J., Brink, A.R., Brake, M.R.W., Quin, D.D.: A review of damping models for structures with mechanial
Joints. Appl. Mech. Rev. Accepted.
Chapter 43
Hyper-Reduction Approaches for Contact Modeling with Small
Tangential Displacements: Applications for a Bolted Joint
Nidish Narayanaa Balaji, Tobias Dreher, Malte Krack, and Matthew R. W. Brake
Abstract The current paper presents a study of two approaches for the hyper-reduction of contact interfaces for non-linear
structural dynamics analyses. The first approach is a reformation of the “whole-joint approach” wherein different regions
of the contact interface are grouped together using a “non-linear patch”; and the second approach is based on a selective
remeshing of the interface. Both are conducted based on some field objectives such as contact traction, etc. A recently
developed quasi-static modal analysis technique is applied to characterize the systems in terms of amplitude-dependent
modal properties, which are compared as a qualification metric.
Keywords Interface reduction · Hyper-reduction · Jointed structures
43.1 Introduction and Motivation
Model order reduction of the non-linear dynamical models of structures with frictional contacts is of high importance in
making simulations of such systems tractable. Even for relatively simple bolted structures, the representational requirements
as well as the amount of non-linear evaluations necessary can make the system very large and very expensive for simulations.
Projection-based techniques for model order reduction (see, for instance [7]), which are based on finding an appropriate
reduced representation of the solution in a lower-rank subspace (and solving the problem there), are a very popular approach
in the literature. One of the main drawbacks with such approaches is that it is generally not trivial to establish a reduced
representation of non-linearities on the chosen subspace itself. In other words, evaluation of non-linearities in such models is
conducted by first transforming the unknowns into the full-order model, evaluating the non-linearities, and then transforming
them back into the reduced domain. This procedure, while beneficial for a lot of cases, becomes computationally cumbersome
when it comes to very large models wherein the evaluation of the non-linear function becomes an important computational
bottleneck. In order to alleviate this issue, there have been several hyper-reduction approaches (see, for instance [8]), which
seek to develop reduced order modeling strategies with the non-linearities completely represented in the reduced domain
itself. One promising approach for this is to develop a data-based representation of the non-linearity on the subspace (see [5]
for an application) based on several non-linear function evaluations on the full-order model. Such an approach, however,
suffers from input-level dependence, i.e., the trained model performs only as good as the training data-set.
The current paper, following previous efforts [4], explores model reduction through the development of reduced
representations of the interface while retaining the physical meaning of the degrees-of-freedom of the reduced model. This
allows for relatively easy definitions of the non-linearities consistently in the reduced domain, thereby avoiding the need
to transform back to the full problem. Two approaches are presented with their merits and shortcomings discussed: (1) an
improved whole-joint formulation (similar to the ones used in [6]) applied to regions on the interface selected based on
a binned field objective; and (2) an interface remeshing approach based on efficient representation of a continuous field
objective.
N. N. Balaji () · M. R. W. Brake

T. Dreher · M. Krack
Institute of Aircraft Propulsion Systems, University of Stuttgart, Stuttgart, Germany

296 N. N. Balaji et al.
43.2 Description of Approaches
As already mentioned, the two approaches are closely tied together in the sense that both are aimed at developing a
representation of the interface that best represents a particular field quantity over the interface (contact pressure, dissipation
fluxes, etc.). Therefore, a scalar-valued field objective, denoted P(x ) in the following, has to first be identified so that the
+
reduced model can be developed.
All of the results shown in the following are based on the interfacial mesh shown in Fig. 43.1, which corresponds to the
bolted joint interface for a finite-element model of the Brake-Reuß Beam (BRB), a three-bolted lap-joint structure [3]. The
field objective that will be used here will be the normal contact pressure from static bolt-prestress simulations (with each bolt
applying a prestress of 11.58 kN) used as the field objective (re-ranged to [0,1] and denoted using color here).
43.2.1 Whole Jointed Approach Based on Binned Field Objective
For the first approach, the field P(x ) is first divided into a set of discrete levels Pmin = P0 < P1 < · · · < PNlev = Pmax .
+
Elements corresponding to each bin i are selected as the elements that have P(x ) ∈ [Pi−1 , Pi ). Following this, disconnected
+
trees of the ensuing graph are identified and separated into separate “patches”. Repeating this procedure for all the chosen
levels yields a set of patches (defined as element-sets).
An improved stiffness-preserving whole-joint formulation1 (see [2] for formulation) is then employed to represent each
of these patches by a single six-Degree-of-Freedom (six-DoF) virtual node (three displacements and three rotations). Finally,
conducting a CMS (Component-Mode Synthesis) procedure on this gives the effective reduced order model. Since this is
a model that is based on a set of physical DoF’s, contact models may be used on these nodes directly (making the model
hyper-reduced).
Figure 43.2 indicates the patches identified for three different binning levels for the interface under consideration.
Fig. 43.1 Reference mesh. The static contact pressure mapped to [0, 1], used as the field objective in the following, is denoted using colours
(a) 2 Levels (3 Patches) (b) 4 Levels (11 Patches)
(c) 7 Levels (21 Patches)
Fig. 43.2 Binned objective patches for the Brake-Reuß Beam (BRB) interface. The colors indicate different patches. (a) 2 Levels (3 Patches). (b)
4 Levels (11 Patches). (c) 7 Levels (21 Patches)
1 Similar, but not identical to the RBE3 elements in ANSYS terminology.

43 Hyper-Reduction Approaches for Contact Modeling with Small Tangential Displacements: Applications for a Bolted Joint 297
43.2.2 Interface Remeshing Approach
The second approach comes from the idea that not all nodes in an interface may be necessary to accurately represent a field
variable. Therefore, information about the local gradients of the field quantity are employed to guide the design of a new
mesh on top of the initial mesh in the interface. Consider a “full mesh”, denoted by T , and a “reduced mesh”, denoted by
Tr . Denoting the nodal DoF vectors in each as u and ur respectively, the vector u may be approximated by interpolating ur
+ +
on the mesh Tr using its shape functions as follows: + +
u Qr ur . (43.1)
+ +
Here, the matrix Qr denotes the interpolation matrix developed using the corresponding shape functions of Tr . Note that the
same relationship (in Eq. (43.1)) may be used to approximate nodal tractions on the original mesh but not forces.
For a dynamical system of the form
Mü + Ku + fnl (u) = fext (t), (43.2)

+ + + + +
a Galerkin-projection-based Reduced-Order-Model (ROM) may be developed if Eq. (43.1) is interpreted as a projection onto
a reduced basis. Equation (43.2), under this projection, becomes

Qr T MQr u¨r + Qr T KQr ur + Qr T fnl (Qr ur ) = Qr T fext (t). (43.3)
+ + + + +
, -. /
fnlr
+
Note here that the non-linear force fnl still has to be evaluated as many times as in the original model, i.e., the model
is not hyper-reduced yet. This is achieved by evaluating the non-linear tractions in Tr (denoted as tr (ur )), interpolating it
+
onto T (becoming Qr tr (ur )), integrating this in on T (computed with some quadrature integration matrix P and denoted
+
as PQr tr (ur )), and finally applying the Galerkin projection to obtain the forcing for the ROM. This procedure may be
summarized + as,
fnlr = Qr T fnl (Qr ur ) = Qr T PQr tr (ur ). (43.4)

+ + + , -. / +
f
Tm
Here, the non-linearities need to be evaluated only in the reduced mesh Tr , and projected onto the ROM using a single
matrix Trm . Since integrals of tractions on Tr directly do not have any meaning for the ROM, the reduced mesh may not be
interpreted as a regular finite element mesh that discretizes the interface in a reduced fashion. It merely represents a reduced
representation of the original mesh with consistent (but approximate) mappings.
Figure 43.3 presents sample reduced meshes developed for the BRB interface using the contact pressure as the objective
field (see Fig. 43.1 for the field on full mesh). The pressure values are re-scaled to range from 0 to 1 for convenience. For the
meshes shown in the figure, nodes were biased to lie in regions with large gradients or changes in the objective field.
43.3 Results
Figure 43.4 shows the results for amplitude-dependent non-linear modal simulations using the technique developed in [1] for
the first mode of vibration. Unilateral linear springs and planar elastic dry friction elements (2D Jenkins models) are used in
the traction-sense as the normal and tangential contact constitutive relationships for the reference model and the remeshing
ROMs. The same contact laws are employed for the whole joint ROMs but since these require phenomenological models
(force-displacement models), the traction-stiffnesses used in the reference are scaled by the areas of each patch to obtain
consistent model parameters for the ROMs.
298 N. N. Balaji et al.
(a) 52 Elements (b) 124 Elements
(c) 300 Elements
Fig. 43.3 Examples of different reduced meshes using approach 2. The colors indicate the field-objective function, with blue indicating 0 and
yellow indicating 1. (a) 52 Elements. (b) 124 Elements. (c) 300 Elements
180 180
160 160
140 140
120 120
100 100
80 80
10 0 10 0
10 -5 10 -5
10 -6 10 -5 10 -4 10 -3 10 -2 10 -1 10 -6 10 -5 10 -4 10 -3 10 -2 10 -1
(a) (b)
Fig. 43.4 Results for the (a) Whole Jointed Approach and the (b) Remeshing approach. Reference results are calculated using the full mesh
(Fig. 43.1); whole joint ROMs are developed based on the patches shown in Fig. 43.2; and remeshing ROMs are developed based on the meshes
shown in Fig. 43.3
In terms of performance, one can see that both demonstrate good convergence in capturing the reference trends in the
stiffness characteristics (amplitude-dependent natural frequency). However, the remeshing approach seems to represent the
dissipative characteristics (amplitude-dependent damping factor) much better than the whole joint approach. This may be
an observation that could be related to the formulations of the ROMs, or just numerical relics of the estimated dissipation
metric. Further investigations are currently being conducted to draw more conclusions.
43.4 Discussions and Conclusions
The paper presents the development and relative comparison of hyper-reduction approaches for structures with small
displacement contacting interfaces. The ROMs developed are consistently hyper-reduced, i.e., simulations of the full non-
linear model are not necessary for the hyper-reduction. An additional advantage of this is that a developed ROM can
potentially be used to conduct cheap simulations over any amplitude range of interest accurately.
In addition to extracting conclusive inferences from observations such as in Sect. 43.3, the current investigation also
explores the use of different choices of objective functions such as modal strains from linear modal analyses, dissipation
fields from non-linear modal analyses, etc., as well as weighted combinations of these.
43 Hyper-Reduction Approaches for Contact Modeling with Small Tangential Displacements: Applications for a Bolted Joint 299
Another aspect that is taken up in the current work is the fact that it may not always be possible to come up with asymptotic
accuracy analyses for the approaches considered here. For the first approach, having a very large number of binning levels
leads to a system where some patches will be constituted with just a single element. These present numerical difficulties since
in this case all the nodes will be shared by the patch with adjacent patches. For the second approach, increasing the number
of required elements starts failing after a point since the elements one can come up with, while following the specified
objective, start losing shape-regularity and yield bad quality elements. In most cases, however, this has been encountered
when the sizes of some elements in the reduced mesh start becoming significantly smaller than those of the reference mesh.
References
dynamical systems. Comput. Struct. 230, 106184 (2020)
2. Balaji, N.N., Dreher, T., Krack, M., Brake, M.R.W.: Reduced Order Modeling for the Dynamics of Jointed Structures Through Hyper-Reduced
Interface Representation. Mech. Syst. Signal Process. Under Review
3. Brake, M.R.W.: The Mechanics of Jointed Structures: Recent Research and Open Challenges for Developing Predictive Models for Structural
Dynamics. Springer, Cham (2017)
4. Dreher, T., Balaji, N.N., Groß, J., Brake, M.R.W., Krack, M.: Gerrymandering for interfaces: modeling the mechanics of jointed structures. In:
Nonlinear Structures and Systems, vol. 1, pp. 81–85. Springer, Cham (2020)
5. Jain, S., Tiso, P.: Hyper-reduction over nonlinear manifolds for large nonlinear mechanical systems. J. Comput. Nonlinear Dyn. 14, 081008
(2019)
Process. 118, 133–157 (2019)
7. Mitra, M., Zucca, S., Epureanu, B.I.: Adaptive microslip projection for reduction of frictional and contact nonlinearities in shrouded blisks. J.
Comput. Nonlinear Dyn. 11(4), 041016 (2016)
8. Ryckelynck, D.: A priori hyperreduction method: an adaptive approach. J. Comput. Phys. 202(1), 346–366 (2005). ISSN 00219991. https://doi.
org/10.1016/j.jcp.2004.07.015
Chapter 44
Fatigue Damage of a Single-Edge Notched Beam Specimen Under
Variable Amplitude Loading with Similar Probabilistic and Cycle
Counting Statistics
Abstract Fatigue life estimation under variable amplitude (VA) loading is still a major unresolved engineering problem. The
linear cumulative damage rule (LDR) based methodology is inadequate to predict fatigue damage or useful remaining life if
the fatigue loading is complex (not sinusoidal). In addition, the LDR based damage estimation methods rely on statistics of
various cycle counting methodologies (e.g., rain-flow counting) where the load interaction effects are being ignored. It has
been shown that the damage estimation fails if two load-time (load-cycle) histories with different temporal dynamics have the
same or similar load spectra. A robust damage estimation should take the temporal dynamics, i.e., the overload/underload
information, stress memory statistics, into consideration. In this extended abstract, the shortcomings of the LDR based
methods will be presented through a comparison where two VA load spectra with similar stress history and cycle counting
statistics were applied to a singled-edge notched beam specimen.
Keywords Fatigue crack propagation · Fatigue damage estimation · Rain-flow counting · Variable amplitude loading ·
Load statistics
44.1 Introduction
Fatigue damage estimation under constant amplitude (CA) loading is a solved engineering problem while damage estimation
under highly irregular loading is still an open field of study among researchers in fatigue structure and materials. Among
existing methodologies in predicting useful remaining life of engineering structures, linear cumulative damage rules (LDR)
and linear fracture mechanics based empirical crack propagation laws are widely used in contemporary industry for their
simplicity and general accuracy. The LDR, or Palmgren-Miner’s rule tries to resolve the damage under VA loading by
consider damage as a summation of the fractions between the number of applied cycles in a stress range and the number of
cycles to failure under CA loading with the same stress range [1], or simply

n
ni
D= (44.1)
Ni
i=1
where D is the estimated damage, ni is the number of cycles applied at the ith stress range, Ni is the number of cycles
to failure at the same stress range under CA loading (this number is obtained from the S-N curve of the material under
investigation). However, these methodologies fails to estimate the damage or of the useful remaining life when the applied
loads are complex or have variable amplitude. This is partly due to the load interaction effect introduced by crack closure
and opening at the wake of the crack [2]. Later on, modifications have made to these models by (1) adding weights to
each binned load level in LDR and (2) adding additional empirical parameters into the Paris law based method in order to
take load interaction effect into consideration [3]. Here, we focus on the modification to the LDR based methodologies for
its simplicity in computation. The accuracy of damage estimation decreases dramatically when various applied loads have
similar statistical characteristics but different temporal dynamics due to its simple dependence on stress range statistics [1, 4].
For example, these applied loading could have similar power spectral density, probability density (or mass), and/or cycle
H. Li () · D. Chelidze
Nonlinear Dynamics Laboratory, Department of Mechanical, Industrial and Systems Engineering, University of Rhode Island, Kingston, RI, USA

1 1
Normalized magnitude
0.5 0.5
x i+25
0 0
-0.5 -0.5
-1
-1
-1 -0.5 0 0.5 1
0 5 10 15 20 25 30 35 40
Time (second) xi
Fig. 44.1 Upper two: pressure-voltage curve by incremental increasing the voltage input to the regulator; upper left: pressure-voltage curve for
the front regulator; upper right: rear regulator. Lower two: deflection-pressure curves and the corresponding curve fitting results
counting statistics. According to previous experimental work [5], the fatigue lives on the same notched beam structure have
two times the difference between chaotic loading and its corresponding random surrogate with nearly identical magnitude
statistics and power spectral density. Additionally, simulation based study indicates that the damage estimation fails if two
load-time (load-cycle) histories with different temporal dynamics have the same or similar load spectra [6]. This indicates
that metric based on simple spectral or statistical parameters tend to fail the purpose of accurately estimate damage. In
the following section, a demonstrative simulation will be presented to show the insufficiency of the current LDR based
methodologies.
44.2 Fatigur Crack Propagation Simulation
44.2.1 Simulation Setup
In order to illustrate the shortcomings of the LDR and its modifications, an illustration based on fatigue crack propagation
simulation was conducted.
In order to keep the stress level statistics and cycle counting statistics of loading consistent, a normalized load time history
was obtained by numerically integrating the Lorenz equation [7]:
⎧
⎪
⎨ẋ = 10 (y − x)
⎪
ẏ = −xy + 28 x − y (44.2)
⎪
⎪
⎩ż = xy − 8 z
3
where (˙) represents derivative with respect to time. And the solution was obtained using a sampling frequency of 250 Hz and
the initial condition of x0 = 0, y0 = −0.01, z0 = 27. For simplicity, the solution to the third variable, z, was selected due
to its wide sense stationarity. The obtained load time history was normalized in its magnitude and its corresponding phase
space representation can be obtained through phase space reconstruction, see Fig. 44.1.
In order to fit the load time history to the LDR, local maximum stress and minimum stress were extracted to form the
fatigue spectrum of the obtained chaotic load time history, and it was name as the chaotic reversals or the chaotic spectrum.
Since the comparison was based on similar stress history statistics and cycle counting statistics, the obtained stress range of
the chaotic reversals can be defined as:
σ i = σmax
i
− σmin
i
(44.3)
44 Fatigue Damage of a Single-Edge Notched Beam Specimen Under. . . 303
Fig. 44.2 Schematics of the load factors that are used in this report. F1: load cycle range; F2: unload cycle range; F3: negative peak difference;
F4: positive peak difference; F5: load cycle range that corresponds to its following overload, F4; F6: unloading cycle range that corresponds to it
following overload, F4
1 1
Noralized Magnitude
Noralized Magnitude
0.5 0.5
0 0
-0.5 -0.5
-1 -1
10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100
Index of fatiuge reversals Index of fatiuge reversals
Fig. 44.3 Load-cycle history between the chaotic spectra and the random spectra. Left: the chaotic spectra; right: the random spectra
where, i indicates the ith load cycle range, please refer to Fig. 44.2 for detailed explanation. The previously defined load cycle
range (or stress range), σ i ,which is considered as a major driving force of macro fatigue cracks was randomly permuted
temporally. For wide sense stationary load time histories, the cycle counting statistics and the stress level statistics should
be identical to its chaotic counterparts. Here, rain flow counting was selected among various cycle counting methods since
it accounts for the load interaction by taking the stress-strain memory effect into consideration [8, 9]. The obtained chaotic
spectrum and its randomly permuted version, called the random spectrum, will be used as input spectra to a widely used
fatigue crack propagation software package, AFGROW 5.3. As a direct comparison, from Fig. 44.3, the two spectra have
different temporal dynamics. However, the stress level statistics, measured by the stress level counting, and the rain flow
cycle counting statistics, measured by cycle range counting are identical among the two cases, see Fig. 44.4.
After the preparation of the two fatigue spectra, crack propagation simulations will be performed in AFGROW. The
specimen is assumed to be made from 2024-T3 aluminum alloy. According to the nature of the applied load from the fatigue
experimental apparatus and the specimen configuration, single-edge through crack was considered under pure bending.
Since both the chaotic and random spectra are normalized in their magnitude, AFGROW considers it as 1 ksi. Therefore,
in order to fulfill the purpose of simulating crack propagation under realistic loading, the load spectra are multiplied with a
spectrum multiplication factor[10] (SMF) of 14 (maximum stress within 14 ksi). Further, since the simulation should take
load interaction effects into account, a simple crack retardation model [11], crack closure model, was utilized. The initial
crack length was set to 0.3 inches and the criteria for stopping the simulation are (1) crack size reaches 0.6 inches and (2)
stress intensity factor reaches the fracture toughness.
1200 600
Loaing type Loaing type
1000 Chaotic 500 Chaotic
Cycle range counting

Stress level counting
Randperm Randperm
800 400
600 300
400 200
200 100
0 0
-1 -0.5 0 0.5 1 0.5 1 1.5 2
Dimensionless stress Dimensionless cycle ranges
Fig. 44.4 Comparison of the stress level statistics and the cycle counting statistics between the chaotic spectra and the random spectra. Left: the
stress level counting histogram; right: the cycle range counting histogram
0.6
Loaing type
0.55 Chaotic
Crack length (inches)
Randperm
0.5
0.45
0.4
0.35
0.3
0 1000 2000 3000 4000 5000 6000 7000
Fatigue life (cycles)
Fig. 44.5 Comparison between the fatigue life (in applied stress cycles) under chaotic spectra and the random spectra. Left: the crack propagation
curve under chaotic spectra loading; right: the crack propagation curve under random spectra loading
44.2.2 Simulation Results
As it can be observed from Fig. 44.5, the predicted results indicates the fatigue life, under two (1) similar stress level statistics
and (2) similar cycle counting statistics yielded different fatigue life. This difference in fatigue life is assumed to be resulted
from the difference in their temporal dynamics since the load sequence has changed considerably after random permutation
of the load cycle ranges. This will contradict the damage estimation if LDR is used since the two load spectra (i.e., the chaotic
spectrum and the random spectrum) are identical.
44.3 Discussion
This work is aimed to illustrate the importance of taking more sophisticated consideration during fatigue life prediction or
damage estimation under VA loading. More specifically, the measures of temporal dynamics By only taking stress level
related statistics and/or whenever cycle-counting method is utilized, LDR based methods tend to produce erroneous results.
44 Fatigue Damage of a Single-Edge Notched Beam Specimen Under. . . 305
44.4 Future Work
For the current work, the load spectrum was selected to be wide sense stationary, this is partly due to the future plan for
experimental verification work on the fatigue experiment apparatus in our laboratory environment. Later on, simulation-
based analysis to a wider range of fatigue load spectra will be conducted to generalize the understanding of the proposed
fatigue estimation method. Aside from the simulation, modification to the current experimental setup will be made to assist
the findings from simulations.
References
1. Schijve, J.: Fatigue of structures and materials. Springer Science & Business Media (2001)
2. Suresh, S.: Fatigue crack deflection and fracture surface contact: micromechanical models. Metall. Trans. A 16(2), 249–260 (1985)
3. Santecchia, E., Hamouda, A., Musharavati, F., Zalnezhad, E., Cabibbo, M., El Mehtedi, M., Spigarelli, S.: A review on fatigue life prediction
methods for metals. In: Advances in Materials Science and Engineering, vol. 2016. Hindawi (2016)
4. Suresh, S.: Fatigue of materials. Cambridge University Press, Cambridge (1998)
5. Nguyen, S.H., Falco, M., Liu, M., Chelidze, D.: Different fatigue dynamics under statistically and spectrally similar deterministic and stochastic
excitations. J. Appl. Mech. 81(4):041004 (2014)
6. Li, H., Chelidze, D.: Identification of variable amplitude fatigue loading based on bivariate probability mass functions. In: IDETC/CIE 2019 –
MSNDC. ASME (2019)
7. Sprott, J.C.: Chaos and Time-Series Analysis, vol. 69. Oxford University Press, Citeseer (2003)
8. ASTM: Standard Practices for Cycle Counting in Fatigue Analysis. e1049-85, ASTM international (2017)
9. Sunder, R., Seetharam, S., Bhaskaran, T.: Cycle counting for fatigue crack growth analysis. Int. J. Fatigue 6(3), 147–156 (1984)
10. Harter, J.A.: Afgrow users guide and technical manual. LexTech Inc., Tech. Rep. (2019)
11. Elber, W.: The significance of fatigue crack closure. In: Damage Tolerance in Aircraft Structures. ASTM International (1971)

Nonlinear Structures & Systems, Volume 1 - Proceedings of The 38th IMAC, A Conference and Exposition On Structural Dynamics 2020

Uploaded by

Copyright:

Available Formats

Nonlinear Structures & Systems, Volume 1 - Proceedings of The 38th IMAC, A Conference and Exposition On Structural Dynamics 2020

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Nonlinear Structures & Systems, Volume 1 - Proceedings of The 38th IMAC, A Conference and Exposition On Structural Dynamics 2020

Uploaded by

Copyright:

Available Formats

Conference Proceedings of the Society for Experimental Mechanics Series

Gaetan Kerschen · Matthew R. W. Brake

More information about this series at http://www.springer.com/series/8922

ISSN 2191-5644 ISSN 2191-5652 (electronic)

© The Society for Experimental Mechanics, Inc 2021

Liège, Belgium G. Kerschen

14 Model Correlation to a Nonlinear Bolted Structure Using Quasi-Static Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . 87

33 Analysis of an Actuated Frictional Interface for Improved Dynamic Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227

Florian Pichler and Wolfgang Witteveen

Keywords Joint contact · Friction · Damping · Model order reduction · Nonlinearity

F. Pichler () · W. Witteveen

© The Society for Experimental Mechanics, Inc 2021 1

M̂ (q) q̈ + K̂ (q) + CTq (q) λ = Qv + Qe + Qnl

Fig. 1.1 Practical computation workflow

(Qnl )f = Φ T fnl . (1.5)

1.3 Model Description

Fig. 1.2 Finite element model of the C-beam

Fig. 1.3 Representation of the screw

1.4 Convergence Study

Abaqus presented in [9] MSC.Adams (20 joint modes)

Abaqus presented in [9] MSC.Adams (20 joint modes)

Fig. 1.7 Displacement of IA for full and half preload

Lukas Koller, Wolfgang Witteveen, and Florian Pichler

L. Koller () · W. Witteveen · F. Pichler

© The Society for Experimental Mechanics, Inc 2021 9

2.2 Model Reduction and Description of the Nonlinear Term

Mẍ + Kx = fLIN + fNL (x, ẋ) (2.1)

σn ≈ ψn1 r1 + · · · + ψnl rl = ψn r, (2.4)

p = −σn ≈ −ψn r ⇒ p = p (r) (2.5)

2.3 Computation of Stress Modes

2.4 Idea of Semi Hyper-Reduction

g (p) = 0 ⇒ g (r) = 0 (2.9)

or a reduced minimization problem

min J (p) . . . min J (r)

2.5 Application of Semi Hyper-Reduction to Reynolds Equation

The proposed method of SHR is applied onto Reynolds equation (RE)

2.6 Numerical Studies

Fig. 2.2 Estimation of the computational time (2D) [9]

3.1 Introduction and Motivation

S. Pöchacker () · T. Lauß · F. Pichler · S. Oberpeilsteiner · W. Witteveen

© The Society for Experimental Mechanics, Inc 2021 15

fast qualitative results

linear substitution variant 3

Fig. 3.1 Graphical representation of the research idea

3.2 Purpose of Research

3.3 Mechanical Model of the Numerical Test Structure

3.3.1 Nonlinear High Fidelity FE-Model

M ẍ + D ẋ + Cx = flin + fnl (3.1)

Preload 1 Preload 2 Preload 3

Fig. 3.2 Draft of the high fidelity model

75mm 75mm 2 x 75mm 2 x 75mm 75mm

3.3.2 Linear Substitution Model

or with the transformation x = q in the modal space

M̃ q̈ + D̃ q̇ + C̃q = f˜lin (3.3)

wherein the matrices and vectors are defined below.

x = q . . . DoF’s, (x are physical DoF’s, q are modalDoF’s) (3.4)

T M = M̃ = I . . . mass matrix (3.5)

T flin = T (fc + fd + fl ) = f˜lin . . . force vector (3.8)