Studies in Systems, Decision and Control 467

Didier Theilliol
Józef Korbicz
Janusz Kacprzyk Editors

Recent Developments
in Model-Based and
Data-Driven Methods
for Advanced Control
and Diagnosis
Studies in Systems, Decision and Control

Volume 467

Series Editor
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences,
Warsaw, Poland
The series “Studies in Systems, Decision and Control” (SSDC) covers both new
developments and advances, as well as the state of the art, in the various areas of
broadly perceived systems, decision making and control–quickly, up to date and
with a high quality. The intent is to cover the theory, applications, and perspectives
on the state of the art and future developments relevant to systems, decision
making, control, complex processes and related areas, as embedded in the fields of
engineering, computer science, physics, economics, social and life sciences, as well
as the paradigms and methodologies behind them. The series contains monographs,
textbooks, lecture notes and edited volumes in systems, decision making and
control spanning the areas of Cyber-Physical Systems, Autonomous Systems,
Sensor Networks, Control Systems, Energy Systems, Automotive Systems,
Biological Systems, Vehicular Networking and Connected Vehicles, Aerospace
Systems, Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power
Systems, Robotics, Social Systems, Economic Systems and other. Of particular
value to both the contributors and the readership are the short publication timeframe
and the world-wide distribution and exposure which enable both a wide and rapid
dissemination of research output.
Indexed by SCOPUS, DBLP, WTI Frankfurt eG, zbMATH, SCImago.
All books published in the series are submitted for consideration in Web of Science.
Didier Theilliol · Józef Korbicz · Janusz Kacprzyk
Editors

Recent Developments
in Model-Based
and Data-Driven Methods
for Advanced Control
and Diagnosis
Editors
Didier Theilliol Józef Korbicz
Faculté des Sciences et Techniques Institute Control and Computation
University of Lorraine Engineering
Nancy, France University of Zielona Góra
Zielona Gora, Poland
Janusz Kacprzyk
Systems Research Institute
Polish Academy of Sciences
Warsaw, Poland

ISSN 2198-4182 ISSN 2198-4190 (electronic)

Studies in Systems, Decision and Control
ISBN 978-3-031-27539-5 ISBN 978-3-031-27540-1 (eBook)
https://doi.org/10.1007/978-3-031-27540-1

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Switzerland AG 2023
This work is subject to copyright. All rights are solely and exclusively licensed 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

This book contains selected papers presented at the 16th European Workshop on
Advanced Control and Diagnosis, ACD 2022 for short, held in Nancy (France) on
November 16–18, 2022.
The conference ACD 2022 (https://acd2022.cran.univ-lorraine.fr/) was organized
by European Advanced Intelligent Control and Diagnosis working group with
the support of the Centre of Research in Automatic Control (CRAN) and Ecole
Polytechnique (Polytech Nancy), Universite de Lorraine. ACD 2022 was techni-
cally co-sponsored by IFAC (the International Federation of Automatic Control)
Technical Committee 6.4. Fault Detection, Supervision & Safety of Technical
Processes—SAFEPROCESS of which Société d’Automatique, de Génie Industriel
et deProductique SAGIP is the National Member Organization for France.
The series of ACD conferences has been an opportunity for high-profile scien-
tists and experts in the subject of advanced control, diagnosis, and health monitoring
of systems to present recent results on focused on advanced fault-tolerant control
strategies, health-aware control design strategies, advanced control approaches,
deep learning-based methods for control and diagnosis, reinforcement learning-
based approaches for advanced control, diagnosis & prognosis techniques applied to
industrial problems, Industry 4.0 as well as instrumentation and sensors.
The series of conferences is an excellent forum for the exchange of knowledge
and experience and sharing solutions in the academic and industrial environment. An
important task of this forum is the integration of scientists and engineers from various
industries, as well as producers of hardware and software for computer control and
diagnostic systems.
This book is divided into six parts:
I. Observers and Estimation,
II. Diagnosis and Prognosis,
III. Advanced Control,
IV. Nonlinear Systems, Localization, and FDI,
V. Machine Learning, Artificial Intelligence, and Data-Driven Methods, and
VI. Robust and Fault-Tolerant Control.

v
vi Preface

We sincerely thank all the participants and the reviewers of the articles from the
International Program Committee for their personal scientific contributions to the
conference. I extend special appreciation to the authors of the accepted articles that
are published in this collective book by Springer, as well as to the speakers of the
plenary:
• Prof. Vicenç Puig: Cyberphysical Security of Critical Infrastructures
• Universitat Politècnica de Catalunya, BarcelonaTech (UPC)—Spain
• Dr. Lorenzo Fagiano (Italy) Dealing with uncertainty in learning-based control
and optimization: the Set membership paradigm
• Politecnico di Milano, Milan—Italy
• Dr. Jean Baptiste Mouret (France) Damage compensation in robotics without
diagnosis INRIA, Nancy—France
We believe that this collective book will become a great reference tool for scientists
and researchers working in the area of fault diagnosis and fault tolerant control.
Readers are kindly encouraged to contact the corresponding authors for further details
concerning their research and presented results.

Nancy, France Editor Chair:

Zielona Gora, Poland Didier Theilliol
Warsaw, Poland Editor co-chairs:
November 2022 Józef Korbicz
Janusz Kacprzyk
In collaboration with
General Chair: J. C. Ponsart
IPC Chair: M. S. Jha
IPC Co-Chair: H. Schulte

Acknowledgments We are grateful to the members of the ACD 2022 National Organizing
Committee, especially to General Chair J. C. Ponsart, International Program Committee Chair
M. S. Jha and Co-Chair H. Schulte for supervising many technical matters, to P. Weber and all
National organizing committee members for organizational and administrative support, and to
L. Suel for the secretariat.
As always, only the great synergistic effort of the organizers makes the conference a successful
scientific event—especially in such difficult times of pandemic and war.
Contents

Observers and Estimation

H∞ Stochastic State-Multiplicative Uncertain Systems-Robust
Luenberger Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
E. Gershon
Anticipating the Loss of Unknown Input Observability for Sampled
LPV Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Gustave Bainier, Jean-Christophe Ponsart, and Benoît Marx
Luenberger Observer Design for Robust Estimation of Battery
State of Charge with Application to Lithium-Titanate Oxide Cells . . . . . 23
Eero Immonen
Fault Detection and Diagnosis of PV Systems Using Kalman-Filter
Algorithm Based on Multi-zone Polynomial Regression . . . . . . . . . . . . . . . 35
Yehya Al-Rifai, Adriana Aguilera-Gonzalez, and Ionel Vechiu
Parity-Space and Multiple-Model Based Approaches
to Measurement Fault Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Ivo Punčochář and Ondřej Straka

Diagnosis and Prognosis

Online Condition Monitoring of a Vacuum Process Based
on Adaptive Notch Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Mohammad F. Yakhni, S. Cauet, A. Sakout, H. Assoum,
and M. El-Gohary
A Study of OBF-ARMAX Performance for Modelling
of a Mechanical System Excited by a Low Frequency Signal
for Condition Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Oscar Bautista Gonzalez and Daniel Rönnow

vii
viii Contents

Pre-localization of Two Leaks in a Water Pipeline Using Hydraulic

and Spatial Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Lizeth Torres and Cristina Verde
Diagnosis and Failure Prognosis of Intermittent Faults: A Bond
Graph Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Wolfgang Borutzky
Remaining Useful Life Estimation Based on Wavelet
Decomposition: Application to Bearings in Reusable Liquid
Propellant Rocket Engines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Federica Galli, Vincent Sircoulomb, Ghaleb Hoblos, Philippe Weber,
and Marco Galeotta

Advanced Control
Adaptive Finite Horizon Degradation-Aware Regulator . . . . . . . . . . . . . . . 123
Amirhossein Hosseinzadeh Dadash and Niclas Björsell
Set-Membership Fault Detection Approach for a Class of Nonlinear
Networked Control Systems with Communication Delays . . . . . . . . . . . . . 133
Afef Najjar and Jean-Christophe Ponsart
Demanded Power Point Tracking for Urban Wind Turbines . . . . . . . . . . . 145
Felix Dietrich, Lukas Jobb, and Horst Schulte
Degradation Simulator for Infinite Horizon Controlled Linear
Time-Invariant Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Amirhossein Hosseinzadeh Dadash and Niclas Björsell
Distributed Observer-Based Leader-Following Consensus Control
Robust to External Disturbance and Measurement Sensor Noise
for LTI Multi-agent Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Jesus A. Vazquez Trejo, Jean-Christophe Ponsart,
Manuel Adam-Medina, Guillermo Valencia-Palomo,
and Juan A. Vazquez Trejo

Nonlinear Systems, Localization and FDI

Contact-Less Sensing and Fault Detection/Localization in Long
Flexible Cantilever Beams via Magnetoelastic Film Integration
and AR Model-Based Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Robert-Gabriel Sultana and Dimitrios Dimogianopoulos
A Comparative Simulation Study of Localization Error Models
for Autonomous Navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Anis Koliai, Stephane Bazeille, Michel Basset, and Rodolfo Orjuela
Contents ix

Study of Testing Strategy for Performance Analysis of Actuator

Layer in Safety Instrumented System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Walid Mechri and Christophe Simon
Localization and Navigation of an Autonomous Vehicle in Case
of GPS Signal Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
S. N. Oubouabdellah, S. Bazeille, B. Mourllion, and J. Ledy
Nonlinear Analysis of SRF-PLL: Hold-In and Pull-In Ranges . . . . . . . . . . 225
Tatyana A. Alexeeva, Nikolay V. Kuznetsov, Mikhail Y. Lobachev,
Marat V. Yuldashev, and Renat V. Yuldashev

Machine Learning, Artificial Intelligence and Data-Driven Methods

Complementary Reward Function Based Learning Enhancement
for Deep Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
George Claudiu Andrei, Mayank Shekhar Jha, and Didier Theillol
Data-Driven Dissipative Verification of LTI Systems: Multiple
Shots of Data, QDF Supply-Rate and Application to a Planar
Manipulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Tábitha E. Rosa and Bayu Jayawardhana
Hybrid Model/Data-Driven Fault Detection and Exclusion
for a Decentralized Cooperative Multi-robot System . . . . . . . . . . . . . . . . . . 261
Zaynab EL Mawas, Cindy Cappelle, and Maan EL Badaoui EL Najjar
Development of a Hybrid Safety System Based on a Machine
Learning Approach Using Thermal and RGB Data . . . . . . . . . . . . . . . . . . . 273
Nicolas Jathe, Hendrik Stern, and Michael Freitag
A Comparative Study on Damage Detection in the Delta Mooring
System of Spar Floating Offshore Wind Turbines . . . . . . . . . . . . . . . . . . . . . 283
Christos S. Sakaris, Anja Schnepf, Rune Schlanbusch,
and Muk Chen Ong

Robust and Fault Tolerant Control

A Cooperative Centralized MPC for Collision Avoidance of a
Fleet of AVs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
Manel Ammour, Rodolfo Orjuela, and Michel Basset
Fault Tolerant Control of Markov Jump Systems Using
an Asynchronous Virtual Actuator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Damiano Rotondo
Robust Safety Control for Automated Driving Systems
with Perception Uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Yan Feng Yu, Kaveh Nazem Tahmasebi, Peng Su, and Dejiu Chen
x Contents

Algorithms and Methods for Fault-tolerant Control and Design

of a Self-balanced Scooter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
Ralf Stetter, Markus Till, and Marcin Witczak
Robust Control of a Customized Lane Change Maneuver . . . . . . . . . . . . . 343
Benoit Vigne, Rodolfo Orjuela, Jean-Philippe Lauffenburger,
and Michel Basset
Robust L2 Proportional Integral Observer Based Controller
Design with Unmeasurable Premise variables and Input Saturation . . . . 355
Ines Righi, Sabrina Aouaouda, Mohammed Chadli, and Said Mammar
Observers and Estimation
H∞ Stochastic State-Multiplicative
Uncertain Systems-Robust Luenberger
Filters

E. Gershon

Abstract Linear, continuous-time state-multiplicative noisy systems are consid-

ered. The problem of H∞ Luenberger filtering for either deterministic norm-bounded
or polytopic-type uncertain systems are solved via a simple LMI(s) condition. In this
problem, a cost function is defined which is the expected value of the standard H∞
performance cost with respect to the stochastic parameters. An illustrative example is
given that demonstrates the tractability of our solution method in the robust uncertain
case.

Keywords H∞ Luenberger Filters · Stochastic systems · Polytopic uncertainty

1 Introduction

We address the problem of H∞ Luenberger-type filtering of continuous-time, state-

multiplicative noisy linear systems. The multiplicative noise appears in the system
model in both the dynamic matrix and in the measurement matrix and the system
may encounter either norm-bounded or polytopic-type uncertainties.
The field of control and estimation of stochastic state-multiplicative noisy systems
has been greatly developed since the onset of the H∞ control theory in the early 80‘s
(see [1, 2] and the references therein). Within a span of more than three decades,
many approaches to the study of the various stochastic control and filtering problems,
including those that ensure a worst case performance bound in the H∞ sense, have
been derived for both: delay-free systems [1–9] and state-delayed, linear, stochastic
systems [10–14].
Delay-free systems with parameter uncertainties that are modeled as white noise
processes in a linear setting have been treated in [1–9] for both the continuous-time
and the discrete-time cases. Such models of uncertainty are encountered in many
areas of applications such as: nuclear fission and heat transfer, population models

E. Gershon (B)
Holon Institute of Technology, Holon, Israel
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 3

D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_1
4 E. Gershon

and immunology. In control theory, such models are encountered in gain scheduling
when the scheduling parameters are corrupted with measurement noise.
The problem of H∞ filtering of the above systems has been addressed via the appli-
cation of a Luenberger filter type or via the general-type filter where in the latter case,
the solution of the corresponding robust problem has been solved for both types of
the above uncertainties. The former robust case where the simple Luenberger filter
is applied has not been tackled before, hence the rational for addressing this type of
filtering. Although the use of the general-type filter accounts also for the Luenberger
filter type (being a special case), the unique solution of the Luenberger-type filtering
may serve as a template in the predictive control case (say, state-feedback control)
where the predictor may itself be of the Luenberger type form. In this study, follow-
ing the problem formulation, we bring the BRL (Bounded Real Lemma) solution for
stochastic systems as a preliminary result needed for the solution of the robust Luen-
berger type filter. We then treat the case of norm-bounded uncertainties followed by
the solution of the robust polytopic case. The theory developed is demonstrated via
a simple example that shows the tractability of our solution method.
Notation: Throughout the paper the superscript ‘T ’ stands for matrix transposition,
Rn denotes the n dimensional Euclidean space and Rn×m is the set of all n × m real
matrices. For a symmetric P ∈ Rn×n , P > 0 means that it is positive definite. We
denote expectation by E{·} and we provide all spaces Rk , k ≥ 1 with the usual inner
product < ·, · > and with the standard Euclidean norm || · ||. The space of vector
functions that are square integrable over [0 ∞) is denoted by L2 and col{a, b}
implies [a T b T ]T . We denote by L 2 (, Rk ) the space of square-integrable Rk −
valued functions on the probability space (, F , P), where  is the sample space,
F is a σ algebra of a subset of  called events and P is the probability measure on F .
By (Ft )t>0 we denote an increasing family of σ -algebras Ft ⊂ F . We also denote by
L̃ 2 ([0, T ); Rk ) the space of nonanticipative stochastic process f (·) = ( f (t))t∈[0,T ]
in Rk with respect to (Ft )t∈[0,T ) satisfying
 T  T
|| f (·)||2L̃ = E{ || f (t)||2 dt} = E{|| f (t)||2 }dt < ∞.
2
0 0

Stochastic differential equations will be interpreted to be of I t ô type [1].

2 Problem Formulation

We consider the following linear system:

d x(t) = Ax(t)dt + Dx(t)dν(t) + B1 w(t)dt,

(1a, b)
dy(t) = C2 x(t)dt + F x(t)dζ (t) + D21 w(t)dt

with the objective vector

H∞ Stochastic State-Multiplicative Uncertain … 5

z(t) = C1 x(t), (2)

where x(t) ∈ Rn is the system state vector, w(t) ∈ R p is the exogenous disturbance
signal, y(t) ∈ Rm is the measured output and z(t) ∈ Rr is the state combination
(objective function signal) to be estimated. The variables ν(t) and ζ (t) are zero-
mean real scalar Wiener processes that satisfy:

E{dν(t)} = 0, E{dζ (t)} = 0, E{dν(t)2 } = dt,

E{dζ (t)2 } = dt, E{dν(t)dζ (t)} = 0.

3 Preliminaries

The solution of the robust Luenberger filter is achieved in the sequel by resorting to
the BRL of stochastic state-multiplicative continuous-time linear systems. This BRL
was obtained by several groups (see, for example, [1], Chap. 2) and it is brought here
as a preliminary result. We consider the system of (1a) and (2) and the following
index of performance;

J B = ||z(t)||2L̃ − γ 2 [||w(t)||2L̃ . (3)
2 2

we obtain the following result:

Lemma 1 Consider the system of (1a) and (2). For a prescribed scalar γ > 0,
a necessary and sufficient condition for J B of (3) to be negative for all nonzero
w(t) ∈ L̃ 2 ([0, ∞); R p ), is that there exist 0 < P ∈ Rn×n such that the following
LMI holds:
⎡ T ⎤
A P + P A + D T P D P B1 C1T
⎣ ∗ −γ 2 I 0 ⎦ < 0. (4)
∗ ∗ −I

In the sequel we will solve the estimation problem under study by applying the
above Lemma 1 to the estimation error as given in the next section.

4 Robust Luenberger-Type Estimator

In this section we treat the problem of designing a robust norm-bounded Luenberger-

type estimator which will be utilized in latter sections. The nominal counterpart of
the latter problem, can be found in [1]. Denoting
6 E. Gershon

A = A − Ā and C2 = C2 − C̄2

where Ā and C̄2 are the nominal values for A and C2 , respectively, we consider:

d x̂(t) = Ā x̂(t)dt + L(dy(t) − C̄2 x̂(t)dt) =

Ā x̂(t)dt + L C̄2 e(t)dt + LC2 x(t)dt + L F x(t)dζ + L D21 w(t)dt, (5a, b)
ẑ(t) = C̄1 x̂(t),

where

e(t) = x(t) − x̂(t).

Denoting z̄(t) = z(t) − ẑ(t), we consider the following cost function:


JF = ||z̄(t)||2L̃ − γ 2 [||w(t)||2L̃ . (6)
2 2

Given γ > 0 , we seek an estimate C̄1 x̂(t) of C1 x(t) over the infinite time horizon
[0, ∞) such that JF is negative for all nonzero w(t) ∈ L̃ 2 ([0, ∞); R p ).
It is readily found that

de(t) = Ax(t)dt + Dx(t)dν(t) + B1 w(t)dt − Ā x̂(t)dt

−L C̄2 e(t)dt − LC2 x(t)dt − L F x(t)dζ (t) − L D21 w(t)dt

or
de(t) = ( Ā − L C̄2 )e(t)dt + Ax(t)dt − LC2 x(t)dt

+ Dx(t)dν(t) − L F x(t)dζ (t) + (B1 − L D21 )w(t)dt. (7)

Denoting η(t) = col{x(t), e(t)} we obtain:

dη(t) = Ãη(t)dt + D̃η(t)dν(t)+ F̃η(t)dζ (t)+ B̃1 w(t)dt

where   
Ā+A 0 D0
à = , D̃ = , (8a-e)
A− LC2 Ā − L C̄2 D0
   
0 0 B1
F̃ = , B̃1 =
−L F 0 B1 − L D21

and
z̄(t) = C̃1 η(t),

where C1 = C1 − C̄1 and C̃1 = [C1 C̄1 ]. We thus arrive at the following result.
H∞ Stochastic State-Multiplicative Uncertain … 7

Theorem 1 Consider the system of (1a, b) and (2). For a prescribed scalar γ > 0,
a necessary and sufficient condition for JF of (6) to be negative for all nonzero
w(t) ∈ L̃ 2 ([0, ∞); R p ), is that there exist 0 < P ∈ R2n×2n , L ∈ Rn×m such that
the following inequality holds:
⎡ ⎤
à T P + P à + D̃ T P D̃ P B̃1 C̃1T F̃ T P
⎢ ∗ −γ 2 I 0 0 ⎥
⎢ ⎥ < 0. (9)
⎣ ∗ ∗ −I 0 ⎦
∗ ∗ ∗ −P

Remark 1 In the case where there is no uncertainty, A = Ā, C2 = C̄2 and C1 = C̄1 ,
the standard result of ([1], see p. 31) is retrieved. We note that in order to extract the
filter gain L, one must put a structure on the 2n × 2n Lyapunov function P, as appears
in the sequel. In case of uncertainty one may consider both types: norm-bounded and
polytopic-type uncertainties.
In the norm-bounded case we assume that:

A = Ā + A, C2 = C̄2 + C2

and [A, C2 ] = H δ(x, t)[E 1 E 2 ], ||δ(x, t)|| < 1. (10a-d)

We then have:    
Ā 0 H
à = + δ(x, t) E 1 0 +
0 Ā − L C̄2 H
 
0
δ(x, t) E 2 0 .
−L H

Using Young’s inequality ([2]) and (10d) the LMI solution readily follows. Note that
for simplicity we took C1 = 0.
Uncertainty of the polytopic type can also be solved for, around nominal values
of A and C2 (i.e. taking A = 0, C2 = 0). We assume that the system matrices
in (1a, b), (2) lie within the following polytope:

¯ = Co{
 ¯ 1, 
¯ 2 , ..., 
¯ N }, (11)

where

¯i =
 A(i) B1(i) C1(i) C2(i) D21
(i)
(12)

and where N is the number of vertices. In other words:

N N
¯ =
 ¯ i fi ,
 f i = 1 , f i ≥ 0. (13)
i=1 i=1
8 E. Gershon

Denoting, for the uncertain polytopic case,

 (i)   (i) 
Ā 0 D 0
Â(i) = , D̂ (i)
= ,
0 Ā(i) − L C̄2(i) D (i) 0
   
0 0 (i) B1(i)
F̂ (i) = , B̂ = (i) ,
−L F (i) 0 1
B1(i) − L D21

i = 1, 2, ..., N , we obtain the following result—the so-called ‘quadratic’

solution—by applying a single Lyapunov function over the whole uncertainty poly-
tope:

Theorem 2 Consider the system of (1a, b) and (2), where the system matrices lie
within the polytope of (11). For a prescribed scalar γ > 0, a necessary and sufficient
condition for JF of (6) to be negative for all nonzero w(t) ∈ L̃ 2 ([0, ∞); R p ), is that
there exist 0 < P ∈ R2n×2n , L ∈ Rn×m such that the following set of LMIs holds:
⎡ (i) ⎤
ϒ11 P B̂1(i) C̃1(i),T F̂ (i),T P
⎢ ∗ −γ 2 I 0 0 ⎥
⎢ ⎥ < 0. (14)
⎣ ∗ ∗ −I 0 ⎦
∗ ∗ ∗ −P

where
(i)
ϒ11 = Â(i),T P + P Ã(i) + D̂ (i),T P D̂ (i)

and where i = 1, 2, N . As in the case of Theorem 1 for the nominal case, in order
to extract the filter gain L one must apply a special structure to the above Lyapunov
function P. Taking the following structure:
 
P̄ α P̂
P= , α > 0,
α P̂ P̂

we obtain the following result, denoting Y = P̂ L,

Lemma 2 Consider the system of (1a, b) and (2) where the system matrices lie within
the polytope of (11). For a prescribed scalar γ > 0 and a given tuning parameter
α > 0, JF of (6) is negative for all nonzero w(t) ∈ L̃ 2 ([0, ∞); R p ), if there exist
0 < P̄ ∈ Rn×n , 0 < P̂ ∈ Rn×n , Y ∈ Rn×m such that the following set of LMIs hold:
⎡ (i) ⎤
ϒ̂11 P B̂1(i) C̃1(i),T F̂ (i),T P D̂ (i),T P
⎢ ∗ −γ 2 I 0 0 0 ⎥
⎢ ⎥
⎢ ∗ ∗ −I 0 0 ⎥<0 (15)
⎢ ⎥
⎣ ∗ ∗ ∗ −P 0 ⎦
∗ ∗ ∗ ∗ −P
H∞ Stochastic State-Multiplicative Uncertain … 9

(i)
for i = 1, 2, ...N where ϒ̂11 =
 (i) 
P̄ Ā + Ā(i),T P̄ α[ P̂ Ā(i) − Y C̄2(i) − Ā(i),T P̂]
.
∗ P̂ Ā(i) − Y C̄2(i) + Ā(i),T P̂ − C̄2(i),T Y T

5 Example

We consider the system of (1a, b) and (2) with

   
0 1±a 0 0.3
A= , D= ,
−1 0.4 0 −0.12
   
1 −0.5 1
B1 = and C1 = ,
−1 0 0

where C2 = [0.1 0.1] and a = 0.2. A Luenberger-type filter is sought that stabilizes
the system and achieves a negative JF for a minimum γ . Considering the nominal
case (i.e. a = 0) and applying the result of Lemma 2, where a single vertex is solved,
we find that the error system can be stabilized for α = 0.1 for a near minimum of
γ = 5.72, where the following block matrices of the Lyapunov function are obtained:
   
3.43 0.52 2.78 0.26
P̄ = , P̂ = .
0.52 2.94 0.26 1.71

Solving for the robust uncertain case, we apply the result of Lemma 2 (for two
vertices) and obtain a near minimum of γ = 7.54.

6 Conclusions

The theory of linear H∞ robust Luenberger-type filtering is developed for continuous-

time systems, where the stochastic uncertainty is encountered in both the dynamic
and the measurement matrices of the state space model of the system. Two uncertain
strategies are accounted for where the matrices uncertainties are either norm-bounded
or of the polytopic-type nature. In both cases, the solution is obtained by applying
the BRL result of stochastic state-multiplicative noisy systems to the error system.
In the norm-bounded case a single LMI is obtained, whereas a simple and readily
tractable set of LMIs is obtained for the polytopic case.
An inherent over-design is admitted to the polytopic-type solution due to the spe-
cial structure imposed on the Lyapunov function. The numerical example demon-
strates the efficiency and tractability of the proposed results.
10 E. Gershon

References

1. Gershon, E., Shaked, U., Yaesh, I.: H∞ control and Estimation of State-multiplicative linear
systems. Lecture Notes in Control and Information Sciences, vol. 318. LNCIS, Springer (2005)
2. Gershon, E., Shaked, U.: Advanced topics in control and estimation of state-multiplicative
noisy systems. Lecture Notes in Control and Information Sciences, vol. 439. LNCIS, Springer
(2013)
3. Chen, B.S., Zhang, W.: Stochastic H2 /H∞ control with state-dependent noise. IEEE Trans.
Autom. Control 49(1), 45–57 (2004)
4. Costa, O.L.V., Kubrusly, C.S.: State-feedback H∞ -control for discrete-time infinite-
dimensional stochastic bilinear systems. J. Math. Syst. Estim. Control 6, 1–32 (1996)
5. Willems, J.L., Willems, J.C.: Feedback stabilizability for stochastic systems with state and
control dependent noise. Automatica 12, 277–283 (1976)
6. Dragan, V., Morozan, T., Halanay, A.: Optimal stabilizing compensator for linear systems with
state-dependent noise. Stoch. Anal. Appl. 10, 557–572 (1992)
7. Dragan, V., Halanay, A., Stoica, A.: A small gain theorem for linear stochastic system. Syst.
Control Lett. 30, 243–251 (1997)
8. Hinriechsen, D., Pritchard, A.J.: Stochastic H∞ . SIAM J. Control Optim. 36(5), 1504–1538
(1998)
9. Dragan, V., Stoica, A.: A γ attenuation problem for discrete-time time-varying stochastic sys-
tems with multiplicative noise. Reprint series of the Institute of Mathematics of the Romanian
Academy, vol. 10 (1998)
10. Florchinger, P., Verriest, E.I.: Stabilization of nonlinear stochastic systems with delay feedback.
In: Proceedings of the 32nd IEEE Conferences on Decision and Control, San Antonio, TX, pp.
859–860, Dec. 1993
11. Chen, W.H., Guan, Z., Lu, X.: Delay-dependent exponential stability of uncertain stochastic
systems with multiple delays: an LMI approach. Syst. Control Lett. 54, 547–555 (2005)
12. Gershon, E., Shaked, U.: Robust H∞ output-feedback control of state-multiplicative stochastic
systems with delay. Int. J. Robust Nonlinear Control 21(11), 1283–1296 (2011)
13. Xu, S., Lam, J., Chen, T.: Robust H∞ control for uncertain discrete stochastic time-delay
systems. Syst. Control Lett. 51, 203–215 (2004)
14. Yue, D., Tian, E., Zhang, Y.: Stability analysis of discrete systems with stochastic delay and its
applications. Int. J. Innov. Comput. Inf. Control 5(8), 2391–2403 (2009)
Anticipating the Loss of Unknown Input
Observability for Sampled LPV Systems

Gustave Bainier, Jean-Christophe Ponsart, and Benoît Marx

Abstract Given a continuous-time Linear Parameter-Varying (LPV) system with

a sampled scheduling parameter and subject to an unknown input, this paper
provides—under some Lipschitz assumptions—an exact discretization of an extended
system which translates the sampled-data unknown input estimation problem into a
discrete-time LPV observer design problem with norm-bounded uncertainties. The
bounds developed in this process account for the inter-sample behavior of the schedul-
ing parameter, and allow for an estimation of some near-future observability Grami-
ans, from which it is possible to lower bound the number of samples for which the
unknown input is guaranteed to remain observable.

1 Introduction

The continuous-time Linear Parameter-Varying (LPV) representation is a powerful

tool to study a large class of nonlinear systems using a linear-like framework, which
generally facilitates controller and observer synthesis [16, 18]. It relies on the con-
struction of a scheduling parameter θ typically accounting for the non-linearities of
the initial system, and which is assumed to be continuously known or estimated in
real time. However, in practice, the values of θ are generally recovered at a fixed
sampling rate and are assumed to be held constant between two samples [20]. This
zero-order hold (ZOH) assumption made on the values of θ creates a sampled-data
problem which has already been discussed in the LPV literature [8, 15, 19].
The present paper considers a continuous-time LPV system subject to an unknown
input. Taking into consideration the sampled-data problem described above, an exact
discretization of an extended version of the system is given with consideration for the
uncertainties due to the ZOH assumption made on θ . This discretization effectively
translates the sampled-data unknown input estimation problem into a discrete-time
LPV observer design problem with norm-bounded uncertainties. To the authors’

G. Bainier (B) · J.-C. Ponsart · B. Marx

Université de Lorraine, CNRS, CRAN, 54000 Nancy, France
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 11

D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_2
12 G. Bainier et al.

knowledge, this is a new approach in the sampled-data literature. The exhibited

bounds on the uncertainties rely on some Lipschitz assumptions, which are similar
to the usual bounds put on the variations of θ that are typically found in the literature
of LPV systems. These bounds constitute a generalization of the ones found in [1],
the idea behind their proofs being similar.
Following the ideas of [1], the norm-bounded uncertainties of the exactly dis-
cretized extended-system allow for some real-time structural analysis. Specifically,
from the last known value of θ , the observability Gramian of the extended discrete-
time LPV system can be estimated between two future instants, as if the values of θ
were known in advance. In practice, this Gramian-estimation can be used to evaluate
in real time a number of samples for which the unknown input is guaranteed to remain
observable. To the authors’ knowledge, this is an uninvestigated question of the LPV
literature. Some results on the robust observability of discrete time varying systems
could arguably be leveraged to obtain similar bounds [14, 17], but the approach of
this paper is more straightforward, the uncertainties being treated here as ”yet to be
known” values.
The paper is organised as follows: Sect. 2 introduces the notations used throughout
the paper. Section 3 provides technical results, allowing for a smoother reading of
the next two sections. Section 4 translates the sampled-data unknown input estima-
tion problem into a discrete-time LPV observer design problem with norm-bounded
uncertainties. Section 5 estimates the near-future observability Gramians of the pre-
viously constructed system in order to lower-bound the number of samples left until
the unknown input observability of the initial system may be lost. Section 6 finally
provides an illustrative example. Some perspectives are discussed in Sect. 7 to con-
clude the paper.

2 Notations

N is the set of natural numbers. K stands for R or C, resp. the set of real and
complex numbers. K p×q denotes the set of matrices with p rows, q columns and
coefficients in K. · stands for the spectral norm. μ : K p× p → R denotes the
logarithmic norm induced by the spectral norm. Given a matrix E ∈ K p× p , E ∗
denotes its conjugate transpose, E positive-definite is denoted E  0, and by defini-
I +h E−1
tion μ(E) = lim h→0+ p h where I p ∈ K p× p stands for the identity matrix [6].
Given the following continuous-time system ẋ(t) = A(t)x(t) with A : R → Kn×n
a continuous function, the state transition matrix from t1 to t2 of such a system is
often abstractly denoted  (t2 , t1 ). However, this paper considers the state-transition
matrix to be a product integral, which is denoted in the following way [1, 4, 21]:


t2
 (t2 , t1 ) = e A(s)ds (1)
t1
Anticipating the Loss of Unknown Input Observability … 13

3 Preliminary Results

Among the many properties of the logarithmic norm and of the product integral,
several of them play a crucial role in this document.
Lemma 1 (Duhamel’s Formula) Let E, F : R → K p× p be two continuous func-
tions. For all t1 , t2 ∈ R
 t   

t2 
t2 t2 2 
u
e E(s)ds
− e F(s)ds
= e F(s)ds
(E(u) − F(u)) e E(s)ds
du (2)
t1 t1 t1 u t1

Proof See Theorem 5.1 of [4]. 

Lemma 2 Let E : R → K p× p be a continuous function. For all t1 , t2 ∈ R such that

t1 ≤ t2 t 
 2   t2
  μ(E(s))ds
 e E(s)ds  ≤ e t1 (3)
t 
1

Proof This result is a consequence of Grönwall’s lemma [3]. 

Lemma 3 If E : R → K p× p is a L E -Lipschitz function, then μ(E(·)) is also a L E -

Lipschitz function.

Proof For t1 , t2 ∈ R

μ(E(t1 )) = μ(E(t2 ) + E(t1 ) − E(t2 ))

≤ μ(E(t2 )) + μ(E(t1 ) − E(t2 )) (4)
⇒ μ(E(t1 )) − μ(E(t2 )) ≤ μ(E(t1 ) − E(t2 )) ≤ E(t1 ) − E(t2 )

interverting the role of t1 and t2 in the previous equations yields the same upper-bound.
Since E is L E -Lipschitz, combining (4) with its interverted counterpart provides

|μ(E(t1 )) − μ(E(t2 ))| ≤ E(t1 ) − E(t2 ) ≤ L E |t1 − t2 | (5)

which concludes the proof. 

Lemma 4 Let E : R → K p× p be a L E -Lipschitz function with for all t ∈ R, μ(E(t))

≤ σ and let τ be defined by τ = t0 + σ −μ(E(t
LE
0 ))
. For all t0 , t1 , t2 ∈ R such that t0 ≤
t1 ≤ t 2
⎧
t  ⎪e(t2 −t1 )σ if τ ≤ t1
2  ⎨
 E(s)ds  L
t2 σ −t1 (μ(E(t0 ))− 2E (2t0 −t1 ))− 2L1 (σ −μ(E(t0 ))+L E t0 )2
 e  ≤ e E if τ ∈ (t1 , t2 ) (6)
t  ⎪ ⎩
e(t2 −t1 )μ(E(t0 ))+ 2 L E (t2 −t1 +2t0 (t1 −t2 ))
1 2 2
1
if τ ≥ t2
14 G. Bainier et al.

Proof Lemma 3 provides

 t2  t2
μ(E(s))ds ≤ min(μ(E(t0 ) + L E (s − t0 ), σ )ds (7)
t1 t1

and since s → μ(E(t0 )) + L E (s − t0 ) is increasing, it follows

1
(7) ≤ min (s − t1 )μ(E(t0 )) + L E (s 2 − 2t0 s − t12 + 2t0 t1 ) + (t2 − s)σ
s∈(t1 ,t2 ) 2 (8)
= min as 2 + bs + c
s∈(t1 ,t2 )

The minimum of as 2 + bs + c being reached for τ = −b/2a, the sharpest upper-

bound of (7) depends on whether τ belongs to (t1 , t2 ) or not. The results are then
applied to Lemma 2, providing (6) and thus concluding the proof. 

Remark 1 The first and last upper-bounds of (6) stay true for all values of τ , contrary
to the second upper-bound which only holds when τ ∈ (t1 , t2 ).

Lemma 5 If E : R → K p×q and F : R → K p×r are L E and L F -Lipschitz functions

respectively, then for all k ∈ N, the following function is (L E + L F )-Lipschitz

G k : R → K(q+r +k)×(q+r +k)

⎡ ⎤
E(t) F(t) 0 (9)
t →⎣ 0 0 Ik ⎦
0 0 0

Proof Given t1 , t2 ∈ R, the following inequalities hold

⎡ ∗ ⎤ 
 E (t1 ) − E ∗ (t2 ) 0 
 ∗   
G k (t1 ) − G k (t2 ) = G k (t1 ) − G ∗k (t2 ) = sup 
⎣ F ∗ (t1 ) − F ∗ (t2 ) 0 ⎦ x 

x=1  0 0 
 ∗    ∗  
≤ sup  E (t1 ) − E (t2 ) 0 x1  + sup  F (t1 ) − F (t2 ) 0 x2 
∗ ∗
x1 =1 x2 =1

= E(t1 ) − E(t2 ) + F(t1 ) − F(t2 ) ≤ (L E + L F ) |t1 − t2 |

(10)
which concludes the proof. 

Lemma 6 Let E : R → K p× p be a L E -Lipschitz function with for all t ∈ R, μ(E(t))

≤ σ . For all t0 , t1 , t2 ∈ R such that t0 ≤ t1 ≤ t2 .
t   1   
 2  L t 2 − t12 + 2t0 (t1 −t2 ) e(t2 −t1 )σ ,
 (t2 −t1 )E(t0 )   2 E 2
 e E(s)ds
−e  ≤ min
e 2 L E (t2 −t1 +2t0 (t1 −t2 )) − 1 e(t2 −t1 )μ(E(t0 ))
1 2 2
 
t1
(11)
Anticipating the Loss of Unknown Input Observability … 15

Proof The proof of the first inequality can be found in [1]. The second inequality only
relies on the Lipschitz assumption on E. Given t0 , t1 , t2 ∈ R such that t0 ≤ t1 ≤ t2 ,
Lemma 1 provides


t2 
t2 
t2
e E(s)ds − e(t2 −t1 )E(t0 ) = e E(s)ds − e E(t0 )ds
t1 t1 t1
 t   u 
t2 2 
E(t0 )ds
= e (E(u) − E(t0 )) e E(s)ds
du
t1 u t1
  u 
t2 
(t2 −u)E(t0 )
= e (E(u) − E(t0 )) e E(s)ds
du
t1 t1
(12)
then, by submultiplicativity of the operator norm and the Lipschitz asumption on E
  u 
t2  
(t2 −u)μ(E(t0 ))  E(s)ds 
 (12)  ≤ L E e (u − t0 )  e  du (13)
t1  
t1

the third upper-bound of Lemma 4 is finally applied, providing

 t2 
(u − t0 )e 2 L E (u −2t0 u)
du e− 2 L E (t1 −2t0 t1 )+(t2 −t1 )μ(E(t0 ))
1 1
2 2
 (12)  ≤ L E
t
 1 
1 1 L E t2 (t2 −2t0 ) 1 1 L E t1 (t1 −2t0 ) − 1 L E t1 (t1 −2t0 )+(t2 −t1 )μ(E(t0 ))
= LE e2 − e2 e 2
LE LE
 1 
= e 2 L E (t2 −t1 +2t0 (t1 −t2 )) − 1 e(t2 −t1 )μ(E(t0 ))
2 2

(14)
which concludes the proof. 

Remark 2 If t1 < τ < t2 , with τ given in Lemma 4, a sharper (but very involved)
upper-bound to (11) can be found.

4 An Exact Discretization of Sampled LPV Systems

Subject to an Unknown Input

This section deals with the state and unknown input estimation of a sampled LPV
system of the following form

ẋ(t) = A(θ (t))x(t) + F1 (θ (t)) f (t)

(15)
y(t) = C(t)x(t) + F2 (t) f (t)
16 G. Bainier et al.

where x(t) ∈ Kn x is the state of the system, y(t) ∈ Kn y is its output, f (t) ∈ Kn f is
its unknown input and θ (t) ∈ Kn θ is a sampled scheduling parameter measured in
real time. Note that the coefficients of C(t) and F2 (t) are assumed to be known at all
time t, hence, these matrices are not θ -dependent. This section provides a discrete
representation of this system which takes into account the intrinsic uncertainties that
come with the sampling of θ .
Polynomial Unknown Input Assumption: f is a polynomial function of degree
q − 1 (hence for all t, the q-th derivative of f (t) is null).
A common way [5, 9, 23] to reconstruct the unknown input f (t) of (15) consists
in finding an observer for the augmented LPV system

ż(t) = A(θ (t))z(t)

(16)
y(t) = C(t)z(t)

⎡ ⎤ ⎡ ⎤
x(t) A(θ (t)) F1 (θ (t)) 0 ... 0
⎢ f (t) ⎥ ⎢ 0 0 In f ... ⎥0
⎢ ⎥ ⎢ ⎥
⎢ f (1) (t) ⎥ ⎢ .
.. .. .. .. ⎥..
with z(t) = ⎢ ⎥ A(θ (t)) = ⎢ . . . ⎥ , (17)
.
⎢ .. ⎥ ⎢ ⎥
⎣ . ⎦ ⎣ 0 0 0 . . . In f ⎦
f (q−1) (t) 0 0 0 ... 0
 
C(t) = C(t) F2 (t) 0 . . . 0 (18)

If the evolution of θ were exactly known between t1 and t2 , it would be possible to

perfectly find z(t2 ) based on the values of z(t1 ) and θ through the relation z(t2 ) =
 t2 A(θ(s))ds
t1 e z(t1 ).
Sampling Assumption: In practice, θ is measured at a sampling period Ts , and
its values are not only unknown in the future, but also unknown in-between two
samples.
For all k, m, n ∈ N with n ≥ m, θ (kTs ) now represents the last known value
of θ . The exact discretization of (16) accounting for these future and inter-sample
uncertainties is given by
 
z k+n = e(n−m)Ts A(θk ) + k,n,m z k+m
(19)
yk+m = C k+m z k+m

where k,n,m ∈ Kn z ×n z stands for a norm-bounded matrix of uncertainties (with n z =

n x + qn f ), and where for all p ∈ N, z p = z( pTs ), y p = y( pTs ), θ p = θ ( pTs ) and
C p = C( pTs ).
Remark 3 As a side note, notice how the peculiar structure of A(θk ) provides for
all s ∈ R  s A(θ )  s (s−u)A(θ )   
s A(θk ) e k 0e k F1 (θk ) 0 . . . 0 eu N du
e = (20)
0 es N
Anticipating the Loss of Unknown Input Observability … 17
 
0 I(q−1)n f
with N ∈ Kqn f ×qn f the nilpotent matrix N = .
0 0

If A(θ (·)) and F1 (θ (·)) are assumed to be L A -Lipschitz and L F1 -Lipschitz

respectively—which is a similar assumption to the usual bounds put on the variations
of θ that are typically found in the literature of LPV systems—then the following
bounds are easily deduced from Lemmas 5 and 6
   
k,n,m  ≤ e 21 ( L A +L F1 )(n 2 −m 2 +2k(m−n))Ts2 − 1 e(n−m)Ts μ(A(θk )) (21)

Moreover, if the logarithmic norm of A(θ (·)) is also upper-bounded by σ ∈ R—

which is a similar assumption to the usual bounds put on the values of θ (see Lemma
1 of [10])—then the following bounds are also verified

    
k,n,m  ≤ 1 L A + L F n 2 − m 2 + 2k(m − n) T 2 e(n−m)Ts σ (22)
1 s
2
Depending on the context, one of the previous upper-bounds might be sharper than
the other. One may also bound k,n,m inside a polytope, or more specifically between
two matrices in a component-wise approach, which can be achieved by shifting the
Lipschitz assumption from A(θ (·)) and F1 (θ (·)) to θ (·) directly.
Finally, to reconstruct the unknown input f (t) while taking into account the
intersample behavior of θ , it is possible to build a robust observer [7, 11–13, 22] for
the discrete-time uncertain LPV system (19) taken with n = 1 and m = 0.

5 Anticipating the Loss of Unknown Input Observability

In this section, the possible loss of unknown input observability of the system (15)
is anticipated using a sufficient condition of finite-time observability for (19) in
a near future. This sufficient condition is based on upper-bounding, in terms of
spectral norm, the difference between the observability Gramian of (19) and the
observability Gramian of (19) taken without its uncertainties (i.e. with k,n,m = 0).
These Gramians are denoted Wk and Wk respectively, where θk still stands for the
last known value of the scheduling parameter. On one hand, Wk evaluated between
(k + m)Ts and (k + n)Ts can be expressed by:

n−1 
  ∗  
∗

Wk,n,m = e(r −m)Ts A (θk ) + ∗k,r,m C k+r C k+r e(r −m)Ts A(θk ) + k,r,m
r =m
⎛ ⎞∗ ⎛ ⎞ (23)

n−1 (k+r
)Ts (k+r
)Ts
∗
= ⎝ e A(θ(s))ds ⎠
C k+r C k+r ⎝ e A(θ(s))ds ⎠

r =m (k+m)Ts (k+m)Ts
18 G. Bainier et al.


hence, Wk,n,m can only be fully computed if θ (t) is perfectly known for all t ∈
[(k + m)Ts , (k + n)Ts ], that is to say: in the future, and without sampling. On the
other hand, Wk evaluated between (k + m)Ts and (k + n)Ts can be easily computed
using:

n−1
∗ ∗
Wk,n,m = e(r −m)Ts A (θk ) C k+r C k+r e(r −m)Ts A(θk ) (24)
r =m

The augmented system (19) is observable between (k + m)Ts and (k + n)Ts if and

 if Wk,n,m  0 [2]. Moreover, following the ideas of [1], if the error ek,n,m =
only
(W − W )k,n,m  is upper-bounded by a constant Mk,n,m ∈ R, then


Wk,n,m − Mk,n,m In z  0 ⇒ Wk,n,m 0 (25)

Hence, Wk,n,m − Mk,n,m In z  0 is a sufficient condition of unknown input observ-

ability for (15) between (k + m)Ts and (k + n)Ts . The upper-bound Mk,n,m is easily
deduced from Lemmas 4 to 6, and the following inequality [1]:
⎛  ⎞
 (k+r 

n−1
 )Ts A(θ(s))ds    
ek,n,m ≤ ⎝ e  + e(r −m)Ts μ(A(θk )) ⎠ k,r,m  C k+r 2 (26)
 
r =m+1 (k+m)Ts 

This approach generalizes the upper-bound found in [1], and extends the resulting
structural analysis [1] to unknown input observability of sampled LPV systems.
Following from the previous results, a lower-bound to the number of samples for
which the unknown input is guatanteed to remain observable is given by

m ∗ (θk ) = max{m ∈ N : ∃n > m, Wk,n,m − Mk,n,m In z  0} (27)

where θk is still assumed to be the last known value of θ .

6 Illustrative Example

The following second order unknown input LPV system is considered

   
−10 0 θ (t)
ẋ(t) = x(t) + 1 f (t)
θ2 (t) −15 0
  (28)
41
y(t) = x(t)
15

where θ ∈ [0, 60] × [−60, 60] is assumed to be a scheduling vector (with L A = 2,

L F1 = 3) sampled at Ts = 0.03s, and the unknown input is assumed to be constant,
hence f (1) (t) = 0. The values of m ∗ (θk ) for (28) with k = 0 are plotted on Fig. 1.
Anticipating the Loss of Unknown Input Observability … 19

Fig. 1 m ∗ (θ), a lower-bound to the number of samples for which the unknown input observability
of (28) is guatanteed

Despite the conservativeness of the bounds used to compute m ∗ (θk ), some values of
θ guarantee the observability of the unknown input for at least the next 11 samples.
These results are encouraging since they were computed without consideration for
the structure of the extended system A(θk ), which tends to have a large logarithmic
norm μ(A(θk )).

7 Conclusion and Perspectives

In this paper, given a continuous-time LPV system with a sampled scheduling param-
eter and an unknown input, and under a Lipschitz assumption, the sampled-data
unknown input estimation problem has been translated into a discrete-time LPV
robust observer design problem. The bounds that were developed in the process
allowed for an estimation of some near-future observability Gramians, from which
a lower-bound to the number of samples for which the unknown input is guatanteed
to remain observable was exhibited. The obtained results could be further enhanced
by taking into consideration the structure of the extended system A(θk ), in particular
by using Eq. (20) in the computation of the Gramian estimation error. Moreover,
the extension of this work to sampled systems with a non-constant sampling period
remains to be investigated.
20 G. Bainier et al.

References

1. Bainier, G., Marx, B., Ponsart, J.-C.: Bounding the trajectories of continuous-time LPV systems
with parameters known in real time. In: 5th IFAC Workshop on Linear Parameter Varying
Systems (LPVS) (2022)
2. Callier, F.M., Desoer, C.A.: Controllability and Observability. The Discrete-Time Case, pp.
265–294. Springer, New York, NY (1991)
3. Desoer, C., Haneda, H.: The measure of a matrix as a tool to analyze computer algorithms for
circuit analysis. IEEE Trans. Circuit Theory 19(5), 480–486 (1972)
4. Dollard, J.D., Friedman, C.N.: Product Integration with Application to Differential Equations.
Cambridge University Press, Dec. 1984
5. Gao, Z., Ding, S.X., Ma, Y.: Robust fault estimation approach and its application in vehicle
lateral dynamic systems. Optim. Control Appl. Methods 28(3), 143–156 (2007)
6. Germund, D.: Stability and error bounds in the numerical integration of ordinary differential
equations. Kungl. tekniska hogskolans Handlingar. Almqvist & Wiksells, Uppsala (1959)
7. Han, W., Wang, Z., Shen, Y.: H-/L∞ fault detection observer for linear parameter-varying sys-
tems with parametric uncertainty. Int. J. Robust Nonlinear Control 29(10), 2912–2926 (2019)
8. Hetel, L., Fiter, C., Omran, H., Seuret, A., Fridman, E., Richard, J.-P., Niculescu, S.I.: Recent
developments on the stability of systems with aperiodic sampling: an overview. Automatica
76, 309–335 (2017)
9. Ichalal, D., Marx, B., Ragot, J., Maquin, D.: Simultaneous state and unknown inputs estimation
with PI and PMI observers for Takagi Sugeno model with unmeasurable premise variables. In:
2009 17th Mediterranean Conference on Control and Automation. IEEE, June 2009
10. Jungers, M., Deaecto, G.S., Geromel, J.C.: Bounds for the remainders of uncertain matrix
exponential and sampled-data control of polytopic linear systems. Automatica 82, 202–208
(2017)
11. Khan, A., Xie, W., Zhang, B., Liu, L.-W.: A survey of interval observers design methods and
implementation for uncertain systems. J. Frankl. Inst. 358(6), 3077–3126 (2021)
12. Li, J., Wang, Z., Shen, Y., Wang, Y.: Interval observer design for discrete-time uncertain Takagi-
Sugeno fuzzy systems. IEEE Trans. Fuzzy Syst. 27(4), 816–823 (2019)
13. Marx, B., Ichalal, D., Ragot, J.: Interval state estimation for uncertain polytopic systems. Int.
J. Control 93(11), 2564–2576 (2019)
14. Moheimani, S.R., Savkin, A., Petersen, I.: Necessary and sufficient conditions for robust observ-
ability of a class of discrete-time uncertain systems. In: Proceedings of 35th IEEE Conference
on Decision and Control. IEEE
15. Ramezanifar, A., Mohammadpour, J., Grigoriadis, K.: Sampled-data control of LPV systems
using input delay approach. In: IEEE Conference on Decision and Control. IEEE, Dec. 2012
16. Rugh, W.J., Shamma, J.S.: Research on gain scheduling. Automatica 36(10), 1401–1425 (2000)
17. Seo, J., Chung, D., Park, C., Jee, G.-I.: The robustness of controllability and observability
for discrete linear time-varying systems with norm-bounded uncertainty. IEEE Trans. Autom.
Control 50, 1039 – 1043, 08 (2005)
18. Shamma, J.S., Athans, M.: Guaranteed properties of gain scheduled control for linear
parameter-varying plants. Automatica 27(3), 559–564 (1991)
19. Tan, K., Grigoriadis, K., Wu, F.: Output-feedback control of LPV sampled-data systems. In:
Proceedings of the American Control Conference. ACC. IEEE (2000)
20. Tóth, R., den Hof, P.V., Heuberger, P.: Discretisation of linear parameter-varying state-space
representations. IET Control Theory Appl. 4(10), 2082–2096 (2010)
21. Volterra, V., Hostinský, B.: Opérations infinitésimales linéaires, applications aux équations
différentielles et fonctionnelles. Gauthier-Villars (1938)
Anticipating the Loss of Unknown Input Observability … 21

22. Vu, V.-P., Wang, W.-J., Zurada, J.M., Chen, H.-C., Chiu, C.-H.: Unknown input method based
observer synthesis for a discrete time uncertain T-S fuzzy system. IEEE Trans. Fuzzy Syst.
26(2), 761–770 (2018)
23. Youssef, T., Karimi, H.R., Chadli, M.: Faults diagnosis based on proportional integral observer
for TS fuzzy model with unmeasurable premise variable. In: 2014 IEEE International Confer-
ence on Fuzzy Systems (FUZZ-IEEE). IEEE, July 2014
Luenberger Observer Design for Robust
Estimation of Battery State of Charge
with Application to Lithium-Titanate
Oxide Cells

Eero Immonen

Abstract The State of Charge (SoC) of a lithium-ion battery cannot be measured

directly, and its estimation is one of the main challenges in electrification of transport,
among others. In this article, we introduce a systematic methodology for designing
Luenberger observers for on-line SoC estimation, based on a given second-order
Equivalent Circuit Model (ECM) for the battery. We show how the proposed method-
ology can be augmented with the Linear Quadratic Regulator (LQR) optimal control
framework, resulting in a computationally lightweight state estimator that also dis-
plays time optimal convergence. An application to Lithium-Titanate Oxide (LTO)
battery cells is presented to illustrate the proposed approach.

Keywords Battery · Equivalent circuit model · Luenberger observer · Linear

quadratic regulator · State of charge estimation · Lithium-titanate oxide cell

1 Introduction

Rechargeable lithium-ion batteries, typically composed of several interconnected

cells, are instrumental for modern energy storage and mobility applications. The
SoC of a battery cell is defined as the remaining capacity (Q r ) available relative to
the capacity available when fully charged (Q 0 ), i.e. SoC = Q r /Q 0 . Thus:

˙ dSoC(t) I (t) I (t)

SoC(t) = =η = , t >0 (1)
dt Q0 Qn

where SoC(0) ∈ [0, 1], I (t) is the current [A] through the battery (which by conven-
tion is negative in discharge conditions), η ∈ (0, 1] is the efficiency and Q n = Qη0 is
the nominal (or effective) capacity [As]. Accurate knowledge of the battery SoC at

E. Immonen (B)
Computational Engineering and Analysis Research Group, Turku University of Applied Sciences,
20520 Turku, Finland
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 23

D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_3
24 E. Immonen

all times is instrumental for efficient and safe operation, as well as for extending the
lifetime of the battery [1, 2]. Unfortunately, SoC cannot be directly measured from
the battery, but must be inferred from voltage and current measurements. The purpose
of this article is to present a method, and its theoretical justification, for designing
computationally lightweight Luenberger observers for on-line SoC estimation in the
LQR optimal control framework.
A simple and well-known open-loop procedure called Coulomb counting relies on
integrating Eq. (1) for estimating the SoC. However, this approach requires knowl-
edge of the initial condition SoC(0), which is reasonable in practice only when the
battery has been charged to full or discharged to empty. Moreover, in practice, the
real efficiency η < 1, that results from thermal losses, is seldom known accurately
and may change in time as the battery ages [3]. Further, the battery cell gradually
loses capacity, i.e. Q n decreases slowly, in both usage and storage [4]. Since Q n is
also slightly different for hot and cold temperatures [5], the SoC estimation problem
becomes a non-trivial one, and requires methods that are robust to the above effects.
The present knowledge on battery SoC estimation is summarized in the many
review articles, e.g. [2, 6–9]. Although sophisticated and remarkably accurate com-
putational methods, including those based on Machine Learning [10], have been
presented, the Battery Management System (BMS) hosting the SoC estimation algo-
rithm in practice is often a relatively simple circuit board with limited memory and
computing resources. Even the recurring matrix inversion operations required in
(time-varying) Kalman filtering may thus be computationally too expensive in prac-
tice [11, 12]. Luenberger-type observers, attempting to estimate the joint state of a
battery ECM (see Sect. 2) and the SoC model (1), have been suggested as computa-
tionally inexpensive alternatives in the literature [11, 13]. For strictly linear systems,
the state estimate would then be obtained via a single stabilizing matrix gain, tuned
offline, which is reasonable for low-resource BMS applications.
To address some controversy about the accuracy and need for trial-and-error tuning
for Luenberger observers [14, 15], in this article, we aim to:
1. Present rigorous yet easy-to-verify sufficient conditions for the Luenberger
observer gains to obtain an SoC estimate;
2. Show how tuning of the Luenberger gains can be carried out systematically in the
LQR framework, resulting in time-optimal SoC estimates;
3. Demonstrate in an application to battery cells with the LTO chemistry that SoC
estimation by the proposed method can be made accurate in the presence of small
modeling errors and nonlinearities.
In the proposed approach, the designed observer gain is a vector of 3 SoC-
dependent components, obtained offline from the solutions of a set of LQR problems.
In contrast to Kalman filtering, during battery operation, the SoC estimate can thus be
obtained by resolving the corresponding plant-observer dynamics without repeated
matrix inversion at every time step (analogous to using a steady state Kalman fil-
ter). We demonstrate the proposed approach for LTO cells (cf. Sect. 4), which are
important in many heavy-duty electric vehicle applications, among others [16].
Luenberger Observer Design for Robust Estimation … 25

2 Equivalent Circuit Model for the Battery

In this article, we consider batteries represented by second-order ECM systems, as

illustrated in Fig. 1. This ECM system is driven by the dynamic current [ A] input
u(t) = I (t). The system output V (t) models the battery terminal voltage [V ]. It
arises from the open-circuit voltage V0 > 0 [V ], the internal resistance Rs > 0 []
and V12 = V1 + V2 , which is the voltage [V ] across the dashed subcircuit consisting
of R1 > 0 [], C1 > 0 [F], R2 > 0 [] and C2 > 0 [F].
The following ECM-SoC system of equations represents that in Fig. 1:

ẋ(t) = Ax(t) + Bu(t), t > 0, x(0) = x0 (2a)

˙ u(t)
SoC(t) = , t > 0, SoC(0) ∈ [0, 1] (2b)
Qn
y(t) = Cx(t), t ≥ 0 (2c)
V (t) = V0 (t) + Rs u(t) + y(t), t ≥ 0 (2d)

with  
V1 (t)
x(t) = , u(t) = I (t), y(t) = V12 (t) (3)
V2 (t)

and
   1   0
− R11C1 0   V1
A= , B= C1
, C = 1 1 , x0 = (4)
0 − R21C2 1
C2
V20

In general, the parameters (4) depend on SoC, i.e. A = A(SoC), B = B(SoC) and
x0 = x0 (SoC). Also the open-circuit voltage V0 and internal resistance Rs depend
on the battery SoC, i.e. V0 = V0 (SoC) and Rs = Rs (SoC). We assume that all these
dependencies on SoC are represented by continuously differentiable functions. In
practice, see Sect. 4, this can be ensured by using a suitable interpolation scheme to

Fig. 1 Battery
representation by a
second-order ECM, with
voltage output V = V (t) and
current input I = I (t) [5,
17]
26 E. Immonen

parameter data reported at discrete SoC points. We emphasize that these parameter
dependencies on SoC can be identified off-line using standard dynamic test data (see
e.g. [5, 17]), and we assume in the remainder of this article that they are known.

3 Luenberger Observer Design

3.1 Preliminaries

For Luenberger observer design, we have to make SoC explicit in the output of
the dynamical system (2). To this end, without loss of generality, we propose to
decompose the open-circuit voltage as:

V0 (SoC) = V0a · SoC + V0b + V0r es (SoC) (5)

where V0a > 0, V0b > 0 and, of course, the residual term V0r es (SoC) can be nonlinear.
Further, we assume that the measured battery current Im (t) and the corresponding
terminal voltage Vm (t) are known at every t ≥ 0. Then we set u(t) = Im (t).
If ŝ(t) is an estimate of battery SoC at time t, then also the synthetic measurement

ym (t) = Vm (t) − V0b − V0r es (ŝ(t)) (6)

can be immediately calculated at every t ≥ 0. This estimate ŝ(t) is obtained from a

full state estimate for the system:

ż(t) = Az(t) + Bu(t), z(0) = z0 (7a)

v(t) = Cz(t) + Du(t) (7b)

with
     
A0 B   x(t)
A= , B = 1 , C = C V0a , D = Rs , z(t) = (8)
0 0 Qn SoC(t)

Note that a state estimate ẑ(t) for the system (7)–(8) yields the SoC estimate:

ŝ(t) = (0 1)ẑ(t) (9)

and, analogous to Eq. (2d), a prediction V p (t) for the battery voltage:

V p (t) = v̂(t) + V0b + V0r es (ŝ(t)) = Cẑ(t) + Du(t) + V0b + V0r es (ŝ(t)) (10)
= V̂0 (t) + Rs (ŝ(t))Im (t) + V̂1 (t) + V̂2 (t) (11)
Luenberger Observer Design for Robust Estimation … 27

If the state estimate was exact, i.e. ẑ(t) = z(t), so that x̂(t) = x(t) and ŝ(t) = SoC(t),
and if the voltage model (2d) was accurate, i.e. V (t) = Vm (t), then the predicted
voltage would coincide with the measurement: V p (t) = V0 (t) + Rs Im (t) + y(t) =
V (t) = Vm (t). We propose to obtain the estimate ẑ(t) by utilizing a Luenberger state
observer, as described in the following Subsections.

3.2 Main Results

Let L = (L 1 L 2 L 3 )T ∈ R3×1 . The Luenberger state observer for the system (7)–(8)
is defined using the synthetic measurement (6) as:
 
˙ = Aẑ(t) + Bu(t) + L(ym (t) − v̂(t)), t > 0, ẑ(0) = x̂(0) ŝ(0) T (12a)
ẑ(t)
v̂(t) = Cẑ(t) + Du(t), t ≥ 0 (12b)

We begin by demonstrating that state estimation errors for the system (12) arise
from modeling inaccuracies in V (t) and nonlinearities in V0 (SoC).
Lemma 1 If the voltage measurement is from (2d), i.e. Vm (t) ≡ V (t), if the residual
V0r es (SoC) ≡ 0 and if L can be chosen such that A − LC ∈ R3×3 is stable, i.e.
 
max{(λ) | det λI − (A − LC) = 0} < 0 (13)
 
then ẑ(t) − z(t) ≤ Me−ωt for some M > 0, ω > 0 and for all t ≥ 0.

Proof A direct calculation shows that, under the stated conditions, we have
ym (t) − v̂(t) = (V1 (t) − V̂1 (t)) + (V2 (t) − V̂2 (t)) + (V0a · SoC(t) − V0a · ŝ(t)) =
C(z(t) − ẑ(t)). The state estimation error e(t) = z(t) − ẑ(t) satisfies the differential
equation:

˙ = Az(t) + Bu(t) − Aẑ(t) − Bu(t) − L(ym (t) − v̂(t))

ė(t) = ż(t) − ẑ(t)
= A(z(t) − ẑ(t)) − LC(z(t) − ẑ(t)) = (A − LC)e(t) (14)

The eigenvalue condition (13) guarantees the existence of M > 0, ω > 0 e(t) ≤
Me−ωt for all t ≥ 0.

Under the best-case conditions of Lemma 1, the battery SoC estimate converges at
an exponential rate regardless of true SoC and the initial guess ŝ(0). In practice, there
are often modeling inaccuracies and nonlinearities in the voltage representations. We
aim to demonstrate in a practical application (Sect. 4) that accurate state estimation
is still possible. The following result shows that the condition (13) is easy to satisfy,
and it is the basis for observer design in the proposed methodology.

Proposition 1 If L 1 ≥ 0, L 2 ≥ 0 and L 3 > 0, then A − LC is stable.

28 E. Immonen

Proof Denote A1 = − R11C1 < 0 and A2 = − R21C2 < 0. A direct calculation shows
that the first column of the Routh array for A − LC consists of the following entries
Ri , i = 1, . . . , 4:

R1 = 1 > 0, R2 = L 1 − A1 − A2 + L 2 + L 3 V0a > 0, R4 = A1 A2 L 3 V0a > 0,

N
R3 = (15)
R2

where the numerator

N = −A21 A2 + A21 L 2 + A21 L 3 V0a − A1 A22 + 2 A1 A2 L 1 + 2 A1 A2 L 2 + 2 A1 A2 L 3 V0a

2
− A1 L 1 L 2 − A1 L 1 L 3 V0a − A1 L 22 − 2 A1 L 2 L 3 V0a − A1 L 23 V0a
2
+ A22 L 1 + A22 L 3 V0a − A2 L 21 − A2 L 1 L 2 − 2 A2 L 1 L 3 V0a − A2 L 2 L 3 V0a − A2 L 23 V0a
(16)

In the numerator expression (16), those terms with a minus sign are non-positive
(hence non-negative when multiplied by the minus sign), and those terms with a
plus sign are non-negative. Additionally, some terms, such as −A21 A2 , are strictly
positive. Hence R3 > 0. The result follows from Routh’s Test [18].

By Proposition 1, the exponential convergence of the SoC estimate does not

depend on B, and specifically on the battery nominal capacity Q n . Consequently, as
Q n slowly decreases during battery use and storage, the same observer designed for
a brand new battery still works, i.e. is robust with respect to Q n .
Although they establish the theoretical foundations for using Luenberger observers
for ECM-based SoC estimation, strictly speaking, the above results guarantee con-
vergence of the state estimate only in the absence of parameter variations in the
battery ECM. However, in practice, the ECM parameters only vary slowly relative
to the rate of convergence of the state estimate. This is demonstrated in Sect. 4 in
successful SoC state estimation for Lithium-Titanate Oxide battery cells. We aim to
complement these results in the future by considering linear time-varying systems
directly; see e.g. [19].

3.3 Designing the Observer Gain in the LQR Framework

Let Q ∈ R3×3 be Hermitian and let R > 0. Clearly the pair (A, C) in the system (7)–
(8) is observable, and we can thus design the gain L by considering the dynamical
system:
ṙ(t) = AT r(t) + CT u(t), t > 0, r(0) = r0 (17)

and the associated infinite-horizon LQR objective:

Luenberger Observer Design for Robust Estimation … 29

∞  
J (r) = r(s)T Qr(s) + Ru 2 (s) ds (18)
0

It is well-known (see e.g. [20]) that the state feedback control for the system (17) that
minimizes the value of the cost (18) is u(t) = −K r(t) where K = R −1 CX and X
is the unique solution of the Continuous-time Algebraic Riccati Equation (CARE):

AX + X AT − X CT R −1 CX + Q = 0 (19)

By duality, we can choose L = K T for the Luenberger state observer (12). Then also
A − LC is stable. Note that the only parameters chosen by hand in this approach
would be Q and R that place emphasis on different parts of the optimal closed loop
response through the objective (18).

3.4 Observer Design and Operation

The steps in the proposed Luenberger observer design procedure are as follows:
1. Identify the ECM model (2) parameter values (4) at a chosen set of points SoCi ∈
[0, 1].
2. Establish continuously differentiable functions V0 = V0 (SoC), Rs = Rs (SoC),
R1 = R1 (SoC), R2 = R2 (SoC), C1 = C1 (SoC), C2 = C2 (SoC) by piecewise
cubic Hermite polynomial interpolation (PCHIP).
3. Identify the open-circuit voltage decomposition (5) by linear regression.
4. Fix Q and R as in Sect. 3.3.
5. For another set of points SoCk ∈ [0, 1], identify the gain parameters Lk =
L(SoCk ) by solving the corresponding CARE (19) and apply PCHIP interpo-
lation to establish a continuously differentiable function SoC → L(SoC).
During operation, a new prediction ŝ(t j ) for SoC(t j ) is obtained via Eq. (9), by
numerically integrating the observer dynamics (12) using a table lookup at ŝ(t j−1 ).

4 Application: Lithium-Titanate Oxide Battery Cells

Battery cells with the LTO chemistry have been reported to withstand high cur-
rents, repeated cycling and operation in hot and cold temperatures without signif-
icant degradation [5]. This makes them prime candidates for Heavy-Duty Electric
Vehicle (HDEV) applications, including harvesters, mining dumpers, race cars and
trucks, among others. Recently, second-order ECMs for LTO cells were presented in
[5, 17], to understand the cells’ heat generation characteristics in dynamic discharge
conditions. Here we utilize those models and design computationally lightweight
parametric Luenberger observers for robust SoC estimation for the LTO cells.
30 E. Immonen

4.1 Model Parameters

In [5, 17], Immonen and Hurri carried out data-based second-order ECM modeling
for 23 Ah prismatic Toshiba SCiBTM LTO cells, and reported model parameter values
identified through Discharge Pulse Power Characterization (DPPC) tests at C-rate 1.
We utilize their published parameter data and PCHIP interpolation to find continu-
ously differentiable functions V0 (SoC), Rs (SoC), R1 (SoC), R2 (SoC), C1 (SoC) and
C2 (SoC) for the ECM. By regression, we further obtain V0a = 0.74 V and V0b = 1.86
V , with a nonlinear residual V0r es (SoC).

4.2 Observer Gain Design

For the present example application, let us choose Q = diag([0.01, 0.01, 0.5]) ∈
R3×3 and R = 1. Then the CARE (19) can be solved numerically in Matlab for any
given SoC ∈ [0, 1]. By discretizing the interval [0, 1] at resolution SoC = 0.01,
we obtain the gains Lk as shown in Fig. 2. Clearly the conditions of Proposition 1
are satisfied for every SoC value. Moreover, the three components of the vectors
Lk remain relatively constant across the entire SoC interval [0, 1]. A continuously
differentiable function SoC → L(SoC) is obtained by PCHIP interpolation.

Fig. 2 Numerical design of 0.4

the Luenberger observer gain L
1

L = (L 1 L 2 L 3 )T ∈ R3×1 by 0.35 L2

solution of the CARE (19) L

3
0.3

0.25
Observer gain

0.2

0.15

0.1

0.05

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
SoC
Luenberger Observer Design for Robust Estimation … 31

1 1
Measured Measured
0.9 Simulated 0.98 Simulated

0.8 0.96

0.7 0.94

0.6 0.92
SoC

SoC
0.5 0.9

0.4 0.88

0.3 0.86

0.2 0.84

0.1 0.82

0 0.8
0 0.5 1 1.5 2 2.5 3 3.5 4 0 100 200 300 400 500 600 700 800 900
Time [s] 104 Time [s]

(a) Full 4C DPPC test. (b) First 830 seconds of the 4C DPPC test.

Fig. 3 A SoC estimate for the 4C DPPC test compared to measurements [5, 17], using the false
initial SoC estimate ŝ(0) = 0.8. The true SoC(0) = 1 (fully charged)

4.3 SoC Estimation in 4C DPPC Test

For out-of-sample validation, we utilize the 4C DPPC test, which has not been used
in model identification [5, 17]. The test involves repeating the following four steps
10 times or until voltage is below 1.5 V: Constant-current (CC) discharge at 92 A for
10 s; Rest 290 s; CC discharge at 23 A for 320 s; Rest 60 minutes. In the laboratory
measurements, the true battery SoC was calculated by Coulomb counting, starting
from a fully charged cell.
Figure 3 displays the performance of the Luenberger observer (12), with the gain
designed in Sect. 4.2, under the false assumption that SoC(0) = 0.8. Clearly, the
SoC estimate remains accurate despite any modeling errors and nonlinearities. If the
initial SoC estimate is correct, i.e. ŝ(0) = SoC(0) = 1, then the root mean square
error (RMSE) of the SoC estimate is 4.32 · 10−3 , indicating good performance.

5 Conclusions

In this article, we have introduced a method for designing a Luenberger observer

for computationally lightweight on-line battery SoC estimation, based on a given
second-order ECM for the battery. We presented both a theoretical justification and
easy-to-verify design conditions for the gain parameters. These conditions do not
depend on the battery’s nominal capacity, and hence the proposed observer displays
some robustness with respect to battery aging. The proposed design methodology
connects naturally to the LQR optimal control framework, thus reducing the need for
trial and error in Luenberger observer gain tuning. As an application, we considered
SoC estimation for LTO chemistry cells which are interesting for heavy-duty electric
vehicle batteries, among others. The presented numerical results show insensitivity
32 E. Immonen

to a poor initial estimate and good accuracy over a standard dynamic discharge test
that has not been used in identifying the ECM parameters.
An important topic for future work is considering (slowly) time-varying ECMs
for batteries, and establishing more general sufficient conditions for Luenberger state
observer estimate convergence than those presented herein. Moreover, for practical
BMS implementation, the proposed methodology should be optimized with respect
to run-time energy consumption.

References

1. Yu, Z., Huai, R., Xiao, L.: State-of-charge estimation for lithium-ion batteries using a kalman
filter based on local linearization. Energies 8(8), 7854–7873 (2015)
2. How, D.N., Hannan, M.H. Lipu, Ker, P.J.: State of charge estimation for lithium-ion batteries
using model-based and data-driven methods: a review. IEEE Access 7, 136 116–136 136 (2019)
3. Gyenes, B., Stevens, D., Chevrier, V., Dahn, J.: Understanding anomalous behavior in coulom-
bic efficiency measurements on li-ion batteries. J. Electrochem. Soc. 162(3), A278 (2014)
4. Rabah, M., Immonen, E., Shahsavari, S., Haghbayan, M.-H., Murashko, K., Immonen, P.:
Capacity loss estimation for li-ion batteries based on a semi-empirical model. Anwendungen
und Konzepte der Wirtschaftsinformatik, no. 14 (2021)
5. Immonen, E., Hurri, J.: Incremental thermo-electric CFD modeling of a high-energy lithium-
titanate oxide battery cell in different temperatures: a comparative study. Appl. Thermal Eng.
197, 117260 (2021)
6. Zhou, W., Zheng, Y., Pan, Z., Lu, Q.: Review on the battery model and SOC estimation method.
Processes 9(9), 1685 (2021)
7. Ali, M.U., Zafar, A., Nengroo, S.H., Hussain, S., Junaid Alvi, M., Kim, H.-J.: Towards a smarter
battery management system for electric vehicle applications: a critical review of lithium-ion
battery state of charge estimation. Energies 12(3), 446 (2019)
8. Shrivastava, P., Soon, T.K., Idris, M.Y.I.B., Mekhilef, S.: Overview of model-based online
state-of-charge estimation using kalman filter family for lithium-ion batteries. Renew. Sustain.
Energy Rev. 113, 109233 (2019)
9. Wang, Y., Tian, J., Sun, Z., Wang, L., Xu, R., Li, M., Chen, Z.: A comprehensive review of
battery modeling and state estimation approaches for advanced battery management systems.
Renew. Sustain. Energy Rev. 131, 110015 (2020)
10. Vidal, C., Malysz, P., Kollmeyer, P., Emadi, A.: Machine learning applied to electrified vehicle
battery state of charge and state of health estimation: State-of-the-art. IEEE Access 8, 52 796–
52 814 (2020)
11. Lievre, A., Pelissier, S., Sari, A., Venet, P., Hijazi, A.: Luenberger observer for soc determina-
tion of lithium-ion cells in mild hybrid vehicles, compared to a kalman filter. In: 2015 Tenth
International Conference on Ecological Vehicles and Renewable Energies (EVER), pp. 1–7.
IEEE (2015)
12. Meng, J., Ricco, M., Luo, G., Swierczynski, M., Stroe, D.-I., Stroe, A.-I., Teodorescu, R.:
An overview of online implementable SOC estimation methods for lithium-ion batteries.
In: 2017 International Conference on Optimization of Electrical and Electronic Equipment
(OPTIM) & 2017 International Aegean Conference on Electrical Machines and Power Elec-
tronics (ACEMP), pp. 573–580. IEEE (2017)
13. Rahimi-Eichi, H., Chow, M.-Y.: Adaptive parameter identification and state-of-charge estima-
tion of lithium-ion batteries. In: IECON 2012-38th Annual Conference on IEEE Industrial
Electronics Society, pp. 4012–4017. IEEE (2012)
14. Tang, X., Liu, B., Gao, F., Lv, Z.: State-of-charge estimation for li-ion power batteries based
on a tuning free observer. Energies 9(9) (2016). https://www.mdpi.com/1996-1073/9/9/675
Luenberger Observer Design for Robust Estimation … 33

15. Manthopoulos, A., Wang, X.: A review and comparison of lithium-ion battery SOC estimation
methods for electric vehicles. In: IECON: The 46th Annual Conference of the IEEE Industrial
Electronics Society, vol. 2020, pp. 2385–2392. IEEE (2020)
16. Turku university of applied sciences e-rallycross car project. https://erallycross.turkuamk.fi/
en/main-page/. Accessed 10 Jan. 2022
17. Immonen, E., Hurri, J.: Equivalent circuit modeling of a high-energy lto battery cell for an
electric rallycross car. In: IEEE 30th International Symposium on Industrial Electronics (ISIE),
vol. 2021, pp. 1–5. IEEE (2021)
18. Ferrante, A., Lepschy, A., Viaro, U.: A simple proof of the routh test. IEEE Trans. Autom.
Control 44(6), 1306–1309 (1999)
19. Rotella, F., Zambettakis, I.: On functional observers for linear time-varying systems. IEEE
Trans. Autom. Control 58(5), 1354–1360 (2012)
20. Sontag, E.D.: Mathematical Control Theory: Deterministic Finite Dimensional Systems, vol. 6.
Springer Science & Business Media (2013)
Fault Detection and Diagnosis of PV
Systems Using Kalman-Filter Algorithm
Based on Multi-zone Polynomial
Regression

Yehya Al-Rifai, Adriana Aguilera-Gonzalez, and Ionel Vechiu

Abstract Faults must be timely diagnosed, detected, and identified to enhance pho-
tovoltaic (PV) system’s dependability. In this context, this paper presents a novel
Fault Detection and Diagnosis (FDD) methodology based on a hybrid combination
of model-based, through Kalman Filter (KF), and a statistical data-driven regression
approach for online monitoring of a PV system’s DC side. This statistical approach
is formulated on Multi-Zone non-linear Polynomial regression (MZP) techniques of
PV characteristics under Global Maximum Power Points (GMPP) at the array level.
In particular, the proposed method effectively detects intermittent soft Short-Circuit
(SC) even at very low irradiation. The performance of the proposed FDD methods
is evaluated via MATLAB/Simulink®considering varying weather conditions.

1 Introduction

With the growing concerns about the climate change, PV technology represents
an outstanding alternative carbon-free and sustainable energy generation to replace
depleted conventional resources.
Despite the fact that PV systems require low maintenance, they are still vulnerable
to faults that can occur without early warning due to the harsh working environment.
As indicated in [1], faults on the DC side have the severest impact that dramatically
deteriorates PV output performance. Thereby, this paper targets fault on the DC side,
in particular intermittent soft SC at extreme weather conditions.
Undoubtedly, protection systems are paramount to prevent catastrophic failures.
However, the conventional protection systems alone fail to preserve the PV required

Y. Al-Rifai (B) · A. Aguilera-Gonzalez · I. Vechiu

University of Bordeaux, Estia Institute of Technology, F-64210 Bidart, France
e-mail: [email protected]
A. Aguilera-Gonzalez
e-mail: [email protected]
I. Vechiu
e-mail: [email protected]
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 35
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_4
36 Y. Al-Rifai et al.

performance in abnormal circumstances due to several factors as: fault at low irra-
diation, hidden fault by the Maximum Power Points Tracking (MPPT), or nonlinear
Power-Voltage (P − V ) characteristics [2]. Thus, an FDD system is essential to
improve PV system dependability, while reducing maintenance costs and downtime.
Consequently, FDD has become a central research field in PV systems. For exam-
ple, a remote monitoring method is used in [3] to detect abnormal power dissipation
based on environmental data from satellite-based observations. This multi-sensor
method is undeniably less accurate than the on-site measurement, and not affordable
for small-scale projects. Besides, the environmental data are subject to noise and
interruption, leading to erroneous power yield estimation.
Recently, electrical-based FDD methods have become more commonly adopted,
as they require only the measurement of the output electrical signatures [4]. An inter-
esting FDD technique is performed in [5] based on Current-Voltage (I − V ) curve
analysis to detect several faults in a PV system. However, this offline method requires
a special tool for I − V curve checking and not suitable for online monitoring.
Another method is proposed in [6] to detect power lowering at the array level. This
method is based on the relationship between PV module temperature, current, and
voltage under GMPP conditions by using Polynomial Regression (PR) at Standard
fixed Temperature (ST) (25 ◦ C). Although, this method was confined to detect Partial
Shading (PS) at high irradiation. However, the I − V characteristics become highly
non-linear at low irradiation, where the simple cubic PR function is definitely not
accurate, resulting in false and missing alarm.
To overcome these issues, this paper proposes a novel FDD approach by com-
bining a statistical data-driven regression analysis and a model-based through KF,
for robustness against high uncertainties and measurement noises. This technique is
derived from [6] and developed to detect faults even at very-low irradiation based on
MZP interpolation according to the current level. Further, to improve the projection
function accuracy into ST, the average value of Temperature Coefficients (TCs) is
divided into Multi-Zone Average value (MZA) accordingly. Aiming to detect soft
faults, an adaptive threshold is proposed based on a PR in state of temperature vari-
able. To evaluate the performance of the proposed FDD approach, virtual intermittent
soft SCs fault are analyzed via MATLAB/Simulink®at various weather conditions.
The rest of this paper is organized as follows: Sect. 2 describes the PV model.
Section 3 outlines the performance of the proposed estimation methods. Then, the
FDD strategy with the KF is presented in Sect. 4. Next, the simulation and results
are addressed in Sect. 5. Finally, conclusion is given in Sect. 6.

2 PV Array Characteristics and Modeling

PV system naturally has nonlinear time-variant P − V characteristics that vary

according to irradiation and temperature [7]. Thus, PV modelling plays a vital role
in FDD to accurately emulate its I − V operating point at various weather condi-
tions. Several models have been addressed in the literature, but the most common
Fault Detection and Diagnosis of PV Systems … 37

Table 1 Soltech-1STH-350-WH PV module parameters

Parameter Value Units
Maximum power ( PM P P ) 349.59 W
Open circuit voltage (VOC ) 51.5 V
Short-Circuit current (ISC ) 9.4 A
Voltage at PM AX (VM P P ) 43 V
Current at PM AX (I M P P ) 8.13 A
Temperature coefficient of VOC (α) −0.36 %/◦ C
Temperature coefficient of I SC (β) 0.09 %/◦ C
Configuration 9 Series—21 Parallel

due to simplicity and efficacy is the single diode model with additional shunt resis-
tance in parallel [8]. This five-parameters model is adopted by ‘Simscape’ library
and employed in this paper to generate the fitting/test-data-set. For the simulation,
Soltech-1STH-350-WH is used, and its electrical parameters are listed in Table 1.
Indeed, a PV system strongly depends on weather conditions. The temperature influ-
ence on I and V under MPP condition is mathematically expressed in Eq. (1) as
follow:

Impp (T ) = ITr e f ,mpp + β(T − Tstc ) (1)

Vmpp (T ) = VTr e f ,mpp + α(T − Tstc )

where ITr e f ,mpp , VTr e f ,mpp denote the output current and voltage projected to ST,
respectively. T is the instantaneous temperature, and Tstc is the ST. α and β represent
the TCs for voltage and current, respectively. This paper employs the perturbation
and observation technique to track the MPP at varying weather conditions [8].

3 Regression Estimation Methods

3.1 Estimation Based on the PR Method

The PR method presented in [6] is able to detect power lowering based on a cubic
PR at ST with variable irradiation, as expressed in Eq. (2):


n
Vr e f,diag (I ) = bk ITn−k
r e f ,mpp
(2)
k=0

where Vr e f,diag is the reference diagnosis voltage that is later estimated by KF, bk
represents the coefficients of the polynomial function, ITr e f ,mpp is the instantaneous
38 Y. Al-Rifai et al.

Fig. 1 Loss function of V = f (I ) PR method under fitting and test-data at different degrees

projected MPP current to ST by Eq. (1). Indeed, to detect faults, Vr e f,diag is compared
with the projected VTr e f ,mpp by Eq. (1). Thereby, if a discrepancy exceeds a predefined
threshold, an alarm is then triggered indicating the occurrence of PS. However, the
method was tested at 600 W/m2 , and not performed against other faults. Further,
the authors had not proven the choice of the third-degree function over higher/lower
degree accuracy. Thus, this method could generate false alarm at low irradiation
where the I − V relationship is strongly nonlinear.
Therefore, this paper performs initially a loss function to evaluate the fitting accu-
racy of PR at different irradiation, and thereby, to extract the best fitting polynomial
degree that covers the whole irradiation scope. For that, a fitting data-set is firstly
generated using 1-single diode model with additional shunt resistance, at wide tem-
perature and irradiation range ([−5 : 50] ◦ C, [0 : 1000] W/m2 ). Subsequently, the
PR is evaluated by using RMSE (root-mean-square error), MAPE (mean absolute
percentage error), and R 2 (R-squared). Then, to avoid overfitting, the different PR
degrees are re-evaluated with another test-dataset. The loss function at different
degrees with the fitting/test-data is analyzed in Fig. 1.
Obviously, the test data reveals that the cubic voltage PR (V (I )) is bad-fitted with
an RMSE higher than 6.3 V. However, the Estimation Error (EE) slightly decreases
while increasing PR’s order (to reach less than 2 V at the 9th degree). Although, this
EE will lessen the effectiveness of FDD for detecting soft faults, beside the high PR
degree poses higher computational complexity.
In order to reach a better prediction accuracy with the least EE possible by accept-
able PR’s degree, an improved method is proposed based on voltage MZP.

3.2 MZP Regression Proposed Method

Aiming to improve the prediction accuracy of the PR, the MPP curve at ST is divided
into three-zones as depicted in Fig. 2. The first zone represents the variation of MPP at
Very Low irradiation (VLG) below 100 W/m2 . The second stands for the non-linear
Fault Detection and Diagnosis of PV Systems … 39

variation at Low irradiation (LG) below 400 W/m2 . The last zone represents the
MPP variation at High irradiation (HG). Since this method is irradiation sensor-less,
the zone is rather split with respect to current magnitude, which is an irradiation’s
image.
Remarkably, comparing with the PR method of [6], the RMSEs based on MZP
method have significantly dropped, as indicated by the legend box in Fig. 2. As
appreciated, the cubic MZP over all zones are more reliable than the 9th PR degree.
To emphasize the MZP’s robustness, likewise, the loss function is employed under
test-data to select the best fitting order of voltage MZP at each zone, as illustrated in
Fig. 3.
Obviously, the test-data confirm that the 3rd order voltage MZP is well-fitted with
trivial RMSE at high and low irradiation. Although, its RMSE at VLG is slightly
high. Whereas, the 4th degree is remarkably more reliable with higher estimation
accuracy over all 3-zones. Indeed, the latter outperform specially at VLG with RMSE
less than (0.9 V) comparing with the 3rd degree of (1.57 V). Thereby, the 4th order
is selected for all three regions. The voltage MZP function is represented in Eq. (3)
as follows:

Fig. 2 MZP of V(I) at ST divided into 3-zones (High, low, and very low irradiation)

Fig. 3 MZP goodness-fitting at different degree in each zone

40 Y. Al-Rifai et al.
⎧
⎪ 4 4−k
⎪ k=0 lk ITr e f ,mpp ∀I ≥ 68 A
⎪
⎪
⎪
⎪
⎨ 4 m k I 4−k 17A <I < 68 A
k=0 Tr e f ,mpp
Vr e f,diag (I ) =  (3)
⎪
⎪ 4 4−k
0 <I ≤ 17 A
⎪
⎪ k=0 n k I Tr e f ,mpp
⎪
⎪
⎩0, I = 0A

where, lk ,m k and n k denote the polynomial coefficients at high, low, and very low irra-
diation, respectively. Briefly, the MZP broadly outperform the PR estimation method,
with notably lower EE over all zones. Thus, due to its perfect prediction accuracy
covering all irradiation scope, the proposed MZP allows setting the threshold closer
to the reference diagnosis indicator, which will certainly increase the likelihood of
detecting soft faults even at low irradiation.
Although, this technique is a two-steps methodology where the instantaneous
MPPs voltage and current are initially projected to the ST by Eq. (1) based on TCs.
However, (α and β) have an average value, where the projection EE rises at extreme
weather conditions. Therefore, to address this issue, the values of TCs are divided
into MZA according to the current level.
Figure 4 illustrates the EE of the projection function Eq. (1) using an average value
of TCs in comparison with the proposed method based on MZA according to the
current level. As appreciated, with the proposed method the EEs for both current
and voltage projection functions have significantly reduced, even at extreme temper-
ature, making the projection to ST curve error-less. Generally, this overall proposed
methods significantly diminish the EE, leading to a higher prediction accuracy at
various weather conditions, that crucially improve the FDD performance.

Fig. 4 EE comparison between projection function using average value and MZA value at different
temperature
Fault Detection and Diagnosis of PV Systems … 41

4 Fault Detection and Diagnosis Method

Model-based fault detection is a simpler and cheaper alternative method than physical
redundancy [9]. This approach consists of two consecutive stages: residual genera-
tion and residual evaluation. The structure of the proposed FDD algorithm is shown
in Fig. 5. The PV configuration represents actually 9-series and 21-parallels connec-
tions.

4.1 Residual Generation by Using KF

For the residual generation, the recursive algorithm KF is employed due to its robust-
ness against high uncertainties and measurement noises in the stochastic framework
[10, 11]. Thus, two KFs are employed in this paper to estimate firstly the projection
variable VTr e f ,mpp and ITr e f ,mpp in Eq. (1), and then applied in FDD to estimate the
independent variable Vr e f,diag in Eq. (3). For that, the proposed MZP equation can
be formulated with the following state Eq. (4):

x(k + 1) = x(k) + ω(k)

y(k) = C x(k) (4)
z(k) = y(k) + υ(k)

where ,  and c are the state, noise process and output matrices, respectively. The
state vector x(k) represents the dependent and independent variable in Eq. (3), and
y(k) is the output vector. z(k) denotes the measured data. For a quick divergence
in faulty case, the dependent variable is estimated and not measured. ω and υ are
respectively the process and measurement noises. Indeed, the KF equation fall into
two phase process. The first represents the time update phase, where the KF algorithm

Fig. 5 General FDD scheme of the proposed diagnosis method based on MZP
42 Y. Al-Rifai et al.

predicts the state by projecting ahead the current state and error covariance to get a
priori estimate for the next time steps as expressed in Eq. (5):

x̂k− = x̂(k−1) (5)

Pk− = P(k−1)  + Q
T

where, x̂k− , Pk− are the priori state and covariance estimate at iteration (k) from
the previous iteration, (k − 1), respectively. Q is the process noise covariance. Sec-
ondly, the next phase represents the measurement update. In this phase, the priori
estimates are then corrected by the Kalman gain K k through the statistical filtered
noisy measurement to obtain the posteriori states as given in Eq. (6).

kk = P − C T (C Pk− C T + R)−1
x̂k = x̂k− + K k (z k − C x̂k− ) (6)
Pk = (ϒη − K k C)Pk−

where, x̂k is the update state, the covariance Pk is a positive definite symmetric matrix
with relatively large initial value P0 = diag(4) that declares low credibility for initial
guess. ϒη is an identity matrix. R is the measured noise covariance. The matrices Q
and R are diagonal initially predefined 10−4 [12].

4.2 Residual Evaluation

This stage examines the system’s health status by checking the consistency between
the projected measurement VTr e f ,mpp and the predicted value Vr e f,diag . If the discrep-
ancy exceeds a defined threshold, a fault alarm is triggered. Aiming to minimize the
false and missing alarms by addressing the projection function and MZP uncertain-
ties, this paper proposes an adaptive threshold based on a quadratic PR in state of
temperature variable, as formulated in Eq. (7):


2
M A P E V (T ) = ¯ k T 2−k (7)
k=1

where, T is the instantaneous-measured temperature and ¯ k is the polynomial coef-

ficient based on the EE of MZP and projection function at different temperature, as
calculated previously. Thus, the health status can be examined by Eq. (8):

(1 − M A P E V (T ))Vr e f,diag (k) ≤ VTr e f ,mpp ≤ (1 + M A P E V (T ))Vr e f,diag (k) (8)

Fault Detection and Diagnosis of PV Systems … 43

Thereby, to trigger a true alarm, a Boolean fault vector f (V ) ∈ {0, 1} is settled

according to Eq. (9):

1 i f |VTr e f ,mpp | > (1 ± M A P E V )Vr e f,diag
fV = (9)
0, other wise

Therefore, if the fault signature is 1, an alarm is generated indicating the occur-

rence of faults. Otherwise, the PV system is considered in healthy state.

5 FDD Simulation Results

To investigate the robustness of the proposed FDD method, simulations under healthy
and faulty cases are performed through MATLAB/Simulink®. The diagnosis test is
performed under test-data to investigate the reliability of the projection function at
variable temperature and the robustness of PR and MZP at different irradiation zone
(high, low, and very low irradiation), as shown in Fig. 6.

5.1 Healthy Scenario

This scenario considers a normal condition of the PV system. A comparison between

PR and MZP FDD methods under healthy state is illustrated in Fig. 7. As appreciated,
the proposed MZP threshold in state of temperature variable is inline and much closer
to the reference value than the PR threshold based on a constant MAPE. Thus, the
latter will undoubtedly conceal soft anomalies. Further, due to the high estimation
reliability along with the threshold robustness, the reference value VTr e f ,mpp is always
within the boundaries (upper and lower MZP threshold) defined in Eq. (8) using MZP

Fig. 6 Temperature and irradiation profiles during simulation in diagnosis test

44 Y. Al-Rifai et al.

Fig. 7 Healthy scenario: comparison between PR and MZP FDD methods

method. Accordingly, the fault vector f v MZP remains 0 during the simulation,
confirming a healthy state.
In contrast, due to the bad fitting of the cubic PR function along with the low
threshold reliability, the discrepancies are relatively large. Consequently, the thresh-
old gets violated in almost all zones using PR methods. In particular, at 0 W/m2 , the
threshold sharply deviates from the reference value, since the PR’s intercept is dif-
ferent from 0. Thereby, false alarms are generated. In general, the MZP FDD method
is more accurate and robust.

5.2 Faulty Scenario: Intermittent Soft Short Circuit

Further, to examine the performance of the two FDD methods under abnormal behav-
ior, artificial intermittent soft SCs are induced at each zone, as depicted in Fig. 8. The
time duration of the intermittent fault is considered 2 s, with a soft magnitude of
Rsc = 50 . The SC current (marked in solid red line) magnitude varies between
([6.7 : 7.5] A) during the diagnosis test, that represents 4.2% at HG and more than
50% at LG of the total MPP current generated. Remarkably, due to its high accuracy
the threshold gets violated each time the SCs faults are activated into the PV plant
using MZP FDD method, despite its soft magnitude at HG, as illustrated in Fig. 8.
Thereby, true alarms are generated accordingly, indicating the presence of abnor-
mal behavior. On the other side, due to its low estimation reliability along with the
threshold uncertainties, the PR method is not able to reveal the soft anomalies in the
PV plant at HG. In addition, at LG and VLG The PR fault signals has the same false
alarm profile as under normal condition due to the bad fitting accuracy. Exception-
ally, a true alarm is generated at t = 85 s where the hard SC current become relatively
large, representing 100% of total MPP current. Generally, it is possible to conclude
that the overall proposed methods are more reliable and robust, that able to detect
intermittent very soft SC faults even at VLG.
Fault Detection and Diagnosis of PV Systems … 45

(a) .

(b) Fault signature.

Fig. 8 Faulty scenario: Comparison between PR and MZP FDD method under soft SCs a SC
current magnitude and fault signals, b voltage diagnosis indicator

6 Conclusion

This paper proposed a novel FDD method based on a KF formulated on MZP

approach to detect faults in PV systems. The regression method uses the relationship
between voltage, current, and temperature under GMPP with irradiation sensorless.
In addition, this paper proposes a MZA TCs, according to the current level, to improve
the accuracy of the projection functions. Further, an adaptive threshold based on a
quadratic PR in state of temperature variable is formulated to precisely examine the
health status of the PV plant. The result reveals the effectiveness of the proposed
FDD method to detect intermittent soft SC faults even at very-low irradiation while
minimizing missing and false alarms.

References

1. Gunduz, H., Jayaweera, D.: Reliability assessment of a power system with cyber-physical
interactive operation of photovoltaic systems. Int. J. Electrical Power Energy Syst. 101, 371–
384 (2018)
2. Singh, B., Goyal, S., Kumar, P.: Solar PV cell materials and technologies: analyzing the recent
developments. Mater. Today: Proc. 43, 2843–2849 (2021)
46 Y. Al-Rifai et al.

3. Drews, A., Keizer, A., Beyer, H., Lorenz, E., Betcke, J.: Monitoring and remote failure detection
of grid-connected PV systems based on satellite observations. Solar Energy 81(4), 548–564
(2007)
4. Li, B., Delpha, C., Migan, A., Diallo, D.: Fault diagnosis of photovoltaic panels using full I–V
characteristics and machine learning techniques. Energy Convers. Manag. 248 (2021)
5. Li, B., Migan-Dubois, A., Delpha, C., Diallo, D.: Evaluation and improvement of IEC 60891
correction methods for I-V curves of defective photovoltaic panels. Solar Energy 216, 225–237
(2021, 3)
6. Kang, B., Kim, S., Bae, S., Park, J.: Diagnosis of output power lowering in a PV array by using
the kalman-filter algorithm. IEEE Trans. Energy Convers. 27, 885–894 (2012)
7. Barth, N., Jovanovic, R., Ahzi, S., Khaleel, M.: PV panel single and double diode models:
optimization of the parameters and temperature dependence. Solar Energy Mater. Solar Cells
148, 87–98 (2016)
8. Jendoubi, A., Tlili, F., Bacha, F.: Sliding mode control for a grid connected PV-system using
interpolation polynomial MPPT approach. Math. Comput. Simul. 167, 202–218 (2020)
9. Blanke, M., Kinnaert, M., Lunze, J., Staroswiecki, M.: Diagnosis and Fault-Tolerant Control.
Springer, Berlin, Heidelberg (2016)
10. Isermann, R.: Fault-diagnosis systems: an introduction from fault detection to fault tolerance.
Springer (2006)
11. Kalman, R.: A new approach to linear filtering and prediction problems. J. Basic Eng. 82,
35–45 (1960)
12. Lynch, C., OMahony, M., Scully, T.: Simplified method to derive the kalman filter covariance
matrices to predict wind speeds from a NWP model. Energy Procedia 62, 676-685 (2014)
Parity-Space and Multiple-Model Based
Approaches to Measurement Fault
Estimation

Ivo Punčochář and Ondřej Straka

Abstract The paper compares two approaches to measurement fault estimation for
linear discrete-time stochastic systems. The first fault estimator utilizes the parity-
space approach. It assumes that at most a limited number of components of the
measurement fault can be non-zero. A fault is detected using the χ 2 test applied to
the parity-space based residuals and then the indices of non-zero components and
fault itself are estimated. The second fault estimator is based on the multiple-model
approach. The space of measurement faults is quantized to construct a set of models
and a variable structure interacting multiple-model estimator is employed to estimate
the state and the measurement fault. Both approaches are compared in a numerical
example.

1 Introduction

Fault diagnosis is an established research field [5, 8] with many applications [3,
6]. The approaches to fault diagnosis differ in the level of the information that is
inferred from the available data. The least information is provided by simple fault
detection (a binary decision regarding fault existence), followed by fault isolation
(determination of which fault has occurred) and fault estimation (estimation of fault
signal time profile) provides the most information about a fault. The approaches to
fault diagnosis can roughly be divided into parity-space based, observer based, and
parameter estimation based. Although the basic fault detectors based on parity-space
and observer approaches do not inherently provide a fault estimate, there are several
extensions in this direction, see e.g. [4, 10–12]. On the other hand, fault detectors
based on parameter estimation inherently provide a fault estimate [5].

I. Punčochář (B) · O. Straka

University of West Bohemia, Univerzitní 8, 301 00 Pilsen, Czech Republic
e-mail: [email protected]
O. Straka
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 47

D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_5
48 I. Punčochář and O. Straka

The aim of this paper is to compare two selected fault estimators. The first one is
based on the parity-space approach and represents an extension of an idea presented
in [9] to simultaneous faults in more that one component of measurement fault. The
second one is based on multiple-model estimator and utilizes a variable structure
interacting multiple-model (VSIMM) [7] to reduce computational demands related
to a large number of models.
The paper is organized as follows. Section 2 deals with the problem formulation.
The parity space based fault estimator is described in Sect. 3. Section 4 deals with
the fault estimator that is based on the VSIMM estimator. Both fault estimators are
compared through a numerical example in Sect. 5. Section 6 concludes the paper.

2 Problem Formulation

Let us assume that a monitored system can be described at each time instant k ∈ T =
{0, 1, . . .} by the following linear time-invariant discrete-time stochastic model1

xk+1 = Axk + Gwk , (1)

yk = Cxk + Ffk + Hvk , (2)

where xk ∈ Rn x is a state, wk ∈ Rn w is a state noise, yk ∈ Rn y is a measurement,

fk ∈ F ⊆ Rn f is a measurement fault, and vk ∈ Rn v is a measurement noise. The
state noise and the measurement noise are assumed to be mutually independent white
noises with the standard Gaussian distributions. The initial condition x0 is assumed
to be independent of both noise processes and it has the Gaussian distribution with
the mean x̂0|−1 and covariance matrix P0|−1 . The matrices A ∈ Rn x ×n x , G ∈ Rn x ×n w ,
C ∈ Rn y ×n x , F ∈ Rn y ×n f , H ∈ Rn y ×n v are assumed to be known. The following three
assumptions are made
A1: n y > rank(C) ≥ 1,
A2: rank(H) = n y ,
A3: rank(F) = n f .
The assumption A1 means that the matrix C differs from the zero matrix and it is
possible to design a residual generator using parity-space approach even if only the
measurement equation is considered. The assumption A2 ensures that no linear com-
bination of the components of the measurement yk is free of the measurement noise
vk . The assumption A3 means that the model of the fault is not overparametrized.
We do not have particular assumption regarding the measurement faults fk except
that fault-free behaviour is represented by zero value of the measurement fault and
non-zero value signifies a measurement fault.
The aim is to design a fault estimator of the form

1 Note that a time-varying model and a known input to the monitored system can also be considered
at the expense of more complicated notation.
Parity-Space and Multiple-Model Based Approaches … 49
 
f̂k = σ k y0k , (3)

where f̂k ∈ Rn f is an estimate of the measurement fault fk , σ k : R(k+1)n y → Rn f is a

function that describes the fault estimator,2 and y0k = [y0T , y1T , . . . , ykT ]T is a sequence
of measurements obtained up to time instant k.

3 Measurement Fault Estimator Based on Parity-Space

Approach

The first fault estimator uses the parity-space approach. It is an extension of an

idea presented in [9] to estimating measurement faults with more than one non-zero
component. A linear residual generator that uses only the current measurement yk to
generate a residuum rk is given as

rk = Wyk , (4)

where rk ∈ Rnr is the residuum and W ∈ Rnr ×n y is a full row rank matrix with
n r = n y − rank(C) that satisfies the following two conditions

WC = 0nr ×n x , (5)
WHH W = Inr ×nr .
T T
(6)

Since only the mean value of the residuum rk depends on the fault fk , a decision
generator can make a binary decision using the following rule

0 if Sk ≤ Sth
dk = , (7)
1 if Sk > Sth

where dk ∈ {0, 1} is a decision about fault existence (dk = 1 indicates a fault), Sk =

rkT rk ∈ [0, ∞) is the χ 2 test statistic, and Sth ∈ [0, ∞) is a selected threshold.3 If
fk = 0n f ×1 , the statistic Sk has the central χ 2 distribution with n r degrees of freedom.
For a non-zero fault fk , the statistic Sk has the non-central χ 2 distribution with n r
degrees of freedom and the parameter of non-centrality λk = fkT FT WT WFfk . The
threshold Sth can be computed as

Sth = F −1 (1 − Pfa , n r ) , (8)

2 Note that a particular fault estimator might not necessarily use the whole sequence of past mea-
surements to generate the fault estimate.
3 Although the condition (6) can be omitted, a more computationally demanding expression for the

statistic Sk = rkT (WHHT WT )−1 rk must be evaluated on-line.

50 I. Punčochář and O. Straka

where F −1 : [0, 1] × N → [0, ∞] is the inverse cumulative distribution function of

the central χ 2 distribution with n r degrees of freedom and Pfa ∈ [0, 1] is a desired
probability of the false alarm.
If the decision dk = 0 is generated, the estimate of the measurement fault is
simply set to zero, i.e., f̂k = 0n f ×1 . If decision dk = 1 is generated, a fault estimate
f̂k is computed with the assumption that at most n ∗ ∈ I = {1, . . . , n f } components
of the measurement fault fk are non-zero. A reasonable value n ∗ is proposed in the
following discussion and the index k is omitted for notational convenience. First, it
should be noted that the subspace of undetectable faults is given as
 
Fus = f ∈ Rn f : Mf = 0 , (9)

where M = WF ∈ Rnr ×n f and rank(M) = m ≤ min(n r , n f ). If the number of non-

zero components of f is greater than m, there exist non-zero undetectable faults.
However, the converse is not true and there can be non-zero undetectable faults even
if the number of non-zero components of f is less than or equal to m.4 Therefore, the
remainder of the analysis considers n ∗ ≤ m. For a particular n ∈ Ired = {1, . . . , m},
let us denote the i-th n-tuple of indices of non-zero components of f as

si = [si,1 , si,2 , . . . , si,n ]T , (10)

where si, j ∈ I, si, j = si, j for j = j , j, j ∈ {1, . . . , n}, and i ∈ Icomb = {1, . . . ,
Nn }. The number of different n-tuples is denoted Nn and it is given by the combination
number
 
nf nf!
Nn = = . (11)
n n!(n f − n)!

The measurement fault for the i-th n-tuple can be written as

n
f= esi, j αi, j , (12)
j=1

where esi, j is the si, j -th column of the identity matrix In f ×n f and αi, j is the signed
magnitude of the fault in the si, j -th component of f. The the mean value of the
residuum can be written as
n
E{r} = Mf = Mesi, j αi, j = Mi α i , (13)
j=1

4 Note that there are no undetectable faults if M has the full column rank, i.e., m = n f .
Parity-Space and Multiple-Model Based Approaches … 51

where Mi = [[M]:,si,1 , [M]:,si,2 · · · , [M]:,si,n ] ∈ Rnr ×n is a matrix5 consisting of

selected columns of M and α i = [αi,1 , αi,2 , · · · , αi,n ]T . If the fault given by the i-th
n-tuple has truly occurred, then the mean value of the residuum lies in the subspace
Ri that is spanned by the columns of Mi . Thus, we are interested in relationships
between subspaces Ri for i = 1, . . . , Nn .
Since the condition Ri ∩ R j = {0} is too strong to be satisfied except for n = 1, it
seems practically reasonable to consider the i−th and j−th n-tuple to be marginally
distinguishable if

dim(Ri ∩ R j ) < min(dim(Ri ), dim(R j )). (14)

This condition admits the existence of the mean value of residuum that can result
from different multiple faults. The greatest n, for which the matrices Mi has the full
column rank and the above condition is satisfied for all pairs i, j ∈ Icomb , is denoted
as n ∗ .
For a selected n ∗ , we can estimate the index j of the n ∗ -tuple of non-zero elements
as the one that minimizes the distance of the residual vector r to the subspaces R j
for j = {1, . . . , Nn ∗ }
 −1 T
ĵ = arg min rT I − M j MTj M j M j r. (15)
j∈{1,...,Nn ∗ }

If a tie occurs in the minimization, it can be interpreted as that some n ∗ -tuples cannot
be distinguished. Once ĵ is determined, the estimate of the fault f can be computed
as

f̂ = M ĵ α̂, (16)

where α̂ k is a least square estimate

−1
α̂ = MTĵ M ĵ MTĵ r. (17)

This fault estimator requires an additional assumption in the form of the maximum
number of simultaneous faults. It should improve the quality of the fault estimates if
the assumption is satisfied. When faults occur in more measurements than assumed,
there are no guarantees regarding the quality of the fault estimate.

5 Notation [M]:,i denotes the i-th column of the matrix M.

52 I. Punčochář and O. Straka

4 Measurement Fault Estimator Based on Multiple-Model

Approach

The multiple-model approach to fault estimation requires a set of models that repre-
sent fault-free and faulty behaviours of the system. A discrete set F̄ ⊂ F of possible
values of the measurement faults can be set up as the Cartesian product of discrete
sets for individual dimensions of the measurement fault

F̄ = F̄1 × F̄2 × . . . × F̄n f , (18)

where F̄i = f¯i(1) , f¯i(2) , . . . , f¯i(Mi ) for each i ∈ {1, 2, . . . , n f } and f¯i(li ) ∈ R denotes
the discrete value of the i-th component of the fault vector. It is assumed that
0 ∈ F̄i and f¯i(i ) < f¯i(i +1) for  each i ∈ {1, . . . , Mi − 1}. The discrete set F̄ =
nf
{f̄ (1) , f̄ (2) , . . . , f̄ (M) } with M = i=1 Mi creates a regular grid of points in the n f -
dimensional space. The selection of the grid is driven by the following two factors.
The set F̄ should cover the part of the space F where the faults are expected to lie
and the spacing of the regular grid should be chosen to provide a desired accuracy
of the fault estimates. Thus, the measurement model is

yk = Cxk + Ff(μk ) + Hvk , (19)

where μk ∈ M = {1, . . . , M} is a model index and f : {1, . . . , M} → F̄ is a bijective

function defined as

f( j) = f̄ j . (20)

We also complement the original dynamical model with a dynamical model that
describes time behaviour of the measurement fault. For each dimension of the mea-
surement fault, we consider an auxiliary Markov chain
⎧
⎪
⎪ p if i = i − 1,
⎪
⎨p
(i ) if i = i + 1,
P(νi,k+1 = f¯i(i ) |νi,k = f¯i ) = (21)
⎪1 − 2 p
⎪ if i = i ,
⎪
⎩
0 otherwise.

If we assume that the switching is independent among the dimensions, the transition
probabilities of a complete Markov chain are given as
Parity-Space and Multiple-Model Based Approaches … 53

( ) ( ) ( )
P(ν1,k+1 = f¯1(1 ) , . . . , νn f ,k+1 = f¯n f f |ν1,k+1 = f¯1 1 , . . . , νn f ,k+1 = f¯n f f ) =
n n


nf
( )
P(νi,k+1 = f¯i(i ) |νi,k = f¯i i ).
i=1
(22)

This transition probability matrix could be far from sparse, which is a key factor
for lowering computational demands of the multiple-model estimator. Therefore, the
transition probabilities that corresponds to a change in more then one component
of the fault vector are forced to zero and the remaining non-zero probabilities are
renormalized. Finally, we define the transition probabilities of the Markov chain for
μk ∈ {1, . . . , M} using the bijective function f as

P(μk+1 |μk ) = P(ν k+1 = f(μk+1 )|ν k = f(μk )). (23)

The initial probability P(μ0 ) of this Markov chain is usually selected such that the
fault-free case gets the probability one. A multiple-model estimator can be used
to compute an estimate the model index μk based on all available measurements
y0k . However, the standard multiple-model estimators like GPB or IMM cannot be
used due to a high number of models. This issue can be avoided to a certain degree
by using the VSIMM estimator [7] that is an extension of the standard interacting
multiple-model (IMM) estimator [2]. It considers the original set M and its subset
Ma,k that includes models assumed to be active at the time instant k. At each time
instant k, the VSIMM performs the following three steps:
• Filtering step—The standard filtering step of the IMM algorithm is used to compute
pdf p(xk , μk |y0k , Ma,k ) using the current measurement yk and the predictive pdf
p(xk , μk |yk−1 , Ma,k−1 , Ma,k ).
• Adaptation step—The filtering model probabilities P(μk |y0k , Ma,k ) and the transi-
tion probabilities are used to design set Ma,k+1 that is an extension or reduction of
the set Ma,k . The parameters that govern the adaptation are the minimum number
of active models Nmin , a lower thresholds PL and an upper threshold PU to which
the filtering model probabilities are compared.
• Predictive step—The standard predictive step of the IMM algorithm is used to
compute the predictive pdf p(xk+1 , μk+1 |yk , Ma,k , Ma,k+1 ) using the filtering
pdf p(xk , μk |y0k , Ma,k ) and dynamical models.
Since the VSIMM is implemented without any modification, the details are omitted
here and an interested reader is referred to [7].
54 I. Punčochář and O. Straka

5 Numerical Comparison of Fault Estimators

The presented fault estimators are compared by means of a numerical example.

Instead of making the comparison on equal ground, the goal is to demonstrate the
main feature of both estimators. Although the intended application is detection of
measurement faults in s satellite navigation system, a model of a lower dimensionality
is used to ease the presentation. The discrete-time version of the continuous white
noise acceleration model [1] for a scalar position is
   3 
1 Ts √ Ts
0
xk+1 = x + q √ 3 √ wk , (24)
0 1 k 3Ts Ts
2 2

where Ts = 1 is the sampling period, q = 0.1 is the intensity of the continuous-time

process noise, and n x = n w = 2. It is assumed that four measurements of the position
are available and the measurement model is
 T
1234
yk = xk + fk + 0.01vk , (25)
0000

where F = In y ×n y and n y = n f = n v = 4. The initial state has the mean value

x0|−1 = [5, 0.02]T and covariance matrix P0|−1 = diag(10, 0.01). The simulated time
profile of the considered measurement fault is given in the left graph of Fig. 1. The first
half of the simulation, i.e., the time interval {0, . . . , 250}, contains the measurement
faults fk that have various combinations of non-zero values in the first three com-
ponents including the case where all three components are non-zero simultaneously
({120, . . . , 149}). Note that it violates the assumption of the parity-space based fault
estimator. The second half of the simulation, i.e., the time interval {251, . . . , 500},
contains the measurement faults fk that are gradually brought up to the fault vec-
tor fk ∈ Fus = span{[1, 2, 3, 4]T } that is undetectable by the parity-space approach.
The gradual rise is simulated to satisfy the assumption encoded by P(μk+1 |μk ) on
the time behaviour of the measurement fault. Note that the sudden disappearance
of the measurement fault at the end of the simulation violates the assumption of
multiple-model based fault estimator.
The parity-space based fault estimator is designed as described in Sect. 3. The
desired probability of false alarm is set to Pfa = 10−3 . The criterion of marginal dis-
tinguishability reveals that the assumed maximum number of non-zero components
of a measurement fault is n ∗ = 2 and all matrices Mi have the full column rank. Since
n f = 4, we have Nn ∗ = 6. The multiple-model based fault estimator is designed as
described in Sect. 4 and parameters are chosen as follows. The grid of discrete mea-
surement faults is chosen as F̄i = {−4, −3, −2, −1, 0, 1, 2, 3, 4} for i = 1, 2, 3, 4
and the total number of models is 6561. The probability for the auxiliary Markov
chain is selected p = 0.2.The adaptation parameters of the VSIMM are chosen as
Nmin = 9, PL = 0.1, and PU = 0.9. The number of models that were kept active by
the multiple-model fault estimator is given in the right graph of Fig. 1.
Parity-Space and Multiple-Model Based Approaches … 55

4 17

3.5 16

15
3

14
2.5

13
2
12

1.5
11

1
10

0.5 9

0 8
0 100 200 300 400 500 0 100 200 300 400 500

Fig. 1 The time profile of measurement faults and the number of models used by the VSIMM

Figure 2 illustrates the estimates of measurement fault computed by the parity-

space based fault estimator (left graph) and by the multiple-model based fault esti-
mator (right graph). It can be seen that both fault estimators provide a decent quality
of estimate in the time intervals where their respective assumptions are satisfied.
Although the estimates of the multiple-model base fault estimator are naturally
impacted by the coarseness of selected grid F̄, this estimator is able to estimate
the measurement fault that is even undetectable by the fault estimator based on the
parity-space approach.
The main features of both fault estimators can be summarized as follows. The
parity-space based fault estimator is theoretically and computationally simple. The
only parameters that must be selected are the required probability of the false alarm
Pfa and the number of the maximum non-zero components of the measurement
fault n ∗ . However, this number cannot be chosen arbitrarily high due to structural
constraints imposed by the model and projection operation performed within the
fault estimator. Moreover, this fault estimator is suitable for estimating the additive
faults only. The multiple-model based fault estimator is more complex and its high
computational demands are partially mitigated by the use of the VSIMM. Besides the
parameters of the VSIMM estimator itself Nmin , PL , and PH , also a dynamical model
56 I. Punčochář and O. Straka

4 4

3 3.5

2
3

1
2.5

0
2
-1

1.5
-2

1
-3

-4 0.5

-5 0
0 100 200 300 400 500 0 100 200 300 400 500

Fig. 2 The measurement faults (dotted lines) and their estimates (solid lines) generated by the
parity-space based (left graph) and multiple-model based (right graph) fault estimators

of the fault needs to be chosen. Due to measurement fault model, this estimator is
able to estimate faults that parity-space based fault estimator cannot and is suitable
for multiplicative faults as well.

6 Conclusion

The paper presented two fault estimators that are based parity-space and multiple-
model approaches. The main advantage of the parity-space base fault estimator is its
simplicity and low computational demands. The multiple-model based fault estimator
can provide better estimation capabilities but its computational demands are higher
and an accurate dynamical model of the measurement fault is needed.

Acknowledgements The work was supported by the Czech Science Foundation under grant 22-
11101S.
Parity-Space and Multiple-Model Based Approaches … 57

References

1. Bar-Shalom, Y., Li, X.R., Kirubarajan, T.: Estimation with Applications to Tracking and Nav-
igation. Wiley, New York, NY, USA (2001)
2. Blom, H.A.P., Bar-Shalom, Y.: The interacting multiple model algorithm for systems with
Markovian switching coefficients. IEEE Trans. Autom. Control 33(8), 780–783 (1988)
3. Chiang, L.H., Russell, E.L., Braatz, R.D.: Fault Detection and Diagnosis in Industrial Systems,
1st edn. Springer, London, UK (2001)
4. Gao, Z., Ding, S.X., Ma, Y.: Robust fault estimation approach and its application in vehicle
lateral dynamic systems. Optim. Control Appl. Methods 28(3), 143–456 (2007)
5. Isermann, R.: Fault-Diagnosis Systems: An Introduction from Fault Detection to Fault Toler-
ance. Springer, Berlin, Germany (2006)
6. Isermann, R.: Fault-Diagnosis Applications. Springer, Heidelberg, Germany (2011)
7. Li, X.R.: Multitarget/Multisensor Tracking: Applications and Advances, chap. Engineer’s
Guide to Variable-Structure Multiple-Model Estimation for Tracking, Artech House (2000)
8. Patton, R.J., Frank, P.M., Clark, R.N.: Issues of Fault Diagnosis for Dynamic Systems, 1st edn.
Springer, London, UK (2000)
9. Punčochář, I., Straka, O., Šimandl, M.: Confidence regions for multi-sensor state estimation
under faulty measurements. In: Proceedings of 2013 Conference on Control and Fault-Tolerant
Systems, pp. 202–207. Nice, France (2013)
10. Sun, X., Patton, R.J.: Robust actuator multiplicative fault estimation with unknown input decou-
pling for a wind turbine system. In: 2013 Conference on Control and Fault-Tolerant Systems
(2013)
11. Wang, Z., Rodrigues, M., Theilliol, D., Shen, Y.: Sensor fault estimation filter design for
discrete-time linear time-varying systems. Acta Automatica Sinica 40(10), 2364–2369 (2014)
12. Zhong, M., Ding, S.X., Han, Q.L., Ding, Q.: Parity space-based fault estimation for linear
discrete time-varying systems. IEEE Trans. Autom. Control 55(7), 1726–1731 (2010)
Diagnosis and Prognosis
Online Condition Monitoring
of a Vacuum Process Based on Adaptive
Notch Filters

Mohammad F. Yakhni, S. Cauet, A. Sakout, H. Assoum, and M. El-Gohary

Abstract Vacuum systems play an essential role in industrial installations. This sys-
tem is subject to various operating conditions, making it susceptible to failures that
can cause real damage in the working environment. Online condition monitoring is
therefore necessary to quickly detect any malfunction and avoid its consequences.
It increases the life of system components while reducing maintenance costs and
downtime. Various techniques can perform this task. In this paper, we have relied on
the transient motor current signature analysis. The system studied is a vacuum system
located in the Municipal Technical Center of Poitier, France. Matlab/Simulink pro-
gram was used to build a digital twin of this system and create several types of faults
at different operating speeds. Four Adaptive Notch Filter techniques were devel-
oped, based different structures, which are: Chambers’ all-pass structure, Regalia’s
all-pass solution, Cho, Choi & Lee’s all-pass method, and M’Sirdi’s structure, in
order to identify the system faults. M’Sirdi’s structure, which is the most recent
technology, was discussed in detail. The comparison between the results of the four
methods was presented, where the simulation results proved their effectiveness in
achieving the desired goal, with the superiority of the M’Sirdi structure.

M. F. Yakhni (B) · S. Cauet

Université de Poitiers, 86000 Poitiers, France
e-mail: [email protected]
S. Cauet
e-mail: [email protected]
M. F. Yakhni · A. Sakout
Université de La Rochelle, 17000 La Rochelle, France
e-mail: [email protected]
M. F. Yakhni · H. Assoum · M. El-Gohary
Beirut Arab University, Beirut, Lebanon
e-mail: [email protected]
M. El-Gohary
e-mail: [email protected]
M. El-Gohary
Alexandria University, Alexandria, Egypt

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 61

D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_6
62 M. F. Yakhni et al.

1 Introduction

There is an increasing demand to enhance the availability and reliability of industrial

processes. Any unexpected malfunction can lead to expensive downtime, destruction
of equipment, and even danger to human life. Autonomous online health monitoring
systems with integrated fault detection algorithms provide early alerts on mechani-
cal and electrical failures to prevent system malfunction. The types of faults differ
according to the system concerned [21]. In this paper, we deal with a unit used
for suctioning sawdust in the Municipal Technical Centre (MTC) of Poitiers. These
units mainly consist of an AC motor, a fan, shafts, bearings, ducts, etc. [22]. In most
applications, the fan is not directly linked to the electric motor, primarily to serve
two purposes. The first one is to protect the driver, as the fan is subject to various
operating and environmental disturbances. Meanwhile, the second objective is to
regulate the rotational speed and torque of the fan by adjusting the gear or pulley
diameter [10]. So, in brief, we can state that the suction system can be separated
into two major parts. The first is the electric motor, which we will consider, in this
paper, as an induction motor (IM) as it is the most widely used in the industrial
fields, including suction systems. The second section is the pure mechanical unit,
which comprises the rest of the components. Thus, we will deal with these systems
as Fan/Motor systems. The IM is affected by various faults, that can be divided into
two types, electrical and mechanical. The most common electrical defercts are: the
stator faults [7] and the voltage unbalance [8], While for mechanical ones, the most
common are: the presence of eccentricity [24], the bearings damage [17] or the bro-
ken rotor bar [1]. Regarding the fan side, it may be affected by various faults, like
any rotating mechanism. The most significant ones are: fan unbalance, belt defects,
shaft misalignment, and bearing problems [11].
Vibration and acoustic emission analysis is still the most used approach, but both
are expensive technologies in terms of complexity and sensors. An efficient method
that may accomplish this task is motor current signature analysis (MCSA) since all
of these defects affect the electric motor either directly or indirectly. Since most
applications require the driver to operate at variable speeds, we go for the transient
MCSA, since the current varies with load and speed [21]. Since the system under
study is in a company and not in an academic laboratory, it is impractical to manipulate
it by creating faults and seeing its results. Therefore, a digital twin, created by [22],
was used to simulate the actual system.
Since the concept was first coined by John Vickers and Michael Grieves [9], many
authors have attempted to define the term Digital Twin, beginning with the aerospace
industry [12], focusing on structural mechanics, materials science, and long-term
performance prediction of air and spacecraft [18]. With the growth of Industry 4.0,
the focus has shifted to manufacturing [12]. It has been adopted in many types of
previous research and has shown satisfactory performance in achieving its goals
[6, 20].
To process the current signals, four different Adaptive Notch Filters (ANFs)
structure were used. ANF approach has been used in many researches to estimate
Online Condition Monitoring of a Vacuum Process … 63

the frequencies of sinusoidal signals. Various structure were adopted to develop the
filter. Chaochao et al. [3] describes a technique for calculating the power system
frequency using the second-order ANF. The suggested method accurately calcu-
lates the power system frequency, according to simulation findings. The suggested
methods make use of Newton’s method and stochastic optimization, are simple for
using, and converge quickly on a global scale. A novel improved ANF with phase
shift was presented by [4]. And used to reduce synchronous vibration for the rotor of
active magnetic bearings. Simulation and experimentation show the suggested ANF’s
efficacy and adaptability in eliminating synchronous regulated current over a wide
operating speed range. A variable step-size ANF utilizing a mixed gradient technique
is provided by [23] for frequency estimation in order to enhance the performance of
ANF. Coriolis mass flow meter application and simulation results both attest to the
efficacy of the suggested approach. A parallel ANF technique was created by [19]
to get a precise estimate of the frequency needed to generate a sinusoidal reference
signal. The performance of the noise reduction system is enhanced by the proposed
algorithm’s accurate estimation.
Punchalard [14] developed a limited IIR ANF for frequency estimation of a single
real tone imbedded in Gaussian noise using a weighted least squares technique.
The suggested approach performs better than several earlier methods in terms of
both convergence rate and steady-state mean square error. Results from computer
simulations are presented to support the claim.
This document is structured into four main sections. Section 2 describes the MTC
vacuum system. It then discusses the development of the digital twin. The four ANF
methods are presented in Sect. 3. The simulation results are presented in Sect. 4 and
discussed. The last section is the conclusion of this article with a portal for future
work.

2 Vacuum Model

The MTC of Poitiers, France, is a public organization that contributes with its dif-
ferent resources to the realization of projects for the town. The carpentry is part of
its workshops. Among the devices in this section, what concerns us is the sawdust
suction device. We are interested primarily in monitoring the operating condition
of the fan/motor unit. The system to be controlled is represented in Fig. 1. Figure 2
presents a detailed view of this arrangement. In this article, the mathematical system
that the one developed by [22] was used. The concept of the free body diagram and
Newton’s second law were used to build it. It was divided into three masses: Pulley
1, Pulley 2 and the Fan, where their three equations of motions are presented below
in Eqs. (1) to (3), respectively:
64 M. F. Yakhni et al.

Fig. 1 The CTM fan/motor unit a fan side b transmission and IM side

Fig. 2 The real system with labels

J1 θ¨1 = Te − 2K b (R p1 θ1 − R p2 θ2 ) (1)

J2 θ¨2 = 2K b (R p1 θ1 − R p2 θ2 ) − K s (θ2 − θ F ) − Br (θ˙2 − θ˙F ) (2)

JF θ¨F = K s (θ2 − θ F ) + Br (θ˙2 − θ˙F ) + Ta (3)

Let θ1 , θ2 , and θ F be the rotational angle of pulley 1, pulley 2, and the fan respectively,
whereas J1 , J2 , and JF be the polar moments of inertia of the three weights respec-
tively. Te is the resulting electromagnetic torque of the IM, and Ta is the load torque,
which will be discussed in the next paragraph. The shaft has a torsional stiffness
K s , while three belts were modeled as flexible components with equivalent linear
stiffnesses K b . The overall friction coefficient of the two bearings is Br . The defect is
added to the model as a pair Td as a sine wave, and all defect values will be added to
Ta . These two parameters are expressed as Eqs. (4) and (5), Where A and f d are the
amplitude and frequency of the defect, that varies with time (t) respectively, while
Online Condition Monitoring of a Vacuum Process … 65

Fig. 3 Simulink model diagram

Q, w and P are the flow rate, the fan speed (rad/s) and the pressure difference on
the fan.
Td = A × cos(2π f d t) (4)

Ta = (Q × P/w) + Td (5)

The value of Ta varies according to two cases: the normal case and in case of the
presence of one or more defects in the system. In the first one, Ta is the aerodynamic
torque produced by the fan motion. When a defect appears, its value has added to
the equation of this torque. In this paper, we discuss the most common faults in the
system, each with a particular frequency proportional to the speed of rotation [21].
Matlab/Simulink software was used to create the model. Figure 3 shows the over
all model that consists of a three-phases variable speed IM, driven with a variable
frequency drive. The desired speed is set by the user. Feedback from the mechanical
system (Tm ), which is the sum of the torques on the fan (JF θ¨F ), was utilized to control
the motor performance.

3 Adaptive Notch Filtering (ANF) Methods

Constant speed operation is not possible in most real-world applications, so it is

necessary to apply an analysis method that accommodates all working environments,
not just stationary signals. To this end, the technique adopted in this research achieves
this goal and can be used for online condition monitoring without the need to store
the data and analyze it at a later stage [21]. One phase of the three-phase current is
processed, and noise is added to this signal to make it as close as possible to reality.
The first step is to filter the current using a variable bandwidth filter because the
amplitude of the target frequency is much smaller than the fundamental frequency,
so we need to isolate the region where the frequency of the defect appears to track it.
66 M. F. Yakhni et al.

Each defect has a particular signature in current, where the frequency is modulated
to appear in the current as expressed by Eq. (6), with f e is fundamental frequency of
the supply power source.
f modulated = f d ± f e (6)

There is a relationship between these frequencies and the rotation speed [22].
For this, variable frequency estimation techniques were used, where four Adaptive
Notch Filter (ANF) structures to track transient frequencies were developed, which
are: Chambers’ all-pass structure [2], Regalia’s all-pass solution [15], Cho, Choi &
Lee’s all-pass method [5] and M’Sirdi’s structure [11]. The ANF structure is adjusted
to track these specific frequencies. ANFs are well suited for predicting the frequencies
of the sinusoidal components as shown in the ANF transfer function (Eq. (7)), where
the i th sinusoidal component is filtered by the second order ANF. Te is the sampling
period.
1 − ai z −1 + z −2
Hi (z) = (7)
1 − rai z −1 + r 2 z −2

with:
ai = −2cos(2π f i Te ) (8)

The transfer function of the second-order notch filter is independent of amplitudes or

phases. The bandwidth of the notch filter is determined by the parameter 0 < r < 1,
while the frequency of the notch filter is determined by the value ai . Since a perturbed
sinusoidal signal has p sinusoidal components, let us consider the transfer function
of the notch filter with p cells of the second-order adaptive filter in cascade, which
is given by Eq. 9.
 p
Hi (z) (9)
i=1

We assume that the bandwidth parameter r is the same for the notches without sacri-
ficing generality. Therefore, frequency independence determines the independence
of the characteristics of each second-order cell. After prefiltering the signal through
each filter, each cell can be modified independently of the others [11, 13]. Correla-
tions between frequencies can be exploited to reduce the computational burden of the
comb filter. If all frequency components are independent, so we can write Eq. (10)
as follows:
j

p
ỹk = Hi (q)yk (10)
i=1
i = j

If we consider that the center-frequency cascade filters f i of Hi (z) (for i = 1... p

and i = j) have converged, the filter Hi (z) will remove one of the jth sinusoidal
components [11, 13]. Then the error εk of prediction is obtained by removing the
jth component, we can thus write:
Online Condition Monitoring of a Vacuum Process … 67

j 1 − ai z −1 + z −2
ỹk = yk (11)
1 − rai z −1 + r 2 z −2

j 1 − a j q −1 + q −2 j
εk = Hi (q) ỹk = ỹ (12)
1 − ra j q −1 + q 2 z −2 k

Therefore, a single global minimum for the mean square error (MSE) criterion will
be created, with each notch removing a frequency component. For the estimation
process, we derive the estimation gradient using the Output Error Prediction Method
(OPEM) approach according to Eqs. (13) [16]:

j dεk (1 − r )(1 − rq −2 ) j
ψk−1 = = ỹ (13)
da j 1 − ra j q −1 + r 2 q −2 k−1

In terms of computational complexity, the RML (Recursive Maximum Likelihood)

algorithm results mainly from the calculation of the gradient and the filtering
(Eqs. (11), (12) and (13)). By switching to an approximate RML method (ARML)
[11, 13], a first approximation can be constructed to reduce the computational com-
plexity. Instead of using the Eq. (13), we can use Eq. (14), shown below:
j
j dεk Sk−1 − r εk−1 j
ψk−1 = = ỹ (14)
da j 1 − ra j q −1 + r 2 q −2 k

j
with Sk is calculated as shown below:

f or j = ( p − 1), ..., 1 do

j

p
yk = Hl (q)yk (15)
l= j+1

j j j−1
Sk = yk − yk (16)

The RML algorithm can be summarized as follows:

⎧
⎨ f or j = ( p − 1), ..., 1 do
j j j j
ã = ã + Fk ψk−1 εk (17)
⎩ j k j k−1 j j j
Fk = Fk−1 /(λ + ψk−1 Fk−1 ψk−1 )

j
where Fk is the matching gain, yk is the input signal to the cascade filter, and rk is the
forgetting factor. The bandwidth or what we call the depolarization factor rk must
be positive. In the application, this parameter varies exponentially in time from 0 or
r0 to 1 or r f , according to the expression (18):
68 M. F. Yakhni et al.

rk = rd rk−1 + (1 − rd ) − r f (18)

These developed techniques have a shortcoming in that, in the absence of mal-

function, the output of the ANF is random. Therefore, it must be adjusted. To do this,
a fake defect whose frequency is close to the frequency of the actual one is added
to the system. By tuning the magnitude of this defect, we can create a threshold
that will determine the severity of the real malfunction. The estimated frequency is
biased towards the frequency with the highest magnitude and constrained between
two values, which are the value of the actual and fake defect frequencies.

4 Results and Discussions

This unit consists mainly of Pulley 1 linked directly to the motor. The power is
delivered to Pulley 2 through three belts, each with a length (L) equal to 2300 mm
of type SPB Vbelt. The fan is coupled to Pulley 2 by a drive shaft with a diameter
of 45 mm and a length of 80 mm. The diameter of pulleys 1 and 2 (R p1 and R p2 ) is
160 and 225 mm respectively. The two bearings supporting the fan shaft are 22210
EK type. The fan is made of 60A steel, like the other elements, and has eight blades.
The other components have little impact on the basic function of the system, so we
ignored their impact on the modeling process.
To test the effectiveness of the developed system, we simulated for 11 s under the
following conditions: the defect occurred at t = 5 s, while the rotor speed varies from
2950, which is the nominal speed, to 2500 rpm at t = 6. s Seven types of defects
were added to the system, which are: fan imbalance, bearing inner ring and balls
defects, belts malfunction, broken rotor bar, notches harmonics, and eccentricity or
load effects. All parameters used to creates these faults are as that used by [22].
Note that fr is the rotational frequency of the rotor, and Nb is the fan blades number.
The initialisation parameters of each ANF structure are as follow, for Chambers’
structure: α = 0.9385, ψ0 = 1, μ = 0.0000127, γ = 0.99 and β = cos(w0 ), with
w0 is the initialization frequency, while the parameters of Reglia’s are: ψ0 = 1,
μ = 0.0000125, γ = 0.99 and β = 0.02. For Cho’s method: α = 0.95, γ = 0.5
and λ = 0.9989, a0 = −cos(w0 × Ts ), where Ts = 0.0001 s is the sampling time
and k1 = 1. Finally the M’Sirdi’s ANF has the parameters estimates aˆi k is initially
set to zero (aˆi 0 = 0). The ANF bandwidth are set at r0 = 0.1, at the beginning
and exponentially increases to r f = 0.8 witha factor of rd = 0.99, as Eq. 18, and
λ0 = 0.99.
In this paper, We will show the performance of ANFs with the most popular
defect in this type of systems, wich is the fan imbalance, that varies with the rotation
frequency as follow: Nb × fr , and what applies to it applies to the other kinds of
malfunctions. Figure 4 shows the desired operating speed with the actual rotor fre-
quency ( fr ). The performance of the four developed techniques in case of this type
of malfunction is shown in Fig. 5.
Online Condition Monitoring of a Vacuum Process … 69

Fig. 4 The Desired and the actual rotor speed

Fig. 5 Estimated frequencies for fan imbalance in case of four different ANF methods
70 M. F. Yakhni et al.

The simulation results proved the effectiveness of the proposed approach in

achieving the desired objective. All four techniques were able to track the frequency
of faults and show their severity, by moving towards or away from the frequency of
the real defect. It should be emphasized that the technology developed by M’Sridi
delivers the best results in terms of response time and stability.

5 Conclusion

Electrical Signals can be used to detect vacuum system faults. Each defect in the
system has particular frequencies. These frequencies are proportional to the motor
rotational speed, so the TMCSA was used as a condition monitoring technique for
this system. Four variable frequency estimation techniques based on ANF structures
have been used to achieve this goal. A numerical defect is added to the system in order
to control the output of the ANF and determine the severity of the malfunction by
adjusting its amplitude. At last, as a proposition of future works, the created approach
can be utilized with other kinds of industrial systems and diverse types of defects.
Other approaches can moreover be adopted in processing the current signals. As well
another theme of research, is the usage of the artificial intelligence, like Artificial
Neural Networks, Fuzzy and Neuro-Fuzzy logic, to make the alert that determine the
severity of the defects.

References

1. Asad, B., Vaimann, T., Belahcen, A., Kallaste, A., Rassolkin, A., Heidari, H.: The low volt-
age start-up test of induction motor for the detection of broken bars. In: 2020 International
Conference on Electrical Machines (ICEM), vol. 1, pp. 1481–1487. IEEE (2020)
2. Chambers, J.A., Constantinides, A.G.: Frequency tracking using constrained adaptive notch
filters synthesised from all pass sections. In: IEE Proceedings F (Radar and Signal Processing),
vol. 137, pp. 475–481. IET (1990)
3. Chaochao, J., Yixin, S., Huajun, Z., Shilin, L.: Power system frequency estimation based on
adaptive notch filter. In: 2016 International Conference on Industrial Informatics-Computing
Technology, Intelligent Technology, Industrial Information Integration (ICIICII), pp. 191–194.
IEEE (2016)
4. Chen, Q., Liu, G., Han, B.: Unbalance vibration suppression for AMBs system using adaptive
notch filter. Mech. Syst. Signal Process. 93, 136–150 (2017)
5. Cho, N.I., Choi, C.-H., Lee, S.U.: Adaptive line enhancement by using an IIR lattice notch
filter. IEEE Trans. Acoust. Speech Signal Process. 37(4), 585–589 (1989)
6. Chowdhury, S., Yedavalli, R.K.: Dynamics of belt-pulley-shaft systems. Mech. Mach. Theory
98, 199–215 (2016)
7. Glowacz, A., Glowacz, W., Glowacz, Z., Kozik, J.: Early fault diagnosis of bearing and stator
faults of the single-phase induction motor using acoustic signals. Measurement 113, 1–9 (2018)
8. Gonzalez-Cordoba, J.L., Osornio-Rios, R.A., Granados-Lieberman, D., de Romero-Troncoso,
R.J., Valtierra-Rodriguez, M.: Thermal-impact-based protection of induction motors under
voltage unbalance conditions. IEEE Trans. Energy Convers. 33(4), 1748–1756 (2018)
9. Grieves, M.: Manufacturing Excellence through Virtual Factory Replication (2015)
Online Condition Monitoring of a Vacuum Process … 71

10. Kaya, D., Kılıç, F., Hüseyin Öztürk, H.: Energy Management and Energy Efficiency in Industry:
Practical Examples. Springer Nature (2021)
11. M’Sirdi, N.K., Monneau, A., Naamane, A.: Adaptive notch filters for prediction of narrow band
signals. In: 2018 7th International Conference on Systems and Control (ICSC), pp. 403–408.
IEEE (2018)
12. Negri, E., Fumagalli, L., Macchi, M.: A review of the roles of digital twin in cps-based pro-
duction systems. Procedia Manuf. 11, 939–948 (2017)
13. Pei, S.-C., Tseng, C.-C.: Real time cascade adaptive notch filter scheme for sinusoidal parameter
estimation. Signal Process. 39(1–2), 117–130 (1994)
14. Punchalard, R.: Frequency estimation based on WLS-constrained adaptive notch filter. In: 2020
17th International Conference on Electrical Engineering/Electronics, Computer, Telecommu-
nications and Information Technology (ECTI-CON), pp. 353–356. IEEE (2020)
15. Regalia, P.A.: An improved lattice-based adaptive IIR notch filter. IEEE Trans. Signal Process.
39(9), 2124–2128 (1991)
16. Stearns, S.: Error surfaces of recursive adaptive filters. IEEE Trans. Circuits Syst. 28(6), 603–
606 (1981)
17. Toma, R.N., Prosvirin, A.E., Kim, J.-M.: Bearing fault diagnosis of induction motors using a
genetic algorithm and machine learning classifiers. Sensors 20(7), 1884 (2020)
18. Tuegel, E.J., Ingraffea, A.R., Eason, T.G., Spottswood, S.M.: Reengineering aircraft structural
life prediction using a digital twin. Int. J. Aerosp. Eng. 2011 (2011)
19. Wang, H., Sun, H., Sun, Y., Ming, W., Yang, J.: A narrowband active noise control system with
a frequency estimation algorithm based on parallel adaptive notch filter. Signal Process. 154,
108–119 (2019)
20. Yakhni, M.F., Ali, M.N., El-Gohary, M.A.: Magnetorheological damper voltage control using
artificial neural network for optimum vehicle ride comfort. J. Mech. Eng. Sci. 15(1), 7648–7661
(2021)
21. Yakhni, M.F., Cauet, S., Sakout, A., Assoum, H., Etien, E., Rambault, L., El-Gohary, M.:
Variable speed induction motors’ fault detection based on transient motor current signatures
analysis: a review. Mech. Syst. Signal Process. 184, 109737 (2023)
22. Yakhni, M.F., Hosni, H., Cauet, S., Sakout, A., Etien, E., Rambault, L., Assoum, H., El-
Gohary, M.: Design of a digital twin for an industrial vacuum process: a predictive maintenance
approach. Machines 10(8), 686 (2022)
23. Yang, H., Tu, Y., Li, M.: A variable step-size adaptive notch filter for frequency estimation
using combined gradient algorithm. J. Phys.: Conf. Ser. 1187, 032085 (2019). IOP Publishing
24. Yassa, N., Rachek, M., Houassine, H.: Motor current signature analysis for the air gap eccen-
tricity detection in the squirrel cage induction machines. Energy Procedia 162, 251–262 (2019)
A Study of OBF-ARMAX Performance
for Modelling of a Mechanical System
Excited by a Low Frequency Signal
for Condition Monitoring

Oscar Bautista Gonzalez and Daniel Rönnow

Abstract A digital twin of a mechanical system (a pair of axial rolls in a ring mill
used in a steel plant) with poles close to the unit circle and the real axis in the discrete
pole-zero map was built. The system was excited by a signal concentrated in the low-
frequency band. For this particular case, it is shown that the ad-hoc combination of
ARMAX and orthonormal basis filter model structures outperform model structures
based on either ARMAX or orthonormal basis functions when estimating the poles
of the basis by analyzing the data in the frequency domain. The followed modelling
methodology of the system is described in detail to help replicate the work for similar
systems in the steel industry. Real production data from a steel plant were used in
contrast to previous studies, where the combination of ARX and ARMAX with
orthonormal basis filter model structures was evaluated using simulated data instead
of real data. We believe that the resultant model can be used when having systems with
poles close to the unit circle and real axis and poor excited input signal concentrated
in the low frequency band. The resultant model can be used for condition monitoring
and failure detection.

1 Introduction

The fourth industrial revolution significantly improves the industry’s efficiency

through its digitalization using technologies based on cloud computing, digital twins,
and IoT. Digital twins are virtual representations of objects, systems, or processes
[1]. Digital twins bring the opportunity of monitoring systems status, real-time com-
munication between the virtual representation and the plant, and failure prediction,
which needs a model of the system’s behaviour [2]. In the steel industry, digital twins

O. B. Gonzalez (B) · D. Rönnow

Department of Electrical Engineering, Mathematics and Science, University of Gävle,
80176 Gävle, Sweden
e-mail: [email protected]
D. Rönnow
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 73

D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_7
74 O. B. Gonzalez and D. Rönnow

are needed at each stage of the product life cycle, but digital twins of more machines
and production processes are still needed [3].
The field of system identification provides a toolkit of methods to describe the
behavior of dynamical systems. ARX, ARMAX and ANN are common model struc-
tures of such methods where arithmetic complexity plays an essential role in condi-
tioning monitoring of real systems. However, other methods are available, like the
ones based on orthonormal basis filters (OBF) [4–7]. The last group presents advan-
tages compared to the first ones regarding the statistical properties of the estimators
by having orthonormal inputs when solving the least squares normal equations, mod-
els linear in the parameters, and the possibility to encode prior information about the
system in the model structure before estimation [8]. A combination of both meth-
ods (ARMAX and OBF) in an ad-hoc manner can be found in the literature for the
case of simulated data. For example, both methods were combined for identifying
systems running in closed-loop where OBF-ARX outperforms ARMAX for control
purposes [9]. OBF-ARMAX has been used for identifying closed-loop systems for
both stable and unstable systems by exciting the system with an extra signal [10].
A high-order ARX system was identified using an OBF model structure, achiev-
ing a lower number of parameters than the ARX case [11]. ARMAX and Laguerre
functions have been used for modelling the implementation of an efficient Model
predictive control (MPC) controller. Laguerre was selected to provide an efficient
number of parameters [12].
In the author’s previous work, a digital twin of a radial-axial ring rolling (RARR)
machine was designed to monitor the machine’s status over time by modelling its
five subsystems. It was found that ARMAX SISO models gave poor model results
for some subsystems [6]. In this work, one of the RARR’s subsystems, i.e., the axial
canonical rolls, which was modelled unsuccessfully in previous work through a SISO
ARMAX model structure, is successfully modelled with a SISO OBF-ARMAX. To
our knowledge, it is the first time that OBF-ARMAX has been used with experimental
data, and for modelling systems with complex conjugate poles close to the unit circle
and the positive real axis in the discrete pole-zero map by using signals of in the low
frequency band with poor excitation. This particular system was investigated and
validated by modelling real data from a steel plant during production and numerical
experiments. The resultant model can be used for condition monitoring of the system
and failure detection.

2 System

The analyzed RARR machine from Banning (1971) is used in an operating steel
plant. A mechanical scheme of the RARR system can be found in [6]. The RARR
can be divided into five different subsystems: (1) the horizontal subsystem with the
mandrel, (2) the canonical rolls, (3) the vertical tower with the canonical rolls attached
to it, (4) and (5) which are the lower and upper arms, repectively [6]. The canonical
rolls’ subsystem plays an essential role in the forming process of the ring and is
A Study of OBF-ARMAX Performance for Modelling of a Mechanical System … 75

studied in this work. The system operates in different batches (cycles), which can be
divided into three stages. (1) The initialization corresponds to the rapid movement
of the tools until a pre-defined position close to the ring. (2) A slow movement of
the tools until touching the ring is defined as preparation. (3) The rolling process
is when the tools are in contact with the ring, and the forming process takes place.
The modelled data correspond to the forming process. The control signal of the axial
rolls is selected as the input and, the position signal is selected as the output signal.
A IBA-DAQ system is used to collect data at 0.01 s sampling time, which is given
by the system design.

3 Methodology

A standard procedure in time series analysis for dynamical systems inspired this
work’s methodology [5]. The data from the RARR’s canonical rolls during production
in a steel plant of one ring were selected as training data for parameter estimation
and data of the following ring as test data for model validation. Simulations were
used to investigate further the selected models with signals and systems based on the
knowledge of the real system obtained in the data analysis. Six steps were followed
and iterated:
1. Data collection: Data from production time in the steel plant were collected. The
data during the forming process was used for the following steps.
2. Data analysis: The collected data was analyzed in the time and frequency domain.
The input and output signal were plotted in the time domain for several cycles of
the process. Relatively similar relations between input and output were observed.
No new frequency components in the output were observed compared to the input.
Hence, a linear time-invariant (LTI) system was assumed after the analysis. The
system’s poles were estimated by analyzing data in the frequency domain. These
are the poles which are used as prior information in the next step.
3. Model structure selection: After establishing that the system was an LTI system,
ARX and ARMAX model structures were considered as candidates. OBF was
considered after analyzing the modelling results provided by ARX and ARMAX.
Finally, OBF-ARX and OBF-ARMAX were studied. A pair of complex conjugate
poles was selected for building the orthonormal basis filters in the case of OBF,
OBF-ARX, and OBF-ARMAX as described in the previous step.
4. Data cleaning: Small noise and no outliers were observed in the signals in the time
domain. The mean value of signals was removed before parameter estimation.
5. Parameter estimation: The least-square criterion was used for estimating the
parameters of each selected model in step 3. The least-square method was used
for estimating the parameters of the ARX, OBF-ARX, and OBF-ARX models
by using the function arx in MATLAB. An iterative method was used for the
estimation of ARMAX and OBF-ARMAX’s parameters with the function armax
[13]. Training data were used for estimation.
76 O. B. Gonzalez and D. Rönnow

6. Model evaluation:
a. Model validation: Cross-validation with the test data, pole-zero map and
residual analysis were considered for validating the models. Cross-validation
is based on
||y − ŷ||
f it = 1 − , (1)
||y − y||

where y is a vector containing the measured output, ŷ is a vector of the model’s

output, and y is the mean value of y. The pole-zero map was studied to check
the overfitting and stability of the model. The error signals were analyzed in
the time domain.
b. Simulations: The same models, which were selected in step 3, were further
investigated by formulating a numerical example with input signals concen-
trated in the low frequency band, and a system with complex conjugate poles
close to the unit circle and the real axis. The modelling of such a formulated
system was done as with the RARR’s canonical rolls, with the difference that
the output signal was obtained by simulation.
The above modelling procedure was considered finished with relatively good
results in the model validation, i.e., similar accuracy levels given by (6) were observed
for the training and test data, no overlapping between zeros and poles in the pole-zero
map, poles inside the unit circle, and residuals with small amplitudes were considered
as desired validation results. In the other case, the procedure was re-started, going
from step 6a back to step 3. In the case of OBF, OBF-ARX and OBF-ARMAX, a
new number of basis functions or a new pair of complex conjugates was selected in
addition. The pole selection was repeated with prior information about the location
of the poles.

4 Models

A LTI set of models for SISO systems with y(t) as the scalar output, u(t) as the
scalar input and e(t) as the white-noise processes of finite variance is defined as

y(t) = G(q, θ )u(t) + H (q, θ )e(t), (2)

where G(q, θ ) and H (q, θ ) are scalar transfer functions with rational functions of
the shift operator q as entries, and θ is the parameter vector which is an element of
a subset of Rd [4]. The scalar transfer functions can be factorized as

B(q) C(q)
G(q, θ ) = , H (q, θ ) = , (3)
A(q) A(q)
A Study of OBF-ARMAX Performance for Modelling of a Mechanical System … 77

where
A(q) = 1 + a1 q −1 + a2 q −2 + ... + ana q −na ,
B(q) = b1 q −1 + b2 q −2 + ... + bnb q −nb , (4)
−1 −2 −nc
C(q) = 1 + c1 q + c2 q + ... + cnc q .

The model structure given by (2)–(4) is identifiable. θ includes all the coefficients
of the polynomials in (4). The one-step ahead predictor is

ŷ(t|θ) = H −1 (q, θ )G(q, θ )u(t) − (1 − H (q, θ )y(t) (5)

as indicated in [4, 5].

The identification of the parameters in θ was performed using the given training
dataset D = {((u(1), y(1)), ..., (..., u(N ), y(N ))}. The least square criterion is used
for estimating θ by
1  2
N
θ̂ = arg min y(t) − ŷ(t|θ) . (6)
θ N
t=1

More information on the parametrization of the model structures of ARX,

ARMAX and OBF models can be found in [6].
In the case of the OBF-ARX ad-hoc framework, (2) contains

G O B F (q) 1
G(q, θ ) = , H (q, θ ) = , (7)
A(q) A(q)

where

p−1
G O B F (q, θ ) = θk Bk (q), H (q, θ ) = 1. (8)
k=0

More information on how to construct Bk (q) can be found in [8]where prior
information can be used in the selection of the poles ξ0 , ξ1 , ..., ξ p−1 [8, 9]. The
one-step ahead predictor can be written as (5) where
 
θ = a1 . . . ana θ0 . . . θ p−1 (9)

is the ((na + p) × 1) parameter vector and ϕ(t) is the ((na + p) × 1) regressor

vector where
T
ϕ(t, θ ) = (−y(t − 1) . . . − y (t − n a ) B0 (q)u(t) . . . B p−1 (q)u (t) . (10)

The optimization problem in (6) gives a closed-form solution when considering (9)
and (10).
78 O. B. Gonzalez and D. Rönnow

For the OBF-ARMAX ad-hoc framework, (2) contains the same transfer functions
as indicated in (7), but H (q, θ ) = C(q)/A(q) where G O B F (q) is the same as in (8)
[10].
The one-step ahead predictor can be written as (5) where
 
θ = a1 . . . ana θ0 . . . θ p−1 c1 . . . cnc (11)

is the ((na + p + nc) × 1) parameter vector and ϕ(t) is the ((na + p + nc) × 1)
regressor vector where

ϕ(t, θ ) = (−y(t − 1) . . . − y (t − n a ) B0 (q)u(t) . . .

T (12)
B p−1 (q)u(t) ε(t − 1, θ ) . . . ε (t − n c , θ ) ,

having ε(t − k, θ ) = y(t − k) − ŷ(t − k|θ ) [4, 5]. The optimization problem in (6)
is solved by an iterative method.

5 Results

This section is divided into the modelling of the RARR’s canonical rolls using real
production data and a numerical example for investigating how OBF-ARMAX com-
pares to the other cases for similar systems and signals. This work compares how
well the different model structures fit the data. A comparison in terms of compu-
tational cost between ARX, ARMAX and OBF of similar systems can be found in
[6], where the reader can find guidelines to analyze the computational complexity of
OBF-ARX and OBF-ARMAX.

5.1 Modelling of the RARR’s Canonical Rolls

The modelling results of the canonical rolls are presented in Table 1. In the case
of the ARX model with 40 parameters, θ and ϕ(t) become (40 × 1) vectors with
na = nb = 20. The ARMAX model contains θ and ϕ(t) as (12 × 1) vectors with
na = nb = nc = 4. The same vectors become (34 × 1) vectors for the OBF model.
The specific expressions of θ and ϕ(t) in the mentioned cases can be found in [6].
OBF-ARX with 12 parameters has θ in (9) and φ(t) in (10) as (12 × 1) vectors
where na = 2 and p = 10. Finally, (11) and (12) become (24 × 1) vectors with
na = nc = 2 and p = 20. The parenthesis in Table 1 for OBF, OBF-ARX and OBF-
ARMAX in the last column corresponds to the parameter ξk and ξi in (14) of [6],
which are equal to each other in our case. The poles must be defined in terms of
magnitude and phase.
A Study of OBF-ARMAX Performance for Modelling of a Mechanical System … 79

Table 1 Modelling results for the RARR’s canonical rolls and the numerical example. First column
indicates if the row is about the canonical rolls or the numerical example. The second column
shows the selected model structure, the third and fourth are the values of (6) for each model for the
training and validation data, respectively, and the fourth one indicates the number of parameters.
“uns.” indicates unstability and “(2)” the required two parameters that define the pair of complex
conjugate poles of the basis functions
Type Model Training Validation #p
RARR’s ARX 0.225 0.224 40
canonical rolls ARMAX (uns.) 0.990 0.987 12
OBF 0.610 0.562 34(2)
OBF-ARX 0.364 0.310 12(2)
OBF-ARMAX 0.983 0.967 24(2)
Numerical ARX 0.018 0.015 2
experiment ARX(uns.) 0.979 0.973 3
ARMAX 0.017 0.015 3
ARMAX (uns.) 0.979 0.962 5
OBF 0.998 0.843 18(2)
OBF-ARX 0.994 0.861 21(2)
OBF-ARMAX 0.966 0.926 22(2)

The ARX models gave poorer results than ARMAX. But, ARMAX gave an unsta-
ble model, which is not reflected in the fit values of the table due to the short time of
the process and its slow dynamics. A stable model and better results were given by
using OBF to create the transfer function between the input and output signals. OBF-
ARX didn’t give any improvement in comparison to previous cases. Finally, the use
of OBF-ARMAX where G O B F (q) is encoded with complex conjugate poles close
to the real positive axis, and the unit circle in the pole-zero map gave the best result.
The number of parameters is smaller than for ARX and OBF. Two models’ outputs
and their corresponding error signals are shown in Fig. 1. Figure 1a shows a small
similarity between the OBF’s output and the real output, whereas OBF-ARMAX’s
output is much closer to the real output. Figure 1b presents information about the
error signals. OBF’s residuals have periodical components of 1 and 2 s and some
high frequency components. OBF-ARMAX’s residuals are in the high frequency
range, but much smaller in amplitude than in the OBF case. The high frequency
components of the OBF’s error signal are caused by the G ob f (q) structure. The func-
tions Bk (q) are built by placing a pair of complex conjugates poles at the opposite
side from (a mirror operation in the imaginary axis) the dominating pole, which is
close to one in the unit circle. High frequency components are amplified as a result of
these poles. The parameters θ of these basis functions are not well identified because
of the small amplitudes of the excitation signal at high frequencies.
80 O. B. Gonzalez and D. Rönnow

(a) (b)
Fig. 1 Modelling of the RARR’s canonical rolls. a Measured real (y real) data and modelled output
signals (y OBF and y OBF-ARMAX) for the investigated models. b Respective error signals (e OBF
and e OBF-ARMAX)

5.2 Numerical Example

The system formulation to generate the data of the numerical example follows (2)
and (3), (4) becomes A(q) = 1 − 3.11q −1 + 3.695q −2 − 2.061q −3 + 0.4755q −4 ,
B(q) = 0.1q −1 and C(q) = 1 − q −1 + 0.2q −2 . The signal u(t) is a filtered white
noise for excitation in the low frequency band as the real input signal, and e(t) is
a white noise signal with variance equal to 1. A(q) was selected according to the
frequency response of the real system, which is studied in the data analysis according
to step 2 in sect. 3, and B(q) and C(q) are chosen based on the experience of modelling
the other subsystems of the machine, reported in [6].
The modelling results of this numerical example are shown in Table 1. In the case
of the ARX model with 2 and 3 parameters, θ and ϕ(t) become (2 × 1) and (3 × 1)
vectors, respectively. The ARMAX model with 3 and 5 parameters contain the θ and
ϕ(t) as (3 × 1) and (5 × 1) vectors, respectively. The same vectors become a (18 ×
1) and (5 × 1) vectors, respectively, for the OBF model. The specific expressions
of θ and ϕ(t) for the mentioned cases can be found in [6]. OBF-ARX with 21
parameters has θ in (9) and φ(t) in (10) as (21 × 1) vectors where na = 3 and p = 18.
Finally, (11) and (12) become (22 × 1) vectors with na = nc = 2 and p = 18. The
parenthesis in Table 1 for OBF, OBF-ARX and OBF-ARMAX in the last column
corresponds to the poles of the scalar transfer functions Bk (q) in (8) which are
complex conjugate poles and must be defined in terms of magnitude and angular
frequency.
According to the numerical example results in Table 1, ARX and ARMAX give
poor fit values when the number of parameters is small and excellent results when
the number of parameters is high. But the models become unstable for the last cases.
The stability problem is not reflected in the metrics due to computing the system’s
output for short time intervals and the slow dynamics of the system. OBF substan-
tially improves model error but with a significant overfitting problem. This overfitting
A Study of OBF-ARMAX Performance for Modelling of a Mechanical System … 81

Fig. 2 Numerical example.

Simulated (y sim) and
modelled output signals for
the investigated models. The
respective error signals are
shown (e)

problem was reduced when applying OBF-ARX and OBF-ARMAX, where the latter
outperformed all other models. OBF, OBF-ARX, and OBF-ARMAX were formu-
lated by setting complex conjugate poles close to the unit circle and the real positive
axis in the pole-zero map when defining G O B F (q). Figure 2 shows the simulated
output and the model’s output for the different ad-hoc model structures and the error
signals corresponding to the difference between the simulated and each models’ out-
put. The residuals corresponding to the OBF model have a periodical component
corresponding to 10 s, and 5 s for the OBF-ARX case. The difference between the
simulated and the OBF-ARMAX’s output is the smallest, with no apparent periodical
components and which offset is bigger in the initial and final stages of the process.
The measured output signal shows no dynamics corresponding to having complex
conjugate poles because the poles are close to the real axis in the discrete pole-zero
map, and its magnitude is small relative to the gain of the system.
Comparing the results from the RARR’s canonical rolls and the numerical example
in Table 1, the results for ARX and ARMAX gave poor modelling results and models
with stability problems for both. The modelling results of OBF in the numerical
example was better than in the real case in terms of the fit metric and the number of
parameters. In the case of OBF-ARX, the numerical example gave better performance
than the real case study, and the OBF-ARMAX case gave similar results for both with
almost the same number of parameters. OBF-ARMAX outperforms other models in
terms of its f it value.

6 Conclusions

A conventional model selection methodology was used and ARX, ARMAX, OBF,
OBF-ARX, and OBF-ARMAX were selected for modelling the RARR’s canon-
ical rolls. More flexible methods, such as ANN, were discarded after the data
82 O. B. Gonzalez and D. Rönnow

analysis step. We conclude that OBF-ARMAX outperforms ARX, ARMAX, OBF

and OBF-ARX for these type of systems, i.e., systems with a pair of complex con-
jugate poles close to the unit circle and real axis, and an exciting signal concen-
trated in the low-frequency band. Simulations were used to corroborate the results
obtained for the real production data. The problem of estimating the parameters of
the ARMAX model structure could be complemented by adding some inequality
constraints that take into account information of the pole positions obtained in the
frequency analysis. Choosing a model structure based on OBF-ARMAX brings the
possibility of encoding prior information of the pole positions into the model. It
also gives good identification properties due to the orthonormal input signals. OBF-
ARMAX is more demanding in terms of computations than the other methods. We
suggest that the obtained OBF-ARMAX can be used for monitoring the system and
predicting anomalies in a way similar to other machines in the industry, such as in
[1, 14].

References

1. Roy, R.B., Mishra, D., Pal, S.K., Chakravarty, T., Panda, S., Chandra, M.G., Pal, A., Misra, P.,
Chakravarty, D., Misra, S.: Digital twin: current scenario and a case study on a manufacturing
process. Int. J. Adv. Manuf. Technol. 107, 3691–3714 (2020)
2. He, B., Bai, K.: Digital twin-based sustainable intelligent manufacturing: a review. Adv. Manuf.
9, 1–21 (2021)
3. Xiang, F., Zhi, Z., Jiang, G.: Digital twins technolgy and its data fusion in iron and steel product
life cycle. In: 2018 IEEE 15th International Conference on Networking, Sensing and Control
(ICNSC), pp. 1–5 (2018)
4. Ljung, L.: System Identification: Theory for the User. Pearson, 2nd edn (1998)
5. Söderstrom, T., Stoica, P.: System Identification. Prentice-Hall, Inc. (1988)
6. Bautista Gonzalez, O., Rönnow, D.: Time series modelling of a radial-axial ring rolling system.
Int. J. Model. Identif. Control.
7. Akouemo, H.N., Povinelli, R.J.: Data improving in time series using ARX and ANN models.
IEEE Trans. Power Systems. 32, 3352–3359 (2017)
8. Ninness, B., Gustafsson, F.: A unifying construction of orthonormal bases for system identifi-
cation. IEEE Trans. Autom. Control. 42, 515–521 (1997)
9. Tufa, L.D., Ramasamy, M.: Closed-loop identification of systems with uncertain time delays
using ARX-OBF structure. J. Process. Control. 21, 1148–1154 (2011)
10. Tufa, L.D., Ramasamy, M.: Closed-loop system identification using OBF-ARMAX model. J.
Appl. Sciences. 10, 3175–3182 (2010)
11. Muddu, M., Narang, A., Patwardhan, S.C.: Reparametrized ARX models for predictive control
of staged and packed bed distillation columns. Control. Eng. Pract. 18, 114–130 (2010)
12. Nia, S.H.H., Lundh, M.: Robust MPC design using orthonormal basis function for the processes
with ARMAX model. In: Proceeding of the 2014 IEEE Emerging Technology and Factory
Automation (ETFA), pp. 1–8 (2014)
13. Ljung, L.: System Identification Toolbox: User’s Guide. MathWorks Incorporated (1995)
14. Mattsson, P., Zachariah, D., Björsell, N.: Flexible models for smart maintenance. In: 2019
IEEE International Conference on Industrial Technology (ICIT), pp. 1772–1777 (2019)
Pre-localization of Two Leaks in a Water
Pipeline Using Hydraulic and Spatial
Constraints

Lizeth Torres and Cristina Verde

Abstract This article presents a method to estimate frequency distributions of the

parameters associated with the presence of two leaks in a main water pipeline: the
locations and the emitter coefficients. Since the method is offline, it can be used to
locate simultaneous or sequential leaks; the only requirement is that both leaks must
be present when the data for diagnosis are acquired. The method requires as input
information pressure and flow rate data to be measured at the pipeline boundaries. In
a first stage, hydraulic quasi-steady state relations are computed by using the input
information. These relations are used to calculate hydraulic and spatial constraints. In
a second stage, Monte Carlo simulations are used together with the previously defined
restrictions to obtain frequency distributions of the parameters associated with both
leaks. To test the feasibility and potentiality of the method, it was implemented and
tested in MATLAB with synthetic data.

1 Introduction

The location of leaks in pipelines has been approached in various ways depending
on the type of leak that is searched. Leaks can be roughly classified as: (i) single,
(ii) sequential and (iii) simultaneous. The localization of the third class of leaks is
one of the most challenging problems since it is not possible to estimate the exact
locations of two or more simultaneous leaks by using flow rates and pressures taken
at the pipeline boundaries when the flow is in steady state [1]. Therefore, the unique
possibility to estimate the exact locations of simultaneous leaks with hydraulic data
is to force the flow to be in unsteady or steady-oscillatory state as done by Navarro
et al. [2].

L. Torres (B) · C. Verde

Instituto de Ingeniería UNAM, Coyoacán, Mexico City, Mexico
e-mail: [email protected]
C. Verde
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 83

D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_8
84 L. Torres and C. Verde

1.1 Related Work

There are few works that present a solution to the problem of locating of two leaks
(simultaneous or sequential leaks). Some of these works have been taken up to
propose the method presented in this work; for this reason, they are briefly following
described. In [3], Verde et al. presented an offline algorithm to identify the parameters
associated with two leaks: two positions and two emitter coefficients. This algorithm
is based on a family of nonlinear lumped models with the same steady state behavior
in leak condition and expressed in terms of an equivalent position. Each member of the
family is parameterized with different positions and emitter coefficients from those
of another member of the family. To determine which of the members is the correct
model, i.e., the model with the closest position values and coefficients to the real ones,
the transient response of each member is compared with the transient response of
the real pipe with two leaks. The parameters of the member that generates a transient
response more similar to the real flow response of the pipeline are the parameters
that are taken as the best estimations. The key to the success of this algorithm is the
generation of a transient to evaluate the responses of the models because it allows
the identifiability of the parameters. For improvement purposes, in [4] Verde et al.
proposed an algorithm similar to that proposed in [3], but with a significant change:
the number of parameters to be identified were reduced to three by using some
hydraulic constraints, which were derived from equations of motion and continuity.
Navarro et al. submitted the most recent solution in [2], where an approach based on
the numerical differentiation of the boundary measurements was proposed. Although
the procedure for inverting the nonlinear map was not presented, the authors claim
that the leak parameters can be recovered from the boundary measurements and their
time derivatives. A requirement for using this approach is that the derivatives of the
boundary measurements must be non-zero, preferably oscillatory, as well as nearly
noise-free.

1.2 The Essence of the Proposed Method

Taking inspiration from the mentioned works, and in order to take a step forward
towards a realistic and implementable solution, this article presents a leak pre-
localization method based on a 2-stage offline algorithm. The main strategy of this
algorithm is the use of Monte Carlo simulations to estimate the frequency distri-
butions of the parameters that are associated with the presence of two leaks: (i)
positions and (ii) emitter coefficients; see Fig. 1. This task is done by replacing each
hydraulic variable that is not exactly known with a range of values (a probability
distribution). The algorithm then computes the results over and over again, each iter-
ation using a different set of random values generated from probability distributions,
which change at each iteration. In a first stage, the parameters of a fictitious leak
are calculated. These fictitious parameters are associated to the parameters of two
real leaks that are actually present in the pipeline. These fictitious parameters will
Pre-localization of Two Leaks in a Water Pipeline … 85

Fig. 1 The goal of the method is the estimation of the probability distributions of the parameters
associated with each leak. Notice that each leak has associated a position and an emitter coefficient

help, in the next stage, to define spatial constraints. In a second stage, frequency
distributions of the parameters associated with both leaks are obtained by means
of Monte Carlo simulations and the spatial constraints. To obtain the parameters of
the fictitious leak, the rigid water column (RWC) theory is used. Models under this
theory can be obtained by assuming that the walls of the pipelines are rigid and the
flowing fluid is incompressible [5–8].

1.3 Disadvantages and Advantages

(i) The proposed method is based on an offline algorithm. An advantage of this fact
is that the leak time is not relevant for the method and, therefore, it can be applied for
both sequential and simultaneous leaks. (ii) Since the method is based on algebraic
equations, the sampling time is not important. (iii) This method assumes the existence
of two leaks, reason by which it is not a complete diagnosis approach, but solely a
localization approach, i.e. it does not detect if there are one, two or more leaks in
the pipeline. Detecting the amount of leaks before their location is still a pending
issue. (iv) This method does not provide a unique and exact solution, but frequency
distributions of the parameters associated with the leaks. (v) The main advantage
of the proposed method is that it does not require forcing the behavior of the flow
in an unsteady state, which is not possible in most real pipelines operating in the
world. The price to pay, however, is to get only likely locations. Section 2 presents
the physics-based mathematical models, as well as auxiliary equations used for the
design and implementation of the algorithm. Section 3 explains with details the main
algorithm of the method. Section 4 presents some simulation results that show the
method feasibility. Section 5 concludes this work.

2 Rigid Water Column Model

The proposed method involves useful mathematical relationships, which can be

derived from the so-called rigid water column model (RWC model). The assump-
tions and physics behind the RCW model can be consulted in [9, 10]. By assuming
86 L. Torres and C. Verde

that the flow rate is unidirectional and the pipeline has not extractions and/or devices
(such as valves or elbows), the RWC model for a pipeline section (denoted by z)
is expressed by the following equation:

H
Q̇ = θ − α Qγ , (1)
z

where Q is the flow rate flowing through the pipeline section z = z L − z 0 . z 0

denotes the coordinate of the upstream boundary of the pipeline and z L the down-
stream boundary. H = Hin − Hout is the pressure head loss along z, where Hin
and Hout are the upstream and downstream hydraulic head,1 respectively. θ = g Ar ,
where g is the gravity acceleration and Ar is the cross-sectional area. α and 1 ≤ γ ≤ 2
are parameters related with the friction losses, and they can be associated to physical
parameters of the pipeline and fluid by means of the formula employed to describe
such losses (e.g. the Darcy-Weisbach equation, the Hazen-Willian equation or the
Gaukler-Strikler formula).

2.1 One-Leak Moment Equations

The two following equations express the flow motion along a pipeline with a leak,
i.e., describe the flow trough two different sections of the pipeline:

θ  
Q̄˙ in =
γ
H̄in − H̄ − α Q̄ in ,
z in
(2)
θ  
Q̄˙ out =
γ
H̄ − H̄out − α Q̄ out .
z out

where Q̄ in and Q̄ out are the flow rates before and after the leak, respectively,
z in = z  − z 0 and z out = L − z  are the sections to the left and the right of the
leak, respectively, z 0 = 0 is the origin coordinate, z L = L is the downstream-edge
coordinate, z  is the leak coordinate (leak position), H̄ is the hydraulic head at the
leak position, Hin and Hout are the hydraulic heads at the upstream and downstream
boundaries, respectively. Check Fig. 2 for a graphic illustration of this description.
The total flow
 rate of the leak can be calculated in two ways: with the equation
Q̄  = C A 2g H̄ , where C is the discharge coefficient and A is the leak area, or
with
Q̄  = Q̄ in − Q̄ out . (3)

Be aware that ¯ over a variable means that the variable is associated to the one-leak
model.

P
1 The hydraulic head is defined as H = + z, where P is pressure, ρ is density, g is the gravity
gρ
acceleration and z is the elevation at the piezometer bottom.
Pre-localization of Two Leaks in a Water Pipeline … 87

Fig. 2 Pipeline with a single

leak

2.2 Two-Leaks Moment Equations

The three equations that the describe the flow motion in a pipeline with two leaks
are the following:

θ  
Q̇ in = Hin − H1 − α (Q in )γ ,
z in
θ  
Q̇ 12 = H1 − H2 − α (Q 12 )γ , (4)
z 12
θ  
Q̇ out = H2 − Hout − α (Q out )γ ,
z out

where Q in , Q 12 and Q out are flow rates trough sections z in = z 0 − z 1 , z 12 =

z 2 − z 1 and z out = L − z 2 , respectively. z 0 , z 1 , z 2 and L are the coordinates of
the upstream edge, the position of the first leak, the position of the second leak and the
pipe downstream edge, respectively. Hin and Hout are the hydraulic heads at the inlet
an outlet ends, whereas H1 and H2 are the hydraulic heads at the leak coordinates.
The flow rates of both leaks can be described by the following equations:
 
Q 1 = λ1 2g H1 , Q 2 = λ2 2g H2 , (5)

where λ1 = C1 A1 and λ2 = C2 A2 are the emitter coefficients. C1 and C2 are
the discharge coefficients, respectively, and A1 and A2 are their areas. A visual
description of the variables involved in the moment equations is given in Fig. 3.

Fig. 3 Pipeline with two

leaks
88 L. Torres and C. Verde

3 Hydraulic Relations

3.1 Parameters of a Fictitious (Equivalent) Leak

An equivalent leak is a fictitious single leak with parameters (a position and an

emitter coefficient) calculated with information acquired at the boundaries of a pipe
that actually has two real leaks. The equivalent leak parameters are very useful
because they can be used to define hydraulic constraints, which in turn can be used
to form probability distributions of the two actual leak parameters. The parameters
of a equivalent leak can obtained by assuming that the flow in a pipeline with two
leaks is in steady state. If this assumption holds, then

Q in = Q̄ in and Q out = Q̄ out ,

which means that the flow rates at the boundaries of a pipeline with a single fictitious
leak are equivalent to the boundary flow rates of a pipeline with two leaks. There
are four parameters associated to the fictitious leak: equivalent position, equivalent
flow rate, equivalent pressure and equivalent emitter coefficient. The constitutive
laws of these parameters are following described. It is worth mentioning that these
constitutive laws are hereinafter called equivalence relations, since they relate the
parameters of a fictitious leak with those of two real leaks.

3.1.1 Equivalent Position

In [11] was shown that the information of the equivalent position helps to delimit
the section where each leak is located without implanting additional sensors to those
already located on the boundaries. The constitutive law of the equivalent position,
which can be obtained from equations (2) in steady state, is given by
   γ
θ H̄in − H̄out − Lα Q̄ out
z =  γ  γ  ; where z 1 < z  < z 2 . (6)
α Q̄ in − Q̄ out

3.1.2 Equivalent Flow Rate

The equivalent flow rate is nothing more than the sum of the two leakage discharges:
Q 1 and Q 2 . The equivalent flow rate, Q  , can be estimated using the leak discharges
if the positions of the leaks are known. On the contrary, it can be estimated with

Q  = Q 1 + Q 2 = Q̄ in − Q̄ out . (7)
Pre-localization of Two Leaks in a Water Pipeline … 89

3.1.3 Equivalent Hydraulic Head

The equivalent hydraulic head results from calculating the pressure at the assumed
coordinate (position) of the equivalent leak by using (2) in steady state, that is

αz  γ α(L − z  ) γ
H = H̄in − Q̄ in ; H = H̄out + Q̄ out . (8)
θ θ

3.1.4 Equivalent Emitter Coefficient

There exist two manners to calculate the equivalent emitter coefficient. The first one
is by using the following relation:

Q
λ = √ (9)
H

The second one is by adding both leak coefficients:

 
λ = C1 A1 2g + C2 A2 2g = λ1 + λ2 . (10)

3.2 Estimation of the Two-Leaks Parameters

The following equations are used to estimate the parameters of both leaks. The
hydraulic heads, H1 and H2 , can be calculated with (4) in steady state, such that (5)
can be rewritten as follows:
  
α(z 0 − z 1 )
Q 1 = λ1 2g Hin − (Q in )γ ,
θ
   (11)
α(L − z 2 )
Q 2 = λ2 2g Hout + (Q out )γ .
θ

where λ1 and λ2 can be calculated from (10). The leak position can be calculated
from sections z in and z out as follows: z 1 = z in and z 2 = L − z out .

3.3 Constraints

The algorithm uses the following constraints to remove (filter) estimates that do not
make physical sense:
• The position of the first leak is positive and it is not located at the upstream edge.
90 L. Torres and C. Verde

z in > 0 (12)

• The position of the first leak is before the fictitious equivalent leak.

z in < z  (13)

• The position of the second leak must be before the downstream edge.

(L − z out ) < L (14)

• The position of the second leak must be after the fictitious equivalent leak.

(L − z out ) > z  (15)

• The emitter coefficients of both leaks should be limited by an upper bound and a
lower bound.
1 < λ1 < λ1 ; λ 2 < λ2 < λ2 ,
λmin max min max
(16)

where the maximum values of λ1 and λ2 can be calculated as follows

Q
λmax
1 = λmax
2 =√ .
H

4 Method Description

The method is based on the following Monte Carlo-based algorithm, which estimate
the frequency distributions of the parameters associated with the two leaks, i.e., λ1 ,
λ2 , z in and z out .

Algorithm

1. Load physical and algorithm parameters: α, γ , θ , N , λmin1 , λ2 , λ1 , λ2 .

min max max

2. Compute z  , Q  , H and λ with (6), (7), (8) and (9), respectively.

3. Start a loop to generate random values for λi1 and z in
i
from uniform distributions.
i is the index used to denote each random value (a particle). This loop ends when
i = N , where N is a predefined value.
for i = initial value : step : f inal value (N )
(a) Generate random sample sets for λi1 from a bounded prior (uniform) probability
distribution:
λi1 ∼ U [λmin
1 , λ1 ]
max
Pre-localization of Two Leaks in a Water Pipeline … 91

(b) Generate random sample sets for z in

i
from a bounded prior (uniform) probability
distribution:
z in
i
∼ U [0, z  ]

(c) Calculate: (1) λi2 from λ and λi1 , (2) Q i1 from λi1 and z in
i
, (3) Q i2 from Q 
and Q i1 , (4) z out
i
from Q i2 and λi2 . It is important to follow the order, since it is
a staggered estimation. The calculated values together with the values generated in
this step compose a i set.
(d) Verify that the set of values i comply with restrictions (12)–(16). If they do not
meet them, the set of values is discarded, if they do meet them, the set of values are
saved in vectors as follows:
j j j j
Zin [ j] ← z in , Zout [ j] ← z out , 1 [ j] ← λ1 , 2 [ j] ← λ2 ;

where j is the index of each element that meets the restrictions.

End
4. From the vectors, generate frequency distributions for the parameters of the two
leaks.
5. Calculate the mean and standard deviation of each distribution.

5 Simulation Tests

The results presented in this section were obtained from simulations executed in
Matlab. For this purpose a simulator was implemented by using the momentum
and continuity equations that describe the flow in a pipeline. These equations were
approximated by using finite differences and were numerically solved by using the
Runge-Kutta method. The parameters used for the simulator were: Ar = 0.0045365
(m2 ), g = 9.81 (m/s2 ), L = 163.72 (m) α = 19.43, γ = 1.8794. The boundary con-
ditions that were imposed for the numerical solutions were Hin = 20.5 (m) and
Hout = 5.8 (m). The solver time step was fixed at ts = 0.01 (s). The algorithm was
launched using the followin parameters: λmin1 = λ2 = 0.1λ and N = 1 × 10 . In
min 5

Table 1, three cases of two leaks in a pipeline are presented.

Table 1 Leak cases

Case number Parameter values
Case 1 z in = 20 (m), z out = 100 (m), λ1 = 6 × 10−4 , λ2 = 3.5 × 10−4
Case 2 z in = 50 (m), z out = 90 (m), λ1 = 5 × 10−4 , λ2 = 3 × 10−4
Case 3 z in = 60 (m), z out = 30 (m), λ1 = 4 × 10−4 , λ2 = 4 × 10−4
92 L. Torres and C. Verde

Table 2 Results of the pre-localization

Case number Pre-localization results
μ σ = 9.33, z μ σ
Case 1 z in = 20.47, z in out = 86, z out = 36.16
μ σ μ σ = 30.56
Case 2 z in = 37.96, z in = 15.30, z out = 62.16, z out
μ σ μ σ
Case 3 z in = 65.90,z in = 21.96, z out = 40.59, z out = 21.31

Fig. 4 Frequency distributions of the leaks’ parameters for Case 3

The results of the prelocalization are summarized in Table 2. The values with the
super-index μ are the means of the frequency distribution of z in and z out . The
values with the super-index σ are the standard deviations of the frequency distribution
of z in and z out . The frequency distributions of the results for Case 3 are illustrated
in Fig. 4.

6 Conclusions

A method for pre-localizing two leaks in a pipelines was presented. This method
does not require oscillating the flow or causing a transient. Some simulation results
of three leaks’ cases were presented. The algorithm is perfectible, so in future works,
some improvements to the method will be presented, such as the incorporation of
changes in the operating points, iterations with cost functions to reduce the standard
deviation of the distributions. In addition, experimental results will also be presented
to show the advantages and disadvantages of the method in the real world.

References

1. Verde, C., Visairo, N.: In: Proceedings of the American Control Conference, Boston, USA,
vol. 5, pp. 4378–4383 (2004)
2. Navarro-Díaz, A., Delgado-Aguiñaga, J.A., Begovich, O., Besançon, G.: Sensors 21(23), 8035
(2021)
3. Verde, C., Visairo, N., Gentil, S.: Adv. Water Res. 30(8), 1711 (2007)
4. Verde, C., Molina, L., Torres, L.: J. Loss Prev. Process Ind. 29, 177 (2014)
Pre-localization of Two Leaks in a Water Pipeline … 93

5. Wood, D., Funk, J., Boulos, P.: In: Proceedings of the 6th International Conferences on Pressure
Surges, pp. 131–142 (1990)
6. Ivanov, K.P., Bournaski, E.G.: Comput. Methods Appl. Mech. Eng. 130(1–2), 47 (1996)
7. Axworthy, D.H.: Water Distribution Network Modelling: from Steady State to Waterhammer.
University of Toronto (1998)
8. Kaltenbacher, S., Steffelbauer, D.B., Cattani, M., Horn, M., Fuchs-Hanusch, D., Römer, K.:
Proc. Comput. Control Water Ind. 2–134 (2017)
9. Cabrera, E., Garcia-Serra, J., Iglesias, P.L.: In: Improving Efficiency and Reliability in Water
Distribution Systems. Springer, pp. 3–32 (1995)
10. Nault, J., Karney, B.: J. Hydraul. Eng. 142(9), 04016025 (2016)
11. Verde, C.: Control Eng. Pract. 9(6), 673 (2001)
Diagnosis and Failure Prognosis
of Intermittent Faults: A Bond Graph
Approach

Wolfgang Borutzky

Abstract Diagnosis of incipient parametric faults has been widely addressed and
various model-based as well as data-based methods have been reported in the lit-
erature. In contrast to incipient faults, it is unknown if, when and for how long
intermittent faults will reappear which makes their diagnosis and prognosis difficult.
This paper addresses intermittent faults with a magnitude that increases over time.
Such a sequence of intermittent faults may reach a failure alarm threshold and may
eventually lead to a component or even a system failure. The proposed approach
uses a diagnostic Bond Graph (DBG) model for an online detection, isolation and
estimation of intermittent faults. Pulses of increasing height and varying width not
equally spaced on the time evolution of residuals of Analytical Redundancy Relations
(ARRs) obtained from an offline developed DBG indicate a degradation trend of a
parameter or a variable. Certain values of this trend are concurrently extrapolated in
a repeated failure prognosis resulting in a sequence of Remaining Useful Life (RUL)
estimates. The proposed BG-based approach to a failure prognosis in the presence
of intermittent faults can be combined with a previously presented BG-approach to
incipient fault prognosis and is applicable to mode switching models. For illustration,
it is applied in a case study to a small electronic circuit in an offline simulation. The
case study also considers intermittent faults of constant magnitude.

Keywords Intermittent faults · Increasing fault magnitude · Diagnostic bond

graph-based online fault diagnosis · Concurrent repeated failure prognosis ·
Remaining Useful Life (RUL) estimates

1 Introduction

Failure prognostic is a constitutive part of Prognostic and Health Management (PHM)

and is of imperative importance beyond fault diagnosis for all safety critical engi-
neering systems and processes, for supervision, automation and condition based

W. Borutzky (B)
Bonn-Rhein-Sieg University of Applied Sciences, Sankt Augustin, Germany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 95

D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_9
96 W. Borutzky

maintenance (CBM) of industrial processes. Fault detection and isolation (FDI) as a

prerequisite for failure prognostic has been a subject of research with regard to var-
ious applications and has become a quite mature discipline in the course of the last
decades. Data-driven as well as physics model-based approaches to FDI are widely
used in industry and in academia. So far, FDI has mostly addressed abrupt as well as
incipient faults. In recent years, research has extended to FDI of intermittent faults.
As to failure prognosis, more recently, combinations of data-driven and model-
based approaches have received increasing attention. In comparison to fault diagno-
sis, the combined use of model-based and data-driven methods for failure prognosis
is a still rather young and is still a developing research subject with contributions
from various fields. A 2016 survey on prognostics may be found in [2]. Combined
bond graph model-based, data-driven approaches have been presented in [3].
Failure prognosis, so far, mostly considers incipient parametric faults so that a
degradation trend can be extracted from measurements in a moving time window
and can be projected into the future in order to obtain estimates of the remaining
useful life (RUL). In contrast, the occurrence of intermittent faults, their duration
and their magnitude are unpredictable. Accordingly, while diagnosis of intermittent
faults has been subject of ongoing research, the effect of intermittent faults on the
RUL of system is still a rather new topic.
With time, the number of intermittent faults may increase and may develop into a
permanent fault, or their magnitude may increase until an admissible alarm threshold
is reached and action is required in order to avoid a failure. The BG approach to a
prognosis of intermittent faults presented in the following assumes that the magnitude
of intermittent faults increases over time and is a modification of the BG model-
based, data driven approach to the prognosis of incipient parametric faults in [1].
This assumption has also been adopted in a recent case study [4].

2 Bond Graph Approach to Intermittent Fault Prognosis

The proposed BG model-based approach assumes that intermittent faults of increas-

ing magnitude happen and consists of two parts. The first part based on the evaluation
of ARRs addresses the detection, isolation and estimation of intermittent faults. The
second part computes a sequence of RUL estimates.

2.1 Bond Graph Model-Based Fault Diagnosis

The first part of the proposed BG model-based approach to failure prognosis in

the presence of intermittent faults, i.e. FDI and fault estimation as a prerequisite
for failure prognosis, is performed concurrently to the continuous monitoring of
a system on an offline developed DBG. The latter one is obtained from a direct
BG model by inverting the causality of detectors representing sensors. The reason
for this causality inversion is that sensors deliver measurements, i.e. known inputs
into the DBG which serves as generator of constraints between measurements and
Diagnosis and Failure Prognosis of Intermittent Faults: A Bond Graph Approach 97

known input command signals. These constraints are called Analytical Redundancy
Relations (ARRs). They may be nonlinear. In the case of BG models with switches,
ARRs are even mode-dependent. Switch states may switch off and on parts of the
set of ARRs. Their numerical evaluation denoted as residuals should result in values
close to zero in case no faults happen. The time evolution of a residual enables to
detect intermittent faults.
The dependencies of ARRs from element parameters can be captured in a struc-
tural fault signature matrix (FSM) which indicates whether a fault can be detected
and, moreover, can be isolated. In the case of a unique fault signature, a potentially
faulty parameter can be isolated just by inspection of the FSM. In case several ele-
ments do have the same fault signature, numerical parameter estimation can identify
the parameters that deviate from their nominal values. Once a faulty parameter has
been isolated, the magnitude of a fault can be estimated and its unknown faulty time
behaviour, i.e. a degradation trend, can be reconstructed. As an example, in the case
study in Sect. 3, the power supply of a small electronic circuit is not stable in the
sense that it does not deliver a constant voltage E as expected. Instead, the voltage
value is affected by intermittent faults with a magnitude that increases with time. The
computation of an ARR residual is used to reconstruct the unknown time behaviour
Ẽ(t) of the faulty voltage supply. If the time evolution of a residual reveals a sequence
of intermittent faults of increasing magnitude, the latter ones can be used to compute
a sequence of RUL estimates.

2.2 Repeated Prediction of RUL Estimates

The second part of the proposed approach uses the ARR residuals obtained in the
first part for a repeated prediction of RUL estimates. Once online monitoring of
the health of a system has resulted in a residual different from zero, the monitoring
has to proceed until the residual value returns to zero before one can conclude that
an intermittent fault happened. If, instead, the residual retains the nonzero value
for quite some time, one may conclude that an abrupt fault happened. In case the
residual returns to zero at some later time instant, it is clear that an intermittent
fault of unforeseeable length has happened. Moreover, even when the residual under
consideration returns to zero, one cannot be sure that another intermittent fault will
take place at any time instant later. Therefore, to be able to identify some degradation
trend and to predict a RUL estimate, two consecutive intermittent faults must be
detected and the latter one must be of higher magnitude. A first estimate of a RUL
may be obtained from a reconstructed fault signal by determining the intersection of
an interpolation line through the points (t1 , v1 ), (t2 , v2 ) with an alarm threshold line.
The time point ti is the time instant where a considered residual returns to zero, i.e
where the length of the ith intermittent fault ends.
If subsequent intermittent faults happen while time moves on, the process can
be repeated. The result is a sequence of RUL estimates that converges to zero with
ongoing time. If the monotonic increase of intermittent fault magnitudes stops at
some time instant, then RUL prediction is no longer possible but can resume when
98 W. Borutzky

intermittent fault magnitudes increase again. As long as there is no intersection of

an interpolating line with an alarm threshold no action is required and a system
may continue its operation. However, the time interval between two consecutive
intermittent faults may become shorter, their number may increase and may result
in a permanent fault.
The approach to a repeated prediction of RUL estimates starting from ARR resid-
uals obtained from a diagnostic BG is illustrated in Sect. 3 by means of an offline
simulation. However, the diagnosis of intermittent faults as well as the subsequent
repeated computation of RUL estimates can be performed online concurrently to the
monitoring of the dynamic behaviour of a system.

3 Offline Simulation Case Study

The approach to a detection of intermittent faults and a prediction of RUL estimates

is illustrated by means of the simple circuit in Fig. 1.
First, it is assumed that the element parameter do not become faulty over time but
keep their nominal value. The power supply, however, is not stable. Intermittently,
the constant voltage E drops to a lower value which becomes smaller over time until
a value is reached at which the circuit does not function properly any more. The time
evolution of the intermittent voltage supply Ẽ(t) is depicted in Fig. 2a. As can be
seen, the faults do not occur periodically and their length is not constant.
In this study, the real circuit is replaced by a BG model. To obtain pseudo mea-
surements, the time evolution of the voltages numerically computed by means of
the open source simulation program Scilab are overlayed with noise. These ‘mea-
surements’ are then filtered before they are input into a DBG for the computation of
ARRs. The offline simulation uses the parameter in Table 1.
The noisy voltages are smoothed by the Scilab function lsq_splin(), which
performs a weighted least squares cubic spline fitting and provides a smoothed signal
together with its time derivative. Fig. 2b shows the time evolutions of the smoothed
capacitor voltages ũ 1 , ũ 2 .

Sw

R1 R2

E C1 V C2 V

Fig. 1 Simple network with a switch and two voltage sensors

Diagnosis and Failure Prognosis of Intermittent Faults: A Bond Graph Approach 99

(a) (b)
Fig. 2 a Voltage supply Ẽ(t) b smoothed capacitor voltages ũ 1 , ũ 2

Table 1 Parameters of the switched circuit in Fig. 1

Parameter Value Units Meaning
E 5 V Voltage supply
E thr 1 V Fault threshold
R1 500  Resistor R : R1
R2 5 k Resistor R : R2
Ron 0.1  ON resistance of switch of Sw
C1 5000 μF Capacitor C : C1
C2 1000 μF Capacitor C : C2
C2crit 400 μF Fault threshold
tsw 10 s Switching point of Sw
t0 5 s Start of the decay of C̃2 (t)
λ 0.0308065 s −1 Rate of the decay

3.1 BG Based Detection and Estimation of Intermittent

Faults

The circuit schematic in Fig. 1 is transformed into the DBG in Fig. 3, in which the
voltages with a tilde denote measurements from the faulty system. It is assumed that
the voltage sensors themselves presented by the detectors De : ũ 1 and De : ũ 2 are
faultless.
From the DBG in Fig. 3 the following two mode-dependent ARRs can be deduced.
100 W. Borutzky

Fig. 3 Diagnostic bond

graph of the network in
Fig. 1

E − ũ 1 b2
01 : ARR1 : r1 = i 1 − i sw − C1 ũ˙ 1 = − (ũ 1 − ũ 2 ) − C1 ũ˙ 1
R1 Ron + b · R2
E − ũ 1 b
= − (ũ 1 − ũ 2 ) − C1 ũ˙ 1 (1)
R1 R2
b
02 : ARR2 : r2 = i sw − C2 ũ˙ 2 = (ũ 1 − ũ 2 ) − C2 ũ˙ 2 , b(t) ∈ {0, 1} (2)
R2

The FSM (Table 2) is a book keeping of which element parameters affect which
residual. The last but one column of the FSM indicates whether a faulty parameter
can be detected given the two sensors. An entry in the last column records whether
a parameter is isolatable.
As can be seen, non of the element parameters is isolatable, except C : C2 . Residual
r1 depends on elements R : R1 , C : C1 and on R : R2 in case the switch is closed. As
a result, by inspection of the FSM it cannot be decided which parameter is faulty and
causes an abnormal behaviour of the circuit. By means of parameter estimation over
a time window it can be shown that the element parameters, in fact, do not deviate
from their nominal values. To that end, the vector of parameter dependent residuals
can be used to define a cost function that is minimised

Table 2 Structural FSM for the switched network in Fig. 1

Element Parameter r1 r2 Db Ib
Voltage source E 1 0 1 0
De : ũ 1 ũ 1 1 1 1 0
De : ũ 2 ũ 2 1 1 1 0
Switch b b b b 0
R : R1 R1 1 0 1 0
R : R2 R2 b b b b
C : C1 C1 1 0 1 0
C : C2 C2 0 1 1 1
Diagnosis and Failure Prognosis of Intermittent Faults: A Bond Graph Approach 101

Given that the element parameters keep their nominal values, then numerical
computation of (1) yields for the residual res1 = R1 · r1 a time evolution that clearly
enables to detect the intermittent faults (Fig. 4).
Moreover, residual res1 can be used to estimate the magnitude of the intermittent
faults and to reconstruct the unknown faulty voltage supply Ẽ. If the constant voltage
E in (1) is replaced by the faulty unknown one Ẽ, then ARR1 reads:

Ẽ − ũ 1 b
0 = − (ũ 1 − ũ 2 ) − C1 ũ˙ 1 (3)
R1 R2

Solving for Ẽ and observing (1) gives

r Ẽ = E − R1r1 = E − res1 = Ẽ (4)

E = res1 (5)

By comparison of Fig. 4a and b one can see that the residual values equal the
magnitude of the intermittent faults.

3.2 RUL Prediction for Intermittent Faults

From Fig. 4 one can see that the intermittent faults return to zero at time instances
t = 10, 30, 40, 48 s. Connecting the values of the reconstructed time evolution of the
faulty voltage supply r E(t) at two consecutive time points yields lines that intersect
with an alarm threshold defined to be 1 V as depicted in Fig. 5.
The time points of the intersections with the alarm threshold yield the decrease of
the RUL over time shown in Fig. 6b. The offline simulation run up to t = 50 s takes
into account that online measurement cannot look into the future. Measurements only
available up to the current time instant can be used for computing ARR residuals

(a) (b)
Fig. 4 a Residual res1 = R1 · r1 b Reconstructed input voltage rtE and faulty input voltage tE
102 W. Borutzky

Fig. 5 Prediction of the degradation trend of the intermittent faults

(a) (b)
Fig. 6 a Determination of a last RUL value different from zero b Decrease of the RUL over time

and for reconstruction of a time behaviour subject to intermittent faults. Although

the latest intermittent fault return to zero at t = 48 s, we cannot be sure at that time
point that another intermittent fault will follow.
Given that continued measurement yields another intermittent fault resulting in a
value of the reconstructed faulty voltage r Ẽ that is close to the alarm threshold value
E thr up to a value  ≥, 0 i.e. |r Ẽ(t) − E thr | ≤  or even below E thr . As an alarm
threshold is chosen well with a distance to the failure threshold, the faulty supply
voltage can still cross the alarm threshold due to an intermittent fault without causing
harm to the system. Let t4 be the last time instant where r Ẽ(t4 ) still has a distance
to the alarm threshold and where r Ẽ returns to the constant value E and let t f be
the time point where r Ẽ(t f ) touches or crosses the threshold for the first time. The
last RUL value different from zero then reads RU L(t4 ) = t f − t4 and RU L(t f ) = 0.
Time instant t4 is known as End-of-Prediction (EoP). Figure 6a shows the case where
the true faulty voltage crosses the alarm threshold due to the last intermittent fault.
The last RUL value different from zero is RU L(t4 = 48 s) = 2 s and t f = 50 s.
Diagnosis and Failure Prognosis of Intermittent Faults: A Bond Graph Approach 103

3.3 A Switch with Intermittent Faults

As intermittent faults are common in electronic interconnection systems, a second

scenario is considered, in which it is assumed that the pass transistor in the circuit of
Fig. 1 has connection problems. Instead of connecting the two capacitors permanently
as of tsw = 10 s, the switch modelling the transistor closes in an unpredictable manner
for only short time spans as depicted in Fig. 7. All other components of the circuit,
including the constant voltage supply are assumed to be faultless.
As can be seen from Fig. 7, the rise of capacitor voltage ũ 1 is very little affected by
the faulty behaviour of the switch. At tsw = 10 s, the switch closes and connects the
capacitors. As of that time instant, capacitor C : C2 is also charged which results in a
slight decline of ũ 1 . However, since the switch immediately opens again after 2 s and
remains open for 2 s, voltage ũ 1 increases again.At t = 14 s, the switch closes again
and remains closed for 8 s which is enough for voltage ũ 1 to reach the steady-state
value of 5 V .
In contrast, capacitor voltage ũ 2 keeps the value it has reached whenever the
switch opens for a time span. As a result, ũ 2 reaches a steady-state value that is
0.22 V below 5 V .
Evaluating (2) and plotting the result yields the time evolution of residual r2 (t)
displayed in Fig. 8a. Figure 8a clearly indicates that residual r2 is different from zero
whenever the switch is open.
In case the faulty switching behaviour b̃(t) were known, residual r2 would vanish.

b̃
0 = (ũ 1 − ũ 2 ) − C2 ũ˙ 2 (6)
R2

Substraction of (6) from (2) yields

b b̃ b − b̃
r2 = (ũ 1 − ũ 2 ) − (ũ 1 − ũ 2 ) = (ũ 1 − ũ 2 ) (7)
R2 R2 R2

Fig. 7 Faulty switching

behaviour b̃(t) of the
transistor and effect of the
faulty switching of the
transistor on the capacitor
voltages u 1 , u 2
104 W. Borutzky

(a) (b)
Fig. 8 a Time evolutions of residual r2 (t) and b̃(t) b Recovery of the time evolutions of the faulty
switching function b̃(t)

In a similar way, subtracting (3) from (1), one obtains for residual r1

b̃ − b
r1 = (ũ 1 − ũ 2 ) (8)
R2

That is, r2 = −r1 .

Solving (7) gives an equation for recovering the unknown faulty switching
behaviour b̃(t).

R2 · r 2
rb = b − (9)
ũ 1 − ũ 2

Figure 8b indicates that r b(t) well recovers the true unknown faulty time evolution
b̃(t).
As a result, from measurements available up to a current time point one can
compute the time evolution of an ARR residual that enables to detect intermittent
faults and to recover an unknown faulty time behaviour of a system variable.
In the considered scenario, the faulty switching of the transistor does not cause a
failure of the circuit. However, the transient behaviour of voltage ũ 2 may be too slow
and may not meet requirements of the system to which the circuit is connected.

3.4 A Sensor with Intermittent Faults

Residuals derived from a DBG can also be used to detect a sensor with intermittent
faults and to reconstruct the faultless signal. Let’s assume that the circuit in Fig. 1
doesn’t have any parametric faults. If the measurements of the capacitor voltages are
correct then the residuals of ARRs (1), (2) are zero or close to zero. In contrast, if
sensor De : ũ 1 provides a faulty reading ũ 1 (t) subject to intermittent faults due to
connection problems then the residuals of both ARRs are different from zero.
Diagnosis and Failure Prognosis of Intermittent Faults: A Bond Graph Approach 105

Assume for simplicity that the second sensor De : ũ 2 provides faultless readings.
ARR (2) then yields for the reconstruction of the faultless signal ũ 1 (t)

R2 · r2 = b (ũ 1 − ũ 1 ) (10)
r ũ 1 = ũ 1 − R2 · r 2 = ũ 1 − res2 for t ≥ tsw (11)

Conclusion

The paper proposes a BG model-based approach to diagnosis and failure prognosis

of intermittent faults with a magnitude that increases with time. Faults are detected
by evaluating ARRs derived from an offline developed DBG. In an offline simulation
case study of a switched electronic circuit, the constant voltage supply is subject to
intermittent faults. Intermittent faults of constant magnitude on the switch as well
as faulty sensor readings due to intermittent faults are also analysed. Simultaneous
intermittent and incipient parameter faults are addressed elsewhere. due to the lack
of space. Parameter uncertainties have been disregarded but can be well taken into
account in the DBG. As intermittent faults are common in electronic interconnec-
tion systems, the approach is applied to further electronic systems in which contact
problems due to corrosion, moisture, or vibration may happen in an unpredictable
manner, for instance, in electromechanical relays.

References

1. Borutzky, W.: Bond Graph Modelling for Control, Fault Diagnosis and Failure Prognosis.
Springer International Publishing, Switzerland (2020)
2. Elattar, H.M., Elminir, H.K., Riad, A.M.: Prognostics: a literature review. Complex Intell. Syst.
2016(2), 125–154 (2016). https://doi.org/10.1007/s40747-016-0019-3, Open access
3. Jha, M., Dauphin-Tanguy, G., Ould Bouamama, B.: Particle filter based integrated health mon-
itoring in bond graph framework. In: Borutzky, W. (ed.) Bond Graphs for Modelling, Control
and Fault Diagnosis of Engineering Systems, pp. 233–270. Springer International Publishing,
Switzerland (2017)
4. Yu, M., Lu, H., Wang, H., Xiao, C., Lan, D.: Compound fault diagnosis and sequential prog-
nosis for electric scooter with uncertainties. Acuators 9(128) (2020). https://doi.org/10.3390/
act9040128, open access
Remaining Useful Life Estimation Based
on Wavelet Decomposition: Application
to Bearings in Reusable Liquid
Propellant Rocket Engines

Federica Galli, Vincent Sircoulomb, Ghaleb Hoblos, Philippe Weber,

and Marco Galeotta

Abstract The increasing trend towards the development of reusable rocket engines
pushes towards the development of Prognosis and Health Monitoring (PHM)
approaches useful to monitor the health of the system and plan its maintenance.
In this perspective, we present a hybrid method for Remaining Useful Life (RUL)
estimation of rocket engines turbo-pump bearings based on Maximum Overlap Dis-
crete Wavelet Packet Transform (MODWPT) and polynomial approximation. The
proposed method calculates the bearing RUL in six main steps: data acquisition,
wavelet decomposition, feature extraction, degradation detection, Health Indicator
(HI) computation and RUL estimation. The obtained results showed that this tech-
nique can be successfully used to separate and isolate vibration trends connected
to progressive degradation. The main contribution of this work consists in propos-
ing a monotonically increasing HI which maintains a physical meaning allowing to
estimate the degradation level. The proposed HI has proven to be able to effectively
quantify the level of degradation and predict it for a certain group of degradation
evolution profiles.

F. Galli (B) · V. Sircoulomb · G. Hoblos

IRSEEM, UNIROUEN, ESIGELEC, Normandie University, 76000 Rouen, France
e-mail: [email protected]
V. Sircoulomb
e-mail: [email protected]
G. Hoblos
e-mail: [email protected]
P. Weber
CRAN, 54506 Nancy, France
e-mail: [email protected]
M. Galeotta
CNES, 75012 Paris, France
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 107
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_10
108 F. Galli et al.

1 Introduction

During the last decade, reusable launchers have become a reality, revealing a wide
range of advantages going from cheaper costs to shorter turnaround times. A precise
monitoring of a system’s health state is at the base of a correct reusability. For this
reason, a strong effort is put nowadays in the development of Health Monitoring Sys-
tems (HMS) and PHM approaches. As far as launchers are concerned, many parts can
be reused going from the first stage (with all its associated equipment) to the fairing.
In particular, the first stage engines are the most expensive and complex. Reusing a
liquid propellant rocket engine (LPRE) introduces a series of additional challenges
such as re-ignition, thrust modulation and maintenance performance. Given the high
degree of system complexity, the development of a HMS system for a rocket engine
has been approached from different angles. For example, Kawatsu et al. [1] conducted
a study on a system-level focusing on fault detection and diagnosis using a model-
based quantitative assessment that considers system-level interactions in a rocket
engine system. Banerjee et al. [2] developed an anomaly detection and prognosis
technique based on particle filtering and machine learning. Other authors focused
just on a single engine component. Figure 1 shows the architecture of the Vulcain
2 rocket engine in which the most critical engine components can be recognised:
turbo-pumps, combustion chamber, secondary combustion chambers.
Kanso et al. [5] and Chelouati et al. [6] studied the Remaining Useful Life (RUL)
of a reusable LPRE combustion chamber. Hotte et al. [7] investigated the degradation
of the walls of a reusable combustion chamber with an experimental study. Concern-
ing turbo-pumps, they can be further divided into sub-components as shown in Fig. 2.

Fig. 1 Vulcain 2 rocket

engine scheme [3]
Remaining Useful Life Estimation Based on Wavelet Decomposition: Application … 109

Fig. 2 Example of Rocket

engine turbopump (hydrogen
turbopump) [4]

As explained by El-Thaji et al. [8] and Luo et al. [9], bearings are among the most
critical components of any rotating machinery. Failure of such component could
result in irreversible damage to the turbo-pump. In the worst case, it could cause the
loss of the system. In the domain of rocket engines, such generic problem results
in explosions and loss of the mission leading to huge economical damages. Bearing
RUL prediction has already been the object of many studies. As summarised by Lei
et al. in [10], PHM prediction approaches can be categorised in three major families:
physics model-based approaches, data driven approaches and hybrid approaches. In
the first case, a theoretical model is used to describe the degradation evolution and
approximate its trend with respect to time. Such methods are helpful when small
quantities of data are available. On the other hand they may have a limited precision
for very complex non-linear systems. In the second case, data driven approaches
rely on statistical models and Artificial Intelligence (AI). In both cases, relevant
information about the degradation is inferred from the available data and used to
train the algorithm. These methods are easier to implement when intricate systems
are studied but may prove to be ineffective if not enough data is available for train-
ing. Finally, hybrid methods are a combination of the two previous categories. By
bringing together different methodologies, a more complete hybrid technique can
be proposed. The domain of rocket engines presents two main issues: (i) lack of
data due to small quantities of specimen produced and used, (ii) high-non-linearities
coming from its strong complexity. The objective of the present work is to develop
a data-driven PHM method for RUL prediction of rocket turbo-pump bearings. The
proposed approach treats bearings vibration signals combining wavelets Multi Res-
olution Analysis (MRA) and polynomial approximation. The HI used to track the
degradation evolution in time is the cumulative Root Mean Square (RMS) which
provides an indication of the signal energy amount accumulated in time and thus an
indication of the degradation stage. The proposed approach was applied to a database
available on the NASA repository [11] in view of when some reusable rocket engine
data will be available. It shows good capabilities of degradation detection and trend
110 F. Galli et al.

estimation. The contribution of this paper consists in proposing a cumulative health

indicator with a physical meaning characterised by a strictly monotonic trend. Indeed
such HI is constructed in such a way to maintain a direct link with the component
physical degradation while enhancing and ensuring an increasing trend with respect
to time.
The remainder of the paper is divided into the following sections: Sect. 2 describes
the methodology. Section 3 shows the experimental data used in this study and the
obtained results. Finally, Sect. 4 draws some concluding remarks and future perspec-
tives.

2 Methodology of the Proposed Approach

RUL estimation of Rolling Elements Bearing (REB) is a difficult task. Indeed, during
its operation, a bearing suffers an inevitable degradation process concerning differ-
ent tribological phenomena such as wear and flaking [12]. Such complex damage
phenomena present many non-linearites and a big variety of outcomes depending
on the degradation environment. The aim of the present work is to define a model
describing the turbopump bearings overall degradation and tracking its evolution in
order to predict the RUL. Figure 3 shows the main steps of the proposed method in
a synthetic way. The retrieved vibration data is pre-processed using the Maximum
Overlap Discrete Wavelet Package Transform (MODWPT) which splits the signal
frequency content into different frequency bands. After that, the RMS is computed
and used as statistical indicator for degradation detection. Once the initial degrada-
tion instant is defined, a second indicator is computed and fitted with a polynomial
model using the least squares method. Finally the obtained model is used to pre-
dict the bearing degradation and to compute the RUL. A detailed description of the
different steps is given hereafter:

Fig. 3 Proposed algorithm steps

Remaining Useful Life Estimation Based on Wavelet Decomposition: Application … 111

Fig. 4 Data collection process

1. Data collection. A vibration signal recorded with accelerometers is retrieved

each time the component completed its operating cycle. During each cycle,
N samples are acquired. The samples are denoted n and the sampling time
ps (Fig. 4). Thus t = n · ps . In this study the duration of one cycle, N · ps , is
assumed to be constant and the transient phases at the beginning and at the end
of each cycle are neglected.
2. Wavelets Decomposition. Each vibration signal is decomposed using the MOD-
WPT [13]. The Discrete Wavelet Package Transform (DWPT) is a signal pro-
cessing technique which allows to capture the signal behaviour at different fre-
quency bands by projecting it on a basis of zero-mean oscillating function called
wavelets. The MODWPT is a more complete technique which allows to process
the signal without down-sampling and time delay making it more suitable for
time series [14] . The mother wavelet db5 was chosen for its orthogonality which
results in the conservation of the signal energy through the decomposition pro-
cess. The decomposition frequency bands can be computed using Eq. 1, where
i is the index of the frequency band, f b, and L is the decomposition level.
 
Fs Fs
f bi = (i − 1) · ; (i) · with i = 1, . . . , 2 L (1)
2 L+1 2 L+1

For example, in our case, these parameters are defined such that : L = 3, Fs =
25.6 kHz, i = 1, . . . , 8.
3. Feature extraction. For each frequency band of the wavelet decomposition, a
statistical indicator is computed. Here, the RMS is computed following Eq. 2.
This indicator can alternatively be the Kurtosis, the energy or the relative energy,
see Eqs. 3 to 5 respectively where xs ( j, i) is a sub-signal of frequency band i,
1 < j < n is the vibration sample index, σ is the standard variation and x̄s is the
mean value of xs ( j, i).
112 F. Galli et al.

1 
n 1/2
R M S(n, i) = xs ( j, i)2 (2) 
n

n En tot (n, i) = |xs ( j, i)|2 (4)

j=1
j=1
n
j=1 xs ( j, i) − x̄s ( j, i) En f bi
K T S(n, i) = E Nr el (n, i) = (5)
nσ 4 En tot
(3)

The RMS and Kurtosis are general statistical indicators and have been widely
used for bearing vibration analysis [10, 15, 16]. The first one provides an indica-
tion of the mean energy of the signal while the second one quantifies the signal
impulsivity. The energy and relative energy, [17, 18], are also of particular inter-
est when MODWPT is applied. Moreover, as pointed out by Luo et al. [9], the
bearing degradation has a strong influence on the signal frequency energy dis-
tribution. Thus, its evolution with respect to time can be used to quantify the
degradation in an indirect way.
4. FPT detection. The first step towards RUL computation is to detect the presence
of a degradation. The indicators computed in step 3 for all frequency bands are
monitored simultaneously using paired contiguous sliding windows that check
for mean variation in each sub-signal with respect to a detection threshold.
5. HI computation. Once the First Predicting Time (FPT), denoted t f pt , is obtained
the analysis is carried on only in the frequency band of interest. A second health
indicator is calculated from the first one. The proposed HI consists in the discrete
cumulative integral of the RMS:
 n
H I (n, i) = R M S(l, i) (6)
l=1

The motivation of this HI is as follows: when the bearing balls pass over a defect
the signal amplitude increases due to the impact between the ball surface and
the ring damaged surface. This magnitude variation results in energy variation.
The classical RMS gives an indication of the instantaneous degradation rate. As
a consequence, the integral of such quantity between the FPT and the current
time quantifies the total bearing surface damage. The proposed approach was
first introduced by Boness et al. in [19] for wear volume estimation of a ball-on-
cylinder test apparatus.
6. RUL estimation. The HI profile is fitted with the polynomial Eq. 7 of order 2:
Hˆ I (t) = a(t − t f pt )2 + b(t − t f pt ) + c with t > t f pt (7)
A polynomial of order 2 was chosen as compromise between approximation
accuracy and monotonicity. Indeed, a polynomial of order 1 is too simple and a
polynomial of order 3, or higher, presents at least a maximum and a minimum.
Writing down tk = k N ps the time when the kth operating cycle is completed and
where the samples records [(k − 1)N + 1, k N ] are available, the coefficients
a, b and c are fitted from H I (n, i) using all the samples n from times t f pt to tk
and the least squares method. Hˆ I (t) is then used to predict the HI for t > tk .
Remaining Useful Life Estimation Based on Wavelet Decomposition: Application … 113

This polynomial model is used for HI forecast. Finally, the RUL is calculated
as the time interval between the current time and the future instant when the HI
exceeds a predefined threshold, T h:
RU L = t such that Hˆ I (tk + t) = T h (8)
Defining the HI failure threshold is a sensitive step. Indeed finding a general limit
is quite tough given the wide range of possible degradation profiles. In some
cases, as in Jia et al. [20], the threshold is defined according to external factors
like operational risk. Other authors set the failure limit using some training data
to define a threshold for each degradation trend. Later the failure limit is chosen
according to similarity between the tested signal and the training signals. This
methodology, used in [9], is considered in the present paper.

3 Results and Discussion

The proposed algorithm is tested with run to failure vibration signals retrieved from
the NASA prognostic repository, [11]. The data were obtained running experiments
on the PRONOSTIA platform at the FEMTO-ST Institute, [21]. The platform, pro-
vided with two high frequencies accelerometers one the x and y axis, was used to test
individually seventeen bearings of type NSK 6804RS. The experiments were con-
ducted in three different operating conditions whose details can be found in Table 1.
Each vibration signal was recorded with a sampling frequency of 25.6 kHz. The
samples are recorded every 10 seconds and have a duration of 0.1 s. As an exam-
ple, Fig. 5 shows the raw vibration signal for Bearing 2-1. The presented signal is
strongly non-stationary and is characterised by amplitude variations at low and high
frequencies.

Table 1 Experimental setting conditions

Condition 1 Condition 2 Condition 3
Speed (RPM) 1800 1650 1500
Load (N) 4000 4200 5000
Bearings Bearing 1_1 to 1_7 Bearing 2_1 to 2_7 Bearing 3_1 to 3_3

Fig. 5 Raw vibration signal

114 F. Galli et al.

Fig. 6 FPT Detection on the 8 sub-signals

Fig. 7 RMS evolution of the decomposed signal

When the data collection phase is complete the vibration signal is treated with the
MODWPT, step 2 in Fig. 3. In this case, the signal was split into 8 frequency sub-
bands. The sub-signal of the frequency band [11.2–12.5 kHz] is reported in Fig. 6
as an example. As it can be seen, the initial amplitude variation are not present and
the signal increases in amplitude starting from about t=0.3 hours. Such sub-signal
reveals a clearer degradation trend. All the obtained sub-signals are used for feature
extraction. The time evolution of the RMS in each sub-band is shown in Fig. 7. The
blue curves represent the RMS values with respect to time. To ease the detection
task, the RMS curves were smoothed.
The green markers show the position of the true FPT. Indeed some sub-signals
present some initial instabilities before entering a steady state. Such phenomenon is
explained by the fact that at the beginning of their life the bearing may present some
fabrication defects. Such defects are reduced thanks to friction during the first period
of functioning and must not be considered as actual bearing degradation. Thanks to
the wavelet MRA, it is possible to separate the different vibration phenomena and to
focus only on the ones provoked by damages on the bearing. The FPT detection, step
Remaining Useful Life Estimation Based on Wavelet Decomposition: Application … 115

Fig. 8 RMS and cumul. 0.7

RMS Bearing 2-1 350

0.6
300

0.5
250

200 0.4

150 0.3

100 0.2
HI integral(RMS)
FTP
50 RMS 0.1
FTP

0 0.5 1 1.5 2 2.5

t [h]

4 in Fig. 3, is a crucial step in the RUL estimation. Not only it needs to be precise
but also it has to detect the presence of a degradation as soon as possible. Given
that the proposed algorithm is capable of estimating the true FPT in each frequency
band, the first FPT to be computed defines the frequency band that is to be used for
RUL estimation. In Fig. 6, it is possible to see that the frequency band number 8,
[11.2–12.5 kHz], is the first one to show a sign of degradation. Figure 8 shows the
RMS and cumulative RMS for frequency band [11.2–12.5kHz]. With a close look
on the graph, it is possible to see that they show quite different profiles.
The RMS used for FPT detection shows a general monotonic increasing trend
affected by fluctuation especially towards the end. This kind of curve is not ideal from
the point of view of HI prediction since it could be interpreted as a reduction of the
degradation level. Oppositely, the degradation process is believed to be irreversible.
The cumulative RMS shows a strictly monotonic increasing trend and for this reason,
it is more suitable for RUL estimation. At last, Fig. 9 shows the results of step 6: the
estimated RUL in terms of life cycles. A life cycle of 5 minutes was considered. The
obtained results look promising: the initial estimation falls within a 30% uncertainty
interval and crosses it only twice during the entire calculation. Figures 10, 11, 12 and
13 show the RUL estimation for four other bearings: 2-2, 2-4, 3-1, 3-3 respectively.
The obtained results show good estimation of the RUL and a fast convergence
towards the real RUL. Bearing 3-1 and 3-3 show a good result as well. In this case,
a different failure threshold was needed to be defined since the degradation profile
was too different from the previous cases. The threshold value was estimated using
signal 3-1 as learning set and was later applied to signal 3-3. Lastly, the dashed line
in Figs. 9, 10, 11, 12 and 13 represents the cycle duration and shows whether the
component is reusable or not. Indeed, if the estimated RUL is less than the cycle
duration, the bearing must be changed or at least repaired. For example, Bearing
2-2 in Fig. 10, could be reused for 9 cycles, Bearing 2-4 in Fig. 11, for 14 cycles and
Bearing 3-3, in Fig. 13, for 2 cycles. This result is an example of how RUL estimation
116 F. Galli et al.

Fig. 9 RUL Estimation 0.9

Bearing 2-1 estimated RUL
0.8 real RUL
RUL+30%
RUL-30%
0.7
cicle duration

0.6

RUL [h]
0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 7 8 9
k

Fig. 10 RUL Estimation 1

Bearing 2-2 estimated RUL
0.9 real RUL
RUL+30%
0.8 RUL-30%
cicle duration
0.7

0.6
RUL [h]

0.5

0.4

0.3

0.2

0.1

0
1 2 3 4 5 6 7 8 9 10
k

can be used as aid to the decision process concerning maintenance planning. The
performance of the algorithm was estimated as in [15] by calculating the percentage
of RUL estimated values that fall in the confidential interval at +/-20% during the
last 500 s of component life. In this case, [15], the bearing RUL was computed using
the RMS of the vibration signal as HI and dynamic regression models (DRM) for
HI prediction. The obtained results were rearranged in order to be comparable to the
ones in literature. Table 2 shows the values of the chose metrics for Bearing 2-4 and
3-3. As it can be seen, the proposed approach has a better prognostic performance.
Remaining Useful Life Estimation Based on Wavelet Decomposition: Application … 117

Fig. 11 RUL Estimation

1.4 estimated RUL
Bearing 2-4 real RUL
RUL+30%
1.2 RUL-30%
cicle duration

RUL [h]
0.8

0.6

0.4

0.2

2 4 6 8 10 12 14
k

Fig. 12 RUL Estimation

1.1 estimated RUL
Bearing 3-1 real RUL
1 RUL+30%
RUL-30%
0.9 cicle duration

0.8

0.7
RUL [h]

0.6

0.5

0.4

0.3

0.2

0.1

1 2 3 4 5 6 7 8 9 10 11 12
k

Fig. 13 RUL Estimation

0.3 estimated RUL
Bearing 3-3 real RUL
RUL+30%
RUL-30%
0.25 cicle duration
RUL [h]

0.2

0.15

0.1

1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3

k
118 F. Galli et al.

Table 2 Performance metrics of the proposed approach and the approach in [15]
Proposed DRM method Proposed DRM method
method method
Bearing 2_4 87.5% 38% Bearing 3_3 87.5% 80%

4 Conclusion
In this paper, an approach for RUL estimation of rolling element bearings was pro-
posed and tested with a set of vibration degradation data. The contribution of this
methodology mainly relies on two aspects. The first one concerns the introduction
of wavelets MRA analysis for degradation detection and feature extraction. The sec-
ond one is the use of the cumulative RMS as health indicator. The proposed HI has
proven to be able to effectively quantify the level of degradation and predict it for a
certain group of degradation evolution profiles. The obtained results showed that this
technique can be successfully used to separate and isolate vibration trend connected
to progressive degradation. Further investigation is foreseen to improve the proposed
algorithm. In particular, future work will focus on defining a more adaptive threshold
tuning technique, real time approximation modelling and uncertainty analysis.

Acknowledgements This work is co-funded by CNES “Centre Nationale d’Etudes Spatiales” and
the Normandy Region.

References

1. Kawatsu, K.: PHM by using multi-physics system-level modeling and simulation for EMAs
of liquid rocket engine. In: IEEE Aerospace Conference, pp. 1–10. Big Sky, USA (2019)
2. Banerjee, B., Kraemer, L., Solano, W.: Particle filtering for diagnosis and prognosis of anoma-
lies in rocket engine tests. In: AIAA Infotech at Aerospace Conf. and Exhibit, St. Louis, MO,
USA (2011)
3. Vulcain, A.H.: ®2. 1. Technical report. https://www.ariane.group/wp-content/uploads/2020/
06/VULCAIN2.1_2020_04_PS_EN_Web.pdf
4. Hiraki, H., Inoue, T., Yabui, S.: Influence of impeller’s elastic deformation on the stability of
balance piston mechanism of rocket engine turbopump. In: IOP Conference Series: Earth and
Environmental Science, vol. 240, p. 052028. IOP Publishing (2019)
5. Kanso, S., Jha, M.S., Galeotta, M., Theilliol, D.: Remaining useful life prediction with uncer-
tainty quantification of liquid propulsion rocket engine combustion chamber. In: SAFEPRO-
CESS, vol. 55, pp. 96–101. Greece, Cyprus (2022)
6. Chelouati, M., Jha, M.S., Galeotta, M., Theilliol, D.: Remaining useful life prediction for liquid
propulsion rocket engine combustion chamber. In: Conference on Control and Fault-Tolerant
Systems, SysTol, pp. 225–230. Saint-Raphael, France (2021)
7. Hötte, F., Sethe, C.V., Fiedler, T., Haupt, M.C., Haidn, O.J., Rohdenburg, M.: Experimental
lifetime study of regeneratively cooled rocket chamber walls. Int J Fatigue (2020)
8. El-Thalji, I., Jantunen, E.: A summary of fault modelling and predictive health monitoring of
rolling element bearings. Mech. Syst. Signal Proc. (2015)
Remaining Useful Life Estimation Based on Wavelet Decomposition: Application … 119

9. Luo, H., Bo, L., Liu, X., Zhang, H.: A novel method for remaining useful life prediction of
roller bearings involving the discrepancy and similarity of degradation trajectories. Comput.
Intell. Neurosci. (2021)
10. Lei, Y., Li, N., Guo, L., Li, N., Yan, T., Lin, J.: Machinery health prognostics: a systematic
review from data acquisition to RUL prediction. Mech. Syst. Signal Proc. (2018)
11. Orzech, G.: Prognostics center of excellence data set repository (2022). http://www.nasa.gov/
content/prognostics-center-of-excellence-data-set-repository
12. El-Thalji, I., Jantunen, E.: A descriptive model of wear evolution in rolling bearings. Eng. Fail.
Anal. (2014)
13. Percival, D.B., Walden, A.T.: Wavelets Methods for Time Series Analysis. Cambridge Univer-
sity Press, Cambridge (2000)
14. Yang, Y., He, Y., Cheng, J., Yu, D.: A gear fault diagnosis using hilbert spectrum based on
modwpt and a comparison with emd approach. Meas.: J. Int. Meas. Confed. (2009)
15. Ahmad, W., Khan, S.A., Islam, M.M.M., Kim, J.: A reliable technique for remaining useful
life estimation of rolling element bearings using dynamic regression models. Reliab. Eng. Syst.
Safety (2019)
16. Wang, D., Tsui, K.L., Miao, Q.: Prognostics and health management: a review of vibration
based bearing and gear health indicators. IEEE Access (2017)
17. Tobon-Mejia, D., Medjaher, K., Zerhouni, N., Tripot, G., Tobon-Mejia, D.A., Medjaher, K.,
Zerhouni, N., Tripot, G.: Estimation of the remaining useful life by using wavelet packet
decomposition and HMMs. In: Aerospace Conference. Big Sky, USA (2011)
18. Ocak, H., Loparo, K.A., Discenzo, F.M.: Online tracking of bearing wear using wavelet packet
decomposition and probabilistic modeling: a method for bearing prognostics. J. Sound Vibr.
(2007)
19. Boness, R.J., McBride, S.L.: Adhesive and abrasive wear studies using acoustic emission
techniques, Wear (1991)
20. Jia, X., Li, W., Wang, W., Li, X., Lee, J.: Development of multivariate failure threshold with
quantifiable operation risks in machine prognostics. In: Annual Conference of the PHM Society,
vol. 12, pp. 9–9. PHM Society (2020)
21. Nectoux, P., Gouriveau, R., Medjaher, K., Ramasso, E., Morello, B., Zerhouni, N., Varnier, C.:
PRONOSTIA: an experimental platform for bearings accelerated degradation tests. In: IEEE
International Conference on PHM, PHM’12. Denver, CO, United States (2012)
Advanced Control
Adaptive Finite Horizon
Degradation-Aware Regulator

Amirhossein Hosseinzadeh Dadash and Niclas Björsell

Abstract Predicting the failure and estimating the machine’s state of health is infor-
mation that supports the production planning and maintenance management systems
to increase productivity and reduce maintenance and downtime costs. However, con-
trolling the degradation in the machines will improve the system’s reliability and
resilience and make high-level decisions more accurate and reliable. To control the
degradation in the machines, time should be included in the cost function as a vari-
able, which alters the markovian properties of the system dynamic. In this article,
we include the degradation cost in the quadratic cost function of the infinite hori-
zon controller and calculate the optimal feedback according to the dynamics of the
degradation using dynamic programming. It will be shown that the infinite horizon
control will convert to the finite horizon, and the controller will be able to control
the degradation according to the desired degradation at the desired time. In the end,
with the help of simulation, we show that the degradation controller can control the
degradation in the MIMO systems.

1 Introduction

In the era of Industry 4.0, the horizons of the definition of “optimal control” can
include more variables than its classical definition. Keeping the system output close to
the desired output, which was the optimal control’s primary goal, is now among many
other goals that must be achieved simultaneously. For instance, the Linear Quadratic
Regulator (LQR) is considered an optimal controller when the only parameters to

A. Hosseinzadeh Dadash (B) · N. Björsell

Department of Electronics, Mathematics and Natural Sciences, University of Gävle,
801 76 Gävle, Sweden
e-mail: [email protected]
N. Björsell
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 123
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_11
124 A. Hosseinzadeh Dadash and N. Björsell

control are the input and output, but being optimal in controlling the degradation
simultaneously needs some new information from the system and modifications to
the method [1, 2].
Fault-tolerant controllers have been the subject of research [3]. However, degra-
dation tolerant controllers, which are more helpful in long-term production planning
and maintenance management systems [4, 5], have not been discussed much. For
instance, one of the methods for reducing the maintenance cost is to synchronize
the machines’ degradation so they can reach the maintenance time simultaneously.
This simultaneity is achieved in [6] using advanced high-level controllers and has
the limitation that it is only applicable to systems with similar machines working on
the same process.
To reduce the calculation and data storage cost and make the degradation control
applicable to machines that are not identical and do not work on the same process,
the degradation controller should be integrated with the controller or work on its side
without a need to exchange much information with high-level controller (production
management) [7, 8]. For doing this integration, the degradation model should be
identified. This identification should be in a way that the model can be related to
the system dynamics. This identification can be made using process-aware neural
networks [9, 10] or sparse regressions [11, 12]. In both cases, the identification of the
degradation might not be as accurate as model identification without the limitation
of considering the process parameters, but it can be used for control purposes which
have their own benefits.
Assuming that the machine and degradation models are known, it is possible to
control the degradation in two ways: including the degradation as the system’s state
or including it in the feedback loop. The traditional way of including degradation
inside the model has been tested successfully [13, 14], but having a closed form
of including the degradation in the feedback loop will be helpful in the future and
can change the controller design process. Also, an adaptive feedback loop can be
configured faster with the degradation model change than the adaptive controller.
This paper will introduce the adaptive feedback loop for controlling the degrada-
tion at the same time as the output. In the first section, the method for calculating
different parts of this feedback loop will be shown, and in the second section, the
method will be validated using a simulation model.

2 Method

The method for calculating the adaptive degradation-aware feedback comprises two
parts, the formulation of the feedback, which is explained first, and mapping from
the recorded degradation into the system’s states.
Adaptive Finite Horizon Degradation-Aware Regulator 125

2.1 Formulation

The state-space model of the system can be written as follows:

  
x(t + 1) = Ax(t) + Bu(t) + Mv1 (t) v1
= , (1)
y(t) = C x(t) + v2 (t) v2

where x(t) includes system state(s) at time t, u(t) is system input(s) at time t. A and
B represent the physical system parameters, and C defines the relationship between
the output(s) and state(s) of the system. v1 and v2 are the input and sensor noise,
respectively.
The quadratic cost that the LQ controller minimizes for the infinite horizon is [15]

J =  e 2Q 1 +  u 2Q 2 , (2)

where e = x − r and r is the reference signal, and e is the error; and Q 1 and Q 2
are penalty matrices for the error and input signal, respectively. However, adding the
third term for controlling the degradation converts (2) to

N N −1
 
N 
J (t = 1 : T ) = E  ek 2Q 1 +  u k 2Q 2 +  G x f ,u f (t) 2Q 3 , (3)
k=0 k=0 k=0

G x f ,u f (t) = Dx f ,u f − Dd (t), (4)

where x f , u f are feature of x and feature of u respectively, Dx f ,u f is the degradation

at state x f with the input u f , and Dd (t) is the desired degradation at time t. To better
understand x f and u f , some explanation is needed. First, the working cycle should
be defined as the interval where the machine starts from its standby state and settles
on the desired output (usually the step response of the machine till it fully settles).
For example, a working cycle of a wood-cutting saw starts when it starts rotating
from the stop position until it reaches the desired speed (full load working). Second,
in reality, the degradation rate is lower than the rate of change in the system state (the
degradation happens over tens or hundreds, or even thousands of cycles of operation
while the controller works on the scale of a fraction of a second). So, in order to
be able to detect the degradation from the recorded signal, a fixed point (in time)
or a feature from one machine’s working cycle (for each state and input) should be
recorded for different cycles. This way, the degradation trend can be detected and
mapped to the system state. These recorded features are defined to be x f and u f .
Figure 1 shows this process more clearly. Finally, as the features are the variables
that are going to be used in the last part of (3), t will be defined as a cycle; this means
that t is not an instance and instead is a time duration. However, as only one feature
from each recorded signal during each cycle is recorded in (3), it can be considered
a time instance.
126 A. Hosseinzadeh Dadash and N. Björsell

Fig. 1 Relation of the features per cycle and recorded signal (red circles are the features e.g., x f
or u f that are recorded at each cycle, and their trend is used to map the degradation to the system
states)

After these adaptations of the cost function with the limitation of the function
estimation methods, the whole optimization becomes a finite horizon optimization
that tries to control the degradation of the machine. Hence, it reaches the maintenance
condition (x N ) at the desired time (T ). Note that the degradation of the machine is not
time-dependent (Markovian). However, the desired degradation is only the function
of the time (because the optimization goal is to make the machine reach the desired
maintenance time while keeping the output within acceptable limits).
Applying the dynamic programming for this optimization, we have

JN (T ) = e N (T ) Q 1 e N (T ) (5)
 
  
Jk=1:N −1 (t = 1 : T ) = min E e Q 1 e + u Q 2 u + G Q 3 G + Jk+1 (Axk + Bu k + v1k ) ,
uk
(6)

where Jk is the cost-to-go from state xk at time t to x N at time T .

Now for calculating the best trajectory, the cost function should be calculated
backward from x N at time T to x1 at time 0 [16]. For now, we assume that the
degradation of the machine can be identified as a function of the machine’s states
and inputs. In this case, we have
 
 xk
Dx,u = Wx Wu . (7)
uk
Adaptive Finite Horizon Degradation-Aware Regulator 127

Now expanding (6) gives

Jk (t) = e Q 1 e+
 
  
min E u Q 2 u + G Q 3 G + (Axk + Bu k + wk ) Q 1 (Axk + Bu k + wk ) .
uk
(8)

Expanding this statement and differentiating it in order to find the u ∗k (optimal input)
and assuming that E{v1 } = 0 will result in

u x (t)∗ = −(Q 2 + B  Q 1 B)−1 (B  Q 1 Axk +(Wx xk (t − 1)+

Wu u k (t − 1) − Dd (t − 1))Wu Q 3 ). (9)

This equation is composed of two parts, the first part

u ∗x = −(Q 2 + B  Q 1 B)−1 (B  Q 1 Axk ), (10)

which is the infinite horizon optimal feedback, and only depends on xk , and second
part

u(t)∗ = −(Q 2 + B  Q 1 B)−1 (Wx xk (t − 1) + Wu u k (t − 1) − Dd (t))Wu Q 3 , (11)

which depends on t, x and u and is the adaptive degradation compensation feedback.

2.2 Relevance Vector Machine

In (7), it was assumed that the degradation of the system could be calculated as a
function of the system’s state and input. There are many methods to do this mapping.
One of these methods used for this research is the Relevance Vector Machine (RVM).
The relevance vector machine (RVM) was first introduced in [17]. The RVM structure
is very similar to the support vector machine, which is given as follows:


L
ŷ(x) = w k(x, x ) + c, (12)
=1

where w is a coefficient in w, which is the vector of coefficients, k(·, ·) is the kernel

function, and c is the bias parameter. RVM defines the conditional distribution for
the target value y given the vector of covariates x, prediction outcome ŷ, regression
coefficients w, and a precision parameter called ψ as
128 A. Hosseinzadeh Dadash and N. Björsell

p(y |x , w, ψ) = N(y | ŷ(x ), ψ −1 ), (13)

ŷ = (X)w, (14)

(X) = [φ(x1 ), φ(x2 ), . . . , φ(x L )]T , (15)

φ(x ) = [1, k(x , x1 ), k(x , x2 ), . . . , k(x , x L )]. (16)

The likelihood function for y can be written as

L
p(y|X, w, ψ) = p(t |x , w, ψ −1 ). (17)
=1

RVM introduces a prior distribution for each w in w as a hyperparameter α

M
p(w|α) = N(wi | 0, αi−1 ), (18)
i=1

where M is the number of covariates (bias included). The hyperparameter α measures

the precision of each wi . Following Bayesian inference, the distribution of the weights
becomes Gaussian and takes the following form:

p(w|y, X, α, ψ) = L(w| m, ), (19)

in which
m = ψT y, (20)

=( α + ψT )−1 , (21)

and α = diag(αi ).

2.2.1 Results

For this proof of idea, the model used is the simple mass and spring model with
two inputs. The reason for considering two inputs is that, for SISO systems, with a
correct configuration of Qs, the optimal input of the degradation controller would
be the same as the optimal input calculated by solving the Ricatti equation for LQR
because the only parameter available to control both output and degradation would
be a single input which prioritizing the output quality will make the system a regular
controller, not a degradation controller. However, having more than one input will
make the controller capable of controlling two outputs (the system’s output and
system degradation).
Adaptive Finite Horizon Degradation-Aware Regulator 129

Fig. 2 Mass and Spring Model

The model used for this simulation is shown in Fig. 2, and its respective state-space
is shown in (22) and (23) [18].
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
z˙1 0 1 0 z1 0 0  
⎣z¨1 ⎦ = ⎣ k2−k1 0 − k1 ⎦ . ⎣z˙1 ⎦ + ⎣ 1 0 ⎦ . f 1 , (22)
m m m f2
z˙2 k1
b
0 − kb1 z2 0 b1

y = I x, (23)

where z 1 and z 2 are displacements shown in Fig. 2, m is the mass, k1 and k2 are
respective spring constants, b is the damping ratio and f 1 and f 2 are respective input
forces. The parameter to control or the desired output for this simulation is z 1 , which is
the position of the mass. Although, according to different conditions, the degradation
might differ in real situations, to reduce the complexity of the result and be able to
focus on the research idea, the degradation parameter chosen to be k1 . The degradation
model for k1 considered to be exponential [19]:

k1 (t + 1) = k1 (t) + 2 × 10−5 × exp(t ∗ 5 ∗ 10−5 ). (24)

Two failure thresholds were defined for this simulation:

|zˆ1 (t) − z 1 (t)| > 0.01, (25)

k1 > 0.02, (26)

where zˆ1 (t) is defined as the desired output at time t, and z 1 (t) is the system output
at time t, and the system is considered as failed (from the maintenance point of view)
when the difference between actual output and desired output passes this threshold.
The closed-loop step response of the system is shown in Fig. 3. It can be seen that the
controller is controlling the system according to the desired output. Figure 4 shows
the same controller under degradation explained in (24) and failure threshold defined
in (25) and (26). The left part of the figure shows the normal controller and on the
right is the response of the degradation controller. Note that the x-axis of the Fig. 4
130 A. Hosseinzadeh Dadash and N. Björsell

Fig. 3 Closed-loop step response of the infinite horizon controller

is in cycles, which means that only features of the output (in this case, maximum)
are recorded from each cycle and plotted in the figure, so each point in the figure is
a feature recorded from a whole cycle as shown in Fig. 1.
The degradation coefficients Wx and Wu are calculated according to recorded data
shown in the left part of Fig. 4.
It can be seen that the regular controller loses control, the system output deviates
from the desired output by an amount more than the failure threshold, and the system
is considered as failed after around 90 cycles. However, the degradation controller
keeps the output inside the acceptable threshold for around 500 cycles. In the end,
the simulation finishes not because of the output deviation but because of the failure
threshold mentioned in (26).

3 Discussion

Controlling the degradation at the same time as the output will impact the industry and
reduce the maintenance cost. Unlike most existing control methods, the degradation
control will depend upon machine learning methods because identifying the physi-
cal laws governing the degradation is costly and environment-dependent. Although
machine learning methods are considered the best solutions for solving modern con-
Adaptive Finite Horizon Degradation-Aware Regulator 131

Fig. 4 Controller without degradation awareness (left) versus degradation controller (right)

trol problems, adapting them to existing methods will be challenging. This article
proved that with the correct choice of machine learning method, it is possible to adapt
the traditional control method with machine learning. The stability of the closed loop
is among the most critical questions that answering it is outside the scope of this arti-
cle. However, the short answer is that another cost function can be used to find the
optimal point if the feedback makes the system unstable.

4 Conclusion

In this article, the degradation-aware adaptive feedback loop was introduced. First,
the formulation of the quadratic cost function of the infinite horizon controller was
updated, and the third term for penalizing the controller was introduced. Then using
dynamic programming, the optimal feedback was calculated, which is a combina-
tion of the infinite horizon optimal feedback and time, state, and output dependent
function. Then the limitation of identifying the degradation was considered in the
formulation and adapted to the limitations, and RVM was used as the machine learn-
ing method compatible with the formulation and limitations. Finally, with the help
of simulation, the functionality of the degradation-aware feedback loop was proved.
132 A. Hosseinzadeh Dadash and N. Björsell

Acknowledgements The research project is financed by the European Commission within the
European Regional Development Fund, Swedish Agency for Economic and Regional Growth,
Region Gävleborg, and the University of Gävle.

References

1. Hua, C., Li, L., Ding, S.X.: Reinforcement learning-aided performance-driven fault-tolerant
control of feedback control systems. IEEE Trans. Autom. Control 67(6), 3013–3020 (2022).
https://doi.org/10.1109/TAC.2021.3088397
2. Chemali, E., et al.: Minimizing battery wear in a hybrid energy storage system using a linear
quadratic regulator. In: IECON 2015 - 41st Annual Conference of the IEEE Industrial Elec-
tronics Society, pp. 003265–003270 (2015). https://doi.org/10.1109/IECON.2015.7392603
3. Abbaspour, A., et al.: A survey on active fault-tolerant control systems. Electronics 9(9), 1513
(2020)
4. Hu, J., Sun, Q., Ye, Z.-S.: Condition-based maintenance planning for systems subject to depen-
dent soft and hard failures. IEEE Trans. Reliab. 70, 1468–1480 (2020)
5. Omshi, E.M., Grall, A., Shemehsavar, S.: A dynamic auto-adaptive predictive maintenance
policy for degradation with unknown parameters. Eur. J. Oper. Res. 282, 81–92 (2020)
6. Björsell, N., Dadash, A.H.: Finite horizon degradation control of complex interconnected sys-
tems. IFAC-PapersOnLine 54, 319–324 (2021)
7. Salazar, J.C., et al.: Health-aware and fault-tolerant control of an octorotor UAV system based
on actuator reliability. Int. J. Appl. Math. Comput. Sci. 30(1), 47–59 (2020)
8. Ahmed, Ibrahim, Quiñones-Grueiro, Marcos, Biswas, Gautam: Fault-tolerant control of
degrading systems with on-policy reinforcement learning. IFAC-PapersOnLine 53(2), 13733–
13738 (2020)
9. Read, J.S., Jia, X., Willard, J., Appling, A.P., Zwart, J.A., Oliver, S.K., Karpatne, A., Hansen,
G.J., Hanson, P.C., Watkins, W., et al.: Processguided deep learning predictions of lake water
temperature. Water Resour. Res. 55, 9173–9190 (2019)
10. Daw, A., Karpatne, A., Watkins, W., Read, J., Kumar, V.: Physics-guided neural networks
(pgnn): an application in lake temperature modeling. arXiv:1710.11431
11. Rudy, S.H., Brunton, S.L., Proctor, J.L., Kutz, J.N.: Data-driven discovery of partial differential
equations. Sci. Adv. 3 (2017)
12. Brunton, S.L., Proctor, J.L., Kutz, J.N.: Discovering governing equations from data by sparse
identification of nonlinear dynamical systems. Proc. Natl. Acad. Sci. 113, 3932–3937 (2016)
13. Sun, J., Zuo, H., Wang, W., Pecht, M.G.: Application of a state space modeling technique
to system prognostics based on a health index for condition-based maintenance. Mech. Syst.
Signal Proc. 28, 585–596 (2012)
14. Storvik, G.: Particle filters for state-space models with the presence of unknown static param-
eters. IEEE Trans. Signal Proc. 50, 281–289 (2002)
15. Glad, T., Ljung, L.: Control Theory. CRC Press, Boca Raton (2018)
16. Bertsekas, D.: Dynamic Programming and Optimal Control: Volume I, vol. 1. Athena Scientific,
Nashua (2012)
17. Tipping, M.E.: Sparse Bayesian learning and the relevance vector machine. J Mach Learn Res
1, 211–244 (2001)
18. Close, C.M., Frederick, D.K., Newell, J.C.: Modeling and Analysis of Dynamic Systems.
Wiley, New York (2001)
19. Gebraeel, N.: Sensory-updated residual life distributions for components with exponential
degradation patterns. IEEE Trans. Autom. Sci. Eng. 3(4), 382–393 (2006)
Set-Membership Fault Detection
Approach for a Class of Nonlinear
Networked Control Systems with
Communication Delays

Afef Najjar and Jean-Christophe Ponsart

Abstract In this paper, a Fault Detection (FD) problem for a class of Nonlinear Net-
worked Control Systems (NNCS) in a set-membership framework is investigated.
Under the assumption of bounded network-induced delays and process uncertainties
(i.e. process disturbances and measurement noises), a residual generator is con-
structed based on a set-membership estimation-based predictor approach. Finally, a
numerical example illustrating the performances of the proposed method is given.

Keywords Fault detection · Nonlinear Networked Control System (NNCS) ·

Unknown network delay · Interval observer · Predictor

1 Introduction

The NCS are systems wherein some or all signals are transmitted among the sys-
tem’s components as information flows through a shared network [7]. Compared
with conventional point-to-point architectures, the advantages of NCS are lighter
wiring, lower installation costs and greater abilities in diagnostic, reconfigurability
and maintenance [7]. Thanks to these distinctive benefits, application of NCS ranges
over various industry’s fields nowadays [7, 13]. However, using a shared network
for data exchange make system control [10], monitoring [9] or diagnostic [11] more
difficult where some communication constraints should be considered such as pack-
ets losses, sampling problems and network-induced delays [7]. The last mentioned is
one of the most common problem in literature [6], especially in NCS FD. Intensive
research addresses this challenging subject, one can see for example [6, 11] and the
references therein. Filtering method, Markovian jump approach and observer-based
approach are ones of the most used approaches dealing with this problem. In this

A. Najjar (B) · J.-C. Ponsart

CNRS, CRAN, Université de Lorraine, F-54000 Nancy, France
e-mail: [email protected]
J.-C. Ponsart
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 133
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_12
134 A. Najjar and J.-C. Ponsart

paper, a FD technique is proposed in a set-membership framework for the NNCS

with unknown communication delays and unknown external disturbances [5] by only
knowing their bounds. The main idea consists in detect faults by generating interval
residual signals. Then, FD is ensured through a belonging test of the zero signal to
the interval delimited by upper and lower residual signals.
This paper is structured as follows. Section 2 introduces some preliminaries.
Section 3 represents the studied system architecture. In Sect. 4, we develop the pro-
posed fault detection technique. Then, Sect. 5 is devoted to simulation results proving
the proposed FD strategy. Conclusions are given in Sect. 6.

2 Preliminaries

R and N represent the sets of real and natural numbers, respectively. The eigenvalues
set of a matrix A ∈ Rn×n is named λ(A) and Re(z) is the real part of the complex
number z. The set of Hurwitz matrices from the set Rn×n is denoted by H, i.e.
R ∈ H ⇔ Re(λ) < 0,∀λ∈ λ(R). We denote M as the set of Metzler matrices from
n
the set Rn×n , i.e. R = ri j i, j=1 ∈ M ⇔ ri, j ≥ 0 for i = j. For a variable x(t) ∈ Rn ,
the upper and lower bounds are denoted by x(t) ∈ Rn and x(t) ∈ Rn , respectively,
such that x(t) ≤ x(t) ≤ x(t) and the relation ≤ should be interpreted elementwise
for vectors and matrices, i.e. A = (ai, j ) ∈ Rn×m and B = (bi, j ) ∈ Rn×m such that
A ≥ B if and only if ai, j ≥ bi, j ∀ i ∈ {1, . . . , n} and j ∈ {1, . . . , m}, i, j ∈ N. For
a matrix R ∈ Rn×m , define R + = max {0, R} and R − = R + − R. The matrix of
absolute values of all elements of a matrix M ∈ Rn×m is |M| = M + + M − . The
vector E p is stated for ( p × 1) vector with unit elements, and In denotes the identity
matrix of n × n dimension. Superscript T denotes the transpose of a matrix or a
vector. K is the set of continuous increasing functions γ : R+ → R+ with γ (0) = 0.
We refer by β ∈ KL if β(·, t) ∈ K for all t ≥ 0 and β(r, ·) is continuous and strictly
decreasing to zero for all r > 0. . is the standard 2-norm.
Lemma 1 ([3]) Let x, x, x ∈ Rn be vectors satisfying x ≤ x ≤ x and A ∈ Rn×m be
a time-invariant matrix. Then, the inequalities below hold:

A+ x − A− x ≤ Ax ≤ A+ x − A− x. (1)

Lemma 2 ([2]) Consider the following system:


ẋ(t) = Ax(t) + ψ(t),
(2)
y(t) = C x(t),

where ψ(t) is a continuous function and A, C are known matrices. Suppose that
there exist two known continuous-time functions ψ(t) and ψ(t) : R → Rn satisfied
ψ(t) ≤ ψ(t) ≤ ψ(t), ∀ t ≥ 0.
Set-Membership Fault Detection Approach for a Class of Nonlinear … 135

If there exists a gain L such that (A − LC) ∈ H ∩ M and x0 , x0 , x0 ∈ Rn , x 0 ≤

x0 ≤ x 0 , then the system:

ẋ(t) = Ax(t) + ψ(t) + L(y(t) − C x(t))
(3)
ẋ(t) = Ax(t) + ψ(t) + L(y(t) − C x(t))

is an interval observer for (2) and x(t) ≤ x(t) ≤ x(t), ∀t ≥ 0.

Definition 1 Consider the nonlinear system

ẋ = f (x, u), (4)

with f (x, u) ∈ Rn , the system (4) is input-to-state Stable (ISS) if for any input
u ∈ Rm and x0 ∈ Rn there exist functions β ∈ KL and γ ∈ K such that

|x(t, x0 , u)| ≤ β(x0 , t) + γ ( u ), ∀t ≥ 0. (5)

3 NCS Architecture Description and Problem Formulation

Consider the following NNCS:


ẋ(t) = Ax(t) + F(u(t), y(t)) + w(t) + H f (t),
(6)
y(t) = C x(t) + v(t),

where x ∈ Rn denotes the state vector, y ∈ R p the measurable output vector u ∈ Rm

the known input vector, where w ∈ Rn , v ∈ R p and f ∈ Rn are the external distur-
bances and the additive faults to be detected. The functions u, v, w are continuous. The
function F(u(t), y(t)) ∈ Rn is a globally Lipschitz nonlinear function. The matrices
A, C and H are known matrices of compatible dimensions.
Before proceeding further, we make some assumptions on the process matrices.
Assumption 1 The pair (A, C) is detectable. 

Assumption 2 w(t) and v(t) are unknown but bounded functions with a priori
known bounds, for w, w ∈ Rn , V ∈ R+ :

w ≤ w(t) ≤ w, |v(t)| ≤ V E p , ∀t ≥ 0 (7)

The network induces two uncertain delays: da (t) and dm (t) which refer to the actu-
ation and the measurement channels delay, respectively. These delay functions are
unknown but bounded:

0 ≤ da ≤ da (t) ≤ da , 0 ≤ dm ≤ dm (t) ≤ dm , ∀t, (8)

136 A. Najjar and J.-C. Ponsart

Fig. 1 NCS architecture

where da , da are respectively the upper and lower bounds of da (t) and dm , dm are
respectively the upper and lower bounds of dm (t). Suppose that d and d are respec-
tively the upper and lower bounds of the communication delays such that:

d = max{da , dm }, d = min{da , dm }. (9)

Considering the communication delay in the actuation channel, the process (6) can
be modeled as an input delayed system:

ẋ(t) = Ax(t) + F(u(t − da (t)), y(t)) + w(t) + H f (t)
(10)
y(t) = C x(t) + v(t).

Assumption 3 We assume that u(t − da (t)) is bounded. 

Let {tk , k ∈ N} be the sequence of sampling instants such that tk+1 − tk = T and
lim tk = ∞, T is the sampling period and tk is an increasing sequence such that
k→∞
tk = kT .In a network environment, data sampling is needed. Therefore, the next
assumptions are required.
Assumption 4 ([1]) The sampling communication delays da (tk ) and dm (tk ) are
unknown but bounded with a priori known bounds and the upper bound d is assumed
to be a multiple of the sampling period T . 

Assumption 5 ([1]) The information on the control signal u(tk ), the information on
the output y(tk ) and the information on the sampling upper and lower bounds z(tk ),
 
z(tk ) could be stored and used ∀ tk ∈ tk − d, tk . 

Each block of the NCS architecture shown in Fig. 1 performs the same function as
described in [5]. However, the added Residual Generator block is implemented in the
Set-Membership Fault Detection Approach for a Class of Nonlinear … 137

calculator for residual generation; after receiving predictor outputs; z(tk ) and z(tk ),
Residual Generator calculates interval residual signals r (tk ) and r (tk ) which will be
used for FD test detailed in Sect. 4.

3.1 Interval Observer Structure

In this section, the interval observer developed in [5] is tackled to estimate the unavail-
able states of the process under unknown input delay da (t) and process uncertainties.
In free faulty case, the system (10) is:

ẋ(t) = Ax(t) + F(u(t − da (t)), y(t)) + w(t)
(11)
y(t) = C x(t) + v(t).

Since it is not always possible to compute a gain L for the system (11) such that
A − LC ∈ H ∩ M, a change of coordinates ξ = Sx with a nonsingular matrix S
such that the matrix S(A − LC)S −1 ∈ H ∩ M is used to relax this restriction [8].
Theorem 1 ([5]) Let Assumptions 1–3 be satisfied and x0 ≤ x0 ≤ x0 . If there exists
a change of coordinates ξ = Sx satisfying E = S(A − LC)P ∈ H ∩ M, P = S −1
so that the following system

x̂˙ + (t) = Ex̂ + (t) + S F(u(t − d), y(t)) + S L y(t) + S + w − S − w + |S L| E p V ,
x̂˙ − (t) = Ex̂ − (t) + S F(u(t − d), y(t)) + S L y(t) + S + w − S − w − |S L| E p V ,
(12)
where ⎧
⎪
⎨ S F(u(t − d), y(t)) = max  {S F(u(t − d − α), y(t))},
α∈ 0,d−d
(13)
⎩ S F(u(t − d), y(t)) = min {S F(u(t − d − α), y(t))},
⎪
α∈ 0,d−d

and the initial conditions are calculated as follows:

x̂ + (0) = S + x 0 − S − x 0 , x̂ − (0) = S + x 0 − S − x 0 , (14)

is input-to-state stable (ISS) interval observer for the system (11) satisfying [4]

x̂ − (t) ≤ ξ(t) ≤ x̂ + (t), ∀t ≥ 0 (15)

where the bounds of the solution x(t) are:

x(t) = P + x̂ + (t) − P − x̂ − (t), x(t) = P + x̂ − (t) − P − x̂ + (t) . (16)

138 A. Najjar and J.-C. Ponsart

such that
x(t) ≤ x(t) ≤ x(t), ∀t ≥ 0. (17)

Proof The proof of the above theorem is detailed in [5].

3.2 State Predictor Design

To compensate the large unknown communication delays dm (tk ), an interval predictor

introduced in [5] is used. Based on the delayed data z( tTk − Td ), z( tTk − Td ), we will
reconstruct z( tTk ), z( tTk ) after a finite time tk = d.
Assumption 6 T is selected such that the matrix  = In + T E is positive. 

Theorem 2 ([5]) If Assumptions 4–6 hold ∀ tk ≥ d, we get:

k2
ẑ + (k) = k2 ẑ + (k − k2 ) + k2 − j T S F(u(k − k1 − k2 + j − 1),y(k − k2 + j − 1))
j=1
k2 k2
+ k2 − j L 1 y(k − k2 + j − 1) + k2 − j β, (18)
j=1 j=1

k2
ẑ − (k) = k2 ẑ − (k − k2 ) + k2 − j T S F(u(k − k1 − k2 + j − 1),y(k − k2 + j − 1))
j=1
k2 k2
+ k2 − j L 1 y(k − k2 + j − 1) + k2 − j β, (19)
j=1 j=1


d
where j = 1, 2, 3, . . . , k2 , k = tk
T
, k1 = T
and k2 = d
T
,  = (In + T E), L 1 =
T S L, β = T (S w − S w + |S L| E p V ), β = T (S w − S − w − |S L| E p V ) and
+ − +

z(k) = S + ẑ + (k) − S − ẑ − (k), z(k) = S + ẑ − (k) − S − ẑ + (k) (20)

are a predictor from the sampling instant of time k − k2 to k for the process (11) i.e.
z → x and z → x ∀ tk ≥ d and the following inclusion holds

tk tk tk
z( ) ≤ x( ) ≤ z( ), tk ≥ d. (21)
T T T
Proof Please see the proof detailed in [5].
Set-Membership Fault Detection Approach for a Class of Nonlinear … 139

4 Fault Detection

In this section, a procedure of fault detection is developed thanks to Residual Gener-

ator block as shown in Fig. 1. The residual evaluation in a set-membership context is
ensured via the following belonging test [12]: If the zero signal is enclosed by upper
and lower bounds of the residual signal, it is a fault-free case. Otherwise, a fault is
occurred. Two steps are required to indicate the presence of faults:
• Step1: Residuals generation: in this first step, the Residual Generator calculates
upper and lower residuals defined as the gap between measured outputs and esti-
mated outputs using stored information.
From Lemma 1, the upper and lower bounds of the estimated output are then com-
puted as follows:

y(tk ) = C z(tk ) + V E p = C + z(tk ) − C − z(tk ) + V E p
(22)
y(tk ) = C z(tk ) − V E p = C + z(tk ) − C − z(tk ) − V E p

Then, upper and lower residuals are :

r (tk ) = y(tk ) − y(tk ), r (tk ) = y(tk ) − y(tk ) . (23)

• Step 2: Residuals evaluation: this second step is detailed as follows. When a fault
is occurred, an inconsistency is detected shown that the estimated outputs are no
more compatible with the measurements where:

y(tk ) ∈
/ [y(tk ), y(tk )] (24)

The above belonging test is rewritten as follows:

⎧
⎨0 ∈
/ [y(tk ), y(tk )] − y(tk )
0∈
/ [y(tk ) − y(tk ), y(tk ) − y(tk )] (25)
⎩
/ [r(tk ), r (tk )]
0∈

Then, the zero signal is enclosed by r and r in the fault free case. Otherwise, a
fault is detected.
From (23) and using the fact that C = C + − C − , r and r are computed as follows:

r (tk ) = y(tk ) − y(tk ) = C + (z(tk ) − z(tk )) + C − (z(tk ) − z(tk )) − v(tk ) + V E p

(26)
r (tk ) = y(tk ) − y(tk ) = −C + (z(tk ) − z(tk )) − C − (z(tk ) − z(tk )) − v(tk ) − V E p
(27)
An augmented system is then defined as:
140 A. Najjar and J.-C. Ponsart
   +    
r C C− ez −v + V E p
= + ,
r −C − −C + ez −v − V E p

ez (tk ) = z(tk ) − x(tk ), ez (tk ) = x(tk ) − z(tk ) (28)

Or we have z → x and z → x ∀ tk ≥ d (see Sect. 3.2), then


ez (tk ) = z(tk ) − x(tk ) x(tk ) − x(tk ) = e(tk )
(29)
ez (tk ) = x(tk ) − z(tk ) x(tk ) − x(tk ) = e(tk )

Known that the measurement noise v is bounded, stability analysis of upper and
lower residual signals is equivalent to assure the stability of the estimation errors.
From Theorem 1, we have upper and lower bound estimation errors e and e are
ISS [5]. Therefore, from (29), one can prove that ez and ez are ISS and then the
residual signals r (tk ) and r (tk ) are ISS ∀ tk ≥ d.

5 Numerical Example

To prove the efficiency of the proposed FD strategy, the next system (6) is considered:

⎡ ⎤ ⎡ ⎤
−3.5000 0 0.5000 sin(u(t))
A=⎣ 0 −2.7540 0 ⎦ , F(u(t), y(t)) = ⎣ 1.5 sin(u(t)y(t)) ⎦ ,
0 0 −1.2000
 2 sin(u(t))
C = 0 0 1 , H = [−3 2 1]T

w(t) = [0.1 cos(2t) 0.1sin(3t) 0.1 cos(4t)]T , w = [0.1 0.1 0.1]T , w = −w and
v(t) = 0.2 cos(t) cos(5t) sin(10t) sin(20t) with V = 0.2.

On can see that the function F(u(t), y(t)) is globally Lipschitz.

Considering communication delays, system (6) will be modeled as (10); its output
signal and the input signal delivered by the I.S.P.G are depicted in Fig. 2.
Network proprieties: The network induces unknown but bounded delays as
described by (8). These bounds are d = 0.7s, d = 0.1s and the distribution of delays
is shown in Fig. 3 with the disturbances and the measurement noise. The sampling
period is given by T = 0.01s.
 T
Interval observer design: A gain L = −10 0 0 is computed satisfying A −
LC ∈ H. However, the matrix A − LC is Hurwitz but is not Metzler. Then, a
transformation of coordinates as described in Sect. 3.1 is needed. We propose
Set-Membership Fault Detection Approach for a Class of Nonlinear … 141

Fig. 2 System input and 3

output
2.5

Input
2

1.5

1
0 5 10 15 20 25 30 35 40

Time(s)
3

SystemOutput
2

0
0 5 10 15 20 25 30 35 40

Time(s)

Fig. 3 Network delay,

NetworkDelay

disturbances and noise 0.5

0
0 5 10 15 20 25 30 35 40
Time(s)
0.1
Disturbances

-0.1
0 5 10 15 20 25 30 35 40
Time(s)

0.1
Noise

-0.1
0 5 10 15 20 25 30 35 40
Time(s)

⎡ ⎤
0.5005 0.0000 −2.2800
S = ⎣ 0.000 1.0064 0.0000 ⎦, and we can easily verify that E = S(A −
−0.0010 −0.0012 5.5701
⎡ ⎤
−3.5000 0.0000 0.0020
LC)S −1 = ⎣ 0.0000 −2.7540 0.0000 ⎦ ∈ H ∩ M. Therefore, an interval
0.0049 0.0019 −1.2000
observer as (12) can be designed. The initial conditions of system (10) are
142 A. Najjar and J.-C. Ponsart
 T
chosen such that x 0 (t) ≤ x0 (t) ≤ x 0 (t); x(0) = 0.5 0.5 0.5 and x(0) =
 T
−0.5 −0.5 −0.5 . Also the estimator (12) is initialized by x̂ (0) and x̂ − (0) which
+

are defined in (14).

State⎡ predictor design: For⎤ the sampling period T = 0.01s, we obtain
0.9650 0.0000 0.0000
= ⎣ 0.0000 0.9724 0.0000 ⎦. The predictor described by (18) and (19) can be
0.00005 0.00002 0.9880
computed with , L˜1 = [−0.0500 0.0000 0.0001]T
and β = [0.0127 0.0010 0.0010]T , β = −β.
Fault evolution: The system is affected by an additive fault f (t):

f (t) = 1 20s ≤ t ≤ 30s
(30)
f (t) = 0 else

The fault evolution is presented in Fig. 4. The evolution of the residual signals in
faulty case is plotted in Fig. 4 where black lines are the zero signal and red, blue lines
represent the upper and lower residuals, respectively. As shown in Fig. 5 upper and
/ [r , r ] when 20s ≤ t ≤ 30s.
lower residual signals are sensitive to the fault f i.e. 0 ∈
Then we can conclude that the fault detection strategy is ensured and validated despite
unknown communication delays and external process uncertainties.

Fig. 4 Fault evolution 1

0.9

0.8

0.7

0.6
( )

0.5
Fault

0.4

0.3

0.2

0.1

0
0 5 10 15 20 25 30 35 40
Time(s)
Set-Membership Fault Detection Approach for a Class of Nonlinear … 143

Fig. 5 Evolution of upper 1.5

and lower residual signals
1

0.5

-0.5

-1

-1.5
0 5 10 15 20 25 30 35 40
Time(s)

6 Conclusion and Future Work

In this work, a set-membership FD technique is developed for NNCS subject to

network delays and uncertainties. These uncertainties are assumed to be unknown
but bounded with a priori known bounds. The main contribution consists in using
upper and lower residuals for FD decision. A belonging test of the zero signal to the
interval delimited by the upper and lower residuals is used to ensure the detection of
faults. Theoretical results have been validated through the numerical example.
Nevertheless, the fault isolation problem for such NNCS is not investigated in this
contribution. It will be the subject of a future work.

References

1. Yang, R., Liu, G.P., Shi, P., Thomas, C., Basin, M.V.: Predictive output feedback control for
networked control systems. IEEE Trans. Ind. Electron. 61(1), 512–520 (2013). https://doi.org/
10.1109/TIE.2013.2248339
2. Gouzé, J.L., Rapaport, A., Hadj-Sadok, M.Z.: Interval observers for uncertain biological sys-
tems. Ecol. Model. 133(1–2), 45–56 (2000). https://doi.org/10.1016/S0304-3800(00)00279-
9
3. Efimov, D., Raïssi, T.: Design of interval observers for uncertain dynamical systems. Autom.
Remote. Control. 77(2), 191–225 (2016). https://doi.org/10.1134/S0005117916020016
4. Dashkovskiy, S., Efimov, D.V., Sontag, E.D.: Input to state stability and allied system
properties. Autom. Remote. Control. 72(8), 1579–1614 (2011). https://doi.org/10.1134/
S0005117911080017
5. Najjar, A., Dinh, T.N., Amairi, M., Raïssi, T.: Interval observer-based supervision of nonlinear
networked control systems. Turk. J. Electr. Eng. Comput. Sci. 30(4), 1219–1234 (2022). https://
doi.org/10.55730/1300-0632.3845
144 A. Najjar and J.-C. Ponsart

6. Ren, C., Park, J.H., He, S.: Positiveness and finite-time control of dual-switching poisson jump
networked control systems with time-varying delays and packet drops. IEEE Trans. Control
Netw. Syst. 9(2), 575–587 (2022)
7. Zhang, X.M., Han, Q.L., Ge, X., et al.: Networked control systems: a survey of trends and
techniques. IEEE/CAA J. Autom. Sin. 7(1), 1–17 (2019)
8. Najjar, A., Dinh, T.N., Amairi, M. et al.: Supervision of nonlinear networked control systems
under network constraints. In: 2019 4th Conference on Control and Fault Tolerant Systems
(SysTol), pp. 270–275. IEEE
9. Keller, J.Y., Sauter, D.: Monitoring of stealthy attack in networked control systems. In: 2013
Conference on Control and Fault-Tolerant Systems (SysTol), pp. 462–467. IEEE
10. Wu, C., Liu, J., Jing, X., et al.: Adaptive fuzzy control for nonlinear networked control systems.
IEEE Trans. Syst. Man Cybern.: Syst. 47(8), 2420–2430 (2017)
11. Gao, Z., Cecati, C., Ding, S.X.: A survey of fault diagnosis and fault-tolerant techniques-part i:
fault diagnosis with model-based and signal-based approaches. IEEE Trans. Industr. Electron.
62(6), 3757–3767 (2015)
12. Chevet, T., Dinh, T.N., Marzat, J. et al.: Robust sensor fault detection for linear parameter-
varying systems using interval observer. In: 31st European Safety and Reliability Conference.
Proceedings of the 31st European Safety and Reliability Conference, Angers, France, pp.
1486–1493. https://doi.org/10.3850/978-981-18-2016-8_380-cd. https://hal-cnam.archives-
ouvertes.fr/hal-03239385
13. Heijmans, S., Postoyan, R., Nešić, D., et al.: An average allowable transmission interval condi-
tion for the stability of networked control systems. IEEE Trans. Autom. Control 66(6), 2526–
2541 (2021). https://doi.org/10.1109/TAC.2020.3012526. https://hal.archives-ouvertes.fr/hal-
02909161
Demanded Power Point Tracking
for Urban Wind Turbines

Felix Dietrich, Lukas Jobb, and Horst Schulte

Abstract In addition to increasing the rated power of wind turbines in the multiple
megawatt range, research is advancing into suitable small compact wind turbines
for urban areas. In particular, the requirements for their flexible automated commis-
sioning and plug-and-play solutions are growing. Up to now, small wind turbines
have been operated in such a way that these provide the maximum achievable power
from the wind. In the future, however, it will be essential to integrate them into a
power system as a flexibly controllable current or voltage source. A demanded power
point tracking algorithm based on a model-free extremum-seeking control scheme
is proposed to get closer to this aim. As a logical extension of the well-known maxi-
mum power point algorithm, the method is presented in detail, applied to a real wind
turbine of a manufacturer, and validated in simulation studies for direct application
in the field.

Keywords Active power control · Wind energy · Extremum-seeking control ·

MPPT · Variable-speed control

1 Introduction

Wind energy has seen rapid market share growth over the past few years. Furthermore,
several countries, such as the U.S., have pledged to develop this field in the coming
years further [1]. Given this development, it has become more important for wind
turbines to be able to react to (sudden) changes in load-side energy demand. For
grid-connected turbines, this means adjusting to the grid’s frequency variations. One

F. Dietrich · H. Schulte (B)

School of Engineering-Energy and Information, Control Engineering Group,
University of Applied Sciences (HTW) Berlin, Wilhelminenhofstraße 75A, 12459 Berlin,
Germany
e-mail: [email protected]
L. Jobb
MOWEA - Modulare Windenergieanlagen GmbH, Storkower Str. 115A, 10407 Berlin, Germany

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 145
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_13
146 F. Dietrich et al.

Fig. 1 Power curve of a

variable speed wind turbine
divided into Region 1 to
Region 4

should note that over the years, variable-speed wind turbines have replaced fixed
wind turbines as the most common newly installed wind turbine type [8].
With a full-frequency converter, variable-speed wind turbines decouple the iner-
tial turbine dynamics from the power grid through their power electronics. Therefore,
a variable speed wind turbine without an additional control scheme has no inertial
response like conventional power generators when the power demand and grid fre-
quency vary [3]. However, an advantage of variable-speed wind turbines is that they
can increase their overall power efficiency by applying a power curve, as depicted
in Fig. 1. When they operate below their rated power (Region 1), a Maximum Power
Point Tracking (MPPT) control methodology can be applied to obtain the highest
possible power output at different wind speeds. It is achieved with a wind turbine
controller using torque control to optimize the turbine’s power output [2]. However,
once the rated wind speed is surpassed, the objective is to maintain the wind turbine’s
power output at the rated power level while minimizing the structural loads on the
turbine (Region 2). Wind turbines in the standard megawatt class commonly achieve
this through pitch control while the torque is constant. The pitch control changes the
rotor blades’ angles to reduce the turbine’s aerodynamic efficiency to operate below
its maximum power point.
This paper focuses on small wind turbines with a rated power of up to 2 kW
without a blade pitch mechanism. Therefore, all control is carried out exclusively by
adjusting the generator torque without pitch control. When the wind speed reaches
the cut-out speed, the turbine is shut down to prevent structural damage. In some
cases, the cut-out is not carried out abruptly, but the power output is continuously
reduced according to the wind speed (Region 3). That way, some energy can still be
harvested while the load on the mechanical parts of the turbine is still manageable. To
follow external power commands in Region 1 and Region 2 and internal reference
for smooth power shutdown in Region 3, this paper proposes a demanded power
point tracking (DPPT).
To achieve DPPT functionality, various control strategies have been developed
and proposed in the literature. First, we give an overview of the wind turbines of the
megawatt-class, which are controlled by the generator torque and rotor blade pitch
Demanded Power Point Tracking for Urban Wind Turbines 147

angle: To cover the whole wind speed spectrum in [4] the authors proposed a linear
parameter variable (LPV) control design method for variable-pitch wind turbines.
DPPT methods, also called active power tracking control (APC), to support grid
integration with pitch-based and torque-based power reserve control strategies are
studied and compared in [9]. One proposed control strategy was to obtain rotor speed
set-points according to demanded power points. It was validated through simulations
and field tests with a 550 kW wind turbine. In [12], a transition region between
the partial load region (Region 1) and full load region (Region 2) is established,
see Fig. 1. Here, mode switches enable or disable the DPPT control ability. The
proposed algorithm is validated through simulation and field tests with a down-
scaled wind turbine in a wind tunnel. In [5], a model-based multi-variable controller
design without switching mechanisms for power following behavior in the range
without and with pitch adjustment is presented and experimentally validated.
In this work, the focus is on small wind turbines. Due to the large model uncertain-
ties in the aerodynamic map, a model-free approach such as a modified extremum-
seeking (ES) algorithm controller is appropriate. On the other hand, the ES method is
unsuitable for megawatt-class turbines because the manipulated variable constraints
are significant, and the ES algorithm often reaches saturation, making this method
not well suited. The extremum-seeking control algorithms belong to the model-free
real-time optimization control strategy class. The literature proposes ES algorithms
to solve the maximum power point tracking (MPPT) problem for small wind turbines
and photovoltaic systems. In [6], the authors used an ES controller and combined it
with field-oriented generator control to optimize the power output of a grid-connected
wind turbine. Logarithmic power feedback as input is proposed in [14] for the ES
algorithm to cover a wider spectrum of wind speeds.
The generalization of the ES algorithm to track a given external slope reference
input has been presented in [11] and extends the results to the multivariable case
of gradient seeking. A similar approach is taken in [13]. Here, instead of using the
extremum seeking to optimize a measurable system output, it will be used to optimize
a suitable auxiliary cost function. This cost function is selected, so the system output
reaches a certain percentage of its optimal value.
To the best of the authors’ knowledge, this paper presents a DPPT algorithm for
a real industrial small wind power plant1 based on a modified ES method using a
similar cost function as presented in [13] for the first time. It is shown that the ES
control can be used for DPPT purposes as an extension of the existing MPPT method.
Therefore, the proposed method provides a simple, industrial robust control strategy
for the entire power range (Region 1-4) shown in Fig. 1 without the need for mode
switching or transition control between two control regions.
The paper is structured as follows: Sect. 2 explains the technical scheme of the
small wind power plant under consideration, derives the mathematical first-principle
model, and presents the internal control structure. In Sect. 3, the operation of an
extremum-seeking algorithm for maximum power point tracking is first described,
and second, as a natural extension, the modified ES is presented to achieve the

1 MOWEA String turbine, Modulare Windenergieanlagen GmbH, Berlin, Germany.

148 F. Dietrich et al.

required power point tracking. Afterward, the simulation results, which demonstrate
the performance of the demanded power point tracking algorithm using the ES
method, are presented in Sect. 4. Section 5 concludes the paper with a discussion
and an outlook.

2 Control Objectives and Process Description

The power Pr captured from the wind by the turbine rotor is defined by

1
Pr = ρ v 3 π R 2 c P (λ) , (1)
2

where ρ denotes the air density [kg/m3 ], v the wind speed [m/s] far in front of the
rotor, and R the rotor radius in [m]. Therefore π R 2 is the area perpendicular to the
airflow swept by the rotor blades. A wind turbine can only harness a fraction of the
wind power. The amount of power the turbine captures depends on the aerodynamic
efficiency of the rotor. The latter is described by the power coefficient c P . This paper
investigates a wind turbine whose power coefficient depends on the tip speed ratio
λ. The power coefficient would also be a function of the pitch angle β [deg] for wind
turbines with pitch angle control. The tip-speed ratio is defined by:

ωr R
λ= , (2)
v
where ωr denotes the rotor speed [rad/s]. The Fig. 2 shows the c P (λ) curve of the
studied wind turbine of the manufacturer MOWEA. In this case, the power coefficient
reaches its maximum of c P,max = 0.4 at a tip-speed ratio of λopt = 7.5 (see the upper
part of Fig. 2), which implies that the power output can be set dynamically by varying
the rotor speed. The control objectives of wind turbines are highly dependent on the
current operating region shown in Fig. 1: In Region 1, the objective would be to reach
λopt to operate at the maximum power point for a given wind speed. In Region 2 and 3,
the objective would be to keep the power output constant or reduce it, with the turbine
operating below its maximum power point. Since the rotor power is Pr = Tr ωr the
rotor torque Tr is described by

1 c P (λ(v, ωr ))
Tr (v, ωr ) = ρ v2 π R3 . (3)
2 λ(v, ωr )

The dynamics of the drive train of a small wind turbine is given by the 1-DOF
equation of motion

1 
ω̇r = Tr (v, ωr ) − Tg , (4)
J
Demanded Power Point Tracking for Urban Wind Turbines 149

Fig. 2 c P (λ) curve of the

MOWEA Champ 400 wind
turbine (upper figure) and the
transformed c̃ P (λ) curve for
a c Pr e f of 0.3 (lower figure)

where J is the total inertia of the drive with the main components, rotor, and generator
related to the rotation around the longitudinal axis of the drive train. Since a gearbox
is usually not required for small wind turbines because the rotor speed is high enough
to drive the generator directly.
This paper studies the MOWEA wind turbine using a brushless DC (BLDC) motor
as a generator. Its power flow is reversed, so the BLDC motor is a generator. At its
output, the AC power generated by the generator is rectified by a controllable inverter
that feeds a DC load. The rectifier is also used as a boost DC-DC converter. This
simplifies the design, and thus the generator model is used as follows:

1
Tg = k M ψ i DC (5)
2π
where k M denotes the motor parameter, and ψ is the constant flux parameter caused
by the rotor’s permanent magnets. The variable i DC represents the output current after
being rectified and passed through the DC-DC converter. The control system structure
is cascaded with an inner rotor speed controller K sc , and an outer control loop using
the ES control method is denoted as K E S . The input variables of the ES controller
are the setpoint resp. tracking power value and the instantaneous electrical power
with PDC = u DC i DC , measured at the load. The extreme value controller passes on
a setpoint speed to the rotor speed controller’s inner control loop. The latter controls
the setpoint by adjusting the inverter’s duty cycle signal u d at its output depending
on the control error. The duty cycle determines the DC-DC voltage ratio. It is thus
inversely proportional to the current ratio, which is used to set the speed via the
generator torque according to (5) and the motion Eq. 4. As quantified by Eqs. (1)
and (2), a change in rotor speed affects the power output of the turbine, closing the
causal chain of the outer loop (Fig. 3).
150 F. Dietrich et al.

Pref

Fig. 3 Drive train, power electronics, and control scheme with Extremum seeking for DPPT used
in small wind turbines

3 Methods

It is now investigated how the ES control algorithm can be used for MPPT control in
Region 1 and DPPT control in Region 2 and 3 without structure switching. Specifi-
cally, in Region 2, the turbine is used in stall mode, and Region 3, it is used for power
control to zero to reach Region 4. The overall objectives are
1. Implementing the ES algorithm as an MPPT control strategy by adapting its con-
trol output n r e f where 2 π n r e f = ωr,r e f so that the system output PDC converges
to the maximum power point of the c P (λ) curve;
2. Transforming the c P (λ) curve so that its transformed maximum lies at the refer-
ence power point Pr e f with a respective c Pr e f (λ).
In the following three subsections, we first discuss the original operation of the ES
optimizer, then present the modified ES optimization method to be used as a power
tracking control scheme.

3.1 Original Extremum Seeking Optimizer

Due to the unknown constantly changing effective wind speed in front of the rotor, the
actual maximum power point is unknown in advance. The maximum power point is
shifting continuously. For these reasons, the extremum-seeking optimization method
is advantageous for finding its maximum power point. The method is a model-free,
real-time optimization with rigorously provable stability and convergence [10].
In Fig. 4, an already modified ES scheme is depicted. The control block D is omit-
ted from the original ES optimization method scheme, and the plant’s output signal
is directly fed to the ES algorithm. ES uses a continuous oscillatory perturbation
signal i(t) = î sin ω p t superimposed onto the ES output, the rotor speed set point
Demanded Power Point Tracking for Urban Wind Turbines 151

Fig. 4 Schematic of the extended extremum-seeking controller for DPPT of small wind turbines
(denoted as a plant) without pitch-control

n r e f . The system response and electrical power output are continuously measured,
and its gradient is extracted using a highpass filter. The gradient is then assessed by
multiplying it by the phase-compensated perturbation signal. Finally, the resulting
gradient is amplified and controlled to zero by an integrator. In this way, the ES
continuously updates its output, the rotor speed reference value in proportion to the
gradient. A unique maximum for a given wind speed is observed, comparable to
the c p (λ) curve depicted in the upper part of Fig. 2. The plants’ output signal PDC
eventually converges to this maximum power point. The convergence time depends
on the Hessian matrix of the underlying plant model of the unknown c p (λ) curve
therefore, it is impossible to determine it analytically. However, there is the possi-
bility of building an estimation by measuring the convergence time at different set
points. Alternatively, the ES algorithm can be changed to a so-called Newton version,
rendering the convergence time to be user-assignable [7]. For the sake of simplicity,
the Newton version has not been implemented.
The parameters of the extremum-seeking algorithm must be well chosen. The
dynamics of the system and the characteristics of the disturbances (i.e., the wind
changes) have to be considered to ensure fast convergence to the maximum power
point and overall robustness. The largest time constant of our system can be approx-
imated to τn ≈ 4s. Therefore, the perturbation frequency is ω p = 1s −1 . It is suffi-
ciently slow so that no dynamic mode of the turbine is excited. The perturbation
amplitude is chosen as small as possible (n̂ = 15 rpm), considering a noisy systems
output signal and requiring a clear system response in the operational space. The
gradient of the system output can be obtained with the help of a high-pass filter. Its
cut-off frequency, ω H = 0.2s −1 is chosen to let the perturbation frequency through
but to cut out lower frequencies.
Additionally, a saturation block is used to limit the effects of disruptive changes
in wind velocity on the algorithm. It is chosen to have a maximum level of ±5 W.
Finally, the gain of the integral controller k is selected to ensure fast convergence for
all available urban wind scenarios.
152 F. Dietrich et al.

3.2 Modified Extremum Seeking Control for DPPT

The objective of demand power point tracking control is not the convergence to the
maximum power point as in MPPT control but rather the convergence to the ref-
erenced power. The c P (λ) curve in the upper part of Fig. 2 is a concave function.
Therefore, two λ1 and λ2 values exist for any power value, except for its maximum
power point. The control objective is to transform this c p (λ) curve so that its max-
imum point lies at the reference value. To achieve this goal, a control block D is
added to the input of the ES optimizer, as shown in Fig. 4. Block D transforms the
input variable PDC according to

Pr2e f − (Pr e f − PDC )2

P∗ = (6)
Pr e f

The transformed c p (λ) curve is shown in the lower part of Fig. 2. Assuming Pr e f
is constant for a period of time and PDC is a concave function of λ, the expression
−(Pr e f − PDC )2 is a quadratic-shaped function with two local maxima at λ1 and λ2
and a local minimum at its original maximum power point at λopt . The expression
Pr2e f is added to this expression to be positive definite over most of its operating
range. This is important because the ES configured in this paper can only operate
with positive definite input functions. Finally, it is divided by the reference value
so that when PDC = PRe f , which is to say when the system output is converging
to its reference value, P ∗ = Pr e f . The lower part of Fig. 2 shows a transformation
of the c P (λ) curve for a reference power equalling c Pr e f = 0.3. The two maxima
can be observed for λ-values of 5.02 and 10.07, respectively. λopt of c P (λ) curve of
the upper figure is now a local minimum in the transformed c P (λ) curve since the
transformed function is concave on the intervals [− inf, λopt ] and [λopt , + inf] the
ES algorithm will converge to one of the two new maxima.

4 Results

To validate our extremum-seeking approach for DPPT, a wind turbine subjected

to different wind speeds and power reference values is simulated. The simulations
were performed using Simulink®Mathworks R2022a. The simulation is visualized
for a total time frame of 120 s. The results are shown in Fig. 5. The upper simulated
time series shows the wind power plant’s power reference value and electrical power
output. The second depicts changes in the turbine speed, and the third shows generator
and rotor torque variations. The lower time series represents the wind speed, which
is the disturbance in our controlled system.
The power plant undergoes a cold start, and its power output convergences after
t = 7s to the reference value. The ripple in the turbine speed values, the power
output signal, and torque values are characteristic of the extremum-seeking optimizer.
Although the turbine is subjected to disruptive wind changes, a step at t = 25 s from
Demanded Power Point Tracking for Urban Wind Turbines 153

Fig. 5 Simulation with artificial wind velocity

vwind = 13 ms−1 to vwind = 26 ms−1 and back to vwind = 16 ms−1 at t = 50 s, the

DPPT-controller can converge to the reference value after a maximum of 8 s. From
t = 75 s, changes in the power reference value are applied, but the controller can
track them in less than 10 s. Note that convergence time varies because we deal with
a strongly nonlinear system.

5 Discussion and Outlook

Wind turbines capable of active power variation by an external reference signal

(i.e., superior control) require different control systems than those that feed in as
much power as possible at partial load. The latter requires an MPPT control scheme,
while the former requires limiting or reducing power output. This paper proposed a
demanded power point tracking control scheme to satisfy the requirements of rated
and derated wind regions as well as the power demands of the load side. To avoid
switching in the controller during operation, an MPPT algorithm was modified as an
extremum-seeking optimizer that works as a power controller. It was shown that the
proposed control scheme provides fast convergence to power reference signals and
robustness when faced with strong disturbances (wind speed changes). As necessary
in practice, the control system can perform a cold start. Furthermore, it does not need
error-prone and costly wind speed sensors to operate. It is computationally efficient
and can be easily implemented in a microcontroller.
154 F. Dietrich et al.

Future work will address its implementation onto a wind turbine using the
Hardware-In-The-Loop methodology. Analysis of the response time of the control
system will be necessary to assess if it complies with the requirements of the exist-
ing wind turbine system. Furthermore, its application for the adaptive control of a
multi-rotor system will be investigated.

References

1. U.S. Department of Energy: 20% Wind Energy by 2030: Increasing Wind Energy’s Contribution
to U.S. Electricity Supply. DOE/GO-102008-2567 (2008)
2. Bossanyi, E.A.: The design of closed loop controllers for wind turbines. Wind Energy 3(3),
149–163 (2000)
3. Callegari, G., Capurso, P., Lanzi, F., Merlo, M., Zaottini, R.: Wind power generation impact
on the frequency regulation: study on a national scale power system. In: 14th International
Conference on Harmonics and Quality of Power (ICHQP), pp. 1–6 (2010)
4. Inthamoussou, F.A., De Battista, H., Mantz, R.J.: LPV-based active power control of wind
turbines covering the complete wind speed range. Renew. Energy 99, 996–1007 (2016)
5. Pöschke, F., Petrovic, V., Berger, F., Neuhaus, L., Hölling, M., Kühn, M., Schulte, H.: Model-
based wind turbine control design with power tracking capability: a wind-tunnel validation.
Control Eng. Pract. 120 (2022). https://doi.org/10.1016/j.conengprac.2021.105014
6. Ghaffari, A., Krstic, M., Seshagiri, S.: Power optimization and control in wind energy conver-
sion systems using extremum seeking. IEEE Trans. Control Syst. Technol. 22(5) (2014)
7. Ghaffari, A., Krstic, M., Nesic, D.: Multivariable Newton-based extremum seeking. Automatica
48, 1759–1767 (2012)
8. Hansen, A.D., Hansen, L.H.: Market penetration of wind turbine concepts over the years (2006).
https://doi.org/10.1002/we.210
9. Jeong, Y., Johnson, K., Fleming, P.: Comparison and testing of power reserve control strategies
for grid-connected wind turbines. Wind Energy 17(3), 343–358 (2014)
10. Krstic, M., Wang, H.-H.: Stability of extremum seeking feedback for general nonlinear dynamic
systems. Automatica 36, 595–601 (2000)
11. Krstic, M., Ariyur, K.B.: Slope seeking: a generalization of extremum seeking. Int. J. Adapt.
Control Signal Process 18(1) (2004). https://doi.org/10.1002/acs.777
12. Kim, K., Kim, H.G., Kim, C.J., Paek, I., Bottasso, C.L., Campagnolo, F.: Design and valida-
tion of demanded power point tracking control algorithm of wind turbine. Int. J. Precis. Eng.
Manufact.-Green Technol. 5(3), 387–400 (2018). July
13. Labar, Christophe, Garone, Emanuele, Kinnaert, Michel: Sub-optimal extremum seeking con-
trol. IFAC-Papers Online 50(1), 7762–7768 (2017)
14. Rotea, M.A.: Logarithmic power feedback for extremum seeking control of wind turbines.
IFAC-Papers Online 50(1), 4504–4509 (2017)
Degradation Simulator for Infinite
Horizon Controlled Linear
Time-Invariant Systems

Amirhossein Hosseinzadeh Dadash and Niclas Björsell

Abstract Diagnosis, fault prediction, and Remaining Useful Life (RUL) estimation
are among the predictive maintenance research subjects used for maintenance cost
reduction. Using the available data with different machine learning methods, espe-
cially deep learning methods, the accuracy of estimation and prediction of faults
and RUL have increased dramatically. However, due to the statistical nature of the
machine learning methods and the limitations of available datasets, physically inter-
preting this information might be impossible. On the other hand, controlling the
degradation and faults in the machines as the optimum predictive maintenance solu-
tion needs the physical interpretation of the method’s outcome. In order to test the
new process-based methods for degradation and fault control, datasets with more
information are required (compared to available datasets). In this article, we intro-
duce an open-source degradation simulator for linear systems. This simulator can
simulate the degradation in closed-loop machines whose dynamics are known. It
is also possible to simulate different degradation models for different system parts
simultaneously by adding different processes and output noise to the system. This
simulator can generate enough data to test new machine learning-based predictive
maintenance methods.

1 Introduction

The classic definition of predictive maintenance premises that monitoring the system
condition using different indicators will provide the necessary data for preventing
unscheduled outages and maximizing the Mean Time To Failure (MTTF) [1]. Using

A. Hosseinzadeh Dadash (B) · N. Björsell

Department of Electronics, Mathematics and Natural Sciences, University of Gävle,
801 76 Gävle, Sweden
e-mail: [email protected]
N. Björsell
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 155
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_14
156 A. Hosseinzadeh Dadash and N. Björsell

different machine learning methods, the amount and quality of the information about
the systems’ state of health have increased dramatically. This has led to a more precise
prediction of the system lifetime [2, 3].
However, in the era of Industry 4.0 and complex industrial processes that exploit
distributed systems, the expectations have improved from mere prediction to the need
for control and synchronization using prediction. Having enough information about
the system and its respective degradation, it is possible to control the complex inter-
connected systems so sub-systems can reach their maintenance time simultaneously
[4]. Controlling the degradation in complex systems makes it possible to save more
downtime costs compared to the cases that only exploit prediction.
Considering that most of the research for failure prediction is focused on machine
learning methods such as deep learning [2], Support Vector Machine [5, 6], and k-
nearest neighbor [7, 8], regardless of the outcome precision, their outcome might not
be interpretable in the real world [9, 10]. On the other hand, from the control point
of view, the structure of the available datasets has many limitations, like the absence
of input and any information about the noise. Although these shortages make these
datasets more applicable for research on machine learning methods, they are not
useful from the control point of view.
In order to exploit the diagnosis and fault prediction information derived using
machine learning methods in the control scheme and improve the system perfor-
mance, it is necessary to adapt the machine learning method and explore system
dynamics [11–14]. For being able to test the new methods for degradation con-
trol, special datasets with different characteristics are needed. These datasets should
include more information compared to available datasets. Firstly, the noise should be
configurable (or at least known), so in the end, the effect of noise can be calculated
precisely. Secondly, controlling the degradation means diverting it from one part to
the other inside the system; these datasets should include enough information so that
this diversion can be traced and its consequences can be studied. Thirdly, the input
to the system (the desired system output and control inputs) should be known. The
input is necessary for the outcome to be interpretable physically and be helpful for
control. Finally, the effect of the feedback loop should be studied. This is because
the closed-loop system always compensates for the degradation (system parameters
deviation) by manipulating the system states using control inputs. This might degrade
the system’s more critical and expensive part(s), which might not be the intention of
the maintenance management system.
This article introduces the open-access simulator for simulating the degradation
in infinite horizon closed-loop linear time-invariant systems. This simulator is fully
configurable and can be used for research on controlling the degradation in a closed-
loop system according to system dynamics. It can also be used for the identification
of the dynamics of the system.
Degradation Simulator for Infinite Horizon … 157

2 Method

The prerequisite for designing the degradation simulator is to study how the degra-
dation happens and how it appears in the system dynamics.

2.1 Degradation in the Closed-Loop System

The state-space model (SSM) of the system can be written as follows:

⎧
⎪    
⎨ẋ = Ax + Bu + N v1 v1 R1 R12
z = Mx , v= , v (w) = , (1)
⎪
⎩ v2 T
R12 R2
y = C x + v2

where x is a vector containing system states, u is a vector of inputs. A and B are

the system parameters, C contains the output parameter, and M defines the states to
control. Also, v1 is the input noise, v2 is the sensor noise, and  is the disturbances
covariance matrix.
The degradation can be detected as a change in single or multiple signals recorded
from the system [15]. The effect of degradation can be detected in variations of input,
output, or both [16], meaning the degradation is a function of monotonic change in
the y, u, or both (or their feature) over time. However, the change in the u or y is
a result of the degradation in the system’s parameters ( A, B or C from (1)), but, as
controllers control the system’s output in a way that it follows the reference signal,
and the system is considered time-invariant the controller continuously manipulates
the system variables (states) using the system’s inputs in order to keep the output close
to the desired output. Eventually, at some point, the deviation between the system’s
parameters’ real values and the system’s parameters’ initial values increases so much
that either the controller becomes unstable or the output will go outside the acceptable
range. This is when the system fails, and maintenance is needed.

2.2 Degradation Simulation Algorithm

The whole process of simulating the degradation is shown in Fig. 1. Four inputs need
to be known: system state-space model, degradation model, disturbance model, and
failure or maintenance criterion.

2.2.1 Degradation and Disturbance Model

As mentioned in Sect. 2.1, although the degradation appears in the systems’ output,
it is the result of the controller actions for compensating for degradation system
158 A. Hosseinzadeh Dadash and N. Björsell

Fig. 1 Degradation simulation

parameters (A,B, or C). In the second step of the simulation, the degradation should
be defined. The degradation can be only time-dependent or depend on different
internal or external parameters, or it might come from the neural network or other
function estimators. Like the degradation models, the disturbances can come from
any source. They can also include process noise, measurement noise, or both.

2.2.2 Failure or Maintenance Criterion

When the degradation increases, the output deviates from the desired output. To
some system dependant tolerance, this deviation is acceptable, and after that, this
will consider a failure. The failure tolerance or the maximum deviation for the desired
output should be defined so the maintenance time can be calculated.
Also, no system can work for an infinite time without needing maintenance. This
maintenance might not be considered a response to a failure, but it is necessary to
keep the system going. For example, the system’s oil change is required when the
oil degrades. This can also be configured as the maximum MTTF of the system in
the simulator.

2.2.3 Controller Design

The final stage before the algorithm is the linear quadratic regulator (LQR) design,
which should be calculated according to system state space [17]. The quadratic
criterion that the LQ controller minimizes is given as follows:

J= e T Q 1 e + u T Q 2 u, (2)

where e = z − r and r is the reference signal, z is the controlled output, and e is

the error; u is the control input, and Q 1 and Q 2 are penalty matrices for the error
and input signal, respectively. The optimal control signal for this controller can be
written as follows:
Degradation Simulator for Infinite Horizon … 159

u = −L x + L r r, (3)
ẋ = Ax + Bu, (4)

where L is the optimal feedback gain and L r keeps the static gain of the closed-loop
equal to the identity matrix.

2.2.4 Algorithm

The following pseudo-code shows how the degradation simulator works. The sim-
ulator is designed based on the working cycles. This means that the parameters are
considered constant during each working cycle and get updated at the end of the
cycle. The working cycle here refers to the time from the beginning of one process
until the time that the process is done. For instance, the wood-cutting machine might
cut thousands of wood pieces until its saw needs to change. Each working cycle will
be from the beginning to the end of cutting one piece of wood.

2.3 Sample Simulation

For demonstration, two models are used for degradation simulation. The first model
is the simple mass and spring model with three states and one input, and the second
model is a more complicated linearized model of a rolling mill with five states and
two inputs. All simulation parameters can be found in the appendix.

Pseudo-code

Inputs: system state-space model

degradation model(s)
disturbance model(s)
failure criterion
Outputs: Time series of system outputs
Time series of system inputs
Record of a configurable feature for each of the working cycles
Calculate L and L r
While the failure criterion or maintenance criterion is not met
For each working cycle:
Update the model parameters according to their degradation
model
Calculate inputs and outputs and add the disturbances while
keeping the controller (L and L r ) the same
Record the control inputs, outputs and the desired features
160 A. Hosseinzadeh Dadash and N. Björsell

2.3.1 Mass and Spring

Figure 2 shows the model used for the first degradation simulation. The model state-
space is ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ẋ 0 1 0 x 0 
⎣ẍ ⎦ = ⎣ k2−k1 0 − k1 ⎦ . ⎣ẋ ⎦ + ⎣ 1 ⎦ . f , (5)
m m m
k1 k1
ż b
0 b
z 0

M = [1 0 0]x, (6)

y = I x + I v2 . (7)

Two degradation models with different disturbances were considered for this system.
The first degradation for k1 and its respective disturbances are modeled according to
⎡ ⎤
N (μ1 , σ1 )
k1 (t + 1) = k1 (t) + C1 ex p(C2 t), v2 = ⎣N (μ2 , σ2 )⎦ . (8)
N (μ3 , σ3 )

The second degradation for k1 is modeled according to

k1 (t + 1) = k1 (t) + N (μ4 , σ4 ) × ex p(C3 t), (9)

with the same disturbances as (8).

The result of the degradation simulation can be seen in Fig. 3.

2.3.2 Rolling Stand

The second degradation model is the linearized version of a rolling stand (as shown
in the left part of Fig. 2) that was introduced in [18]. The state-space model of the
mill is

Fig. 2 (Left) mass and spring model, (right) rolling stand model
Degradation Simulator for Infinite Horizon … 161

Fig. 3 Degradation simulation in mass and spring model (left) degradation model 1, (right) degra-
dation model 2

⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ḣ 0 1 0 0 0 h 0 0 0
⎢ ⎥ ⎢ 0 ⎥ ⎢
⎥ ⎥ ⎢ −1 ⎥ ⎢ ⎥
⎢ ḧ ⎥ ⎢ 0 − Fmc A1
− Am2
⎢ ⎥ ⎢ m ⎥ ⎢ ḣ ⎥ ⎢
⎢ ⎥ ⎢m
0 ⎥   ⎢1⎥
⎥ FL ⎢ ⎥
⎢ ⎥ ⎢ 0 − βVA11 − β(CVi +C e) βCi β Ps ⎥ ⎢ ⎥ ⎢0 0⎥ ⎢ ⎥
⎢ Ṗ1 ⎥ = ⎢ ⎥ . ⎢ P1⎥ + ⎢ ⎥. + ⎢0⎥ .v1 ,
⎢ ⎥ ⎢ ⎢
1 V1 Cl V 1 ⎥
⎥ ⎢ ⎥ ⎢ ⎥ uv ⎢ ⎥
⎢ Ṗ ⎥ ⎢ β A2 βCi β(Ci +Ce ) 2β Ps ⎥ ⎢ P ⎥ ⎢0 0⎥ ⎢0 ⎥
⎣ 2 ⎦ ⎣0 V 2 V2 − V2 C V 2 ⎦ ⎣ 2⎦
l
⎣ ⎦ ⎣ ⎦
kv
x˙v 0 0 0 0 − τ1v xv 0 τv 0
(10)

y = I x + I v2 . (11)

where h is the displacement of the rollers, ḣ is the vertical velocity of the rollers
displacement, Ps is the valve’s supply pressure, P1 and P2 are the pressures inside
the hydraulic cylinder; xv is the hydraulic valve spool displacement, FL is the input
force, Fc is the viscous force, m is the mass, A1 and A2 are the areas inside hydraulic
cylinder; V1 and V2 are the volumes of the primary or secondary side of the hydraulic
cylinder, respectively; β is the oil effective bulk modulus, Ci is the internal oil leakage,
Ce is the external oil leakage, Cl is the constant of linearization, τv is the hydraulic
valve time constant, kv is the gain of the hydraulic valve, and u v is the input voltage
to the valve.
The first degradation for Ce and its respective disturbances are modeled according
to 
Ce (t + 1) = Ce (t) + C4 ex p(C5 t), v2 = N (μi , σi ), , i = 1 . . . 5, (12)
162 A. Hosseinzadeh Dadash and N. Björsell

For the second degradation model, while the disturbances are kept the same, the
degradation is defined as:

ρ(t + 1) = ρ(t) − (C6 ex p(ρ(t)C7 t) − 0.1 − Ce , (13)

Ce (t + 1) = Ce (t) + C8 ex p(Ce (t)C9 t) + N (C10 ex p(Ce (t)C11 t), C12 ex p(Ce (t)C13 ),
(14)

Fc (t + 1) = Fc e(t) + C14 ex p(Fc (t)C15 t), (15)

β(t + 1) = β(t) − C16 ρ(t) + Ce (t)C17 ρ(t) − Fc (t)C18 , (16)

where ρ is the hydraulic oil density. The result of the degradation simulation for the
two mentioned degradation models on the rolling stand can be seen in Fig. 4.

Fig. 4 Degradation simulation in rolling mill model (left) degradation model 1, (right) degradation
model 2
Degradation Simulator for Infinite Horizon … 163

3 Conclusion

In this article, the degradation simulator, which is accessible through [19], was intro-
duced. Detailed information about the degradation simulation and its relation to the
system’s state space was discussed, and the algorithm used in this simulator was
explained. Finally, with the help of two models number of degradation was simu-
lated, and the results were shown. This simulator can be used to further investigate
the degradation control problems and degradation detection and RUL estimation in
closed-loop systems according to the system’s model.

Acknowledgements The research project is financed by the European Commission within the
European Regional Development Fund, Swedish Agency for Economic and Regional Growth,
Region Gävleborg, and the University of Gävle.

Appendix

Degradation models parameters:

A1 = 100e−3 A2 = 100e−3 Ps = 1e9 Fc = 10 β= 2e5 V1 = 0.5 V2 = 0.65

Ci = 1e−6 Ce = 1e−6 m= 1e2 Cl = 1.1 τv = 1e−3 kv = 1
C1 = 0.005 C2 = 0.005 C3 = 0.005 μ1 = 1e−5 μ2 = 1e−4 μ3 = 1e−5 μ4 = 5e−3
σ1 =1e−5 σ2 = 1e−4 σ3 =1e−5 σ4 =1e−2
C4 = 1e−8 C5 = 5e−7 C6 = 1e−20 C7 = 1e−4 C8 = 5e−11 C9 = 1e−7 C10 =5e−11
C11 = 1e−7 C12 = 5e−11 C13 = 1e−8 C14 = 5e−4 C15 = 1e−4 C16 = 1e−4 C17 =2
C18 = 50 μ5 = 1e−3 μ6 = 1e−1 μ7 = 1e4 μ8 = 1e4 μ9 = 1e4 σ5 =1e−3
σ6 =1e−1 σ7 = 1e4 σ8 = 1e−4 σ9 = 1e−5

References

1. Mobley, R.K.: An Introduction to Predictive Maintenance. Elsevier (2002)

2. Vollert, S., Theissler, A.: Challenges of machine learning-based RUL prognosis: a review
on NASA’s C-MAPSS data set. In: 2021 26th IEEE International Conference on Emerging
Technologies and Factory Automation (ETFA). IEEE (2021)
3. Ge, M.-F., et al.: A review on state of health estimations and remaining useful life prognostics
of lithium-ion batteries. Measurement 174, 109057 (2021)
4. Björsell, N., Dadash. A.H.: Finite horizon degradation control of complex interconnected sys-
tems. IFAC-PapersOnLine 54(1), 319–324 (2021)
5. Han, H., et al.: Least squares support vector machine (LS-SVM)-based chiller fault diagnosis
using fault indicative features. Appl. Therm. Eng. 154, 540–547 (2019)
6. Zhu, X., Xiong, J., Liang, Q.: Fault diagnosis of rotation machinery based on support vector
machine optimized by quantum genetic algorithm. IEEE Access 6, 33583–33588 (2018)
7. Madeti, S.R., Singh, S.N.: Modeling of PV system based on experimental data for fault detection
using KNN method. Solar Energy 173, 139–151 (2018)
164 A. Hosseinzadeh Dadash and N. Björsell

8. Xiong, J., et al.: An information fusion fault diagnosis method based on dimensionless indica-
tors with static discounting factor and KNN. IEEE Sens. J. 16(7), 2060–2069 (2015)
9. Kahana, A., et al.: Obstacle segmentation based on the wave equation and deep learning. J.
Comput. Phys. 413, 109458 (2020)
10. Jin, K.H., et al.: Deep convolutional neural network for inverse problems in imaging. IEEE
Trans. Image Process. 26(9), 4509–4522 (2017)
11. Read, J.S., Jia, X., Willard, J., Appling, A.P., Zwart, J.A., Oliver, S.K., Karpatne, A., Hansen,
G.J., Hanson, P.C., Watkins, W., et al.: Process-guided deep learning predictions of lake water
temperature. Water Resour. Res. 55, 9173–9190 (2019)
12. Daw, A., Karpatne, A., Watkins, W., Read, J., Kumar, V.: Physics-guided neural networks
(PGNN): an application in lake temperature modeling. arXiv:1710.11431
13. Rudy, S.H., Brunton, S.L., Proctor, J.L., Kutz, J.N.: Data-driven discovery of partial differential
equations. Sci. Adv. 3 (2017)
14. Brunton, S.L., Proctor, J.L., Kutz, J.N.: Discovering governing equations from data by sparse
identification of nonlinear dynamical systems. Proc. Natl. Acad. Sci. 113, 3932–3937 (2016)
15. Isermann, R.: Process fault detection based on modeling and estimation methods-a survey.
Automatica 20, 387–404 (1984)
16. Zagorowska, M., Wu, O., Ottewill, J.R., Reble, M., Thornhill, N.F.: A survey of models of
degradation for control applications. Annu. Rev. Control 50, 150–173 (2020)
17. Glad, T., Ljung, L.: Control Theory. CRC Press (2018)
18. Yan, J., Li, B., Guo, G., Zeng, Y., Zhang, M.: Nonlinear modeling and identification of the
electro-hydraulic control system of an excavator arm using BONL model. Chin. J. Mech. Eng.
26, 1212–1221 (2013)
19. Dadash, A.H.: Degradation Simulator (2022). Available via Github. http://github.com/ahhzd86/
degradation_simulator
Distributed Observer-Based
Leader-Following Consensus Control
Robust to External Disturbance
and Measurement Sensor Noise for LTI
Multi-agent Systems

Jesus A. Vazquez Trejo, Jean-Christophe Ponsart, Manuel Adam-Medina,

Guillermo Valencia-Palomo, and Juan A. Vazquez Trejo
Abstract A robust observer-based leader-following consensus control for linear
multi-agent systems subject to external disturbance and measurement noise is devel-
oped. The H∞ criterion is implemented to guarantee stability and robustness of the
synchronization error and estimation error through the Lyapunov approach. A set
of linear matrix inequalities are obtained to compute the control and observer gain
matrices. To show its effectiveness, the proposed strategy is carried out to solve the
leader-following formation control consensus-based problem in a fleet of unmanned
aerial vehicles under the effect of wind turbulence and measurement noise.

1 Introduction

In recent years, multi-agent systems have attracted the attention of researchers [1].
Multi-agent systems in this work are handled as found in the control community,
where despite the name being shared with the computer science community, the
meanings, the objectives, and the tools of common use are adapted to the control
area [2]. An agent is defined as an autonomous dynamical system, e.g., Unmanned

J. A. Vazquez Trejo (B) · J.-C. Ponsart · J. A. Vazquez Trejo

Université de Lorraine, CNRS, CRAN, 54000 Nancy, France
e-mail: [email protected]
J.-C. Ponsart
e-mail: [email protected]
J. A. Vazquez Trejo
e-mail: [email protected]
M. Adam-Medina
TecNM/CENIDET, Cuernavaca, Morelos 62490, Mexico
e-mail: [email protected]
G. Valencia-Palomo
TecNM/IT Hermosillo, Av. Tec y Per Poniente S/N, 83170 Hermosillo, Mexico
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 165
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_15
166 J. A. Vazquez Trejo et al.

Aerial Vehicles (UAVs), or satellites, among others. Nevertheless, multi-agent sys-

tems can be subject to external disturbance and measurement noise which compro-
mise the synchronization and coordination of the agents. It is therefore necessary to
consider these exogenous inputs in the dynamic modeling of the multi-agent system
in order to be able to develop a robust control of these external disturbances and
measurement noise. This is why it is essential to add robustness to the distributed
control and observer because external disturbance (and sensor noise) affect their
performance [3]. Different works in the literature have addressed this problem. A
robust finite-time problem for a second-order multi-agent system is addressed in [4]
for rejecting external disturbances. Also, a sliding mode leader-following control for
multiple UAVs is proposed in [5], where the system remains insensitive to external
disturbances. Alternatively, in [6], a leader-following sliding mode formation con-
trol approach for underactuated surface vehicles considering model uncertainties and
external disturbance is developed. In [7], a robust controller using H2 performance
is implemented in a system with communication and input time delays presented
in the frequency domain minimizing the error and the disturbance effect. In [8], the
robust optimal formation control problem for heterogeneous multi-agent systems
considering external disturbances is addressed based on reinforcement learning. An
adaptive semi-global bipartite consensus assuming a connected switching topology
graph under input saturation and external disturbance is proposed in [9].
In this paper, the main contribution is the design of the distributed leader-following
control law based on state estimates of the neighboring agents to coordinated linear
multi-agent systems, where the robust control and observer gains are computed simul-
taneously based on linear matrix inequalities (LMIs). The main difference concern-
ing previous work is the distributed robust observer-based leader-following control
for linear multi-agent systems under external disturbance and measurement noise,
where the main objective is to follow the trajectories described by a virtual leader
and maintain consensus between followers despite the external disturbance and the
sensor noise.

2 Problem Statement and System Description

Consider a linear homogeneous multi-agent system, which means that A, B, and C

matrices are identical for each agent, as follows:

ẋi (t) = Axi (t) + Bu i (t) + Du du i (t),

(1)
yi (t) = C xi (t) + D y d yi (t),

where i = 1, 2, . . . , N ; N is the total number of agents; xi (t) ∈ Rn , u i (t) ∈ Rm , yi (t) ∈

R p , du i (t) ∈ Rm , d yi (t) ∈ R p , are the state, input, output, external disturbance and
noise vectors respectively; A ∈ Rn×n , B ∈ Rn×m , C ∈ R p×n , Du ∈ Rn×m , and D y ∈
R p×n are constant matrices of the system. Graph theory [2] can be used for the
Distributed Observer-Based Leader-Following Consensus … 167

communication description of multi-agent systems. In this way, the Laplacian matrix

L describes the dynamics of the multi-agent system through the adjacency matrix
A. Let us consider a directed graph G(V, E, A) where V = {v1 , v2 , . . . , v N } is
a set of nodes (agents), E = {(i, j) : i, j ∈ V} ⊆ V × V is a set of edges. The
adjacency matrix A = [ai j ] ∈ R N ×N is defined as aii = 0, ai j = 1 if and only if
the pair (i, j) ∈ E otherwise ai j = 0. When the graph is undirected, also, a
i j = a ji ,
∀i = j and A = AT . The Laplacian matrix L ∈ R N ×N is defined as Lii = j=i ai j
and Li j = −ai j . The neighbors of i-th agent are denoted as j ∈ Ni . The following
assumptions are held in this paper:

Assumption 1 The pair (A, B) is stabilizable.

Assumption 2 The pair (A, C) is observable.

Assumption 3 The graph G is undirected and connected.

Lemma 1 ([10]) The matrix L̄ has non negative eigenvalues. The matrix L̄ is posi-
tive definite if and only if the graph is connected and undirected.

The main objective of this work is the design of a robust leader-following control for
multi-agent systems consensus-based, where the virtual leader dynamics is presented
as follows:
ẋl (t) = Axl (t), (2)

where xl (t) ∈ Rn is the state vector of the virtual leader. Let δi = xi − xl , then, the
dynamics of the synchronization error between each agent i and the leader is:

δ̇i (t) = Aδi (t) + Bu i (t) + Du du i (t). (3)

In order to estimate the states of each agent, the following distributed observer is
proposed:
x̂˙i (t) = A x̂i (t) + Bu i (t) + L(yi (t) − C x̂i (t)),
(4)
ŷi (t) = C x̂i (t),

where x̂i (t) ∈ Rn is the estimated state vector, L ∈ Rn× p is the observer gain to
be designed, and ŷi (t) ∈ R p is the estimated output vector. The dynamics of the
estimation error ei = xi − x̂i is computed as follows:

ėi (t) = (A − LC)ei (t) + Du du i (t) − L D y d yi (t). (5)

The considered observer-based leader-following consensus control law is [11]:

⎡ ⎤

u i (t) = K ⎣ ai j (x̂i (t) − x̂ j (t)) + αi (x̂i (t) − xl (t))⎦ , (6)
j∈Ni
168 J. A. Vazquez Trejo et al.

where K ∈ Rn× p is the control gain to be designed, ai j are the elements of the adja-
cency matrix, x̂i (t) is the estimated state vector of the i-th agent, and x̂ j (t) is the
estimated state vector of the neighboring agents, αi represents the communication
between the leader and the followers, where αi > 0 if there is a directed edge from
the leader to the i-th agent, otherwise αi = 0. The considered problem is the design of
a robust control gain K and a robust observer gain L to steer the multi-agent system
(1) subject to external disturbance and sensor noise, to follow the virtual leader’s
trajectories.

3 Observer-Based Leader-Following Consensus Robust

Controller

In this section, the LMI conditions to guarantee the existence of the robust control
and observer gains based on the Lyapunov stability analysis are presented. Con-
sidering the following new variables e(t) = [e1 (t)T , e2 (t)T , . . . , e N (t)T ]T , the dis-
turbance vector du (t) = [du 1 (t)T , du 2 (t)T , . . . , du N (t)T ]T , and noise vector d y (t) =
[d y1 (t)T , d y2 (t)T , . . . , d yN (t)T ]T , then, using the Kronecker product ⊗ the dynamic
estimation error can be expressed as:

ė(t) = (I N ⊗ (A − LC))e(t) + (I N ⊗ Du )du (t) − (I N ⊗ L D y )d y (t). (7)

 T
Replace (6) in (3) and let δ(t) = δ1 (t)T , δ2 (t)T , . . . , δ N (t)T , and L = L + 
where L is the Laplacian matrix and  = diag(α1 , α2 , . . . , α N ) is the communica-
tion exchange between the virtual leader and the followers. Then, the synchronization
error of the multi-agent system is rewritten as:

δ̇(t) = (I N ⊗ A + L ⊗ B K )δ(t) − (L ⊗ B K )e(t) + (I N ⊗ Du )du (t). (8)

Let z(t) = [δ(t)T , e(t)T ]T then, system (1) can be expressed by:

ż(t) = Az(t) + Du du (t) − D y d y (t), (9)

where

IN ⊗ A + L ⊗ B K −L ⊗ B K I ⊗ Du 0
A= , Du = N , Dy = .
0 I N ⊗ (A − LC) I N ⊗ Du IN ⊗ L Dy
(10)
The distributed H∞ criterion is considered as in [11], in order to guarantee the
existence of robust control and observer gains able to reject the effect of the external
disturbances and sensor noise, then, the following theorem is proposed.

Theorem 1 Consider the closed-loop system (9). Given the eigenvalues λ j (L̄), j =
1, 2, . . . , N ; if there exist symmetric matrices P̄1 > 0 ∈ Rn×n and P2 > 0 ∈ Rn×n ,
Distributed Observer-Based Leader-Following Consensus … 169

the tuning scalar variable μ > 0, and minimizing γ by the H∞ criterion satisfying
the following condition:
⎡ ⎤
η1 j 0 Du 0 −λ j B Nc 0
⎢∗ η2 P2 Du −Mo D y 0 I ⎥
⎢ ⎥
⎢∗ ∗ −γ 2 I N 0 0 0 ⎥
⎢ ⎥ < 0, (11)
⎢∗ ∗ ∗ −γ 2 I N 0 0 ⎥
⎢ ⎥
⎣∗ ∗ ∗ ∗ −μ−1 P 1 0 ⎦
∗ ∗ ∗ ∗ ∗ −μP 1

the synchronization error (8) and estimation error (7) are stable; where η1 j =
 
H e A P 1 + λ j B Nc , and η2 = H e {P2 A − Mo C} + I N , then, the robust control
−1
law can be computed with K = Nc P 1 , and the observer gain with L = P2−1 Mo .

Proof Consider the candidate Lyapunov function as follows:

I N ⊗ P1 0
V (t) = z(t)T z(t), (12)
0 I N ⊗ P2

where, P1 ∈ Rn×n = P1T > 0, and P2 ∈ Rn×n = P2T > 0. The derivative of the can-
didate Lyapunov function along the trajectories of (9) is given by:

V̇ (t) = 2δ(t)T (I N ⊗ P1 A + L ⊗ P1 B K )δ(t) − 2δ(t)T (L ⊗ P1 B K )e(t)

+ 2e(t)T (I N ⊗ P2 (A − LC))e(t) + 2δ(t)T (I N ⊗ P1 Du )du (t) (13)
+ 2e(t) (I N ⊗ P2 Du )du (t) − 2e(t) (I N ⊗ P2 L D y )d y (t).
T T

Let us perform a spectral decomposition of the matrix L, such that L = T J T −1 with

an invertible matrix T ∈ R N ×N and a diagonal matrix J = diag(λ1 , λ2 , . . . , λ N ) ∈
R N ×N . Defining the change of coordinates as ξ = (T −1 ⊗ I N )δ, = (T −1 ⊗ I N )e,
ωu = (T −1 ⊗ I N )du , and ω y = (T −1 ⊗ I N )d y . For reasons of space, the notation of
the time dependence of certain variables is omitted. Replacing the new coordinates
in (13) leads to:

V̇ = 2ξ T (I N ⊗ P1 A + J ⊗ P1 B K )ξ − 2ξ T (J ⊗ P1 B K ) + 2 T (I N ⊗ P2 (A − LC))
+ 2ξ T (I N ⊗ P1 Du )ωu + 2 T (I N ⊗ P2 Du )ωu − 2 T (I N ⊗ P2 L D y )ω y . (14)

Using Lemma 1, (14) can be rewritten as follows:

N    
V̇ = ξ jT H e P1 A + λ j P1 B K ξ j − 2 Nj=1 ξ jT λ j P1 B K j + Nj=1
j=1
T
j H e {P2 A − P2 LC} j
N     (15)
+ 2 j=1 ξ jT H e {P1 Du } ωu j + 2 Nj=1 Tj H e {P2 Du } ωu j + 2 Nj=1 T
j H e P2 L D y ω y j ,

then, considering (15) with H∞ criterion, ϕ j (t) = [ξ j (t)T , j (t)T , ωu j (t)T , ω Tyj ]T ,

and JT ≤ Nj=1 ϕ(t)Tj j ϕ(t) j , where j is defined as follows:
170 J. A. Vazquez Trejo et al.
⎡ ⎤
Q 1 j + I −λ j P1 B K P1 Du 0
⎢ ∗ Q 2 + I P2 Du −P2 L D y ⎥
=⎢
⎣ ∗
⎥ < 0,
⎦ (16)
j
∗ −γ 2 I 0
∗ ∗ ∗ −γ I
2

 
where Q 1 j = H e P1 A + λ j P1 B K , and Q 2 = H e {P2 A − P2 LC}. Then, (16) is
T
pre and post multiplied by P 1 , where P 1 = P 1 = P1−1 > 0
⎡ ⎤ ⎧⎡ ⎤ ⎫
ϒ1 j 0 Du 0 ⎪
⎪ −λ j B K ⎪
⎪
⎢ ∗ ⎨⎢ ⎬
⎢ ϒ2 P2 Du −P2 L D y ⎥
⎥ 0 ⎥
⎣ ∗ ⎦ + He ⎢⎣
⎥ 0 I 0 0 < 0,
⎦ (17)
∗ −γ 2 I 0 ⎪
⎪ 0 ⎪
⎪
⎩ ⎭
∗ ∗ ∗ −γ I2
0
 
where ϒ1 j = H e A P 1 + λ j B K P 1 + P 1 P 1 , and ϒ2 = H e {P2 A− P2 LC} + I .
Applying the Young relation [12], the following inequality is obtained:
⎡ ⎤ ⎡ ⎤
ϒ1 j
0 Du 0 −λ j B K
⎢ ∗ ϒ2 P2 Du −P2 L D y ⎥ ⎢ ⎥ 
⎢ ⎥ + μ ⎢ 0 ⎥ P −1 (−λ j B K )T 0 0 0
⎣ ∗ ∗ −γ 2 I 0 ⎦ ⎣ 0 ⎦ 1
(18)
∗ ∗ ∗ −γ 2 I 0
 −1 
+ μ−1 I 0 0 0 P 1 0 I 0 0 < 0,
T

where μ > 0. Using the Schur complement in (18), choosing Nc = K P 1 , and Mo =

P2 L, the proof is complete.

4 Simulation Results: Application to a Team of UAVs

In this section, the simulation results for an application to leader-following robust

consensus control of a team of homogeneous UAVs are presented. The dynamic
model of each UAV is:
 
1 Jy − Jz 1
˙
ẍi = (cψi sθi cφi + sψi sφi ) Ti , φ̈i = θi ψ̇i + Ri ,
mi Jx Jx
 
1 Jz − Jx 1
ÿi = (sψi sθi cφi − cψi sφi ) Ti , θ¨i = φ̇i ψ̇i + Pi , (19)
mi Jy Jy
 
1 Jx − Jy 1
z̈ i = −g + (cθi cφi ) Ti , ψ̈i = θ˙i φ̇i + Yi ,
mi Jz Jz

where the shorthand notation c{·} and s{·} is cos(·) and sin(·) respectively; xi , yi , and
z i are the position of the ith UAV in the euclidean space, φi , θi , and ψi , are the Euler
angles roll, pitch, and yaw respectively, Ri , Pi , and Yi are the rotor torques, m i is the
Distributed Observer-Based Leader-Following Consensus … 171

total mass of the UAVs, g is the gravitational acceleration, Jx , Jy , and Jz represents

the moments of inertia in their respectively axis, Ti is the thrust of the UAVs rotors.
The parameters used in the simulations were taken from [13].
The main objective of the multi-agent system, where each UAV is associated to
one agent, is to follow the trajectories of a virtual leader and maintain a desired shape,
where H = [h 1T , h 2T , . . . , h TN ]T contains the desired distance column vectors h i of
every agent. To solve the formation control problem, a second order representation
of each agent [14] is used to find a solution for the LMIs presented in Theorem 1:

ṗi (t) = vi (t), v̇i (t) = u i (t). (20)

However, it is noted that for the simulations the complete nonlinear model (19) is
used. Considering pi (t), vi (t), u i (t) the position, velocity and acceleration respec-
tively of agent i, s̄ˆi (t) = [ p̂i (t)T − h iT , v̂i (t)T ]T the estimated states of agent i,
s̄l (t) = [ pl (t)T , vl (t)T ]T the virtual leader states, h i is a vector of matrix H of the
respective agent, then, the control law (6) is rewritten as:

u i (t) = K ai j (s̄ˆi (t) − s̄ˆ j (t)) + αi K (s̄ˆi (t) − s̄l (t)), (21)
j∈Ni

where K is the control gain to be designed. Considering δ̄ˆi (t) = s̄ˆi (t) − s̄l (t) Eq.
(21) is rewritten as follows:
 ˆ ˆ ˆ
u i (t) = K ai j (δ̄i (t) − δ̄ j (t)) + αi K δ̄i (t). (22)
j∈Ni

Since the dynamic of each UAV is considered as a particle, the desired Euler angles
θdi , φdi , and ψdi are computed as in [15] to control the position of the UAVs, where
the consensus protocol (22) and the desired Euler angles are related as follows:
⎛ ⎞
 
−u 2i u 1i
φdi = arctan ⎝  ⎠ , θdi = arctan , ψdi = 0, (23)
u 21i + (u 3i + g)2 u 3i + g

and the thrust of the UAVs rotors is computed by Ti = m i (u 21i + u 22i + (u 3i + g)2 )1/2 .
The communication topology of the multi-agent system is depicted in Fig. 1 (where a
UAV represents an agent and the arrows the flow of information), and the Laplacian
matrix L:
⎡ ⎤
5 −1 −1 −1 −1 −1
⎢−1 2 −1 0 0 0 ⎥
⎢ ⎥
⎢−1 −1 2 0 0 0 ⎥
L=⎢ ⎢ ⎥ (24)
⎥
⎢−1 0 0 2 −1 0 ⎥
⎣−1 0 0 −1 3 −1⎦
−1 0 0 0 −1 2
172 J. A. Vazquez Trejo et al.

Fig. 1 Directions of the

communication links
between agents

H and  matrices are the following:

⎡ ⎤
1 0 0 0 0 0
⎡ ⎤ ⎢0 0 0 0 0 0⎥
04 √6 4 0
√ √ √ −2 ⎢ ⎥
⎢0 0 0 0 0 0⎥
H = ⎣0 0 2 3 4 3 4 3 2 3 ⎦ ,  = ⎢ ⎥. (25)
⎢0 0 0 0 0 0⎥
00 0 0 0 0 ⎣0 0 0 0 0 0⎦
0 0 0 0 0 0
The first simulation result in Fig. 2, shows that the formation control problem is solved
and the multi-agent system follows the trajectories of the virtual leader. All the UAVs
reach the trajectories of the virtual leader while maintaining the rigid formation of
a hexagon. This first result does not include external disturbance or measurement
sensor noise.
The considered external disturbance is modeled as the Dryden wind turbulence
presented in [16], and the sensor noise was added as a random function with normal
distribution, mean value of zero, and standard deviation of 0.8 ms , in order to verify
the robustness of the control law (22). If the measurement noise or the external
disturbance magnitudes increase, the control law (22) is not able to maintain the
desired formation or to follow the virtual leader. To overcome the disturbance and
noise problem, the LMI presented in (11) is computed with the Sedumi Toolbox
[17], where the LMI variable γ value obtained with μ = 0.1 is γ = 1. The computed
control and observer gains are:

Fig. 2 Formation without disturbance or sensor noise and distances between followers
Distributed Observer-Based Leader-Following Consensus … 173

Fig. 3 Formation with disturbance and sensor noise using the control law (22) and observer (4)
with gains K , and L (26) computed using Theorem 1

⎡ ⎤
3.208 0 0
⎡ ⎤
⎢ 0 3.208 0 ⎥
−0.756 0 0 −2.651 0 0 ⎢ ⎥
⎢ 0 0 3.208⎥
K =⎣ 0 −0.756 0 0 −2.651 0 ⎦, L = ⎢ ⎥ . (26)
⎢0.836 0 0 ⎥
0 0 −0.756 0 0 −2.651 ⎣ 0 0.836 0 ⎦
0 0 0.836

The robustness of the distributed control and estimation implemented in the team of
UAVs has been verified by solving the leader-follower formation control problem.
The effectiveness of the proposed work is shown in Fig. 3, where each UAV achieves
its desired position in the rigid formation of a hexagon while following the trajectory
of the virtual leader because the external disturbance and the measurement sensor
noise are rejected.

5 Conclusions

A distributed robust leader-following consensus control observer-based for LTI

multi-agent systems was presented. The objective of the multi-agent system is to
follow the trajectories described by a virtual leader’s dynamics and to maintain a
desired rigid formation between followers. This was exemplified with the simulation
of a team of UAVs where the agents were able to reject external disturbance and
measurement sensor noise. As further work, this approach will be extended to LPV
multi-agent systems.

References

1. Skorobogatov, G., Barrado, C., Salamí, E.: Multiple UAV systems: a survey. Unmanned Syst.
8(02), 149–169 (2020)
2. Li, Z., Duan, Z.: Cooperative control of multi-agent systems: a consensus region approach.
CRC Press (2017)
174 J. A. Vazquez Trejo et al.

3. Trejo, J.A.V., Adam-Medina, M., Garcia-Beltran, C.D., Ramírez, G.V.G., Yolanda lópez Zap-
ata, B., Sanchez-Coronado, E.-M., Theilliol, D.: Robust formation control based on leader-
following consensus in multi-agent systems with faults in the information exchange: application
in a fleet of unmanned aerial vehicles. IEEE Access 9, 104940–104949 (2021)
4. Tian, X., Liu, H., Liu, H.: Robust finite-time consensus control for multi-agent systems with
disturbances and unknown velocities. ISA Trans. 80, 73–80 (2018)
5. Redrovan, D.V., Kim, D.: Multiple quadrotors flight formation control based on sliding mode
control and trajectory tracking. In: 2018 International Conference on Electronics, Information,
and Communication (ICEIC), pp. 1–6. IEEE (2018)
6. Sun, Z., Guoqing Zhang, Y.L., Zhang, W.: Leader-follower formation control of underactuated
surface vehicles based on sliding mode control and parameter estimation. ISA Trans. 72, 15–24
(2018)
7. Ahmed, Z., Khan, M.M., Saeed, M.A., Zhang, W.: Consensus control of multi-agent systems
with input and communication delay: a frequency domain perspective. ISA Trans. 101, 69–77
(2020)
8. Lin, W., Zhao, W., Liu, H.: Robust optimal formation control of heterogeneous multi-agent
system via reinforcement learning. IEEE Access 8, 218424–218432 (2020)
9. Haichuan, X., Housheng, S., Wang, Q., Chengjie, X.: Semi-global adaptive bipartite output
consensus of multi-agent systems subject to input saturation and external disturbance under
switching network. Int. J. Control Autom. Syst. 19(9), 3037–3048 (2021)
10. Ni, W., Cheng, D.: Leader-following consensus of multi-agent systems under fixed and switch-
ing topologies. Syst. Control Lett. 59(3–4), 209–217 (2010)
11. Chen, J., Zhang, W., Cao, Y.-Y., Chu, H.: Observer-based consensus control against actuator
faults for linear parameter-varying multiagent systems. IEEE Trans. Syst., Man, Cybern.: Syst.
47(7), 1336–1347 (2016)
12. Boyd, S., El Ghaoui, L., Feron, E., Balakrishnan, V.: Linear matrix inequalities in system and
control theory. SIAM (1994)
13. Luis, C., Le Ny, J.: Design of a trajectory tracking controller for a nano quadcopter. Poly-
technique Montreal, Technical report, Mobile Robotics and Autonomous Systems Laboratory
(2016)
14. Yang, H., Zhang, Z., Zhang, S.: Consensus of second-order multi-agent systems with exogenous
disturbances. Int. J. Robust Nonlinear Control 21(9), 945–956 (2011)
15. Guerrero-Castellanos, J.-F., Vega-Alonzo, A., Marchand, N., Durand, S., Linares-Flores, J.,
Mino-Aguilar, G.: Real-time event-based formation control of a group of VTOL-UAVs. In: 2017
3rd International Conference on Event-Based Control, Communication and Signal Processing
(EBCCSP), pp. 1–8. IEEE (2017)
16. Rodríguez-Mata, A.E., Flores, G., Martínez-Vásquez, A.H., Mora-Felix, Z.D., Castro-Linares,
R., Amabilis-Sosa, L.E.: Discontinuous high-gain observer in a robust control UAV quadrotor:
real-time application for watershed monitoring. Math. Probl. Eng. (2018)
17. Sturm, J.F.: Using SeDuMi 1.02, a MATLAB toolbox for optimization over symmetric cones.
Optim. Methods Softw. 11(1–4), 625–653 (1999)
Nonlinear Systems, Localization and FDI
Contact-Less Sensing and Fault
Detection/Localization in Long Flexible
Cantilever Beams via Magnetoelastic
Film Integration and AR Model-Based
Methodology

Robert-Gabriel Sultana and Dimitrios Dimogianopoulos

Abstract The integration of magnetoelastic film into a long, flexible polymer beam
for obtaining sensing and fault diagnosis capabilities is investigated. Metglas®
2826 MB film is attached onto the clamped end of the beam (itself fixed as cantilever),
with its free end connected to a mini-exciter. Vibrating the structure causes emis-
sion of variable magnetic flux by the integrated magnetoelastic film, intercepted in a
contact-less manner by a low-cost reception coil suspended above the film. The result-
ing voltage is linked to the beam’s vibratory motion and is recordable via conventional
oscilloscopes without sophisticated/dedicated circuitry (e.g. power-amplifiers), mak-
ing for a cost-efficient and low-complexity setup. Hence, the long, flexible beam may
act as a sensing device with contact-less data transmission. The innovation is that
different structural faults (represented by loads of various magnitudes and locations
on the beam) are shown to affect accordingly the recorded signal’s spectral char-
acteristics. Furthermore, fault-induced frequency shifts and changed damping can
be reliably estimated and traced back to specific faults via the currently proposed
stochastic model-based algorithmic methodology. Tests involving two different loads
applied on three distinct positions on the beam successfully conclude on the proposed
setup’s potential for vibration sensing and detection/localization of different faults
(loads) on the beam.

1 Introduction

The need for wireless monitoring of hazardous substances in often hostile environ-
ments is one of the main reasons that magnetoelastic sensors marked a net progress
over the past years. The concentration of dangerous chemical, environmental and bio-

R.-G. Sultana · D. Dimogianopoulos (B)

Department of Industrial Design and Production Engineering, University of West Attica,
Campus of Ancient Olive Grove, 250 Thivon Av., 12241 Athens, Greece
e-mail: [email protected]
R.-G. Sultana
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 177
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_16
178 R.-G. Sultana and D. Dimogianopoulos

logical factors could be remotely monitored without human intervention by means

of magnetoelastic (magnetostrictive) sensors. These are made of (ferromagnetic)
materials which alter their shape when subjected to external magnetic fields, or
emit magnetic flux in accordance with externally imposed loading [1–3]. Variable
changes of shape of a suitably fixed (for instance, clamped) magnetoelastic element
(or resonator) will set it to vibrate. Vibratory dynamics change when biological [4]
or environmentally polluting [5] substances accumulate onto the resonator’s surface,
due to the resulting change in its mass distribution. Again, a suitably coated resonator
will exhibit changes in its mass distribution due to the coating’s reaction with volatile
organic compounds [6], or its adsorption of H2 O [7] or H2 O2 [8]. Observing changes
in the resonator’s vibratory dynamics (shifted resonant frequencies) directly points
to increased concentration of substances on it and, hence, the environment.
The majority of these resonators are set into vibration via an interrogation coil
which, under suitable electrical excitation, produces variable magnetic flux towards
the magnetoelastic material (usually in form of a ribbon or thin film), thus driving it
to resonance. A reception coil is placed above the vibrating resonator and transforms
the flux emitted into electrical signal in a contact-less manner. This signal is, then,
monitored for frequency shifts enabling detection of the substance under considera-
tion. The operational concept of these applications, based on the use of two coils and
referred to as the active setup, is comprehensively reviewed in [9]. Obviously, effort
has been invested in optimizing the resonator’s sensitivity to the mass accumulated
on its surface by analyzing relevant characteristics (for instance, length-to-width
ratio) of the resonator’s initial shape [10], or even suggesting particularly shaped (as
hourglass [11], or rhomboid [12, 13]) resonators.
Another class of resonators of particular interest (also reviewed in [9]) involves
those that are parts of a mechanical system (structure, machinery or other) and receive
vibration from the system’s operation or an external mechanical excitation, rather
than an external varying magnetic field [14–18]. These are referred to as passive
setups, with the vibrating magnetoelastic part producing variable magnetic flux (of
lower intensity compared with active setups) in accordance with the external vibra-
tion. Again, voltage is induced in a reception coil in a contact-less manner, with
its frequency analysis allowing for contact-less sensing/measuring the dynamics of
the magnetoelastic part and, hence, of the system. Interestingly, any faults/failures
affecting the system’s vibration dynamics (resonant shifts, for instance), will show
up in the recorded voltage because, as demonstrated in [14, 16, 17], specific reso-
nants of the system dynamics (as estimated via Finite Element Analysis) are present
in the voltage. Based on this principle, fault diagnosis results have been achieved
in polymer slabs with built-in magnetoelastic ribbons via 3D printing [14], or joints
of structures assembled with such slabs [15]. Again, diagnosis of cracks has been
achieved for cantilever metal beams with magnetoelastic ribbon attached onto their
surface [16] or rotating metal beams [18]. The possibility of using such passive setups
as sensing devices has been demonstrated for short, thin polymer beams (fixed as
cantilevers) in [19], and for metal beams also fixed as cantilevers in [17].
This work aims at demonstrating the feasibility of a two-fold extension of pre-
vious fault detection and severity estimation results [14, 15] obtained for single
Contact-Less Sensing and Fault Detection/Localization … 179

(or structures assembled by) slabs with magnetoelastic elements: First, by addition-
ally achieving localization of faults of different magnitudes; and second by using
long, highly flexible beams as structures upon which the magnetoelastic elements
are attached, instead of being 3D-printed into relatively short rigid slabs (or slab
assemblies) [14, 15]. Such flexible beams may also serve as sensing elements when
connected to existing (metallic or polymer) vibrating structures, rather than being
used as load-bearing structural parts (i.e. slabs). Faults are simulated as loads placed
on different positions on the beam surface. Two different loads have been tested, with
each one alternatively placed at one of three available positions. The beam retains
the cantilever arrangement, but unlike [16–18], the magnetoelastic film attachment
position is at the clamped end, with excitation provided via a mini-exciter at its free
end. As in [19], voltage is induced in a contact-less manner in the reception coil and is
recorded via a conventional oscilloscope. The additional absence of excitation coils
(as found in traditional applications [2, 3]) means that complexity and costs are both
minimal. For enhanced cost-effectiveness (and unlike [16–18]), effort is invested in
optimizing the algorithmic framework used for fault diagnosis, rather than optimiz-
ing the hardware (coils, associated circuitry, amplifiers) itself. Hence, faults (loads)
are detected-localized by using stochastic model-based procedures optimized for
fault detectability via a novel criterion. Even though the excitation is always at the
free-end, structural changes (loads) of different intensity and positions on the beam
surface are detectable and localizable, meaning that no special excitation position is
necessary for obtaining global fault diagnosis results. Hence, long and highly flex-
ible beam arrangements suitably receiving excitation from vibrating (mechanical)
systems may, indeed, be used for sensing and diagnostic purposes.

2 Materials and Methods

A 425 mm long, thin and highly flexible beam is formed by tandem connection of
two shorter beams with glue. The beams are 3D-printed in Fused Deposition Mod-
eling (FDM) mode with PET-G filament. The final beam is clamped on one end,
with the opposite end fixed onto the vibrating rod of a mini-exciter. The Metglas®
2826 MB magnetoelastic film is attached onto the beam surface near the clamp with
cyanoacrylate glue. A low-cost Vishay IWAS coil, normally used for wireless charg-
ing, is placed 5 mm above the film and receives the emitted magnetic flux (Fig. 1).
The SMARTSHAKERTM K2004E01 mini-exciter vibrates the beam as defined by
the external waveform generator (SIGLENT, SDG 5122). The reception coil is con-
nected to a conventional oscilloscope, used for voltage data acquisition. The latter is
the only signal used for detecting and localizing faults. This output-only approach is
realistic because in similar real-life applications (bridges, or highly flexible structures
of similar shape), the excitation is often unavailable or hard to measure. Moreover,
such flexible beam setups may also serve as low-cost sensing elements connected to
mechanical systems and driven by their vibrations (see Sect. 3.1).
180 R.-G. Sultana and D. Dimogianopoulos

Fig. 1 Experimental setup with a load (coin) at B, and the contact-less sensing concept (zoom)

Table 1 Test scenarios (fault cases), their designations and positions

Test scenario Load object Mass (g) Position (from free end)
(fault case)
N-1C No load (healthy) 0 n/a
A-1C 1 Euro cent 2.3 A (at 35 mm—nearby excitation
point)
B-1C 1 Euro cent 2.3 B (at 185 mm—middle)
C-1C 1 Euro cent 2.3 C (at 360 mm—nearby clamp)
N-BN No load (healthy) 0 n/a
A-BN Bolt+nuts 6 A (at 35 mm—nearby excitation
point)
B-BN Bolt+nuts 6 B (at 185 mm—middle)
C-BN Bolt+nuts 6 C (at 360 mm—nearby clamp)

Faults are simulated as loads at three key positions on the beam (Fig. 1 and Table 1),
namely near the excitation (A), in the middle (B) and near the clamp (C). The testing
scenarios (Table 1) involve cases without load (prefix N) or with a load at position A,
B or C. Suffix -1C indicates a “small” load of 1 Euro-cent coin, which is a compact
unit of precise weight easily fixable at A, B or C (see prefix) with tape. Suffix -BN
indicates a “bigger” load of a bolt fixed with nuts, again at the designated (by prefix)
position. Bolts are 2.6 times heavier than 1 Euro-cent coins and easier to fix than the
20 Euro-cent coins of equivalent weight, which are too thick for being fixed with
tape on the beam. The aim is to investigate the feasibility of distinguishing faults of
different magnitudes and at various locations of a vibrating structure (for instance, a
flexible bridge), with magnetoelastic material in only one location.
Contact-Less Sensing and Fault Detection/Localization … 181

2.1 Feasibility of Detecting and Localizing Different Faults

As concluded in [14, 15, 19], in passive setups the time history of recorded voltage
signals rarely exhibit visible patterns related to faults affecting the (data-generating)
associated system. But frequency domain characteristics (natural frequencies ωn
and damping ζ ) do change with fault occurrence, since they are physically linked
to (and capable of quantifying) fault-induced alterations of the system’s structure.
Thus, the recorded voltage from a vibrating system is modeled via discrete-time
stochastic output-only AutoRegressive (AR) representations, capable (among others)
of accurately estimating values for ωn and ζ , associated to the system’s current health
state (Table 1). Model-free (see Sect. 3.2) or nonlinear (e.g. Neural Network) fault
diagnosis approaches, although highly effective, cannot easily provide results for
both ωn and ζ . In an AR modeled signal, the current value of a process depends on
a set of past values up to n previous lags and a series of random1 values e[t]:

y[t] = a1 y[t − 1] + a2 y[t − 2] + · · · + an y[t − n] + e[t] (1)

with n the AR model order and e[t] a sequence of independent (i.e. uncorrelated)
and identically distributed random values (white noise). Estimating n along with
parameters a1 , a2 , . . . , an (via Least-Squares criteria) is a task referred to as model
identification. The feasibility of detecting and localizing system faults of different
magnitudes is demonstrated by comparing characteristics of such AR models, each
identified on voltage data from a case in Table 1. The steps for demonstrating the
feasibility of detecting and localizing faults are as follows:
1. A recorded voltage dataset from a system in some state (see Table 1) is filtered
and subsampled, as shown in Sect. 3.2.
2. Stochastic AR time-series representations are identified on the processed dataset,
thus capturing properties of the signal (and the underlying system) dynamics.
3. Natural frequencies ωn and damping ratios ζ are computed for specific frequency
ranges (see Sect. 3.2), and the associated poles are plotted on the s-plane.
The procedure is executed for all available voltage sets. Fault detectability is achieved
when pole groups from AR models identified on N-1C data are distinguishable from
pole groups corresponding to A-1C, B-1C and C-1C data, inside a given frequency
range. Feasibility of localizing faults is demonstrated when pole groups from models
identified on A-1C, B-1C and C-1C data are mutually distinguishable in the given
frequency range. The same reasoning applies to cases with -BN suffix. Finally, if
pole groups from -1C cases may be distinguished from their counterparts resulting
from -BN cases, then faults of different magnitudes are detectable and localizable.

1x[t] denotes values of x in discrete time with t being the normalized discrete time, t × t the
absolute time and t the sampling period; x(t) denotes values of x at continuous time instant t.
182 R.-G. Sultana and D. Dimogianopoulos

2.1.1 Optimizing Model Order for Enhanced Fault Detectability

As stated in Sect. 2.1, faults are detectable by means of delimiting suitable areas
on the complex plane, so that poles from AR models identified on healthy data may
only be found inside these areas. Then, enhancing fault detectability means that these
areas must be as small/compact as possible. This task may be carried out by suitably
optimizing the AR model identification process as follows.
Let an AR model (1) of order n be available for a dataset, with residuals r [t]:

r [t] = y[t] − ŷ[t] (2)

where ŷ[t] is the sequence of predicted output values from the given AR model
and y[t] the measured output ∀ t of the experiment. The AR model of order n is
validated if the sequence r [t] is uncorrelated, as statistically checked via the sample
autocorrelation function. With larger values for n yielding less correlated sequences
r [t], there exist several criteria (Akaike or Bayesian Information Criterion- AIC
or BIC, respectively [20]) which propose an optimal value for n, so that r [t] is
quite uncorrelated and the AR model does not become too long. In the current fault
detection context, however, once a value for n leading to relatively uncorrelated
sequence r [t] is found, one has to make sure that using that n for identifying AR
models for all healthy system datasets leads to all resulting poles (for the frequency
range of interest) lying inside the smallest possible area in the complex plane. This,
in turn, means that the risk of this area potentially including poles from AR models
identified on faulty system data is minimized.
Consider a set of poles from AR models of a specific order n (identified on a given
number of healthy datasets) for the frequency range of interest on the complex-plane.
Let the quantity rsum |n be calculated by summing the absolute distance of every pole
from the barycenter of the pole set in the specific frequency range


p
rsum |n = ri (3)
i=1

where p is the number of datasets (or poles in the selected frequency range) and ri is
the L2-norm (absolute distance) of the i-th pole from the barycenter. Then, one has
to simply minimize rsum |n with respect to n, starting with an initial value for n equal
to the smallest AR order that yields uncorrelated r [t].

3 Results and Discussion

The current section features the application of the fault detection and localization
scheme presented in Sect. 2.1 on data from vibration testing experiments with the
setup in Sect. 2. Two issues are addressed: First, the setup sensing abilities are opti-
Contact-Less Sensing and Fault Detection/Localization … 183

mized with respect to noise (induced by contact-less data transmission), by suitably

defining the coil height (Sect. 3.1). Then, fault detection and localization results from
the optimal setup are presented and commented upon (Sect. 3.2).

3.1 Sensing Properties of the Passive Setup

The setup’s sensing properties are assessed and optimized via two testing scenarios:
The first one considers signals recorded with the beam at rest, which essentially
involve the underlying noise. The second scenario involves data from a vibrating
beam with the coil placed at either 20 or 5 mm above the film on the beam surface.
A triangular force 160 Hz (supplied by the mini-exciter) sets the beam into a low
amplitude (invisible) vibration. Tests at frequencies as low 30 Hz revealed that the
power line value (50 Hz) was always present in data, even for frequencies under 50 Hz.
Data transmission notably improved for frequencies over 80–100 Hz, as had also
been noted for passive setups in [15]. Obviously, multiple excitation locations could
probably improve data transmission, but at increased setup complexity. Triangular
waveforms were preferred to sinusoids or pulses, since they resemble more to real-life
vibration sources due, for instance, to machinery reciprocating parts [15].
For both scenarios, three sets of tests were executed. Figure 2a shows time histories
of signals with and without excitation (noise), with both signals being, as expected,
almost identical (see Sect. 2.1). Figure 2b shows the Fast Fourier Transform (FFT)
plot of the signals in Fig. 2a, with the reception coil placed 20 mm above the beam.
Focusing on the highest peaks around 1300 Hz, signals from vibrating beam are
clearly different from simple noise, meaning that sensing is indeed feasible. Similar
tests with the coil at 5 mm above the beam in Fig. 2c, further highlight the difference
between noise and signals from the vibrating beam. Moreover, the latter exhibit peaks
around 1350 Hz which are more scattered in Fig. 2b than their counterparts in Fig. 2c.

Fig. 2 a Time history of signals with and without excitation showing no visible differences b
Frequency content of signals with and without excitation showing small frequency shifts- coil at
20 mm above the beam. c Frequency content as in b but with coil at 5 mm above the beam
184 R.-G. Sultana and D. Dimogianopoulos

Thus, sensing results are more consistent when the coil is located as close to the beam
surface as possible. With suitable further development, the passive setup could be a
cost-effective alternative to pricey sensors (for instance, load cells) used in devices
for recording and frequency analysis of (force) signals. A relevant example (that
initiated this study) is the Fish Texture Evaluation Tool developed in work package
6 of “FutureEUAqua” project [21].

3.2 Fault Detection and Localization Results

Since the setup sensing properties are validated, testing with fault cases in Table 1
may be carried out. Three tests (each covering two seconds of the beam’s vibration)
are executed per case, making for twenty-four datasets. Figure 3a shows the signal
frequencies from N-1C and B-1C cases, with clear shifting due to load (B-1C) around
1300 Hz. Error bar plots for frequency peaks from all tests in Fig. 3b indicate a clear
distinction of fault classes. Signals obtained from passive setups in a contact-less
manner are, hence, suitable for detecting loads (faults) and denoting their locations.
Even better, the 3-step algorithm in Sect. 2.1 allows for estimating both frequen-
cies ωn and damping ratios ζ for fault (load) detection/ localization purposes. The
dataset of each experiment is split in half, resulting in six (smaller) sets per case
(see Table 1). Hence, transient non-stationary data points are better accounted for
during modeling, since they now spread across smaller datasets. Datasets are mildly
filtered via 5th order Butterworth filters (with –3 dB cutoff at 3 KHz) and subsampled
at 20 kHz (from 1 MHz). The optimal AR model order is estimated (second step in
Sect. 2.1.1) along with the resulting discrete-time poles. Figure 4a shows that rsum |n
is minimal for AR orders of 94 and 75 for the N-1C and N-BN cases, respectively.
Thus AR(94) and AR(75) models are identified on data from cases with suffix -1C
and -BN, respectively. Note that uncorrelated residuals are already obtainable for
n > 45 for all cases. Pole plots on the z-plane for -1C and -BN cases are presented
in Fig. 4b and c, respectively. Around 1300 Hz, poles from N-1C and N-BN cases do

Fig. 3 a Frequency content of signals from N-1C versus B-1C tests, showing shifts due to fault
(load). b Error bars of frequency peaks showing the same shift pattern for -1C and -BN cases
Contact-Less Sensing and Fault Detection/Localization … 185

Fig. 4 a Values of rsum |n for increasing order n, with optimal values pointed out. b Poles corre-
sponding to -1C cases and c poles corresponding to -BN cases on the z-plane

Fig. 5 AR poles corresponding to a -1C cases, and b -BN cases at 1300 Hz on the s-plane, with
horizontal lines indicating that faults (except for those in circles) are correctly localized. c Pole
areas for -1C and -BN cases showing that localized -1C faults cannot be mistaken for -BN ones

not mix with poles from other -1C/-BN cases, thus ensuring fault (load) detectability.
Fault localization is harder, since poles corresponding to data from A-, B- and C-
cases seemingly share the same z-plane area. Nonetheless, converting discrete-time
poles to their continuous time counterparts and plotting them in Fig. 5, shows one
delimited s-plane area for each case, hence ensuring feasibility of fault localization.
More importantly, localized -1C faults cannot be mistaken for -BN ones, since their
respective s-plane areas do not overlap in Fig. 5c. Thus, the current passive setup
may achieve detection and localization of different faults (loads) using magnetoelas-
tic film in only one position of the beam.

4 Conclusions

The feasibility of inducing sensing and fault diagnosis properties to a long flexible
cantilever beam by attaching magnetoelastic film on its clamped side was studied.
The setup is passive (i.e. without excitation coils), driven from beam vibrations.
186 R.-G. Sultana and D. Dimogianopoulos

The reception coil was optimally placed for contact-less sensing of data produced
by beam vibrations and an algorithm for optimal fault detectability was postulated.
Testing showed that all faults (loads on the beam) were detected, and almost all
of them localized. Most importantly, localized faults of given magnitudes cannot
be mistaken for faults of other magnitudes. Results are delivered at minimal cost
and complexity, since no special equipment (accelerometers, signal conditioners) is
required and the beam is excited at one position for all possible fault locations. Future
work will focus on optimizing vibration sensing and fault identifiability by applying
excitation to different locations, and assessing performance by extended testing.

Acknowledgements Work supported in part by the project FutureEUAqua-National Matching

2019 Funding. The authors wish to thank Prof. D. Mouzakis for providing the Metglas® film.

References

1. Le Bras, Y., Greneche, J.M.: Magneto-elastic resonance: principles, modeling and applications.
In: Awrejcewicz, J. (ed.) Resonance, Chap. 2, IntechOpen (2017). https://doi.org/10.5772/
intechopen.70523
2. Grimes, C.A., et al.: Theory, instrumentation and applications of magnetoelastic resonance
sensors: a review. Sensors 11(3), 2809–2844 (2011). https://doi.org/10.3390/s110302809
3. Gonzalez, J.M.: Magnetoelasticity and magnetostriction for implementing biomedical sensors.
In: López-Dolado, E., Serrano, M.C. (eds.), Engineering Biomaterials for Neural Applications,
pp. 127–147. Springer, Cham (2022). https://doi.org/10.1007/978-3-030-81400-7_6
4. Ren, L., Yu, K., Tan, Y.: Applications and advances of magnetoelastic sensors in biomedical
engineering: a review. Materials 12(7), 1135 (2019). https://doi.org/10.3390/ma12071135
5. Grimes, C.A., et al.: Magnetoelastic microsensors for envinronmental monitoring. In: 14th
IEEE International Conference Micro Electro Mechanical Systems (MEMS), pp. 278–281.
Interlaken, Switzerland (2001). https://doi.org/10.1109/MEMSYS.2001.906532
6. Baimpos, T., et al.: A polymer-Metglas sensor used to detect volatile organic compounds.
Sens. Actuators A: Phys. 158(2), 249–253 (2010). https://doi.org/10.1016/j.sna.2010.01.020
7. Atalay, S., et al.: Magnetoelastic humidity sensors with TiO2 nanotube sensing layers. Sensors
20, 425 (2020). https://doi.org/10.3390/s20020425
8. Samourgkanidis, G., et al.: Hemin-modified SnO2 /Metglas electrodes for the simultaneous
electrochemical and magnetoelastic sensing of H2 O2 . Coatings 8(8), 284 (2018). https://doi.
org/10.3390/coatings8080284
9. Dimogianopoulos, D.G.: Sensors and energy harvesters utilizing the magnetoelastic principle:
review of characteristic applications and patents. Recent Pat. Electr. Electron. Eng. 5(2), 103–
119 (2012). https://doi.org/10.2174/2213111611205020103
10. Skinner, W.S., et al.: Magnetoelastic sensor optimization for improving mass monitoring.
Sensors 22(3), 827 (2022). https://doi.org/10.3390/s22030827
11. Ren, L., Cong, M., Tan, Y.: An hourglass-shaped wireless and passive magnetoelastic sensor
with an improved frequency sensitivity for remote strain measurements. Sensors 20(2), 359
(2020). https://doi.org/10.3390/s20020359
12. Saiz, P.G., et al.: Enhanced mass sensitivity in novel magnetoelastic resonators geometries for
advanced detection systems. Sens. Actuators B: Chem. 296, 126612 (2019). https://doi.org/
10.1016/j.snb.2019.05.089
13. Saiz, P.G., et al.: Influence of the magnetic domain structure in the mass sensitivity of magne-
toelastic sensors with different geometries. J. Alloy. Compd. 863, 158555 (2021). https://doi.
org/10.1016/j.jallcom.2020.158555
Contact-Less Sensing and Fault Detection/Localization … 187

14. Dimogianopoulos, D.G., Charitidis, P.J., Mouzakis, D.E.: Inducing damage diagnosis capa-
bilities in carbon fiber reinforced polymer composites by magnetoelastic sensor integration
via 3D printing. Appl. Sci. 10(3), 1029 (2020). https://doi.org/10.3390/app10031029
15. Dimogianopoulos, D.G., Mouzakis, D.E.: Nondestructive contactless monitoring of damage
in joints between composite structural components incorporating sensing elements via 3D-
printing. Appl. Sci. 11(7), 3230 (2021). https://doi.org/10.3390/app11073230
16. Samourgkanidis, G., Kouzoudis, D.: A pattern matching identification method of cracks on
cantilever beams through their bending modes measured by magnetoelastic sensors. Theor.
Appl. Fract. Mech. 103, 102266 (2019). https://doi.org/10.1016/j.tafmec.2019.102266
17. Samourgkanidis, G., Kouzoudis, D.: Characterization of magnetoelastic ribbons as vibration
sensors based on the measured natural frequencies of a cantilever beam. Sens. Actuators A:
Phys. 301, 111711 (2020). https://doi.org/10.1016/j.sna.2019.111711
18. Samourgkanidis, G., Kouzoudis, D.: Magnetoelastic ribbons as vibration sensors for real-time
health monitoring of rotating metal beams. Sensors 21(23), 8122 (2021). https://doi.org/10.
3390/s21238122
19. Dimogianopoulos, D.G., Mouzakis, D.E.: A versatile interrogation-free magnetoelastic res-
onator design for detecting deterioration-inducing agents. Lecture Notes in Civil Engineering,
vol. 110. Springer, Singapore (2020). https://doi.org/10.1007/978-981-15-9121-1_9
20. Ljung, L.: System Identification: Theory for the User, 2nd edn. Prentice Hall PTR, Upper
Saddle River, NJ (1999)
21. FutureEUAqua project (WP6), Horizon2020 grant 817737. https://futureeuaqua.eu/
A Comparative Simulation Study
of Localization Error Models
for Autonomous Navigation

Anis Koliai, Stephane Bazeille, Michel Basset, and Rodolfo Orjuela

Abstract The navigation and guidance of autonomous vehicles require precise and
accurate localization. However, in some situations the use of dead reckoning local-
ization is unavoidable, despite the fact that it introduces an accumulation of errors
known as drift. This paper focuses on a comparison in simulation between three dif-
ferent localization drift models based on position measurement, including a proposed
model. A guidance and control scheme is implemented, including the proposed drift
model. Finally, results from closed loop simulation are exploited to highlight the
discontinuity phenomenon caused by erroneous position measurements.

1 Introduction

Over the past few decades, the development of self-driving vehicles has proven to be
a challenging research topic for many laboratories and industries. Several researches
and applications have been done, however achieving full autonomy still remains a
significant task. Autonomous vehicles architecture is structured on three functional
levels: perception, planning and control [11]. The perception level aims to localize the
vehicle using various sensors as well as to collect information about the surrounding
environment (e.g. obstacles, traffic signals, etc.). From the perception information,
the planning system computes the desired reference trajectory. The last level is the

A. Koliai (B) · S. Bazeille · M. Basset · R. Orjuela

Institut de Recherche en Informatique, Mathématiques, Automatique et Signal (IRIMAS),
12 rue des Frères Lumière, 68 093 Mulhouse, France
e-mail: [email protected]
S. Bazeille
e-mail: [email protected]
M. Basset
e-mail: [email protected]
R. Orjuela
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 189
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_17
190 A. Koliai et al.

control, which produces suitable actuator signals (e.g. steering angle, throttle and
brake) ensuring that the planned actions are carried out safely.
Knowing the precise and accurate localization of the vehicle is critical for
safe autonomous navigation. The localization system must reach 95% of accu-
racy standards and maintain safety integrity in all situations where operations are
expected [13]. Several localization techniques are proposed in the literature. The
most well known methods are dependent on the infrastructure like global navigation
satellites system (GNSS). Most of the time, GNSS is widely used for precise vehicle
localization. Nevertheless, the availability and accuracy are not always guaranteed
due to the natural blockage of the signal or malicious spoofing and jamming. Another
method to localize the vehicle is to use onboard sensors (e.g. camera, lidar, inertial
measurement unit (IMU), wheel encoders, etc.). There are two cases to be considered,
when onboard sensors are used: with prior knowledge and without prior knowledge.
With prior knowledge, usually predefined maps are used and the methods use sensor
data to localize the data in the map. Without prior knowledge, the localization is
estimated via a method known as dead reckoning [4].
Each case has advantages and drawbacks. In this research project, we chose to
use the dead reckoning method because it does not need any knowledge on the
environment and it is easier to implement. This latter exhibits an unbounded drift
induced by the accumulating of small errors. The source of the errors differs with
the method used.
In this context, several studies are done to propose mathematical models describ-
ing the unbounded drift. Among them, Quinchia et al. [12] propose a model to com-
pensate IMU bias drift by considering as input the IMU acceleration and attitude
angles rate. Consequently, a mechanization stage must be added to obtain vehi-
cle position, as recently experimented in [14]. New neural network techniques are
also used to compensate bias-drift [3, 9] but a training step is unavoidable and the
drift model equations cannot be simply used. An empirical formula to model the
unbounded cumulative error is proposed in [15]. This model is based on the Ack-
ermann geometry and consider as input the side slip angle, the heading angle and
the velocity. A drift model related to visual odometry (VO) drift is developed by
Jiang et al. [8]. The model parameters are identified based on the camera estimated
position. In the same context using VO, a recent work is done in our lab to propose a
drift model which uses the GNSS positions to estimate the drifted position of the vehi-
cle [1]. More advanced techniques can be used to improve visual odometry estimation
[5, 6], but here the emphasis is on drift models.
There are many localization techniques integrated in the vehicle to localize it.
However, when the localization performances change over time, some undesirable
phenomena arise (e.g. position drift, discontinuity). As an example of localization
performances changes, let us consider that the localization uses GNSS fused with
IMU. When the GNSS is lost, the localization is only based on the IMU integration,
which degrades the localization accuracy.
One contribution of this paper is to highlight these kind of phenomena through
numerical simulations. For that purpose, two drift models ([8] and [1]) are compared.
Based on the obtained results, some improvements to the model [1] are proposed
A Comparative Simulation Study of Localization Error Models … 191

to reduce the added intrinsic disturbance during turn maneuvers. It should be noted
that the comparison is given by considering the vehicle in open loop. However, the
variation in the localization performances affect the closed loop performances. To
this aim, the closed loop behavior in the presence of erroneous localization is further
investigated through a simulation scenario. The contributions of this work can be
summarized as follows:
1. Proposition of improvements to the drift model developed in our lab [1].
2. Comparison of three different models: the model presented in [8], in [1] and the
proposed one.
3. Closed loop simulation of varying localization performances to highlight the effect
of discontinuity phenomenon.
The rest of the paper is organized as follows: In Sect. 2 the retained models are
exposed and compared by performing open loop simulations. The closed loop sce-
narios, with simulation and discussion, are presented in Sect. 3. Finally, Sect. 4
concludes the paper.

2 Localization Error Modeling

Many different drift models are proposed in the literature. Generally, a drift model
can be constructed by combining a set of equations and a type of noise. An important
aspect of the drift modeling concerns the inputs used. Some models use directly the
positions as input, while others use other quantities like velocity, acceleration, side
slip angle, etc.
In this paper, the choice is made to study the models with positional input.
Therefore, the input of all the models is the vehicle true pose PT (t) = [X T (t), YT (t),
T (t)]T , where X T (t), YT (t) are the coordinates of the vehicle in the global frame
and T (t) is the yaw angle. The model’s output pose PM (t) = [X M (t), Y M (t),
 M (t)]T is calculated such that it drifts over time.
In the following sections, three models are presented and compared. The first
model combines a white noise and a first-order Gauss-Markov (GM) process. The
second and third models rely on the vehicle displacement equations. For the sake of
comparison, the Gaussian noise is applied to the three models. The considered drift
models are schematized as in Fig. 1.

2.1 First Drift Model: Jiang et al. [8]

The authors in [8] have established a model to represent the drift behavior of visual
odometry position estimation. The model has a deterministic part and a stochastic
part. The model input is the camera estimated pose. In this work, the estimated pose
(drifted pose) is unknown and needed to be found, thus the model can not be used as
192 A. Koliai et al.

Position drift
models

Type
Structure
of noise

Proposed Second First Gaussian Brownian ...

No noise
model model [1] model [8] noise noise

Fig. 1 General scheme for the building of the drift model scheme. In gray, the type of noise used
in simulations

it is. Meanwhile, this model is modified slightly and used in the navigation toolbox
of Matlab® to illustrate an example of visual-inertial odometry (VIO) [10]. In the
example, just the zero mean stochastic part is used to model the uncertainty in the
scaling between consecutive frames of the monocular camera. The equations are
given as follows:
 
1
b(t) = 1 − b(t − 1) + Wa (t), (1a) drift(t) = b(t) + Wn (t), (1b)
τ
   
X M (t) X T (t)
= scale + drift(t), (1c)  M (t) = T (t), (1d)
Y M (t) YT (t)

 Wa (t) ∈ R is a white noise of the GM process with zero mean and σa =

2×1
where
σb τ + τ 2 . The process noise Wn (t) ∈ R2×1 is white noise with zero mean and
2 1

variance
√ σn . The model numerical parameters are given as follows: τ = 232, σb =
1.34, σn = 0.139, scale = 1.2. As shown in the equation (1c), the actual position
is used to estimate the drifted position which is ideal in our context. Note that the
yaw angle is not affected by the noise.
Remark: As in [10], the sampling frequency of the model 25 Hz. However, in the
simulations (see Sect. 2.4) the frequency is increased 100 Hz.

2.2 Second Drift Model: Bazeille et al. [1]

A simple VO localization model is proposed in [1]. This model is able to characterize

the erroneous localization given by the VO. This model is based on vehicle position
integration, where a reliable system is used to calculate the real current position
(X T (t), YT (t)). The displacement (x (t),  y (t)) between the current and previous
positions is then deduced, using [m(t), o(t)]T as model input.
The algorithm uses the last position added with the displacement m(t) to retrieve
the current position. However, the displacement is changed by adding noise. For the
A Comparative Simulation Study of Localization Error Models … 193

Fig. 2 Vehicle displacement

yaw angle also a noise is added. This noise is integrated while the vehicle is moving
and generates an unbounded drift. Variables nomenclature are summarized in Fig. 2
and model equations are presented below:

X M (t) = X M (t − 1) + (m(t) + K n m ) cos( M (t − 1)), (2a)

Y M (t) = Y M (t − 1) + (m(t) + K n m ) sin( M (t − 1)), (2b)
 M (t) =  M (t − 1) + o(t) + K n o , (2c)

with

x (t) = X T (t) − X T (t − 1), (2d) m(t) = x (t)2 +  y (t)2 , (2g)
 y (t) = YT (t) − YT (t − 1), (2e) o(t) = atan2(sin( (t)), cos( (t))), (2h)
 (t) = T (t) − T (t − 1), (2f)

where m(t) is the Euclidean distance between the poses PT (t) and PT (t − 1) in the
2D plan. However, o(t) is a recalculated  (t), for most of the cases o(t) =  (t).
The equation (2h) is added to avoid trigonometrical orientation problem when the
vehicle turns more than 360 ◦ C. The gain factor K adjusts noise amplitude. When
K = 0, the pose PM (t) does not accumulate errors. Finally, n m et n o are two random
variables where, n m stands for noise on displacement (movement), n o for noise on
orientation:

n m : α ∈   → n m (α) ∈ R, n o : α ∈   → n o (α) ∈ R, with  ∈ R. (2i)

Remark: The model parameters must be adapted according to the chosen frequency.
In the comparison Sect. 2.4 the frequency is fixed 100 Hz.
194 A. Koliai et al.

2.3 Third Drift Model: Proposed Model

In this paper, a third model is proposed based on the algorithm stated in Sect. 2.2. The
main change deals with the direction of movement angle employed in (2a) and (2b).
The angle T (t) is considered as the direction of movement and it is used to calculate
X M (t) and Y M (t). However, as shown in Fig. 2 the yaw angle T (t) is not always the
direction of movement (t) = θT (t). Therefore, small errors are induced without
the interference of any noise. These small errors are considered as model intrinsic
disturbances. For this reason, the angle θT (t) is introduced in the third model, and
consequently the model input is [θ (t), m(t), o(t)]T . The equations become as
follows:

θ M (t) = θ M (t − 1) + θ (t) + K n o , (3a)

X M (t) = X M (t − 1) + (m(t) + K n m ) cos(θ M (t)), (3b)
Y M (t) = Y M (t − 1) + (m(t) + K n m ) sin(θ M (t)), (3c)

with

θT (t) = atan2( y (t), x (t)), (3d) θ (t) = θT (t) − θT (t − 1). (3e)

It is worth to mention that the calculation of θT (t), given by (3d) is feasible, because
the PT (t) is known.
Remark: The model parameters must be adapted according to the chosen frequency.
In the comparison Sect. 2.4 the frequency is fixed 100 Hz.

2.4 Open Loop Comparison

Three open loop tests are performed in order to compare the outputs of the three drift
models presented in Sect. 2. In the first test no noise is taken into consideration. In
the second one, a gaussian noise is considered in the models. In the two tests, the
same steering angle δ f (t) = 5◦ and the same velocity (vx = 10 m/s) are transmitted
to the CarMaker® [7] (a realistic car simulator) vehicle model. The applied δ f (t)
provides a turning maneuver. In the third test, the same gaussian noise is applied
with a constant steering angle δ f (t) = 0 deg. The applied δ f (t) provides a straight
line maneuver.
In all simulations, the vehicle actual pose PT (t) is assumed well known and it is
directly given by the CarMaker vehicle dynamics equations. The pose PT (t) is then
employed as input signals for the three drift models previously presented.
 In all simulations, the Euclidean position error e p (t) =
(X T (t) − X M (t))2 + (YT (t) − Y M (t))2 and the yaw angle error eψ (t) =  M (t) −
T (t) are shown.
A Comparative Simulation Study of Localization Error Models … 195

Fig. 3 Vehicle position and drift models outputs (without added noise) in open loop at vx = 10 m/s

Remark: The sampling frequency is chosen 100 Hz which is suitable for a con-
trol loop. While increasing the frequency of the model 1 (25 Hz 100 Hz), the ampli-
tude of the added noise becomes higher than the displacement of √
the vehicle. For
this reason, the parameters have been modified as follows: σb = 1001.34
, σn = 0.139
10
.
First test: without adding noise. For the first model, the scale = 1 and dri f t (t) = 0
are considered. For the second and third models the gain K = 0 is considered.
The trajectories outputted by the drift models are depicted in Fig. 3. The Euclidean
position error e p (t) and the yaw angle error eψ (t) are also plotted. The noise is
normally set to zero for all models and the expected position error must be zero. This
last behavior is verified for model 1 and 3. However, the model 2 presents a non null
position error e p (t), that is due to the intrinsic disturbances added by the model.
Second test: with gaussian noise. Parameters of the first model are taken as described
in Sect. 2.1 with the added modifications. The second and third model parameters are
considered as follows: K = 0.1, n m ∼ N (μ1 , σ12 ) where σ1 = 0.14, μ1 = 0.025
and n o ∼ N (μ2 , σ22 ) where σ2 = 0.003, μ2 = − 0.001. The choice of the this
parameters is justified in [1].
The trajectories of the three models exhibit a position drift with respect to the
vehicle true position, as illustrated in Fig. 4. The position error e p (t) shows that this
drift increases over time. Note that the error e p (t) of model 1 is decreasing towards
the end. This is due to the fact that the trajectory is a circle and the cumulative error
is compensated. For the same reason, models 2 and 3 show a slowing down of the
error growth between 8 s and 15 s. The model 1 shows no yaw drift eψ (t) = 0, unlike
the other models which show an increasing yaw drift. As a result, it is important to
note that the proposed model has no intrinsic error.
Third test: with gaussian noise. The parameters of the three models are the same
as the second test. The steering angle is different witch gives an other trajectory. As
196 A. Koliai et al.

Fig. 4 Vehicle position and drift models outputs (with white noise) in open loop at vx = 10 m/s

Fig. 5 Vehicle position and drift models outputs (with white noise) in open loop at vx = 10 m/s.
The three drift models are enabled at 5 s

depicted in Fig. 5, the first 5 s no drift is noticed because the models are disabled. At
5 s, the 3 models are enabled. A large discontinuity is noticed in the results of model
1. In fact, in Eq. (1c) the position is directly multiplied to a scale constant, witch
gives such discontinuity. Meanwhile, the model 2 and 3 start to accumulate errors as
expected. For this trajectory, the model 2 presents unnoticeable intrinsic errors.
A Comparative Simulation Study of Localization Error Models … 197

3 Closed Loop Simulation and Results

While the first model presents undesirable discontinuity when its enabled, the second
model suffers from undesirable intrinsic error. For this reasons, only the proposed
model is chosen to perform closed loop tests.
A simulation scenario is built with the following conditions: at first, the measured
pose of the vehicle is the true position. This latter is used to follow a trajectory
at constant speed of vx = 10 m/s. After some time, the drift model is enabled. In
this way the localization system is deteriorated and the measured pose drifts. After
a while, the measured and true pose becomes identical. The guidance and control
diagram is shown in Fig. 6.
The controller main objective is to produce the right steering wheel angle to follow
the desired path. The lateral controller is provided by a robust state feedback coupled
to a feed-forward based on the bicycle model as proposed in [2]. The longitudinal
controller is provided by CarMaker. In general, the front steering angle δ f (t) is
calculated to reduce both lateral error y (t) and orientation error  (t), defined as
follows:

˙ y (t) = v y (t) + vx (t)  (t) ,  (t) = (t) − r e f (t) (4)

where vx and v y represent longitudinal and lateral velocities at the center of gravity
(CoG) and r e f is the desired yaw angle.
Three path following simulations are performed using the dataset of the track of
the Mulhouse Automobile Museum. The initial vehicle position includes a lateral
offset of ≈2.1 m. This lateral offset is added to see the ability of the guidance and
control system to generate the appropriate steering angle to join the desired path.
For the simulation 1, the switch in Fig. 6 is at position 1, which means that the
measured PM (t) and true PT (t) positions are identical. The same experiment is
repeated in simulation 2. However, during the time 9–15 s the switch is in position 2,
which means that the measured position is now affected by the drift and is no longer

Fig. 6 Bloc diagram of closed loop guidance and control system integrating a localization drift
model (DM)
198 A. Koliai et al.

Fig. 7 Trajectory of the vehicle while following the track of the Automobile Museum of Mulhouse.
vx = 10 m/s

Fig. 8 Control loop signals: steering angle δ f (t), lateral error y (t), orientation error ψ (t) and
lateral acceleration a y (t)

the actual position. The simulation 3, two switching actions are applied, during the
time 9–15 s and from 16 s to 19 s the switch is in position 2. The simulations results
are depicted in Figs. 7 and 8.
The simulation 1 results show that the measured and true signals are identical
and the path following is carried put as expected. In simulation 2, the results are the
same as in simulation 1 except when the drift model is activated. In this last case, the
controller generates a steering angle to ensure that the vehicle is on the trajectory,
with wrong position measurements. As a consequence, the vehicle deviates from the
desired trajectory. Finally, an undesirable measurement discontinuity happens when
the true position is recovered. As shown in Fig. 8, the steering angle is saturated due
to the large lateral and orientation error caused by the discontinuity. Also, high lateral
A Comparative Simulation Study of Localization Error Models … 199

accelerations are measured. In the simulation 3, these discontinuities happen twice,

witch cause longer saturation and larger following errors. Note that, the performances
of the controller are deteriorating under this discontinuity.

4 Conclusion

This paper presents a comparison between three different localization drift models
based on position measurements. The first model combines a Gauss-Markov process
with a white noise. The second one is a reformulation of vehicle displacement equa-
tions combined with a white noise. Finally, a third model is proposed here based on
the improvements made to the second model to eliminate its intrinsic errors. A guid-
ance and control scheme is implemented, including the third drift model. Then, the
closed loop simulation results highlight the discontinuity phenomenon caused by
erroneous position measurements when varying the performance of the localization.
In future work, the proposed model will be used to test and validate control strategies
that take into account measurement discontinuity.

Acknowledgements The authors gratefully acknowledge the National Automobile Museum of

Mulhouse (www.musee-automobile.fr) for the access to its test track.

References

1. Bazeille, S., Josso-Laurain, T., Ledy, J., Rebert, M., Al Assaad, M., Orjuela, R.: Characterization
of the impact of visual odometry drift on the control of an autonomous vehicle. In: IEEE
Intelligent Vehicles Symposium (IV), p. 2020. Las Vegas, NV, USA, October (2020)
2. Boudali, M., Orjuela, R., Basset, M.: A comparison of two guidance strategies for autonomous
vehicles. IFAC-PapersOnLine 50, 1 (2017)
3. Brossard, M., Bonnabel, S.: Learning wheel odometry and IMU errors for localization. In:
International Conference on Robotics and Automation (ICRA), Montreal, Canada (2019)
4. Chirca, M.: Perception pour la navigation et le contrôle des robots mobiles. Application à un
système de voiturier autonome. Ph.D. thesis, Université Blaise Pascal - Clermont II (2017)
5. Cioffi, G., Scaramuzza, D.: Tightly-coupled fusion of global positional measurements in
optimization-based visual-inertial odometry. In: IEEE/RSJ International Conference on Intel-
ligent Robots and Systems (IROS), Las Vegas, NV, USA (2020)
6. Forster, C., Carlone, L., Dellaert, F., Scaramuzza, D.: On-manifold preintegration for real-time
visual-inertial odometry. IEEE Trans. Robot. 33(1), 1–21 (2017). February
7. IPG Automotive GmbH. Reference Manual Version 10.1 (www.ipg-automotive.com)
8. Jiang, R., Klette, R., Wang, S.: Modeling of unbounded long-range drift in visual odometry. In:
Fourth Pacific-Rim Symposium on Image and Video Technology (PSIVT), p. 2010. Singapore,
Singapore (2010)
9. Mao, N., Jiangning, X., Li, J., He, H.: A LSTM-RNN-based fiber optic gyroscope drift com-
pensation. Math. Prob. Eng. 2021, 1636001 (2021). July
10. Matlab MathWorks. Visual-Inertial Odometry Using Synthetic Data - MATLAB & Simulink
- MathWorks
200 A. Koliai et al.

11. Pendleton, S., Andersen, H., Xinxin, D., Shen, X., Meghjani, M., Eng, Y., Rus, D., Ang, M.:
Perception, planning, control, and coordination for autonomous vehicles. Machines 5(1), 6
(2017). February
12. Quinchia, AG., Falco, G., Falletti, E., Dovis, F., Ferrer, C.: A Comparison between different
error modeling of MEMS applied to GPS/INS integrated systems. Sensors 13(8) (2013)
13. Reid, T.G.R., Houts, S.E., Cammarata, R., Mills, G., Agarwal, S., Vora, A., Pandey, G.: Local-
ization requirements for autonomous vehicles. SAE Int. J. Connect. Autom. Veh. 2(3) (2019)
14. Vieira, D., Orjuela, R., Spisser, M., Basset, M.: Positioning and Attitude determination for
Precision Agriculture Robots based on IMU and Two RTK GPSs Sensor Fusion. Munich,
Germany (2022)
15. Zhang, F., Wang, Z., Zhong, Y., Chen, L.: Localization error modeling for autonomous driving
in GPS denied environment. Electronics 11(4), 647 (2022)
Study of Testing Strategy
for Performance Analysis of Actuator
Layer in Safety Instrumented System

Walid Mechri and Christophe Simon

Abstract This paper analyzes the degradation of actuator layer in Safety Instru-
mented System (SIS) by considering an intermediate degraded state between the
working-and failed states. Sometimes, the current system states are not distinguished
perfectly during proof tests. The developed approach consists in using the knowledge
of both the system’s functioning and, proof testing nature and its parameters. These
parameters increase the complexity of the model when considering the unavailabil-
ity due to the proof test. This approach has been applied to a study case in order to
illustrate the effect of proof testing parameters on the SIS performance.

Keywords Dynamic Bayesian network · Unavailability · Safety instrumented

system · Proof testing

1 Introduction

Safety Instrumented System (SIS) are usually installed in process industries to protect
either people or the environment from real risks involved by dangerous situations
such as overpressure, toxic gas, fire, etc.) [2]. The aim of SIS is to bring suitable
functions that allow to maintain a safe state of an equipment under control (EUC),
in case of a dangerous event occurrence [2]. A SIS is a three layers-system which is
made of a sensor layer, a logic solver layer and an actuator layer. This structure aims
to take the EUC to a safe state when predetermined conditions are violated.
The requirements of safety function exhibited in [2] introduces a probabilistic
measure for the quantitative assessment of the SIS performance. In this context, IEC

W. Mechri
Laboratoire de Recherche MACS, LR16ES22, Université de Gabès, Ecole Nationale d’Ingénieurs
de Gabès, 6029 Gabès, Tunisie
e-mail: [email protected]
C. Simon (B)
Université de Lorraine, CRAN CNRS UMR 7039, 54000 Nancy, France
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 201
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_18
202 W. Mechri and C. Simon

61508 [2] recommends the average probability of failure on demand (PFDavg ) as a

performance index in the low demand mode. In this demand mode, some failures
remain hidden until a proof test is executed or an undesired event occurs on the EUC
[7]. These hidden failures can lead to dangerous events with severe consequences.
To improve the availability and to increase the failure tolerance, redundant structures
are often considered. For example, N redundant shutdown valves are installed on the
same pipeline to stop the flow and relieve pressure in case the downstream pressure
is too high. When at least one of the N valves works, the EUC is still safe because
the SIS works. This redundancy is named 1-out of-N (1ooN) [10].
Modeling a SIS unavailability can be carried out using suitable methods. These
methods include fault trees, Markov chains [5], Petri Nets [9], and Dynamic Bayesian
Network (DBN)[1]. Indeed, the parameters related to multi-states components in a
SIS and the proof test strategy, can be cumbersome when dealing with Markov chain.
In this context, the DBN method is an interesting formal tool to model dynamic
structure of SIS by considering the events met (failure, degradation, repair) as well
as their reliability parameters [7]. The DBN has the ability to also model the dynamic
effects associated with the proof test strategy which sometimes induces a modeling
complexity and impacts the performance. The DBN model can thus be a powerful tool
for multi-state SIS unavailability assessment considering test as exogenous variables.
The remainder of this paper is organized as follows: Sect. 2 illustrates the char-
acteristics of actuator layer in a SIS, as well as the proof test strategy. Section 3
investigates the computation of instantaneous unavailability of a 1oo3 layer struc-
ture of a SIS with the proposed test strategy. Section 4 conducts a numerical example
to present the performance variation of actuator layer in safety system according to
the proof test parameters.

2 SIS Structure and EUC

2.1 SIS Structure

Figure 1 shows the whole structure of a typical SIS that monitors an EUC (chemical
reactor). Several architectures of SIS sub-layers are considered. Actuators (FC) and
Logic Solvers (LS) are in 1oo3 voting architectures. The sensor layer is divided in
two parts (temperature and pressure) each in 1oo2 voting architecture.
For the sake of illustration, this paper considers only the actuator part as a 1oo3
architecture. It is constituted of three identical valves whose goal is to stop the flow
and relieve pressure in case the downstream pressure is too high. The process (EUC)
is safe when at least one of the three valves works on demand.
Usually, the actuators are subject to degradation phenomena (wear out) given
their mechanical property which are the research objective of degradation in SISs
and deserve more attentions [10]. The analysis in this paper refers the test function
Study of Testing Strategy for Performance Analysis of Actuator Layer … 203

Fig. 1 Example of a typical SIS

for the redundant actuator layer. Therefore, only the actuator subsystem has been
considered in this study without loss of generality.
The PFDavg concept is used to qualify SIS operated in low demand mode. In the
PFDavg computation to take into account several characteristics like : failure rates,
diagnostic coverage (DC), common cause failure (CCF) factor. In addition, testing
strategy must be included in the unavailability assessment process of SIS. Proof
testing can be executed through many test strategies [6].

2.2 Proof Test Strategies

A test is a repetitive activity that can be classified as partial or complete, i.e. applied to
all or part of the elements of a SIS. A partial test allows to test some of the components
while keeping the availability of the SIF, but with a temporarily degraded performance
[3]. For the verification of SIS several proof tests strategies have been defined, in this
context Torres-Echeveria et al. [8] proposes a classification of test strategies:
• Simultaneous test: All the components are tested together. This requires having a
sufficient number of repairmen to test all components. The SIS is made unavailable
during simultaneous test [4].
• Sequential test: All the components are tested consecutively one after the other.
Just after a component is tested and put into service, the next component is tested
and so on. This strategy considers that other components are always working [7].
• Staggered test: All the components of SIS are tested with their own period of time.
This is the most common staggering test. This strategy of test is called uniform
staggered test. It is assumed that the other components are working.
204 W. Mechri and C. Simon

• Random test: In this strategy, the time interval of test of a component does not
follow a specific program, the time interval between two tests of components is
randomly chosen [8] or computed according to the current state.
Considering all elements or a part of them for a test defines the strategy. For a
redundant layer’s elements, testing all or testing one of them are the opposite solutions
and allows or not the SIF availability [10]. Moreover, a test has some effectiveness,
harmless and duration that impacts the performance index and the modeling effort.
The test is not always able to detect all failures, its ineffectiveness is modeled by
ξ . This parameter represents the inability of the test to reveal undetected failures.
Other characteristic parameters can be considered like the harmless: the probability
of failure due to the test γ [7] and the test duration π .
Considering a non null test duration π and also component redundancy, alternate
test strategies can be proposed. Better than testing the layer which renders the layer
unavailable, one or more component but less than all of them can be tested. The
components under test are taken out of service but the layer remains available. So,
the layer structure changes during the test and the unavailability modeling should
take it into account to compute the performance index.

3 Performance Analysis Model

Dynamic Bayesian Network (DBN) is an interesting approach that can be used to

assess the performance of a SIS [7]. When using DBN, it is possible to make a
dynamic analysis of the SIS in each of its functioning phases i.e. by considering
the structure modifications. The state of the tested layers are determined given the
functioning phase (Normal, Under test) and the test strategy chosen. The probability
distribution on the layer states should be reallocated given the structure modification.
It represents the main modeling effort.

3.1 Assumptions

For the unavailability modeling the following assumptions are considered:

• All failures follow the exponential distribution.
• After repair, all the SIS components are considered in a state “As Good As New”.
• After proof test, the SIS components are “As Good As New” or in undetected
failure consequently to the test.
• The β-factor model is used to express CCFs.
• The safe failures are not treated. All failures are dangerous.
These assumptions have for goal to develop a detailed model for having an accurate
assessment of the SIS performance (PFDavg ).
Study of Testing Strategy for Performance Analysis of Actuator Layer … 205

Fig. 2 DBN for

unavailability modelling

3.2 Performance Analysis Modelling

Considering the system architecture, the proof test state and its characteristics, it is
possible to define a DBN model which can be easily simulated, as described in Fig. 2.
The gray node Uk represents the unavailability of the modeled layer according to the
state probability of X k representing the layer state at time kT and the state of the
proof test Tk .
Depending on the state of the test there may be several ways to redistribute the
probability distribution between states of two consecutive phases due to state of the
proof test (Tk =0, Tk = 0 → 1, Tk = 1, Tk = 0 → 1).
The instantaneous unavailability is computed by summing the probabilities on the
states where the layer cannot fulfill its mission when the EUC demand occurs [5].
The performance index (PFDavg ) is determined by a discrete time integration of the
instantaneous unavailability Uk .

3.3 Study of a 1oo3 Structure

The 1oo3 structure is composed of three channels functionally connected in parallel,

operating in active redundancy. This structure remains operational until at least one
channels is operational [3]. In this study, an alternate testing strategy is adopted, at a
fixed frequency. In fact, during test, only one component out of three is tested. The
component under test is taken out of service.
In this case, the 1oo3 structure goes to a 1oo2 structure and goes back to a 1oo3
structure after the proof test. In a DBN model, the stochastic process represented by
the relation between X k−1 and X k is a Markov model given by Fig. 3a. Considering
3 components and 3 states by components (OK, DD, DU), 27 states should be con-
sidered. To reduce the complexity, states grouping can be operated. By the way, the
number of states is reduced to 12 but none of the components can then be identified.
When testing a particular component, it becomes impossible to allocate specifically
the state of the tested component. Thus, three 1oo2 structures that share the previous
probability distribution should be considered (cf. Fig. 3b).
The change of the structure from 1oo3 to three 1oo2 through state of proof testing
Tk , is realized as follows:
206 W. Mechri and C. Simon

Fig. 3 Cycle of model structures from 1oo3 to a set of three 1oo2

Table 1 CPT of 1oo3 system when (Tk = 0 → 1)

Tk X k−1 Xk
{2OK} {1OK,1DD} {1OK,1DU} {2DD} {1DD,1DU} {2DU}
0→1 {3OK} 1/3 0 0 0 0 0
{2OK,1DD} 1/6 1/6 0 0 0 0
{2OK,1DU} 1/6 0 1/6 0 0 0
{1OK,2DD} 0 1/6 0 1/6 0 0
{1OK,1DD,1DU} 0 1/9 1/9 0 1/9 0
{1OK, 2DU} 0 0 1/6 0 0 1/6
{3DD} 0 0 0 1/3 0 0
{2DD,1DU} 0 0 0 1/6 1/6 0
{1DD,2DU} 0 0 0 0 1/6 1/6
{3DU} 0 0 0 0 0 1/3

• If the test state is not effective the 1oo3 structure and the associated Conditional
Probability Table is determined by using the Markov model of Fig. 3a.
• When the test is launched (Tk = 0 → 1), the state probabilities before the test are
reallocated to three anonymous 1oo2 structures where one component is under the
test and the others not. The relative CPT is given by Table 1.
• When the test is effective (Tk = 1), the 1oo3 is divided in three 1oo2 architectures.
The CPT where the 1oo3 structure is under test as described in Table 2 using the
Markov model of Fig. 3b.
Study of Testing Strategy for Performance Analysis of Actuator Layer … 207

Table 2 CPT of 1oo3 structure when (Tk = 1)

Tk X k−1 Xk
s1 ... s6 s7 ... s12 s13 ... s18
0 s1 – – – 0 ... 0 0 ... 0
.. .. .. .. .. .. ..
. – 1oo2 – . . . . . .
s6 – – – 0 ... 0 0 ... 0
s7 0 ... 0 – – – 0 ... 0
.. .. .. .. .. ..
. . . 0 – 1oo2 – . . .
s12 0 ... 0 – – – 0 ... 0
s13 0 ... 0 0 ... 0 – – –
.. .. .. .. .. .. ..
. . . . . . . - 1oo2 –
s18 0 ... 0 0 ... 0 – – –

• When the test ends (Tk = 1 → 0), the probabilities are reallocated from three 1oo2
to one 1oo3 architecture by applying the CPT of Table 3. The CPT to apply at time
k is given in Table 3.
Once the CPTs are defined (cf. Tables 1, 2 and 3) by using the testing strategy
adopted, the DBN model of Fig. 2 is used to compute the instantaneous unavailability
of 1oo3 structure according the states proof testing Tk .

4 Numerical Example

To illustrate the proposed model and testing strategy, a numerical example is pro-
posed. The actuator part of the system presented in Fig. 1 is treated. For the simula-
tions, the numerical values of the example are given on Table 4. The reader should
remind that the paper focuses on the actuator layer for the sake of illustration without
loss of generality. The actuator layer is made up of three valves and structured in
1oo3 architecture.
The proposed alternate testing strategy i.e.. one of three redundant valves is com-
pared with the simple test strategy where the 3 components are tested at the same
time. In both cases the same fixed test period is considered.
Figure 4 shows the variation of actuator layer unavailability and its average value
PFDavg , for different values of proof testing characteristics ξ , γ and π .
As shown in Fig. 4a, the PFD of layer increases to 1 for all test periods, this is
explained by the fact that the 1oo3 architecture is unavailable along the test duration π
where the all components of layer are tested simultaneously. As presented in Fig. 4b,
the unavailability is less than in the previous case because of the non-simultaneous
tests in the layer. This decrease is mainly due to the change of structure of the actuator
 208

Table 3 CPT of 1oo3 structure when (Tk = 1 → 0)

Tk X k−1 Xk
s1 s2 s3 s4 s5 s6 s7 s8 s9 s10
1→0 s1 (1 − γ ) γ .(1 − ξ ) γ .ξ 0 0 0 0 0 0 0
s2 0 (1 − γ ) 0 γ .(1 − ξ ) γ .ξ 0 0 0 0 0
s3 0 0 (1 − γ ) 0 γ .(1 − ξ ) γ .ξ 0 0 0 0
s4 0 0 0 (1 − γ ) 0 0 γ .(1 − ξ ) γ .ξ 0 0
s5 0 0 0 0 (1 − γ ) 0 0 γ .(1 − ξ ) γ .ξ 0
s6 0 0 0 0 0 (1 − γ ) 0 0 γ .(1 − ξ ) γ .ξ
W. Mechri and C. Simon
Study of Testing Strategy for Performance Analysis of Actuator Layer … 209

Table 4 Reliability parameters values

λ D (×10−6 / h) = 7; DC = 0.4; MTTR = 10 h ; Ti (h) = 2190;
ξ = 0.4; γ =0.06; π = 20 h
μD D = 1
MT T R

100

10-2
PFD

10-4

10-6
0 2000 4000 6000 8000 10000 12000 14000
Time

(a) PFD variation with simultaneous test

10-2

10-3

10-4

10-5

10-6
0 2000 4000 6000 8000 10000 12000 14000

(b) PFD variation with sequential test

Fig. 4 PFD and PFDavg of actuator layer with proposed testing strategy

layer. Indeed, by applying the proposed test strategy, the actuator layer goes from
a 1oo3 structure to a 1oo2 structure. The variation of the PFDavg due to a variation
of test strategy is clearly visible. The value of PFDavg varies from 1.0134 × 10−2
(Fig. 4a) to 0.10276 × 10−2 (Fig. 4b).
Here, discussions center on the effects of different proof test parameters on system
performance and the strategy testing with parameters in Table 4. The variation PFDavg
is computed based on developed model, as presented in Fig. 5. This figure depicts
210 W. Mechri and C. Simon

Fig. 5 PFDavg variation of actuator layer according to proof test parameters

3D-representations of the PFDavg according to the parameters γ , ξ and π . Three

cases of simulation are considered.
Obviously, the layer performance is sensitive to the characteristic values of the
proof test and also the testing strategy. On one hand, γ and ξ each have a slight
effect on the performance. On the other hand, their combined effect is particularly
deleterious (cf. Fig. 5a). According to Fig. 5b the test duration π have less effect than
the test inefficiency ξ or test harmless γ at the fixed test frequency. In third case (cf.
Fig. 5c) the variation of PFDavg , is presented according to γ and π with ξ constant.
Considering the previous remarks, it seems clear that the test strategy should
focus more on detecting faults and not deteriorating the tested components better
than working on reducing the test duration. Nevertheless, the test periods are critical
because the structure modification is less fault tolerant and the most influent on the
unavailability.

5 Conclusion

This paper studied the sensitivity of a SIS layer unavailability to proof test strategy
and characteristics. For this purpose a DBN is proposed for its ability to model the
impact of structure modification and considering test test as an exogenous variable.
Study of Testing Strategy for Performance Analysis of Actuator Layer … 211

A numerical example has been provided to illustrate the effect of the proposed
testing strategy on the SIS performance. Based on the assumption, for a 1oo3 struc-
ture, we found that proposed strategy can contribute to a better system performance
compared to simultaneous tests.
The DBN offers a simplified construction of the overall model for the assessment
of the SIS performance. The integration of the proof tests parameters allowed the
ability to formulate the behavior of the component block during the test and therefore
configure the effect of the test on the SIS components.
The results obtained dependent on the assumptions made for the construction of
DBN model. Currently, the failure rate were considered constant which is a usual
assumption. However, it is necessary to develop a more precise DBN with the abil-
ity to model ageing effect on unavailability which is usually modeled by Weibull
distributions. DBN remains a satisfying tools for this purpose.

References

1. Cai, B., Liu, Y., Fan, Q.: A multiphase dynamic bayesian networks methodology for the deter-
mination of safety integrity levels. Reliab. Eng. Syst. Saf. 150(Supplement C), 105 – 115
(2016). https://doi.org/10.1016/j.ress.2016.01.018
2. IEC61508: Functional safety of Electrical/Electronic/Programmable Electronic Safety related
systems. part 1-7 (2010)
3. Innal, F., Lundteigen, M.A., Liu, Y., Barros, A.: Pfdavg generalized formulas for sis subject
to partial and full periodic tests based on multi-phase Markov models. Reliab. Eng. Syst. Saf.
150, 160–170 (2016)
4. Mechri, W.: Evaluation de la performance des systèmes instrumentés de sécurité à paramètres
imprécis. University of Tunis El-Manar, Tunisie, Thèse (2011)
5. Mechri, W., Simon, C., Ben Othman, K.: Switching Markov chains for a holistic modeling of
sis unavailability. Reliab. Eng. Syst. Saf. 133, 212–222 (2015)
6. Mechri, W., Simon, C., Bicking, F., Ben Othman, K.: Probability of failure on demand of safety
systems by multiphase Markov chains. In: Conference on Control and Fault-Tolerant Systems
(SysTol), pp. 98–103 (2013). https://doi.org/10.1109/SysTol.2013.6693839
7. Simon, C., Mechri, W., Capizzi, G.: Assessment of safety integrity level by simulation of
dynamic bayesian networks considering test duration. J. Loss Prevent. Process Ind. 57, 101–
113 (2019)
8. Torres-Echeverria, A., Martorell, S., Thompson, H.: Multi-objective optimization of design
and testing of safety instrumented systems with moon voting architectures using a genetic
algorithm. Reliab. Eng. Syst. Saf. 106, 45–60 (2012)
9. Wu, S., Zhang, L., Lundteigen, M.A., Liu, Y., Zheng, W.: Reliability assessment for final
elements of siss with time dependent failures. J. Loss Prevent. Process Ind. 51, 186–199 (2018)
10. Zhang, A., Wu, S., Fan, D., Xie, M., Cai, B., Liu, Y.: Adaptive testing policy for multi-state
systems with application to the degrading final elements in safety-instrumented systems. Reliab.
Eng. Syst. Saf. 221, 108360 (2022)
Localization and Navigation
of an Autonomous Vehicle in Case of GPS
Signal Loss

S. N. Oubouabdellah, S. Bazeille, B. Mourllion, and J. Ledy

Abstract In recent years, autonomous vehicles have become an axis of academic

and industrial research. Localizing these vehicles without a GPS signal represents a
challenge for researchers, because the other sensors are usually less accurate, less fast
and require more computation. Among localization methods, dead reckoning ones
do not need prior knowledge as they are easier to implement for real time purposes.
However, their biggest flaw is the accumulation of errors over time. In this work,
we present an onboard localization method dedicated to autonomous vehicles for
short time navigation without GPS. We developed a method with a high rate inertial-
visual data fusion module that allows locating the vehicle in real-time. This method
has been validated offline and tested online in a path following control loop on an
experimental vehicle.

1 Introduction

The Global Positioning System (GPS) is one of the main localization methods when
it comes to autonomous vehicles. However, there are situations such as tunnel, forest,
dense city, interference problem, satellite incidents where GPS is no longer reliable
or available.

S. N. Oubouabdellah (B) · S. Bazeille · B. Mourllion · J. Ledy

IRIMAS, Université de Haute-Alsace, Brunstatt-Didenheim, France
e-mail: [email protected]
S. Bazeille
e-mail: [email protected]
B. Mourllion
e-mail: [email protected]
J. Ledy
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 213
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_19
214 S. N. Oubouabdellah et al.

1.1 Context

Several researches have been conducted in the field of onboard sensors based local-
ization for autonomous vehicles [1]. The most used technique is the Visual-Inertial-
Odometry (VIO) [2, 3] because it relies on two sensors, camera and Inertial Mea-
surement Unit (IMU), that have complementary properties. VIO can be applied with
monocular cameras [4, 5] or stereo cameras [6, 7]. There are two types of data fusion:
tight coupling [8] and loose coupling [9]. The tight coupling exploits raw data mea-
sured by sensors and is known for its robustness while the loose coupling works with
processed data and is known for its simplicity to implement. Data fusion using the
Kalman filter is widely used in the VIO context [10]. Mainly, the Extended Kalman
Filter (EKF) [11, 12] but also the Multi State Constraint Kalman Filter (MSCKF)
[13, 14].

1.2 Motivation

The IRIMAS institute works since many years on many topics linked to autonomous
vehicles. One of its contributions is an efficient lateral controller based on the GPS.
It was developed and validated on the experimental platform of the institute [15, 16].
Then, a Visual Odometry (VO) algorithm based on a monocular camera dedicated
to smalls robots was proposed [17] to self-localize without any dependence on the
infrastructure. This algorithm did not require unrealistic computation time and did
not require prior knowledge on the environment. Considering these two working
modules we decided to try to apply onboard localization to a bigger vehicle in a high
rate control loop. The compatibility of the lateral controller linked to the drifting
algorithm was studied [18] and tested in simulation. In this work, we show the
results of the implementation of the onboard localization system in our experimental
platform in real time in a control loop. We choose to implement a loosely coupled
data fusion method based on an IMU, a monocular camera, and wheel velocity
measurements. A Kalman Filter has been implemented to allow localization at the
same rate of GPS. It does not need any prior knowledge and can replace the GPS for
a short time when it is not available. This method has been tested offline on real data
and on the experimental autonomous vehicle of the IRIMAS called ARTEMIPS.

1.3 Summary

The structure of the paper is the following: in the second section, the localization
methods allowing the exploitation of IMU and camera measurements are presented
as well as the proposed Kalman Fusion. In the third section, the offline evaluation
of the method is presented and in the fourth section, the real-time experimentation
results are displayed and discussed.
Localization and Navigation of an Autonomous Vehicle … 215

2 Localization Method

Dead reckoning is the process of calculating current position of moving vehicles by

using a previously determined position, and then incorporating estimates of speed,
heading direction over elapsed time. Dead reckoning is subject to cumulative errors
so it is not accurate for long time localization compared with a GPS. To be able to
calculate the vehicle’s position using an IMU and a camera, we implemented inertial
and visual odometry algorithms, and then we combined them with a Kalman filter.
In this work, the rate of inertial measurement 100 Hz, the rate of the camera (1–10
frames per second), and the rate of the wheel 25 Hz.

2.1 Inertial Odometry (IO)

This technique [19, 20] allowed us to estimate the 3D position and orientation of a
vehicle on North-East-Down (NED) frame by using IMU measurements which are
the linear accelerations and the angle rates. Its principle consists first in calibrating
the sensor, then projecting the angle rates from body to NED frame in order to obtain
the vehicle’s orientation. This orientation permits to project the linear accelerations
to NED frame then to reduce the gravity components. Finally, the position of the
vehicle is obtained by double integration.

2.2 Visual Odometry (VO)

The VO algorithm used in this work has been developed by Rebert et al. [17]. Its
working principle consists first in detecting features from the pre-processed images
using an improved version of Harris corners detector [21]. Then these features are
matched between each two successive images and homography or a fundamental
matrix is searched from the matched features. This matrix is then used to estimate
the relative movement of the vehicle that allows obtaining its absolute position by
adding to the previous position. In this work, we use a monocular camera with an
image size of 2464 × 1300 pixels. As we are in a monocular case, we have to estimate
our position up to a scale factor. In order to obtain this scale factor, we use the wheel
velocity from the vehicle.

2.3 Visual-Inertial Data Fusion Using a Kalman Filter

A discrete Kalman filter has been implemented in order to combine the inertial and
visual data. This filter allows us to estimate the state of a system and its variance
216 S. N. Oubouabdellah et al.

thanks to its linear model by taking control and measurement as inputs. It works in
two steps, the prediction step and the correction step:
• The prediction step uses the estimated state at the previous time and the control
signal at the current time to estimate the current state. In our case, we take the
inertial data as the control signal.
• The correction step uses the measurement at the current time to correct the pre-
dicted state in order to obtain a more accurate estimation. In our case, we take the
visual data as the measurement.
As a reminder, in this work, the rate of inertial measurement (100 Hz) is faster than
the rate of the camera (1–10 frames per second). Since the acquisition rates of inertial
and visual data are different, the filter will repeat the prediction step using the inertial
data until a visual data is available to perform the correction step.
First, a linear system of n = 4 order has been chosen to describe the vehicle
dynamics. It is defined by the following state representation:
⎡ ⎤ ⎡ ⎤
1 0 dt 0 0 0
⎢0 1 0 dt ⎥ ⎢ ⎥
Z k+1 =⎢ ⎥ Zk + ⎢ 0 0 ⎥ uk
⎣0 0 1 0 ⎦ ⎣ dt 0 ⎦
0 0 0 1 0 dt
 
Ak Bk

0010
Sk = Z
0001 k

Ck

where:
• Z represents the state of the system:
 T
Z = X Y VX VY

With:
– (X , Y ) is the position of the vehicle on NED frame.
– (VX , VY ) are the linear velocities of the vehicle on NED frame.
• u represents the control which is defined as follows:
 T
u = a X aY

– a X et aY are the linear accelerations on NED frame obtained by IO.

• S represents the measured output. It is the linear speed of the vehicle on NED
deduced from VO results.
• dt is the acquisition period of the inertial data.
Localization and Navigation of an Autonomous Vehicle … 217

In order to implement a Kalman filter on the system, we proceed as follows:

Initialization Step Initialization of the state and covariance matrix of the estimated
state P and definition of the covariance matrices Q and R that express respectively
the confidence in the model and the confidence in the measurement.
Prediction Step
- State prediction: Ẑ k|k−1 = Ak Ẑ k−1|k−1 + Bk u k
- Predicted state’s covariance estimation: Pk|k−1 = Ak Pk−1|k−1 AkT + Q
Correction step
- State error computation: ε = Sk − Ck Ẑ k|k−1
- Innovation covariance matrix computation: s = Ck Pk|k−1 CkT + R
- Kalman gain computation: K fk = Pk|k−1 CkT s −1
- Estimated state update: Ẑ k|k = Ẑ k|k−1 + K fk ε
- Estimated covariance matrix update: Pk|k = (I − K fk Ck ) Pk|k−1 .
The proposed Kalman fusion allows estimating the position and the linear velocity
of the vehicle in the NED frame but it does not give any information about the
heading. For the real-time implementation, we used the heading estimated by inertial
odometry. It was chosen because it is a bit more accurate than the one estimated by
the visual odometry algorithm.

3 Offline Evaluation

The proposed localization methods have been evaluated using recorded inertial,
visual and GPS data from the experimental vehicle of the IRIMAS institute. These
measurements were recorded at the circuit of the automobile museum of Mulhouse.

3.1 Experimental Platform

ARTEMIPS, depicted in Fig. 1, is the autonomous vehicle of the IRIMAS institute.

It is a Renault Grand Scenic 3 equipped with multiple sensors and actuators. The
sensors used in this work are:
– High precision Oxford IMU RT-3002 with DGPS technology. This sensor will be
considered as the reference position (ground truth).
– Manta G507C camera. Visual data are provided at the rate 10 Hz at best.
– Xsens Mti IMU. Inertial data are provided at the rate 100 Hz.
– Wheels velocity provided 25 Hz.
218 S. N. Oubouabdellah et al.

Fig. 1 Experimental vehicle of the IRIMAS institute: ARTEMIPS

ARTEMIPS is also equipped with 3 actuators: 2 integrated servo motors MAC141

to pilot the steering wheel and the brake of the vehicle, as well as a NI multi-function
DAQ system to pilot the engine. In this work, only the steering wheel actuator is
used; the engine and brake are manually controlled.

3.2 Validation of the Localization Methods

Figure 2 presents the results of the different localization methods tested on data
recorded with ARTEMIPS. The reference trajectory of the vehicle is represented in
blue, inertial odometry is depicted in red, visual odometry is shown in green and the
Kalman localization is displayed in yellow. The trajectory begins at the Start point
(in green), then makes a whole turn of the track to end at the Finish point (in red). We
notice that the IO accumulates drift from the beginning to the end of the tests, while

Fig. 2 Comparison of the localisation algorithms on a trajectory

Localization and Navigation of an Autonomous Vehicle … 219

Table 1 Representation of the localization errors

Errors Inertial odometry Visual odometry Kalman fusion
Mean error (m) 8.49 4.58 4.31
Mean heading error 1.60 4.75 1.60
(deg)
Maximal error (m) 19.21 10.49 10.54
Maximal heading error 5.15 11.57 5.15
(deg)
Final error (m) 19.21 7.47 7.36
Final heading error 1.26 11.57 1.26
(deg)

the VO accumulates more errors in turns than in lines. Let us also remark that IO
works 100 Hz while VO works at a lower frequency (10 Hz at best). Finally, results
presented in Fig. 2 shows that the Kalman fusion was able to estimate the vehicle
trajectory with an accuracy that can be compared to the VO; the yellow and green
layouts are almost overlaid. It is important to note that this precision is fulfilled at
the same rate of the IO (100 Hz).
Results presented in Table 1 confirm the accuracy of the Kalman fusion. We notice
that the Kalman fusion allows reducing the estimation errors of the visual and inertial
odometries. Thus, in case of not favorable environment (bad weather conditions
for example), the inertial estimations prevent high trajectory drift. Therefore, the
proposed method provides 100 Hz estimation of the vehicle’s localization without
GPS while combining the accuracy obtained by VO and IO. Although this accuracy
is not sufficient for long time autonomous car navigation (i.e. drift increases over
time), it can be used in the context of an emergency stop when the GPS signal is lost.

4 Real-Time Implementation

After performing the validation process described in Sect. 3.2, the Kalman localiza-
tion method has been implemented specifically for ARTEMIPS vehicle to be linked
to the autonomous lateral controller [15], instead of the GPS one. The reference
trajectory for real-time experimentation has been chosen as shown in Fig. 3 (the pro-
posed method has been tested with a different reference trajectory and a different
start point). Concerning the camera setup, an acquisition frequency of 3 images per
second was adjusted. This configuration was required to allow the visual odometry to
perform its task, because the average computation time of the real-time visual odom-
etry is 0.193 s and higher frame rate would cause delays for image processing. The
vehicle was forced to drive at lower speed than usual to reduce the distance between
consecutive images. This is particularly critical during turns. To sum up, the speed
profile has been adapted to the situation so we had a speed of 15 km/h in the straight
220 S. N. Oubouabdellah et al.

Fig. 3 Localisation results while following a recorded trajectory

Fig. 4 Images from the camera

line and 7 km/h in turns. Finally, in term of weather, the sky was partially cloudy
which means that the conditions were favorable for the VO algorithm. Examples of
images are shown in Fig. 4, they correspond to the magenta dots in Fig. 3.
Figure 3 presents the experimental results of the real-time implementation. We
observe that the trajectory obtained thanks to Kalman fusion localization is accurate
considering the real position of the vehicle (GPS) until moment 4. At that time,
the accumulated drift was too important for the vehicle to stay on the road and the
localization error was no longer acceptable. However, we let the vehicle continue its
navigation after moment 4 in order to see the effect of the drift accumulation. This
Localization and Navigation of an Autonomous Vehicle … 221

Table 2 Representation of the localization errors

Errors Inertial odometry Visual odometry Kalman fusion
Mean error (m) 370.00 3.87 3.72
Mean heading error 1.89 3.26 1.89
(deg)
Maximal error (m) 726.15 11.68 11.37
Maximal heading error 5.84 22.51 5.84
(deg)
Final error (m) 726.15 11.37 11.34
Final heading error 0.97 7.61 0.97
(deg)

resulted an important final error shown in Table 2. We also remark that the precision
is better in the straight line than in turns which is due to the VO performances. It
should also be noted that the IO accumulates a significant drift when the speed of
the vehicle is low because its acquisition frequency is high (100 Hz) so the vehicle
appears stationary and the information will be hard to distinguish from the noise.
Above all, this result demonstrates that the ARTEMIPS vehicle was able to drive the
whole circuit of the automobile museum of Mulhouse with the use of the proposed
localization and of the lateral controller. ARTEMIPS left the road during the ‘S’
maneuver of the circuit (image 4 of Fig. 4), at t = 171 s after travelling more than
500 m without GPS localization. To summarize, the proposed fusion method allowed
the ARTEMIPS vehicle to drive for about 3 min in autonomous mode without using
the GPS signal to locate itself. It has to be noted that this result was obtained with
a not optimized version of the VO algorithm, so it will be improved in terms of
frequency and accuracy in the close future.
In Table 2 we can see that the proposed Kalman fusion allowed us to obtain
less estimation errors than VO and IO. These results support previously discussed
ones in Sect. 3.2. Table 2 as they confirm that the IO position error is extremely
important. Nevertheless, the heading obtained by IO is accurate with respect to the
GPS. However, given the amount of error accumulated over time, this method is not
accurate enough for long-time autonomous navigation (over 171 s in this case). It
can be used for short term applications such as emergency stop when the GPS signal
is lost or to replace this signal if it is disturbed and not available 100 Hz.

5 Conclusion

In this work, we have proposed a loosely coupled visual-inertial data fusion that
has been evaluated offline and then validated with a real-time implementation. This
method allowed improving the estimation accuracy obtained by visual odometry
and inertial odometry as it permitted to get a localization frequency 100 Hz for
the autonomous control mode of the ARTEMIPS vehicle. During this work, we also
222 S. N. Oubouabdellah et al.

reached a total independence from the GPS signal. To conclude, this work constitutes
a new step toward autonomy for the autonomous vehicle ARTEMIPS. In future
works, the fusion method will be improved by using, for example, an Extended
Kalman Filter on a non-linear system that describes better the vehicle dynamics.
We will also optimize the VO algorithm in order to be able to increase the images
acquisition frequency thus the vehicle speed.

Acknowledgements Thanks to ANR for financing the Evi-Deep support project. Thanks to
Sébastien Jung who worked on the implementation of IO and VO during his internship in 2021.
Thanks to the automobile museum that made the experimentation’s circuit available.

References

1. Marden, S., Whitty, M.A.: Gps-free localisation and navigation of an unmanned ground vehicle
for yield forecasting in a vineyard (2014)
2. Gui, J., Dongbing, G., Wang, S., Huosheng, H.: A review of visual inertial odometry from
filtering and optimisation perspectives. Adv. Robot. 29(20), 1289–1301 (2015)
3. Leutenegger, S., Lynen, S., Bosse, M., Siegwart, R., Furgale, P.: Keyframe-based visual-inertial
odometry using nonlinear optimization. Int. J. Robot. Res. 34(3), 314–334 (2015)
4. Qin, T., Li, P., Shen, S.: Vins-mono: a robust and versatile monocular visual-inertial state
estimator. IEEE Trans. Robot. 34(4), 1004–1020 (2018)
5. Tanskanen, P., Naegeli, T., Pollefeys, M., Hilliges. Semi-direct ekf-based monocular visual-
inertial odometry. In: 2015 IEEE/RSJ International Conference on Intelligent Robots and Sys-
tems (IROS), pp. 6073–6078 (2015)
6. Usenko, V., Engel, J., Stückler, J., Cremers, D.: Direct visual-inertial odometry with stereo
cameras. In: 2016 IEEE International Conference on Robotics and Automation (ICRA), pp.
1885–1892 (2016)
7. Sun, K., Mohta, K., Pfrommer, B., Watterson, Mi., Liu, S., Mulgaonkar, Y., Taylor, C.J., Kumar,
V.: Robust stereo visual inertial odometry for fast autonomous flight. IEEE Robot. Autom. Lett.
3(2), 965–972 (2018)
8. He, Y., Zhao, J., Guo, Y., He, W., Yuan, K.: Pl-vio: tightly-coupled monocular visual-inertial
odometry using point and line features. Sensors 18(4) (2018)
9. Al Bitar, N., Gavrilov, A.I.: Comparative analysis of fusion algorithms in a loosely-coupled
integrated navigation system on the basis of real data processing. Gyroscopy and Navigat.
10(4), 231–244 (2019)
10. Delmerico, J., Scaramuzza, D.: A benchmark comparison of monocular visual-inertial odom-
etry algorithms for flying robots. In: 2018 IEEE International Conference on Robotics and
Automation (ICRA), pp. 2502–2509 (2018)
11. Forster, C., Carlone, L., Dellaert, F., Scaramuzza, D.: On-manifold preintegration for real-time
visual-inertial odometry. IEEE TRO 33(1), 1–21 (2017)
12. Huai, Z., Huang, G.: Robocentric visual-inertial odometry. In: 2018 IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), pp. 6319–6326 (2018)
13. Li, Mingyang, Mourikis, Anastasios I.: High-precision, consistent EKF-based visual-inertial
odometry. Int. J. Robot. Res. 32(6), 690–711 (2013)
14. Ma, F., Shi, J., Yang, Y., Li, J., Dai, K.: Ack-msckf: Tightly-coupled ackermann multi-state
constraint kalman filter for autonomous vehicle localization. Sensors 19(21) (2019)
15. Boudali, M.T., Orjuela, R., Basset, M.: A comparison of two guidance strategies for autonomous
vehicles. IFAC-PapersOnLine 50, 12539–12544 (2017)
16. Laghmara, H., Boudali, M.T., Laurain, T., Ledy, J., Orjuela, R., Lauffenburger, J.P., Basset,
M.: Obstacle avoidance, path planning and control for autonomous vehicles. In: 2019 IEEE
Intelligent Vehicles Symposium (2019)
Localization and Navigation of an Autonomous Vehicle … 223

17. Rebert, M., Monnin, D., Bazeille, S., Cudel, C: Parallax beam: a vision-based motion estimation
method robust to nearly planar scenes. JEI 28(2) (2019)
18. Bazeille, S., Josso-Laurain, T., Ledy, J., Rebert, M., Al Assaad, M., Orjuela, R.: Characterization
of the impact of visual odometry drift on the control of an autonomous vehicle. In: 2020 IEEE
IV, pp. 2037–2043 (2020)
19. Titterton, D., Weston, J.L., Weston, J.: Strapdown Inertial Navigation Technology, vol. 17. IET
(2004)
20. Vieira, D., Orjuela, R., Spisser, M., Basset, M.: Positioning and attitude determination for
precision agriculture robots based on IMU and two RTK GPSs sensor fusion. In: 7th IFAC
Conference Sensing, Control and Automation for Agriculture (2022)
21. Sánchez, J., Monzón, N., Salgado De La Nuez, A.: An Analysis and Implementation of the
Harris Corner Detector, vol. 8 (2018)
Nonlinear Analysis of SRF-PLL: Hold-In
and Pull-In Ranges

Tatyana A. Alexeeva, Nikolay V. Kuznetsov, Mikhail Y. Lobachev,

Marat V. Yuldashev, and Renat V. Yuldashev

Abstract A Synchronous Reference Frame Phase-Locked Loop (SRF-PLL) is a

nonlinear control circuit widely used in power engineering for the synchronization
and control of three-phase grid-connected converters. New applications of power
converters in sustainable energy generators have led to the problem of nonlinear
analysis of SRF-PLL stability. In this paper, a continuous nonlinear model of SRF-
PLL with a first-order loop filter is studied and the hold-in range of the model is
analysed. Using the qualitative theory of dynamical systems and classical methods
of control theory, we conduct a nonlinear analysis of the SRF-PLL model and suggest
an analytical estimate for the pull-in range of the considered model. The obtained
estimate is compared with known engineering estimates of the pull-in range. MAT-
LAB Simulink is used to perform computer simulation which confirms theoretical
results.

1 Introduction

Control problems related to synchronization in electrical networks are important for

developing green energy and sustainable future. One of these problems is connection
of a solar or wind generator, which produces direct current (DC) to a grid [1, 2]. The
grid is usually an alternating current (AC) network, therefore it is necessary to use
DC-AC power converter as part of the generator. It is important to measure grid AC
voltage and synchronize the output voltage of the power converter to the main grid
before connection to avoid power surges and related damage to the equipment. After

T. A. Alexeeva
HSE University, Kantemirovskaya st., 3A, Saint Petersburg, Russia
N. V. Kuznetsov (B)
Institute for Problems in Mechanical Engineering RAS, V.O., Bolshoy pr., 61, Saint Petersburg,
Russia
e-mail: [email protected]
N. V. Kuznetsov · M. Y. Lobachev · M. V. Yuldashev · R. V. Yuldashev
St. Petersburg State University, University Embankment, 7/9, Saint Petersburg, Russia

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 225
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_20
226 T. A. Alexeeva et al.

connection, monitoring of the frequency and phase of the grid allows controlling the
balance of active and reactive power supplied to the network by the generator [3]. The
most frequently used solution is Synchronous Reference Frame Phase-Locked Loop
(SRF-PLL). This nonlinear control device monitors the grid’s phase and frequency
to control the inverter.
In the present paper, we study SRF-PLL model with a first-order loop filter. There
are also several modifications of the SRF-PLL, which are based on some pre-filtering
(adding filter before the loop) and in-loop filtering (adding filter inside the loop), such
as the DDSRF-PLL [4], moving average filter (MAF) PLLs [5, 6], delayed signal
cancellation [7], and numerous other methods (see, e.g., [8–10]). Adding pre-loop
filter does not affect the performance (speed, noise, and stability) of the SRF-PLL
itself, therefore it is omitted for simplicity. However, additional in-loop filtering
requires to redo the whole study for every modification. Nonlinear stability analysis
is especially difficult to carry over to SRF-PLLs with additional in-loop filtering and
is beyond the scope of this article.
The paper is organized in the following way. In Sect. 2, a continuous time mathe-
matical model of SRF-PLL is derived. Synchronization properties and stability of the
corresponding system is studied in Sect. 3. An analytical estimate of global stability
is obtained and plotted on a two-dimensional diagram for all SRF-PLL parameters.
In Sect. 4, the analytical formula is compared with engineering estimates.

2 Mathematical Model of the SRF-PLL

Consider a simplified model of an inverter (Fig. 1) connected to a three-phase elec-

trical grid [2, 3, 11] and corresponding phase (line-to-neutral) voltages

u a (t) = u sin(ωref t),

2
u b (t) = u sin(ωref t − π ), (1)
3
2
u c (t) = u sin(ωref t + π ).
3
Here ωref is the grid (reference) frequency and u is the Root-Mean-Squared volt-
age.1 Denote the reference phase ωref t by θref (t) = ωref t. The measured line voltages
 T  T
u abc = u a u b u c are converted into dq0 reference frame u dq0 = u d u q u 0 by
the Park transformation u dq0 = P(θPLL )u abc where

1 In the present work, we consider a SRF-PLL model, which assumes that the reference signal has
constant frequency and amplitude. When the reference signal is unbalanced or has DC component or
is distorted, some modifications of SRF-PLL can be considered, e.g., a three-phase EPLL (3EPLL)
[2].
Nonlinear Analysis of SRF-PLL: Hold-In and Pull-In Ranges 227

Fig. 1 Nonlinear model of SRF-PLL [2]

⎛ ⎞
sin(θPLL ) sin(θPLL − 2π ) sin(θPLL + 2π )
2⎝ 3 3
P(θPLL ) = cos(θPLL ) cos(θPLL − 2π3
) cos(θPLL + 2π
3
)⎠
3 1 1 1
2 2 2

and θPLL = θPLL (t) is the phase of the inverter output voltage.
The q component of the transformed set of signals is the error signal
 
2u  2 2
u q (t) = sin(θref ) cos(θPLL ) + sin θref − π cos θPLL − π
3 3 3
 
2 2  (2)
+ sin θref + π cos θPLL + π
3 3
= u sin(θref − θPLL ).

Denote by θe (t) the phase error:

θe (t) = θref (t) − θPLL (t). (3)

The error signal u q (t) is an input of the first-order loop filter which transfer function
has the form [12–14]

1 + τ2 s
F(s) = , τ1 > 0, τ2 > 0. (4)
1 + (τ1 + τ2 )s

Denote by x(t) ∈ R the state of the loop filter. The output of the loop filter vF (t) =
1
τ1 +τ2
x + τ1τ+τ
2
2
u sin θe is used to control the frequency ωPLL (t) of the inverter output
voltage, which is proportional to the control voltage:

ωPLL (t) = θ̇PLL (t) = ωbase + K vF (t), (5)

where K > 0 is a gain and ωbase is a base frequency.

228 T. A. Alexeeva et al.

Combining (2)–(5) we get equations of the SRF-PLL:

1 τ1
ẋ = − x+ u sin θe ,
τ1 + τ2 τ1 + τ2
(6)
1 τ2
θ̇e = ωe − K x+ u sin θe ,
τ1 + τ2 τ1 + τ2

where ωe = ωref − ωbase is a frequency error.

3 SRF-PLL Stability Analysis

3.1 Local Stability (Small-Signal Analysis)

Observe that system (6) is 2π -periodic in θe . If ωe < u K , then it has an infinite

eq
number of equilibria (x eq , θe ) which satisfy
τ1 ωe ωe
x eq = , sin θeeq = .
K uK

Equilibria of system (6) correspond to locked states of SRF-PLL (the phase error θe (t)
is constant when the PLL is locked). In engineering literature, a hold-in range concept
is widely used in order to characterize the ability of the loop to maintain phase-locked
conditions when the frequency error ωe varies slowly. Strict mathematical definition
of the hold-in range can be found in [15–17].  
System (6) has asymptotically stable equilibria τ1Kωe , arcsin uωKe + 2π m and
 τ1 ωe 
unstable equilibria K , π − arcsin uωKe + 2π m , and the hold-in range is [0, ωh ) =
[0, u K ).

3.2 Global Stability (Large-Signal Analysis)

To study the hold-in range it is sufficient to analyse the PLL system in vicinities
of equilibria (local analysis). In PLL literature, concept of pull-in range is widely
used in order to characterize global stability properties of the corresponding system.
A pull-in range is the largest interval of frequency errors |ωe | ∈ [0, ω p ) from the
hold-in range for which any trajectory of system (6) tends to an equilibrium (for
brevity, we shall call such systems globally stable).
Notice that analysis of global stability requires consideration of the whole phase
space and nonlinearity cannot be neglected. For global analysis of PLL models and
estimation of the pull-in range various nonlinear methods can be used such as the
direct Lyapunov method (see, e.g., [18, 19]), the method of two-dimensional com-
Nonlinear Analysis of SRF-PLL: Hold-In and Pull-In Ranges 229

parison systems (see, e.g., [20]), phase portrait analysis (see, e.g., [21]), the harmonic
balance method (see, e.g., [22, 23]), and others [24–26].

Theorem 1 Pull-in range of model (6) satisfies the inequality ω p ≥ ωest

p where
ωest
p > 0 is the unique solution of equation

 2
ωest
p  uK π τ1
arcsin + −1= √ . (7)
uK ωp
est 4( τ2 (τ1 + τ2 ) − τ2 )

Proof To analyse the pull-in range of system (6), we apply the direct Lyapunov
method and the corresponding theorem on global stability for the cylindrical phase
space (see, e.g., [27, 28]). If there is a continuous function V (x, θe ) : R2 → R such
that
(i) V (x, θe + 2π ) = V (x, θe ) ∀x ∈ R, ∀θe ∈ R,
(ii) for any solution (x(t), θe (t)) of system (6) the function V (x(t), θe (t)) is nonin-
creasing,
(iii) if V (x(t), θe (t)) ≡ V (x(0), θe (0)), then (x(t), θe (t)) ≡ (x(0), θe (0)),
(iv) V (x, θe ) + θe2 → +∞ as |x| + |θe | → +∞,
then any trajectory of system (6) tends to an equilibrium.
For system (6) we consider the following Lyapunov function satisfying condi-
tion (iv):

θe
1 τ1 ωe 2 τ1 ωe ωe
V (x, θe ) = (x − ) + sin σ − + β0 | sin σ − | dσ
2 K K uK uK
0

where  2π ωe
(sin σ − ) dσ
β0 = −  02π uK
ωe
> 0.
0 | sin σ − u K | dσ

The special form of coefficient β0 allows the Lyapunov function to be 2π -periodic

and, hence, to satisfy the first condition of the theorem. Computation of the Lyapunov
function derivative along the trajectories of system (6), considering it as a quadratic
form, and providing its negative-definiteness lead to fulfilling conditions (ii), (iii)
and, hence, to global stability of the system. 

Notice that the left-hand side of (7) is a monotonous function and ωest p can be
evaluated numerically. Moreover, Eq. (7) can be considered as an equation in two
ωest
variables: ττ21 and u Kp . Figure 2 shows stability regions observed in the second-order
SRF-PLL. The shaded area bounded by curve (7) corresponds to global stability
estimate by the Lyapunov function approach. Above the line uωKe = 1 no equilibria
exist in system (6). For SRF-PLL parameters from the gap between the line uωKe = 1
and curve (7), system (6) has asymptotically stable equilibria, but the global stability
is not guaranteed.
230 T. A. Alexeeva et al.

Fig. 2 Stability regions

observed in the second-order
SRF-PLL. The solid line
ωe
u K = 1 corresponds to
equilibria existence, the
curve which bounds the
global stability domain is
Eq. (7) from Theorem 1

Remark 1 Shaded area in Fig. 2 is a domain in the parameters space such that
system (6) with parameters from this domain is globally stable, i.e., a trajectory with
initial data from any point of the phase space tends to an equilibrium. Notice that
many works studying PLL-based systems conduct local analysis only and estimate
equilibria’s domains of attraction, not analyzing the whole phase space (see, e.g.,
recent papers [29–33]). For global analysis of systems in the cylindrical phase space
the corresponding theorem from [27, 28] should be used.

4 Comparison with Known Results

In this section, we compare estimate (7) for the pull-in range of SRF-PLL with
engineering results from [34–38]. Whereas conservative estimate (7) from Theorem 1
is strictly mathematically justified, the rest rely on approximations.
One of the first equations used in engineering design was derived by D.Richman
[34]:   2
2τ2 τ2
ωpRichman
≈ uK − . (8)
τ1 + τ2 τ1 + τ2

In [34], Richman uses phase plane descriptions and derives an approximate formula
for the pull-in time TF assuming u K τ2 1. Equation (8) for the pull-in frequency
was introduced in [38], [13, p. 168] by setting TF → +∞.
Another formula based on estimations was derived by Viterbi [35, 38, 39] for
u K τ2 1: 
2τ2
ωp
Viterbi
≈ uK > ωRichman . (9)
τ1 + τ2 p
Nonlinear Analysis of SRF-PLL: Hold-In and Pull-In Ranges 231

Fig. 3 Comparison of the 1

pull-in range conservative
estimate ωest 0.9
p from
Theorem 1 with the 0.8
Richman’s estimate (8)
ωRichman ≈ 0.7
p
2τ2 τ2
2 0.6
uK τ1 +τ2 − τ1 +τ2
0.5
estimate by Lyapunov approach
0.4 Richman’s estimate
0.3
0.2

0.5 1 1.5 2 2.5 3 3.5 4

Later Lindsey and Mengali derived approximate formulae for the case of arbitrary
periodic nonlinearity in equation (6) [36, 37]. Their formulae for sinusoidal nonlin-
earity are identical to each other and to the Viterbi’s formula (9).
Notice that formula (9) is not valid for τ2 > τ1 : this case leads to uωKe > 1, whereas
equilibria
 do not exist in this case and the system is not globally stable. In contrast,
2
τ2
2τ2
τ1 +τ2
− ∈ (0, 1) and the pull-in range estimate (8) is less than the hold-in
τ1 +τ2
range.
The global stability estimate ωest
p provided by Theorem 1 is conservative, i.e.,
ω p ≤ ω p . Although the global stability of SRF-PLL model (6) is guaranteed for
est

any ωe < ωest

p , the exact global stability domain can be wider. Figure 3 shows that
the approximate formula (8) provides the wider global stability domain for system (6)
than estimate (7). Consequently, the following inequality is valid for the considered
estimates:
p < ωp
ωest Richman
< ωViterbi
p .

To analyse estimates (8) and (9), we simulate SRF-PLL model (6) in MATLAB
Simulink, which is widely used for the study of PLL-based circuits [14, 40]. Simu-
lation of PLL models in SPICE can be found in [41] and [42].
Let us consider SRF-PLL model (6) in MATLAB Simulink with the following
parameters:
τ1 = 0.0448, τ2 = 0.4, K = 2500, u = 1.

Theorem 1 guarantees that the pull-in frequency ω p ≥ ωest p ≈ 2208, whereas equa-
tion (8) provides the following pull-in range estimate ωRichman
p ≈ 2487.3.
In Fig. 4, simulation of trajectory of system (6) with initial data x(0) = −τ1 , θe (0) =
0 is shown. In the left subfigure, the trajectory tends to an equilibrium point, what
232 T. A. Alexeeva et al.

0.04
0.03 0.03
0.02 0.02 0.0405
0.01 0.01 0.04045

0 0.0404
0
-0.01 0.04035
-0.01 0.0403
-0.02
-0.02 0.04025

-0.03 2.18 2.182 2.184 2.186

-0.03 104
-0.04
-0.04
0 50 100 150 200 0 0.5 1 1.5 2
104

Fig. 4 Simulation of system (6). Initial data: x(0) = −τ1 , θe (0) = 0. Parameters: τ1 =
0.0448, τ2 = 0.4, K = 2500, u = 1. In the left subfigure, ωe = 2208 ≈ ωest p ≤ ω p and the trajec-
tory tends to an equilibria, what corresponds to theoretical results (the system is globally stable). In
the right subfigure, ωe = 2487.3 ≈ ωRichman
p and the system experiences a persistent oscillation (see
the magnified domain), which indicate that the Richman’s estimate may yield an unstable behavior

fits the theoretical results. However, oscillations appear in the right subfigure (see
the magnified domain), hence, there is no global stability. The simulation shows that
estimate (8) should be used carefully.

5 Conclusion

In this work, global analysis of SRF-PLL stability was conducted. As a result, the
analytical formula for the pull-in range for SRF-PLL was derived and compared with
estimates of Viterbi, Richman, Lindsey, and Mengali, which rely on approximations.
Simulation results show that those estimates known from the literature may lead to
an unstable behaviour and should be used carefully.

Acknowledgements The work is supported by the Leading Scientific Schools of Russia project
NSh-4196.2022.1.1 (Sects. 1, 3) and the Russian Science Foundation project 22-11-00172
(Sects. 2, 4).

References

1. Burton, T., Jenkins, N., Sharpe, D., Bossanyi, E.: Wind Energy Handbook. Wiley (2011)
2. Karimi-Ghartemani, M.: Enhanced Phase-Locked Loop Structures for Power and Energy
Applications. Wiley (2014)
3. Teodorescu, R., Liserre, M., Rodriguez, P.: Grid Converters for Photovoltaic and Wind Power
Systems, vol. 29. Wiley (2011)
Nonlinear Analysis of SRF-PLL: Hold-In and Pull-In Ranges 233

4. Rodríguez, P., Pou, J., Bergas, J., Candela, J.I., Burgos, R.P., Boroyevich, D.: Decoupled double
synchronous reference frame PLL for power converters control. IEEE Trans. Power Electron.
22(2), 584–592 (2007)
5. Golestan, S., Ramezani, M., Guerrero, J.M., Monfared, M.: dq-frame cascaded delayed signal
cancellation- based PLL: analysis, design, and comparison with moving average filter-based
PLL. IEEE Trans. Power Electron. 30(3), 1618–1632 (2015)
6. Ghoshal, A., Vinod, J.: Performance evaluation of three phase SRF-PLL and MAF-SRF-PLL.
Turk. J. Electr. Eng. Comput. Sci. 23(6), 1781–1804 (2015)
7. Messo, T., Sihvo, J., Yang, D., Wang, X., Blaabjerg, F.: Improved delayed signal cancellation-
based SRF-PLL for unbalanced grid. In: 2017 IEEE Energy Conversion Congress and Expo-
sition (ECCE), pp. 3103–3110 (2017)
8. Young, K., Dougal, R.A.: SRF-PLL with dynamic center frequency for improved phase detec-
tion. In: 2009 International Conference on Clean Electrical Power, pp. 212–216 (2009)
9. Reza, S., Ciobotaru, M., Agelidis, V.G.: Single-phase grid voltage frequency estimation using
teager energy operator-based technique. IEEE J. Emerg. Sel. Top. Power Electron. 3(4), 1218–
1227 (2015)
10. Shakerighadi, B., Ebrahimzadeh, E., Taul, M.G., Blaabjerg, F., Bak, C.L.: Modeling and adap-
tive design of the SRF-PLL: nonlinear time-varying framework. IEEE Access 8, 28 635-28
645 (2020)
11. Zhong, Q.-C., Hornik, T.: Control of Power Inverters in Renewable Energy and Smart Grid
Integration, vol. 97. Wiley (2012)
12. Gardner, F.: Phaselock Techniques, 3rd edn. Wiley, New York (2005)
13. Egan, W.: Phase-Lock Basics, 2nd edn. Wiley, New York (2007)
14. Best, R.: Phase-Locked Loops: Design, Simulation and Application, 6th ed. McGraw-Hill
(2007)
15. Kuznetsov, N., Leonov, G., Yuldashev, M., Yuldashev, R.: Rigorous mathematical definitions
of the hold-in and pull-in ranges for phase-locked loops. IFAC-PapersOnLine 48(11), 710–713
(2015)
16. Leonov, G., Kuznetsov, N., Yuldashev, M., Yuldashev, R.: Hold-in, pull-in, and lock-in ranges
of PLL circuits: rigorous mathematical definitions and limitations of classical theory. IEEE
Trans. Circuits Syst.-I: Regular Pap. 62(10), 2454–2464 (2015)
17. Best, R., Kuznetsov, N., Leonov, G., Yuldashev, M., Yuldashev, R.: Tutorial on dynamic analysis
of the Costas loop. IFAC Ann. Rev. Control 42, 27–49 (2016)
18. Kuznetsov, N., Lobachev, M., Yuldashev, M., Yuldashev, R.: The Egan problem on the pull-in
range of type 2 PLLs. Trans. Circuits Syst. II: Express Briefs 68(4), 1467–1471 (2021)
19. Kuznetsov, N., Lobachev, M., Yuldashev, M., Yuldashev, R., Volskiy, S., Sorokin, D.: On the
generalized Gardner problem for phase-locked loops in electrical grids. Dokl. Math. 103(3),
157–161 (2021)
20. Kuznetsov, N., Lobachev, M., Yuldashev, M., Yuldashev, R., Tavazoei, M.: The Gardner prob-
lem on the lock-in range of second-order type 2 phase-locked loops. IEEE Trans. Autom.
Control (2023), (accepted)
21. Kuznetsov, N., Arseniev, D., Blagov, M., Lobachev, M., Wei, Z., Yuldashev, M., Yuldashev,
R.: The Gardner problem and cycle slipping bifurcation for type-2 phase-locked loops. Int. J.
Bifurcat. Chaos 32(9) (2022). art. num. 2250138
22. Homayoun, A., Razavi, B.: On the stability of charge-pump phase-locked loops. IEEE Trans.
Circuits Syst. I: Regular Pap. 63(6), 741–750 (2016)
23. Kuznetsov, N., Lobachev, M., Yuldashev, M., Yuldashev, R., Kolumbán, G.: Harmonic balance
analysis of pull-in range and oscillatory behavior of third-order type 2 analog PLLs. IFAC-
PapersOnLine 53(2), 6378–6383 (2020)
24. Kuznetsov, N.: Theory of hidden oscillations and stability of control systems. J. Comput. Syst.
Sci. Int. 59(5), 647–668 (2020)
25. Kuznetsov, N., Matveev, A., Yuldashev, M., Yuldashev, R., Bianchi, G.: Stability of charge-
pump phase-locked loops: the hold-in and pull-in ranges. IFAC-PapersOnLine 53, 2022–2026
(2020)
234 T. A. Alexeeva et al.

26. Kuznetsov, N., Matveev, A., Yuldashev, M., Yuldashev, R.: Nonlinear analysis of charge-pump
phase-locked loop: the hold-in and pull-in ranges. IEEE Trans. Circuits Syst. I: Regular Pap.
68(10), 4049–4061 (2021)
27. Leonov, G., Kuznetsov, N.: Nonlinear Mathematical Models of Phase-Locked Loops. Stability
and Oscillations, Cambridge Scientific Publishers (2014)
28. Kuznetsov, N., Lobachev, M., Yuldashev, M., Yuldashev, R., Kudryashova, E., Kuznetsova, O.,
Rosenwasser, E., Abramovich, S.: The birth of the global stability theory and the theory of
hidden oscillations. In: 2020 European Control Conference Proceedings, pp. 769–774 (2020)
29. Dai, Z., Li, G., Fan, M., Huang, J., Yang, Y., Hang, W.: Global stability analysis for synchronous
reference frame phase-locked loops. IEEE Trans. Ind. Electron. (2021)
30. Chen, Y., de Oliveira, M.C.: Revisiting stability analysis of phase-locked loops with the Popov
stability criterion. IFAC J. Syst. Control (2021). art. num. 100177
31. Mansour, M.Z., Me, S.P., Hadavi, S., Badrazadeh, B., Karimi, A., Bahrani, B.: Nonlinear
transient stability analysis of phase-locked loop based grid-following voltage source converters
using Lyapunov’s direct method. IEEE J. Emerg. Sel. Top. Power Electron. (2021)
32. Zhang, Y., Zhang, C., Cai, X.: Large-signal grid-synchronization stability analysis of PLL-
based VSCs using Lyapunov’s direct method. IEEE Trans. Power Syst. 37(1), 788–791 (2021)
33. Peng, Q., Buticchi, G., Tan, N.M.L., Guenter, S., Yang, J., Wheeler, P.: Modeling techniques
and stability analysis tools for grid-connected converters. IEEE Open J. Power Electron. 3,
450–467 (2022)
34. Richman, D.: Color-carrier reference phase synchronization accuracy in NTSC color television.
Proc. IRE 42(1), 106–133 (1954)
35. Viterbi, A.: Principles of Coherent Communications. McGraw-Hill, New York (1966)
36. Lindsey, W.: Synchronization Systems in Communication and Control. Prentice-Hall, New
Jersey (1972)
37. Mengali, U.: Acquisition behavior of generalized tracking systems in the absence of noise.
IEEE Trans. Commun. 21(7), 820–826 (1973)
38. Greenstein, L.: Phase-locked loop pull-in frequency. IEEE Trans. Commun. 22(8), 1005–1013
(1974)
39. Viterbi, A.: Acquisition and tracking behavior of phase-locked loops. Jet Propulsion Laboratory,
California Institute of Technology, Pasadena, External Publ, vol. 673 (1959)
40. Tranter, W., Bose, T., Thamvichai, R.: Basic Simulation Models of Phase Tracking Devices
Using MATLAB. Synthesis lectures on communications, Morgan & Claypool (2010)
41. Bianchi, G., Kuznetsov, N., Leonov, G., Seledzhi, S., Yuldashev, M., Yuldashev, R.: Hid-
den oscillations in SPICE simulation of two-phase Costas loop with non-linear VCO. IFAC-
PapersOnLine 49(14), 45–50 (2016)
42. Kuznetsov, N., Leonov, G., Yuldashev, M., Yuldashev, R.: Hidden attractors in dynamical mod-
els of phase-locked loop circuits: limitations of simulation in MATLAB and SPICE. Commun.
Nonlinear Sci. Numer. Simul. 51, 39–49 (2017)
Machine Learning, Artificial Intelligence
and Data-Driven Methods
Complementary Reward Function Based
Learning Enhancement for Deep
Reinforcement Learning

George Claudiu Andrei, Mayank Shekhar Jha, and Didier Theillol

Abstract Several complex sequential decision-making problems have been suc-

cessfully implemented using reinforcement learning (RL) for continuous optimal
control. However, the sample efficiency of data collection process during the learn-
ing phase is still not well addressed. The convergence rate to the optimal policy as
well as the time of the learning process are strongly influenced by the efficiency
of the data collected by the agent during the learning phase. This paper proposes
a method to generate efficient sample data which allows the agent to collect high
reward trajectories more frequently, decreasing the learning phase time. The pro-
posed method consists of Complementary reward (CR) function augmented to the
traditional reward function. The CR tends to infinity when the control input leads
to the system performance that meets the given requirements very accurately. Con-
sequently, the control policy which maximizes the reward function can render the
system to optimal performance. The main contributions of this study include the
following aspects: (1) a new proposed Complementary reward that is augmented
to the reward function which improves performance of the reinforcement learning
based controller in terms of system requirements; (2) speed-up of training phase via
generation of more efficient data resulting in a better learned policy.

1 Introduction

Reinforcement Learning (RL) has become one of the most important and useful
approach in control engineering. RL uses a trial-and-error learning process to maxi-
mize a decision-making agent’s total reward observed from the environment. Com-

G. C. Andrei · M. S. Jha (B) · D. Theillol

Centre de Recherche en Automatique de Nancy (CRAN), UMR 7039, CNRS, Université de
Lorraine, 54506 Vandoeuvre-lès-Nancy Cedex, France
e-mail: [email protected]
G. C. Andrei
e-mail: [email protected]
D. Theillol
e-mail: [email protected]
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 237
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_21
238 G. C. Andrei et al.

pared to optimal control theory, this maximization problem can be viewed as mini-
mizing the cost function since the reward function is designed based on the control
problem and system requirements. Recently, significant progress has been made by
combining advances in deep learning (DL) with RL methods for solving optimal
control problems resulting in the “Deep Q Network” (DQN) algorithm [1, 2]. DQN
solves RL problems with high-dimensional observation spaces but it can deal only
with discrete action spaces. Deep Deterministic Policy Gradient (DDPG) method
was introduced in 2015 in order to overcome the curse of dimensionality problem as
well as to deal with continuous spaces [1]. DDPG requires an actor-critic framework
making the algorithm easy to implement and applicable to optimal control problems
characterized by continuous spaces. A key mechanism used in the DDPG algorithm
is the use of replay buffers to store trajectories of experience in order to extract
a batch of samples during training phase as well as break the correlation between
data. A prioritized experience replay buffer was introduced as improvement of the
previous technique in order to make the samples characterized by high rewards be
selected more frequently [1]. The efficiency of the stored trajectories in terms of
reward defines training phase duration and convergence of the policy: more effi-
cient collected trajectories in terms of high reward are and the faster the agent will
learn the optimal policy. This process could make the training phase very slow and
increases the agent’s ability to learn a rather sub-optimal policy. This paper proposes
an approach to generate efficient sampled data during the learning phase which con-
sists of a Complementary reward (CR) function augmented to the traditional reward
function in order to guide control learning for speeding it up and favouring the
convergence to the optimal policy. Similar to the Barrier function concept [3], the
proposed CR tends to infinity when the control input leads to the system performance
meet its requirements very accurately. The main contributions of this study include
the following aspects: (1) a new proposed Complementary reward that is augmented
to the reward function which improves performance of the reinforcement learning
based controller in terms of system requirements; (2) speed-up of training phase via
generation of more efficient data resulting in a better learned policy.
The rest of the paper is organized as follows. Section 2 provides a general state-
of-the-art of reinforcement learning including DDPG algorithm. Section 3 presents
problem formulation. Section 4 provides more details on RL-based control for bal-
ancing a rotary inverted pendulum formulation problem. The proposed method using
Complementary function-based reinforcement learning will be discussed in Sect. 4.
Finally, the results of simulation and its analysis in comparison with traditional
reward function method is discussed in conclusions are outlined in Sect. 5.

2 Preliminaries

RL involves an agent exploring an unknown environment to achieve a goal: the

agent builds up its knowledge of the environment by gaining experience. Beyond
the agent and the environment, it is required to identify four main sub-elements that
Complementary Reward Function Based Learning Enhancement … 239

need to be defined in a RL problem: a policy, a reward signal, a value function

and optionally a model of the environment [2]. A policy π(st ) defines the learning
agent’s way of behaving at a given time t. A reward signal rt+1 = r (st , at ) is sent to
the agent at any time t by the environment. It defines the goal in a RL setting and it is
used for measuring the performance of the agent based on the same concept as cost
function in control theory. The objective of the agentT is i−t
to learn an optimal policy by
maximizing the accumulated discounted reward i=t γ r (st , u t ) from current time
step t to a future time step T , where γ ∈ (0 1) is the discount factor which discounts
the value of future rewards. A value of a state usually denoted as Vπ (st ) is the
expected accumulated reward over the future following the policy π and starting from
a particular system state st : Vπ (st ) = E π [rt+1 + γ rt+2 + γ 2 rt+3 + ...|st ], where E π
is the expected value under policy π . By considering how good it is to be in a state
by taking into account the action taken by the agent, it is necessary to refer to the
action-value function, also known as Q-function or Q-value function as Q π (st , at ).
Q π (st , at ) gives the expected return for performing action at in state st at time
step t, under the policy π : Q π (st , at ) = E π [rt+1 + γ rt+2 + γ 2 rt+3 + ...|st , at ]. The
fourth and final element is a model of the system to be controlled. Methods for solving
reinforcement learning problems that use models are called model-based methods,
as opposed to simpler model-free methods that are explicitly trial-and-error learners.
Even in model-free methods, the system model may be required in order to simulate
the system making the agent able to collect data if the hardware is not available.
Generally, RL algorithms are based on Markov decision process (MDP) frame-
work for modeling the environment, the policy and the agent. It is defined by the
tuple (S, A, r, P):
1. A set of states, S, which contains all possible states of the environment. In other
terms, all the possible discrete or continuous measurable outputs.
2. A set of possible actions, A(s), which contains all possible actions or control
inputs which can be applied to the system in each state.
3. A reward function rt+1 = r (xt , u t ), which is the immediate reward perceived after
transition from state st to st+1 as consequences of taking an action at .
4. A transition model, P(st+1 |st , at ), which denotes the probability of reaching state
st+1 when performing action at in state st in case of stochastic environments. It
can be assumed equal to one in deterministic environments.

2.1 Deep Deterministic Policy Gradient Method

DDPG is a model-free, off-policy and actor-critic framework based algorithm with

continuous action spaces, proposed by Dr. Lillicrap et al. in 2015. This method
is an extension of two other algorithms, Deep Q-Networks (DNQ) and Deter-
ministic Policy Gradient (DPG) which uses the experience replay and target net-
work as main techniques. DDPG uses two neural networks in the actor-critic and
it employs the use of off-policy data and the Bellman equation to learn the Q-
240 G. C. Andrei et al.

function Q(st , at ) which is then used to derive and learn the policy π : Q π (st , at ) =
E st+1 ∼E [rt+1 + γ E at+1 ∼π [Q π (st+1 , at+1 )], where rt+1 is the reward observed from
the environment after the action at at time step t, st+1 ∼ E means that the tran-
sition is sampled from the environment defined as E, and at+1 ∼ π means that
the action is sampled from policy π . If the policy is deterministic, it is usually
denoted as μ, and the inner expectation of the Bellman equation can be avoided:
Q μ (st , at ) = E st+1 ∼E [rt+1 + γ Q φ (st+1 , μ(st+1 )]. Q μ can be learned off-policy, by
using transitions generated by a different policy β since the expectation only depends
on the environment. The function approximator is represented by the critic neu-
ral network parameterized by θφ and it is used to approximate the Q-function.
The mean-squared error can be used as loss function to be minimized in the opti-
mization process: L(θ Q ) = E st ρ β ,at ∼β,rt+1 ∼E [(yt − Q(st , at )|θφ )2 ]. yt is the target
value: yt = rt+1 + γ Q(st+1 , μ(st+1 )|θφ ), where ρ β is the discounted state transi-
tion for the policy β, and μ(st ) = argmax Q(st , at ) is the policy defined by acting
a
greedily. The deterministic policy μ(st ) is represented by the actor neural network
parameterized by θμ . The Bellman equation is then used as in a conventional RL
method to update the critic network. The actor neural network is updated using the
following sampled policy gradient to maximize the expected discounted reward:
∇θμ J ≈ E[∇θμ Q(st , μ(st |θμ )|θφ )]. The off-policy data defined as experience tuple
et = (st , at , rt+1 , st+1 ) is generated during the training phase and then is stored in a
replay buffer in order to break the inter-correlations between experiences which are
sampled from the memory during the updating phase of the networks. This proce-
dure makes the training phase easier to converge by leading the target value as well
as Q-function prediction be independent from each other resulting in a more stable
learning. The idea is that the weights of the target networks are initialized as copies
of the actor and critic networks weights, but updated more slowly to keep them fixed
for some time steps. Finally, the parameters of the target network are updated using
moving averages for both actors and critics:

θφ  ← τ θφ + (1 − τ )θφ 

θμ ← τ θμ + (1 − τ )θμ

where τ << 1 is the soft target update rate which makes the target networks change
slowly improving learning stability.
An important problem in RL is the trade-off between exploration and exploitation.
To achieve high rewards, the agent has to choose actions by applying a control
input it has tried before and knows to be a good one. Exploration refers to the RL
agent exploring the environment to collect unseen trajectories with high rewards.
Exploitation means simply following the current better policy from the experiences
collected in the past to gain as much reward as possible. Thus, DDPG uses a noise
sampled from a noise process N is added to the actor policy when selecting an action
during training as exploration law: π(st ) = μ(st |θμ ) + N . This is called action
noise. Different noise processes can be used. The original DDPG paper suggest a
Complementary Reward Function Based Learning Enhancement … 241

time correlated one known as Ornstein-Uglenbeck process [1]. It is common practice

to employ a variance of the action noise between 1% and 10% of the maximum control
input action to reduce the training failure rate.

3 Problem Formulation

In this paper, the balance control of a rotary inverted pendulum system based on the
DDPG algorithm has been implemented. The control of a pendulum has been one
of extensively applied problems in control field and its choice is justified by the fact
that balancing an inverted pendulum in up-right position from the its stable position
is not a trivial task. The rotary inverted pendulum consists of a motor arm, which is
actuated by a DC servo motor, with a swinging pendulum arm attached to its end. The
measured output coming from the system is defined by the motor angle θ ∈ [− π2 π2 ]
and by the inverted pendulum angle α ∈ [−2π 2π ]. The set of all possible values
of the two angles among with their rate of change defines the continuous observation
space: S ∈ [θ α θ̇ α̇]. The unstable equilibrium point defines the reference point
positioned at the origin of the reference system: α0 = 0 and θ0 = 0. The control input
u t is the applied voltage by the DC Motor and the set of all possible values defines the
continuous action space: A ∈ [−10 10] V. The reward function indicates whether
the task is successful or not during the training phase and it is one of the most critical
part of the RL design. 
The reward function is then designed: rt+1 = − q11 θt2 + q22 αt2 + q33 θ̇t2 +
q44 α̇t2 .
The quadratic reward function indicates the agent how far or close his performance
is from the reference position. It has been designed in such a way as to explicitly
emphasizes the importance of minimizing the control error with respect to the up-right
position of the inverted pendulum (αr e f = 0 and θr e f = 0) positioned at reference
system origin. Consequently, the learned policy has to make the control error et
converge to zero as close as possible at each time step t as shown in Eq. 1.

et = |eα |  0 with eα = |αr e f − α| = |0 − α|. (1)

The absolute value is used to not differentiate between positive (counter clockwise)
or negative rotations (clockwise) of the inverted pendulum.
Additionally, the motor angle’s position needs to be controlled as addition system
requirement: θ ∈ [−30 30] [deg] when balancing the inverted pendulum.
242 G. C. Andrei et al.

4 Complementary Reward Function Method

In RL problems, the agent learns the optimal control policy which maximizes the
reward function by exploring the range of control strategies and selecting the optimal
control strategy which maximizes the reward function. In this context, a new method
is proposed to speed-up the control learning process and favouring the convergence
to the optimal policy by generating more trajectories with high rewards. Such trajec-
tories are generated by a CR included in the reward function which tends to infinity
when the system response leads to optimal performance making the agent collect
a higher reward during the learning phase. In this sense, in the case of balancing
control problem defined in Sect. 3, the CR defined in Eq. 2 is proposed:

q55 e( γ |αr e f −α| ) + q66 e( γ |α̇| ) if θ < ±Aand α < ±B andα̇ < ±D (2)
1 1

C(st , eα , α̇) :
0 otherwise

where γ is the damping coefficient defining the speed divergence of the function
with respect to the control error et , q55 and q66 are weighted parameters while A
and B and D are CR conditions to be fixed based on the control problem or system
requirements.
As shown in Fig. 1, the proposed CR is inversely proportional control error (see
Eq. 1) tending to infinity when eα is very close to zero. The rate of convergence of
the CR is dictated by the damping coefficient as shown in the zoom sub-figure. The
choice of γ is very important and must be chosen according to system requirements
and control problem.
The CR is included in the DDPG agent reward function defined in Sect. 3 as shown
in Eq. 3:
Rt+1 = rt+1 + C(st , eα , α̇) (3)

where rt+1 is the reward function defined in Eq. 3, C(st , α, α̇) in Eq. 2 respectively.

Fig. 1 Complementary function with different damping coefficient

Complementary Reward Function Based Learning Enhancement … 243

The traditional reward function becomes negligible when at least eα is close to

zero: Rt+1 = rt+1 + C(st , eα , α̇)  C(st , eα , α̇) if at least eα  0 and α̇  0.
The corresponding pseudo-algorithm concerning the proposed method imple-
mented with DDPG algorithm is given below Algorithm 1 with M, T and n being
the total number of episodes, episode duration and number of samples respectively.

Algorithm 1 DDPG using Complementary reward function

Step 1. Initialization
Randomly initialize actor network μ(st |θμ ) and critic network Q(st , at |θφ )
with weights θμ and θφ respectively
 
Copy the weights to target network μ and Q , θμ ← θμ , θφ  ← θφ
Initialize replay buffer
Step 2. Exploration
for episode = [1:M]
Initialize the noise process N ∼ OU for exploration
for t = [1:T]
Obtain the current system state st from the environment
Generate control input action at = μ(st |θμ ) + N based on the current actor policy
exploration action noise
Execute action at , then if: eθ < A and eα < B, receive reward Rt+1 (3) otherwise rt+1
Obtain the resulting system state st+1
Store experience (st , at , Rt+1 or rt+1 , st+1 ) into the replay buffer
Step 3. Update
Extract a mini-batch of n experiences (st , at , Rt+1 or rt+1 , st+1 ) from replay buffer
Compute the target value: yt = rt+1 + γ Q(st+1 , μ(st+1 )|θφ )

Update critic network by minimizing the loss L = n1 i (yi − Q(si+1 , ai+1 |θφ ))2
Update the actor policy by using the policy gradient
∇θμ J ≈ n1 ∇θμi Q(si , μ(si |θμi )|θφ )
Update the target networks
θφ  ← τ θφ + (1 − τ )θφ 
θμ ← τ θμ + (1 − τ )θμ
end
end

To demonstrate the effectiveness of the proposed approach while using DDPG

algorithm for optimal control, a comparison between traditional method and the pro-
posed one has been carried out. The training settings, hyper and weighted parameters,
neural networks configuration remain the same. The training phase duration is set
to 600 episodes whose duration is defined by 2500 number of steps each resulting
in 25 s since the sample time is set to 0.01. The learned policy by the agent is then
tested on the same simulink model that was used to train the agent for both methods
with simulation time of 10 [sec]. The position of the pendulum and motor angles
are always initialized in the same way at the beginning of each training episode:
α0 = ±π [rad] and θ0 = 0 which consist on vertically down position for the pen-
dulum and motor angle positioned at the origin of the reference frame respectively,
244 G. C. Andrei et al.

During each episode, the agent will generate online trajectories by observing the
measured output and reward coming from the environment and will subsequently
use them to improve and update their policy.

4.1 Traditional Reward Function Approach

The trained agent performance without using CR function is first analyzed. Figure 2
shows the output response of the inverted pendulum angle α with respect to the
unstable equilibrium point αr e f = 0. As it can be noticed, the trained agent is not
able to balance the pendulum in up-right position but leads to oscillations around the
stable equilibrium point which is positioned at α = +180 [deg] as well as α = −180
[deg] depending on the rotation sense. On the other hand, Fig. 3 shows that θ output
response is respecting the system requirements of keeping the motor angle in between
±30 [deg]. These performances demonstrate that the agent learnt a sub-optimal policy
and it would need to be trained for more than 600 episodes for converging to a better
policy. The sub-optimal convergence can be observed in the episode reward function
as shown in Fig. 6 where the steady state response has an offset of −500 with respect
to zero episode reward which is the maximized one.

Fig. 2 Inverted Pendulum

angle response with no CR

Fig. 3 Motor angle response

with no CR
Complementary Reward Function Based Learning Enhancement … 245

Fig. 4 Inverted Pendulum

angle response with CR and
γ = 0.7

Fig. 5 Motor angle response

with CR and γ = 0.7

Fig. 6 Episode reward

function with no CR

4.2 Complementary Function Approach

The implementation of the proposed method via CR is carried out with the same
training and structure of the neural networks and a damping coefficient γ = 0.7.
Figure 4 shows the inverted pendulum angle α with respect to the reference position
αr e f = 0. As it can be noticed, the trained agent is able to keep the inverted pen-
dulum in up-right position in less than 1 [sec]: α ± 5 [deg]. However, the trained
agent is not able to balance the pendulum for more than one 1.5 [sec] since the
learned policy is not optimal. Figure 5 shows that θ output response is respecting
246 G. C. Andrei et al.

Fig. 7 Episode reward

function with CR and
γ = 0.7

the system requirements of keeping the motor angle in between ±30 [deg]. This
improvement in terms of performance and convergence to a better policy in the same
amount of episodes took place thanks to the introduction of the CR which allowed
the agent to collect higher rewards with respect to the traditional approach resulting
in a better performance. From episode reward shown in Fig. 7 is possible to notice
that at the beginning of the training phase the reward function response is similar to
the traditional approach until the agent is within the CR conditions. As soon as the
agent collects high trajectories, he focuses on generating others with higher reward
by updating the policy. Consequently, CR speed-up the convergence rate to a better
policy with respect to the traditional approach.

5 Conclusions and Future Work

The proposed approach improves the policy learning by speeding up its convergence
rate via Complementary reward function making the agent able to learn a better
policy in less amount episodes thanks to its capability of rewarding the agent with
high values when the control task achieves the required performance levels. The
proposed approach can be implemented in any kind of RL problems since it acts at
the level of reward function design which is a step inherent to all the RL algorithms.
Thus, the proposed CR function concept can be viewed as a general contribution to
RL based approach. Numerical issues that might occur when the CR tends to infinity
represent a mathematical constraint of this method. For this reason, the CR must be
up bounded accurately based on the weighted parameter and damping coefficient so
the CR function diverges to the up bound which can only be reached when the control
error is very close to zero. However, the comparison between the traditional method
Complementary Reward Function Based Learning Enhancement … 247

and the proposed one demonstrates that CR function based approach can provide a
better performance under similar conditions. The next step will be to tune the hyper
parameters of the CR function, in particular, the damping coefficient in order to
speed-up the convergence of the optimal policy in as few episodes as possible.

References

1. Timothy, P., Lillicrap, et al.: Continuous control with deep reinforcement learning (2015). https://
doi.org/10.48550/ARXIV.1509.02971
2. Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. Adaptive Computation
And Machine Learning Series, 2nd edn. The MIT Press, Cambridge, Massachusetts (2018).
ISBN: 9780262039246
3. Long Vu, T. et al.: Barrier Function-based Safe Reinforcement Learning for Emergency Control
of Power Systems (2021). https://doi.org/10.48550/ARXIV.2103.14186
Data-Driven Dissipative Verification
of LTI Systems: Multiple Shots of Data,
QDF Supply-Rate and Application
to a Planar Manipulator

Tábitha E. Rosa and Bayu Jayawardhana

Abstract We present a data-driven dissipative verification method for LTI systems

based on using multiple input-output data. We assume that the supply-rate functions
have a quadratic difference form corresponding to the general dissipativity notion
known in the behavioural framework. We validate our approach in a practical example
using a two-degree-of-freedom planar manipulator from Quanser, with which we
demonstrate the applicability of multiple datasets over one-shot of data recently
proposed in the literature.

1 Introduction

In the design and control of complex high-tech systems, the provision of accurate
models of the systems is often a time-consuming and costly process. Correspond-
ingly, the data-driven approach has recently gained attention to complement the
model-based methods.
In recent years, several approaches have been presented for analyzing systems
properties directly from one-shot of input-output data that is sufficiently rich in
information, or typically referred to as the persistently exciting property [1–4]. The
systems properties that can be verified directly from data include dissipativity [2,
5–7], the persistence of excitation properties [3], and data informativity [8]. In these
cases, it is of utmost importance to have enough and suitable data to verify the
information we desire. However, obtaining a single trajectory that provides such
information on the system can be problematic in many real-life applications. In
many cases, performing new or several experiments to get the proper data shot can

T. E. Rosa (B) · B. Jayawardhana

Engineering and Technology Institute Groningen, University of Groningen, Nijenborgh 4, 9747
Groningen, The Netherlands
e-mail: [email protected]
B. Jayawardhana
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 249
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_22
250 T. E. Rosa and B. Jayawardhana

be costly and time-consuming. In other cases, we can obtain missing or corrupted

data in one-shot of data or get one-shot of poorly exciting data.
Rather than obtaining systems properties directly from the data, one can first
identify linear time-invariant (LTI) models before analyzing the systems. In this
context, several techniques in systems identification that can deal with missing data
or multiple datasets have been proposed, among many others, in [9, 10]. These
approaches have been applied, for instance, in [11, 12].
Inspired by these results from systems identification literature, there are recent
results on data-driven analysis using multiple-shots of input-output data, see for
instance [3, 13, 14]. In [14], the authors propose, among others, a method to verify
the persistence of excitation of a trajectory that recovers a set of multiple trajectories
aligned over at least a certain number of steps. In [3], the datasets can come from
different batches of measurements or experiments without the need to overlap.
As our main contribution, we extend the data-driven dissipative analysis using
one-shot data to multiple shots of data following the approach pursued in [3]. In
particular, we generalize the data-driven verification of dissipativity with quadratic
input-output supply-rate function as presented in [1, 2, 5], which is extended recently
in [5] to the quadratic differential form (QDF) of supply-rate function. The use of
QDFs generalizes the well-known concept of QSR dissipativity and broadens the
applicability of the dissipativity notion beyond passive systems and 2 stable systems.
For instance, it encompasses negative-imaginary systems [15, 16] and clockwise
systems [17, 18]. Apart from the study and control of this class of systems, QDFs
have also found application, for instance, in the stability analysis of mass-spring-
damper systems with hysteresis [17] and in fault detection systems [19]. For the rest
of this work, we will consider the time-difference form of the QDF.
In recent data-driven verification literature, it is common to consider a finite-
time horizon in the summation of the supply-rate function, which can be verified
easier than the infinite horizon case employing Willems’ fundamental lemma. In
this context, when the time horizon of dissipativity is limited to an integer L > 0, a
data-driven verification method of L-QSR dissipativity is proposed in [2] using only
one-shot of data which is convex in its formulation and whose data are required to
be persistently excited. This method resolves the non-convexity issue in [1], which
proposes a method that also incorporates the time-difference form in the QDF. When
the largest time-difference in the QDF is given by an integer N > 0, the authors in [1]
study the data-driven verification of the so-called (L , N )-QSR dissipativity via the
behavioural framework. One crucial aspect in the verification of L-QSR dissipativity
in [2] is that the condition must be checked not only for the time horizon L but also
for the time horizon L − ν for all admissible ν, which we refer to in the work as
the (L , ν)-QSR dissipativity. Motivated by these results, the concept of (L , ν)-QSR
and L-QSR dissipativity are extended to the time-difference formulation of QDF
in [5] using the notion of (L , ν, N )-QSR and (L , N )-QSR dissipativity. In Sect. 2,
we present a precise definition of these notions. In these results, the need to test all
admissible ν and the extension to the infinite-time horizon case is not trivial.
Correspondingly, the contributions of this work are as follows. Firstly, we propose
a data-driven verification method employing multiple datasets whose combination
Data-Driven Dissipative Verification of LTI Systems: Multiple Shots of Data, … 251

is collectively persistently exciting. Secondly, we consider the Q S R-dissipativity

using the time-difference form of the supply-rate. Thirdly, using a single computable
criterion, the proposed method can verify (L , N )-QSR dissipativity without the need
for verification at different time horizons L − ν for all admissible ν.
Notation: The set of vectors (matrices) of order n (n × m) with real entries is repre-
sented by Rn (Rn×m ) and correspondingly, that with integer entries is denoted by Zn
(Zn×m ). Similar notation is applied to denote a vector (matrix) with zeros and ones
by 0n and 1n (or 0n×m and 1n×m ), respectively. The n × n identity matrix is denoted
by I n . The set of positive (or negative) integers is denoted by Z+ (or Z− , respec-
tively). The Kronecker product is indicated by ⊗. The space of square-summable
discrete-time signals is denoted by 2 (R• ). Given e ∈ 2 (R• ), we denote
 
e[i, j] = e(i) e(i + 1) · · · e( j) . (1)

i
For a set of q vectors e[0,Ti −1]
, i = 1, . . . , q, we define
 
 2  q 
e[0,T−1] := e[0,T
1
1 −1] e[0,T2 −1] . . . e[0,Tq −1] . (2)

A Hankel matrix with L ∈ Z+ block rows of a finite sequence e[0,T −1] is given by
⎡ e(0) e(1) ··· e(T −L)
⎤
e(1) e(2) ··· e(T −L+1)
HL (e[0,T −1] ) = ⎣ .. .. .. .. ⎦. (3)
. . . .
e(L−1) e(L−2) ··· e(T −1)

Lemma 1 (Finsler’s Lemma [20]) If there exist w ∈ Rn , Q ∈ Rn×n , B ∈ Rm×n with

rank(B) < n and B ⊥ is a basis for the null space of B, that is, B B ⊥ = 0, then
all the following conditions are equivalent: (i). w Qw < 0, ∀w = 0 : Bw = 0;

(ii). B ⊥ Q B ⊥ < 0; (iii). ∃μ ∈ R : Q − μB  B < 0; and (iv). ∃X ∈ Rn×m : Q +
X B + B  X  < 0.

2 Problem Formulation

We consider the following discrete-time linear time-invariant (LTI) system

x(k + 1) = Ax(k) + Bu(k), x(0) = x0 ,

: (4)
y(k) = C x(k) + Du(k),

where x(k) ∈ Rn corresponds to the state vector, u(k) ∈ Rm is the control input
and y(k) ∈ R p is the output, with u ∈ 2 (Rm ). We also assume that the state-space
matrices are the minimal realization of the system . We consider that we have
access to the inputs and outputs for all instants of time k = 0, . . . , T f , where T f is
any arbitrary given time.
252 T. E. Rosa and B. Jayawardhana

In this work, we address the problem of verifying the (L , N )-dissipativity prop-

erties of  following [1, 5]. For a given ν ≥ n, the system  in (4) is said
to be (L , ν, N )-dissipative with respect to a supply-rate w(y[k,k+N ] , u k,k+N ] ) if
L−ν−1
k=0 w(u [k,k+N ] , y[k,k+N ] ) ≥ 0 holds for all trajectories (u [0,L+N −1] , y[0,L+N −1] )
with x0 = 0. Furthermore, it is called (L , N )-dissipative if  is (L , ν, N )-dissipative
for all n ≤ ν < L. Following the QDF formulation in [5, 21], we consider a quadratic
difference form of supply function w as follows

N  
y(k + i) y(k + j)
w(y[k,k+N ] , u k,k+N ] ) = i j (5)
u(k + i) u(k + j)
i, j=0

Q 
i j Si j
for every k ≥ 0, where each i j is the usual Q S R matrix given by Sij Ri j with
Q i j = Q ij and Ri j = Rij . Note that the N in (L , N )-QSR dissipativity refers to the
largest time differences in the QDF of the supply function. The matrix i j above 
gives the dissipativity relationship between a pair of data y(k + i), u(k + i) and
y(k + j), u(k + j) for a given time shift i and j. For passive and 2 systems,
they satisfy the supply-rate (5) with N = 0. We note here that in the formulation
of quadratic 
 programming, later on, we use the combination of all i j via  N =
00 ··· 0N
.. . . .. where  ji = ij for all i, j = {0, . . . , N }. Specifically, for every ν,
. . .
 N 0 ···  N N
we will refer the (L , ν, N ) dissipative property with the supply function w given by
(5) as (L , ν, N )-QSR dissipativity.
An important assumption used in the literature (as in [1, 2, 5]) for verifying L-
Q S R-dissipativity is the persistent excitation of the input data. The main idea of
having the input measurements persistently exciting is that by using a single shot of
data, we can obtain the rest of the admissible trajectories of (4), which is known as
Willems’ fundamental lemma. In practice, where missing data, corrupted data or an
insufficient amount of data can take place, the available one-shot of data z [0,T −1] may
not satisfy the central hypothesis in this lemma (namely, the persistence of excitation
condition). Consequently, we cannot obtain information on the rest of the admissible
trajectories. To deal with such cases, the following collectively persistently exciting
notion is introduced in [3], which plays an essential role in the present work.
Definition 1 ([3]) Consider a set of q measured trajectories given by e[0,T−1] as
in (2). This set of trajectories is collectively persistently exciting of order L with
0 < L ≤ Ti for all i = 1, 2, . . . , q, if the following mosaic-Hankel matrix
 q 
H L (e[0,T−1] ) = HL (e[0,T
1
1 −1]
) HL (e[0,T
2
2 −1]
) · · · HL (e[0,Tq −1] ) , (6)

where e[0,T−1] is the set of q shots of trajectories, and has full row rank.
Using this notion, a set of measured trajectories with possible different lengths
can be used together to obtain every possible trajectory of  with a time horizon L,
as given in the following lemma.
Data-Driven Dissipative Verification of LTI Systems: Multiple Shots of Data, … 253
 
Lemma 2 ([3]) Suppose that y[0,T−1] , u[0,T−1] , for q snapshots, are trajectories
of . If the set of inputs u i is collectively
 persistently exciting
 of order L + n with 0 <
L ≤ Ti for all i = 1, 2, . . . , q, then ȳ[0,L−1] , ū [0,L−1] is an admissible trajectory
q
of (4) if and only if there exists a vector α ∈ Rq(1−L)+ i=1 Ti such that
  
H L y[0,T−1]  ȳ
α = [0,L−1] . (7)
H L u[0,T−1] ū [0,L−1]

Given the information mentioned above, in this work, we address the problem of
verifying the dissipativity with the QDF supply-rate function of a system  based on
multiple shots of measured trajectories which are collectively persistently excited.
Multiple shots (L , ν,N )-QSR dissipativity verification problem: Given q mul-
tiple shots of trajectories y[0,T+N −1] u[0,T+N −1] , of , verify whether  is (L , ν, N )-
QSR dissipative for some ν ≥ n with 0 < L ≤ Ti for all i = 1, . . . q.

3 Main Result

In the following theorem, we present our main result where a data-driven verification
method is given to check the (L , N )-Q S R-dissipativity of . For any time k ≥ 0,
 
Z (k) := y(k) u(k) · · · y(k + N ) u(k + N ) . (8)

i
Using this windowed data vector of size N , the ith batch data Z [0,T i −1]
is understood
as in (1) and the expression of Hankel matrix HL (Z [0,Ti −1] ) follows the composition
i

in (3). Following the expression of mosaic-Hankel matrix in Definition 1, we define

the mosaic-Hankel matrix H L (Z[0,T−1] ) as in (6) using q snapshots of data Z[0,T−1] ,
following . Additionally, we denote
   
U = Uaux 0(m+ p)ν×(m+ p)(L−ν)(N +1) , Uaux = I ν ⊗ I m+ p 0(m+ p)×(m+ p)N . (9)
 
Theorem 1 Let the integer L > 0 and y[0,T−1] , u[0,T−1] , be a set of q trajectories
q
of (4) with n be the order of the system, i=1 Ti ≥ (L − 1)(m + p + q) and n + 1 <
L ≤ Ti for all i = 1, . . . , q. Suppose that the set of q snapshots of inputs u[0,T+N −1] , is
collectively persistently exciting of order L + N + n and there exists ν s.t. n ≤ ν < L
and
U⊥  H L (Z[0,T−1] )  L H L (Z[0,T−1] )U⊥ 0, (10)

holds, where  L = I L ⊗  N , and U⊥ = (U H L (Z[0,T−1] ))⊥ with U given in (9).

Then (4) is (L , λ, N )-Q S R dissipative for all ν ≤ λ < L.
Proof First, by the hypotheses of the theorem, we have that the set of q snapshots
of inputs u[0,T+N −1] , is collectively persistently exciting of order L + N + n holds
with L + N ≤ Ti , for all i = 1, . . . , q, such that (10) holds for an integer ν in the
254 T. E. Rosa and B. Jayawardhana

interval [n, L). Following the main results in [3], it follows from Definition
 1 and
Lemma 2 that for a given set of q trajectories y[0,T+N −1] , u[0,T+N
 −1] , if u is collec-

tively persistently exciting then other admissible trajectory ȳ[0,L+N −1] , ū [0,L+N −1]
of  satisfies    
H L+N y[0,T+N −1]  ȳ
α = [0,L+N −1] (11)
H L+N u[0,T+N −1] ū [0,L+N −1]

for some α. Likewise, if (11) holds then we can rewrite it using Z such that

H L (Z[0,T−1] )α = Z̄ [0,L−1] (12)

holds with the same α as in (11) and with Z̄ [0,L−1]  be constructed through the
rearrangement and stacking of the elements in ȳ[0,L+N −1] , ū [0,L+N −1] as used in
Z[0,T+N −1] . In other words, verifying (12) is the equivalent of verifying (11). Thus,
the matrices multiplying α in (11) and (12) share the same rank, meaning that we can
verify if the persistence of excitation conditions hold using any of these matrices.
From the hypotheses of the theorem, we have that U⊥ is a null space of
U H L (Z[0,T−1] )), e.g., U H L (Z[0,T−1] )U⊥ = 0 holds. Combining this fact with (12)
implies that U H L (Z[0,T−1] )α = 0, for any α that satisfies (12) with zero initial


condition [ ȳ[0,ν−1] ū 
[0,ν−1] ] = 0. Note that this happens due to the definition of U
as in (9), as the non-zero elements are in place to ensure that the initial con-
ditions are null. Using conditions 1) and 2) from Finsler’s lemma in Lemma 1,
and having in mind the aspects above of U⊥ , the inequality in (10) is equiva-
lent to α  H L (Z[0,T−1] )  L H L (Z[0,T−1] )α 0, for all α as before (which results
in admissible trajectories with zero initial conditions). It follows immediately that
 
L−1  L−1 N ȳ(k + i) ȳ(k + j)
Z̄ (k)  Z̄ (k) = i, j=0 ū(k + i)  ≥ 0, which is
k=0 N k=0 i j
ū(k + j)

the equivalent of verifying k=0 L−1
Z̄ (k)  N Z̄ (k) = Z̄ [0,L−1]  L Z̄ [0,L−1] ≥ 0 for all
  
trajectories ȳ[0,L+N −1] , ū [0,L+N −1] with initial conditions [ ȳ[0,ν−1]

ū 
[0,ν−1] ] = 0. Con-
sequently, considering the ν initial conditions, the previous results are equivalent to
verifying whether

L−ν−1 L−1
Z̃ (k)  N Z̃ (k) = Z̄ (k)  N Z̄ (k) ≥ 0 (13)
k=0 k=0

 
holds for any trajectory ỹ[0,L+N −ν−1] , ũ [0,L+N −ν−1] with initial conditions x̃0 = 0,
where
 x is the state of an  arbitrary minimal realization of , and for any trajectory

ȳ[0,L+N −1] , ū [0,L+N −1] with initial conditions [ ȳ[0,ν−1]

ū 
[0,ν−1] ] = 0. Additionally, we
have that if (13) holds for some value of ν, than we recover the definition of (L , ν, N )-
Q S R-dissipativity, as presented in [2, 5].
Now it remains to prove that (L , λ, N )-Q S R dissipativity holds for all λ > ν.
For this purpose, let us take ν < λ < L. We will now show that the solvability
of (10) also implies that we have (L , λ, N )-Q S R dissipativity,
 i.e., (10) holds with

U and Uaux defined in (9) to be replaced by Uλ = Uaux 0(m+ p)λ×(m+ p)(L−λ)(N +1) ,
Data-Driven Dissipative Verification of LTI Systems: Multiple Shots of Data, … 255
 
Uaux,λ = I λ ⊗ I m+ p 0(m+ p)×(m+ p)N , respectively. First we note that the null space
Uλ,⊥ := (Uλ H L (Z[0,T−1] ))⊥ with the above choice of Uλ exists if i=1 Ti ≥ λ(m +
q
q
p) + (L − 1)q. By the hypothesis of the theorem where we assume that i=1 Ti ≥
q
(L − 1)(m + p + q) and λ < L, it follows immediately that i=1 Ti ≥ λ(m + p) +
(L − 1)q holds. Thus the null space Uλ,⊥ is non-empty. By the construction of
U in (9) and Uλ with ν < λ, it follows that Im(U ) ⊂ Im(Uλ ). This implies that
Uλ,⊥ ⊂ U⊥ . Consequently, if (10) holds for some value of ν then (10) also holds
with U and Uaux be replaced by Uλ and Uaux,λ for all λ > ν. This proves the
claim. 

One direct implication from Theorem 1 is that when (10) holds with ν = n then
we obtain the (L , N )-QSR dissipativity of . Consequently, we can use ν to upper-
bound the order of the system n as discussed in [2, 5].
Theorem 1 covers existing results in [2, 5], which is restricted to the one-shot
data case. On the one hand, by taking q = 1, we immediately recover the results in
[5] with the main difference on the assumption of T . When we apply Theorem 1
in [5] using the hypotheses in Theorem 1 above, then we immediately verify the
(L , N )-Q S R dissipativity of the system without the need to check for each ν as
posited in [5]. On the other hand, by setting N = 0 and taking q = 1, we recover the
L-Q S R-dissipativity, which is also studied in [2]. The idea in Theorem 1 and that
in [5] share many commonalities with the approach in [2]. Note, however, that the
main difference is in how we construct the main inequality and arrange the data used
in the numerical verification.
By modifying the matrices Q i j , Si j and Ri j in i j , one can directly verify various
systems properties of  that involve particular dissipative inequality as discussed
in the Introduction. This includes passive systems, 2 -stable systems, negative-
imaginary/counter-clockwise systems, and positive-imaginary/clockwise systems.
Additionally, we can potentially extend this result to the case where the system is
affected by noise by applying, for instance, the results in [22].

4 Example—2 Degree-of-Freedom Planar Manipulator

For validating our technique experimentally, we consider experiments performed

in a two-degree-of-freedom (DoF) planar manipulator from Quanser [23], using
a rigid joint configuration as detailed in [24]. As presented in [24], the robot is
modelled by state equations with four state variables representing the generalized
positions and momenta of the two joints, with input variables of electrical cur-
rent (in Amperes) to the actuation motors and with output variables of the gen-
eralized position of the joints. From the property of Euler-Lagrange systems [25],
it is well-known that the open-loop behaviour of any robotic manipulator without
damping (lossless) is dissipative with respect to the QDF supply-rate function of
w(y(t), u(t)) = ẏ(t) u(t) (in the continuous-time case) or its associated discrete-
time version w(y[k,k+N ] , u k,k+N ] ) = (y(k + 1) − y(k)) u(k + 1). In our data-driven
256 T. E. Rosa and B. Jayawardhana

0.5

-0.5

-1

0 5 10 15

0.9

0.8

0.7

0.6

0.5

0 5 10 15

Fig. 1 One set of experimental data of a 2-DoF planar manipulator from Quanser. The closed-loop
0
system (robot + controller) operates at an equilibrium point (q , p ) = ( 0. 6
0. 8 , 0 ) while perturbed
by an external signal with a normal distribution, a standard deviation of 0.05 and zero-mean

dissipative verification, we consider a closed-loop system using the controller pro-

posed in Case 2 of Sect. 5.1 from [24], while the robot is operated in the neigh-
bourhood of the generalized positions q ∗ = [ 0.6 0.8 ] and generalized momenta
p ∗ = [ 0 0 ] so that the dynamics can be approximated by an LTI system (4). In
this case, the QDF supply-rate function of the closed-loop system will also contain
dissipation terms introduced by the friction damping and feedback controller.
For generating the input-output data, we excite the system with an external signal
with a normal distribution, a standard deviation of 0.05 and a zero-mean that is
calculated and applied separately to each joint. The time-series input-output data is
collected with a sampling time of Ts = 0.005 s. Figure 1 shows the data from one
of the experiments where we can see that the robots operate in the neighbourhood
of (q ∗ , p ∗ ). Note that the input for the second motor is saturated. A video of this
experiment can be found at youtu.be/2kg4Tp3qp3Y.
We consider five snapshots (q = 5) of data obtained from five different batches of
experiments, representing the case where a batch of measurement data is unavailable.
The size of the snapshots is given by Ti ∈ {10, 9, 13, 10, 12}, where each of them
is taken from a different point of the experiment and each snapshot has the form
(u[0,T+N −1] , y[0,T+N −1] ). Additionally, we take L = 4 and N = 1, where the choice
of L = 4 comes from the knowledge of the order of the system and N = 1 is from the
QDF supply-rate function of open-loop robotic manipulators, as discussed before.
We use the numerical software Matlab (R2020a) in conjunction with the parser
YALMIP [26] and the solver Mosek [27] for the optimization procedures of finding
feasible solutions   0 in Theorem 1 via linear matrix inequality (LMI) conditions.
Data-Driven Dissipative Verification of LTI Systems: Multiple Shots of Data, … 257

Fig. 2 The plot of ϒ(T f ) :=

Tf 104
k=0 w(u(k), y(k)) as a
function of the time T f ,
where the supply-rate
function w is given by (5)
103
with the identified 1

100

102
50

0
5 10 15 20

101

500 1000 1500 2000 2500 3000

Using the aforementioned multiple datasets, the obtained matrix 1 is given by

  1.463 −0.098 
1 = diag 1, 1, 1, 1, 1.999, 2, −0.098 0.674 . (14)

For comparison purposes, we apply the methods in [2, 5] to check whether they
can find a supply function to which the system is dissipative. Both of these methods
are used for verifying the dissipativity, however, they consider using one single data
shot with the input persistently excited. Thus, we test each of the small snapshots
separately, using N = 1 and N = 0 for the method in [5] and N = 0 for [2]. As
expected, the searches using all the snapshots of data separately cannot provide
results since each dataset individually is not persistently excited.
We can verify numerically whether the identified supply function (14) holds for all
Tf
batches of data. Particularly, we evaluate whether ϒ(T f ) := k=0 w(u(k), y(k)) ≥
0 holds for arbitrary T f ≥ 0 with the identified 1 as in (14). This is presented in
Fig. 2, which shows numerically that the system is indeed dissipative with respect
to 1 for an extended time horizon beyond the prescribed time horizon of L.

5 Conclusions

This work presents a data-driven method to verify dissipativity properties using a

quadratic difference form of the supply-rate function and multiple data sets that
collectively satisfy the persistent excitation condition. The method is validated on
experimental data from a 2-DoF robotic manipulator. Future work includes a fault
detection method using the concept of data-driven dissipativity analysis, in which
we identify the dissipativity inequality using a data-driven approach similar to the
one in this work.
258 T. E. Rosa and B. Jayawardhana

References

1. Maupong, T., Mayo-Maldonado, J.C., Rapisarda, P.: On Lyapunov functions and data-driven
dissipativity. IFAC-PapersOnLine 50(1), 7783–7788 (2017)
2. Romer, A., Berberich, J., Köhler, J., Allgöwer, F.: One-shot verification of dissipativity prop-
erties from input-output data. IEEE Control Syst. Lett. 3(3), 709–714 (2019)
3. van Waarde, H.J., De Persis, C., Camlibel, M.K., Tesi, P.: Willems’ fundamental lemma for
state-space systems and its extension to multiple datasets. IEEE Control Syst. Lett. 4(3), 602–
607 (2020)
4. Markovsky, I., Dörfler, F.: Behavioral systems theory in data-driven analysis, signal processing,
and control. Annu. Rev. Control 52, 42–64 (2021)
5. Rosa, T.E., Jayawardhana, B.: On the one-shot data-driven verification of dissipativity of LTI
systems with general quadratic supply rate function. In: Proceedings of the 19th European
Control Conference, (Rotterdam), pp. 1291–1296. IEEE (2021)
6. van Waarde, H.J., Camlibel, M.K., Rapisarda, P., Trentelman, H.L.: Data-driven dissipativity
analysis: application of the matrix S-lemma. IEEE Control Syst. Mag. 42(3), 140–149 (2022)
7. Koch, A., Berberich, J., Allgower, F.: Provably robust verification of dissipativity properties
from data. IEEE Trans. Autom. Control (2021)
8. Van Waarde, H.J., Eising, J., Trentelman, H.L., Camlibel, M.K.: Data informativity: a new
perspective on data-driven analysis and control. IEEE Trans. Autom. Control 65(11), 4753–
4768 (2020)
9. Markovsky, I.: Exact system identification with missing data. In: 52nd IEEE Conference on
Decision and Control, pp. 151–155. IEEE (2013)
10. Liu, Z., Hansson, A., Vandenberghe, L.: Nuclear norm system identification with missing inputs
and outputs. Syst. Control Lett. 62(8), 605–612 (2013)
11. Ding, T., Sznaier, M., Camps, O.: A rank minimization approach to fast dynamic event detection
and track matching in video sequences. In: 2007 46th IEEE Conference on Decision and
Control, pp. 4122–4127. IEEE (2007)
12. Noh, S., Zoltowski, M.D., Sung, Y., Love, D.J.: Pilot beam pattern design for channel estimation
in massive MIMO systems. IEEE J. Sel. Top. Signal Process. 8(5), 787–801 (2014)
13. Markovsky, I.: A missing data approach to data-driven filtering and control. IEEE Trans. Autom.
Control 62(4), 1972–1978 (2016)
14. Berberich, J., Allgöwer, F.: A trajectory-based framework for data-driven system analysis and
control. In: 2020 European Control Conference (ECC), pp. 1365–1370. IEEE (2020)
15. Lanzon, A., Petersen, I.R.: Stability robustness of a feedback interconnection of systems with
negative imaginary frequency response. IEEE Trans. Autom. Control 53(4), 1042–1046 (2008)
16. Mabrok, M.A., Alyami, M.A., Mahmoud, E.E.: On the dissipativity property of negative imag-
inary systems. Alex. Eng. J. 60(1), 1403–1410 (2021)
17. Ouyang, R., Andrieu, V., Jayawardhana, B.: On the characterization of the Duhem hysteresis
operator with clockwise input-output dynamics. Syst. Control Lett. 62(3), 286–293 (2013)
18. Ouyang, R., Jayawardhana, B.: Absolute stability analysis of linear systems with Duhem hys-
teresis operator. Automatica 50(7), 1860–1866 (2014)
19. Lei, Q., Munir, M.T., Bao, J., Young, B.: A data-driven fault detection method based on dissi-
pative trajectories. IFAC-PapersOnLine 49(7), 717–722 (2016)
20. de Oliveira, M.C., Skelton, R.E.: Stability tests for constrained linear systems. In: Perspectives
in Robust Control, pp. 241–257. Springer, Berlin (2001)
21. Trentelman, H., Willems, J.: Every storage function is a state function. Syst. Control Lett.
32(5), 249–259 (1997)
22. De Persis, C., Tesi, P.: Formulas for data-driven control: stabilization, optimality, and robust-
ness. IEEE Trans. Autom. Control 65(3), 909–924 (2019)
23. Quanser.: 2 DOF Serial Flexible Joint, Reference Manual. Doc. No. 800, Rev 1 (2013)
24. Chan-Zheng, C., Borja, P., Scherpen, J.M.: Passivity-based control of mechanical systems with
linear damping identification. IFAC-PapersOnLine 54(19), 255–260 (2021)
Data-Driven Dissipative Verification of LTI Systems: Multiple Shots of Data, … 259

25. Ortega, R., Perez, J.A.L., Nicklasson, P.J., Sira-Ramirez, H.J.: Passivity-Based Control of
Euler-Lagrange Systems: Mechanical, Electrical and Electromechanical Applications. Springer
Science & Business Media (2013)
26. Löfberg, J.: YALMIP: a toolbox for modeling and optimization in MATLAB, pp. 284–289.
Taipei, Taiwan (2004)
27. ApS, M.: The MOSEK optimization toolbox for MATLAB manual. Version 9.0 (2019)
Hybrid Model/Data-Driven Fault
Detection and Exclusion
for a Decentralized Cooperative
Multi-robot System

Zaynab EL Mawas, Cindy Cappelle, and Maan EL Badaoui EL Najjar

Abstract In this paper, a decentralized cooperative localization system is developed,

with a hybrid (model/data driven) diagnosis phase. The fault detection and exclusion
(FDE) is based on the usage of two approaches: model based, establishing the fault
detection indicators based on information theory, followed by a data-driven method,
deducing the behaviour of the system by generated indicators thresholding. Based on
the indicators, a Multi-Layer Perceptron (MLP) is effectively used for each phase: the
detection of the occurrence of fault in one hand and then the isolation and exclusion
of the faulty sensor in the system in the other hand. The approach is tested on real
data acquired by three Turtlebot3 equipped with wheel encoders (for odometry), a
gyroscope (for the yaw angle), a Marvelmind localization system (for the position),
and an Optitrack system (for ground truth).

1 Introduction

The advances in the field of autonomous vehicles requires the design and establish-
ment of a technology panel in which security and communication are key factors.
As vehicles become smarter, they ought to validate multiple requirements as their
awareness evolves. They also need to gain the trust of the pedestrians and passengers

Z. E. Mawas (B) · C. Cappelle · M. E. B. E. Najjar

University Lille, CNRS, Centrale Lille, UMR 9189 CRIStAL, Villeneuve D’Ascq, France
e-mail: [email protected]
C. Cappelle
e-mail: [email protected]
M. E. B. E. Najjar
e-mail: [email protected]
C. Cappelle
CIAD UMR 7533, University Bourgogne Franche-Comté, UTBM, 90010 Belfort, France

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 261
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_23
262 Z. E. Mawas et al.

alike, which increases the focus on the creation of a diagnosis layer leading to much
tolerant systems. This is done by studying problematic scenarios where the fault can
cause severe effects.
In the vehicle environment, the tasks are carried out cooperatively when it is
divided among several agents who carry it out individually. This means that every
vehicle navigates in its own trajectory and tends to localize themselves and other
vehicles that they come across sharing the same environment. This “Cooperative
Positioning System (CPS)” has been introduced by Kurazume et al. in [1] to resolve
the accumulation problem of the dead reckoning approach in a multi-robot system, by
founding the concept of “portable landmarks” for the navigation in an unknown envi-
ronment. Roumelioutis et al. [2] adopted this technique and integrated the Kalman
filter to process the available positioning information shared from the members of
the team and produce a pose estimate for every one of them.
Along the use of Kalman filter, and in multi-robot multi-sensor system, the natural
sensors redundancy makes the fault detection and exclusion (FDE) algorithms more
feasible. In [3, 4], this redundancy was used to exclude the faulty sensors from the
fusion procedure. It is an observer-based FDE approach that uses information theory
indicators to simultaneously localize a group of robots and detect their faulty sensors.
In [5], a decentralized version was introduced, in order to increase the robustness
of the system by deploying an architecture where each robot processes its own infor-
mation and only exchanges (point-to-point) the informational vector and matrix of
correction towards an observed neighboring robot [6].
The main motivation behind this work is weakness that lies within this method.
They require a prior initializing value of the threshold, which can be far from the
actual case of the robot. Moreover, the error’s module can vary given the situation,
and we cannot always presume having a non-faulty case for the trajectory to base
the study on for the calculation of the probabilities of false alarm and detection, as
real acquisition errors can occur. This requires merging the technique with a tool
that can deduces patterns from data: a Machine Learning (ML) algorithm that learns
multiple error cases with various modules to detect abnormal behavior of the system.
This hybrid method is feasible, and proved by [7] in which they used multiple model
adaptive estimation technique based on Kalman filter to recognize failure patterns.
The residuals between the estimators and the real system result are used as input
to a Neural Network (NN) that detects and identifies the faults. ML techniques are
used in various other applications and cases. For example, in [8], a single back-
propagation (BP) neural network is used to detect malfunctions based on the current
and past states of a multi-robot system and for monitoring the behaviour of the system.
Moreover, a fault detection framework for autonomous robots through sensory data
was presented in [9]. It also uses BP NN to synthesize fault detection components
based on the data collected in the training runs and Time-Delay Neural Networks
(TDNNs) as classifiers.
This paper presents a decentralized multi-sensor data fusion approach for coop-
erative multi-vehicle system, with hybrid fault detection and exclusion (H-FDE),
using model-based and data-driven methods. It is organized as the following: this
introduction is followed in Sect. 2 by the proposed approach presentation, where the
Hybrid Model/Data-Driven Fault Detection … 263

various tools are explained, for position estimation as well as the fault tolerant fusion
application and the ML use. The experimental platform is described in Sect. 3 as
well as the results of the method and the final section deals with conclusion and
perspectives.

2 Proposed Approach

The proposed approach includes a hybrid fault detection and isolation step that
merges the use of model based with data-driven techniques. It uses information
theory and Multi-Layer Perceptron (MLP) to learn and detect the pattern of the fault
in order to exclude it. It is presented in Fig. 1.

2.1 Description of the Approach

The approach has 6 main parts, and start by an initialization of the size and value of
the state/observation vectors and matrices.
1. Data Acquisition: The data is acquired on board of the various types of sensors:
encoder for model evolution, “intra” measurement on the robot itself and “inter”
measurements providing pose estimation that is computed and sent by the the
neighboring robot’s sensors whenever it perceives a robot nearby.
2. Information Filtering (2.2): It consists of two parts: prediction and correction.
The prediction model is based on encoder data providing the odometric model. The
correction’s model integrates a Batch Covariance Intersection Informational Filter
(B-CIIF) which is obtained by summing the weighted informational contributions

Fig. 1 Hybrid Fault tolerant system

264 Z. E. Mawas et al.

of the sensors in the configuration. The weights are integrated with the dependent
measurement whose cross correlation is not negligible.
3. Fault detection: The approach then proceeds to the diagnosis phase on each
robot: a generalized informational residual g D J S , describing the dissimilarities
between joined sensor measurements and the predicted solution is generated (2.3).
The divergence used in this work for dissimilarity measurement is the Jensen
Shannon divergence. The occurrence of a default is determined by the use of a
MLP that establishes the limit above which the value of the divergence implies
the passage to an abnormal situation. This MLP takes the generalized residual
along with the prior probability of the no fault hypothesis (P0 ).
4. Fault Isolation: Once a faulty case is detected, an isolation step is applied, by
calculating the divergence between the prediction and corrections: joined correc-
tions solution deprived of each of the corrections once then twice (i.e. Generalized
Observer Scheme [10]). They are then fed to the isolation’s MLP, along with the
prior probability of the no fault hypothesis (P0 ) for each sensor. The choice of
residuals is based on the operation and tolerance requirements, ensuring that all
the sensor-faults are detectable.
5. Fault exclusion: The localization solution excludes the faulty sensors from each
iteration and fuses the rest of the measurements, resulting in a more accurate
positioning.
6. Communication: The receiving/transmission of “inter” measurements data
requires the establishment of a connection between the inter-seeing robots in
order to grant the access, and obtain/transfer the informational contribution of
the Lidar. Any fault detected from an “inter” measurement correction must be
reported back to the emitting robot as a feedback of the communication step
between the robots.

2.2 Information Filtering

In order to estimate the pose of the robots, B-CIIF is used. It consists of two steps:
(a) Prediction and (b) Update where the “intra” measurements and “inter” measure-
ments between the robots are used to correct the robots position prediction.
(a) Prediction: The state vector of each robot of the team is composed of the position
and the orientation of the robot “a” with respect to a fixed frame:

X a = [x a y a θ a ]T (1)

The propagation model is non linear, it is applied as shown below:

X ka = f (X k−1
a
, u ak−1 ) + Wka (2)

where: u ak−1 is the input vector,

Hybrid Model/Data-Driven Fault Detection … 265

and Wka is the process state noise modeled as uncorrelated white noise with
covariance matrix Q a .
The covariance matrix is built as the following:
 a 
Ra
Pk|k−1 = Pk|k−1 . . . Pk|k−1
a←b
. . . Pk|k−1
a←n
(3)

where the notation a ← b is the interaction between robot “a” and robot “b”.
The information matrix is then calculated to be used in the correction step. It is
obtained by the inverse of the covariance matrix of the global system:
Ra
ϒk|k−1 Ra
= (Pk|k−1 )−1 (4)

(b) Correction: The multi-sensor data fusion is accomplished by the summation of

the multiple informational contributions of the sensors used in the configuration.
We consider the case where robots are able to localize themselves relatively to
others using the “inter” sensors (Lidar) providing the position of robot “b”
relatively to the frame of robot “a” and “intra” sensors: gyroscope that provides
a
a correction for the orientation angle [θ g ], and marvelmind that provides a
a a
correction for the position of the robot [x m , y m ].
a
The observation vector Z kobs holds the same components as the state vector:
a a a a
Z kobs = [x obs y obs θ obs ]T (5)

Where obs a ∈ {m a , g a , L a←b , . . .}. In order to deal with the data incest problem,
the (B-CIIF) separates the observations into dependent and independent values
in the correction step.
The update of the information matrix and of the information vector are deduced
obtained as follows [11]:

RC I F Ra
 obs Ian
 a
obs D
ϒk|ka = (ϒk|k−1 + Ik + ωobs Da Ik ) (6)
obs Ian a
obs D

obs a
Where Ik I n denotes the information contribution (Eq. (7)) of the independent
obs a
observations (in our case, marvelmind and gyroscope), and Ik D denotes the
dependent values “inter” observation that use the odometry value to calculate
the absolute value of the pose from a relative
 one. ωobs Da are the weights assigned
to the dependent observation such as ωobs Da ∈ [0, 1].
a
obs D

Ikobs = (Hkobs )T (Rkobs )−1 Hkobs

a a a a
(7)
266 Z. E. Mawas et al.

2.3 Hybrid Fault Detection and Exclusion

After the information filtering step, the diagnosis step aiming to check the reliability
of the obtained positioning solution is applied. It is applied by fusing the model-
based and data-driven methods, and learning the behaviour of the system from the
divergence obtained.

2.3.1 Information Theory and Divergence

This approach is statistical and non deterministic. Each sensor observation is con-
sidered as a distribution, and their comparison is done by calculating the divergence
between them. In this paper, the Jensen Shannon divergence D J S is used [12]. It
is a symmetric form of the kullback-Leibler divergence. For a two d-dimensional
Gaussian distributions, the kullback-Leibler divergence is the sum of three terms
[13]: compactness, mutual information and Mahalanobis distance between two dis-
tributions.
Two kinds of indicators are derived from this divergence: detection and isola-
tion residuals. The detection residual (g D aJ S ) obtained by calculating the divergence
Ra Ra
between the prediction Pk|k−1 and joined correction Pk|k , obtained as follows:

Ra Ra
g D aJ S = D J S (N (X k|k−1
a
, Pk|k−1 )||N (X k|k
a
, Pk|k )) (8)

Once a fault is detected, the isolation residuals are calculated in order to localise the
a
faulty sensor. These residuals J S obs , obs a ∈ {m a , g a , L a←b , . . .} are obtained by
Ra
calculating the divergence between the prediction Pk|k−1 and the joined correction
Ra
obs
Pk|k deprived of one correction once then twice:

a Ra a Ra
J S obs = D J S (N (X k|k−1
a
, Pk|k−1 )||N (X k|k
obs
, Pk|k
obs
)) (9)

2.3.2 Machine Learning for Thresholding

In order to learn the behaviour of the system and delimit the behaviour of the system
in a non faulty case, a MLP is used. It is a feed-forward artificial neural network
that deals with cases where the mapping between inputs and output is non-linear
[14]. In this work, the input of the MLP is the residual of the system, along with
the prior probability of the no fault hypothesis (P0 ) calculated by ratio of the non-
faulty cases count over the overall iterations. P0 = C0 /(C0 + C1 ) where C0 denotes
the accumulated sum of the non-faulty cases and C1 denotes that of the faulty case.
The output is the sensor fault cases presented in the system at each iteration. These
detected faults are then excluded from the fusion process to lead to a much precise
solution.
Hybrid Model/Data-Driven Fault Detection … 267

The models are optimized and trained on a central processing unit, with data
from the 3 identical robots, to generate one model that is used on each of the robot.
In order to decide the architecture chosen for the detection and isolation (MLP), an
optimization step is applied, using scikit-learn’s GridSearchCV on the Learning rate,
number of hidden layers, optimizer, activation function and Loss function.
For the detection MLP, it consists of (150, 100, 50, 50) hidden layers, using the
‘relu’ activation function, of constant learning rate of 1e-06, for a maximum iteration
of 100.
As for the isolation MLP, input data PCA was applied reducing from 8 parameters
to 2 with variance ratios of 80.903 and 18.699%. it consists of (200, 100, 100, 50,
25) hidden layers, using the ‘tanh’ activation function, of constant learning rate of
1e-06, for a maximum iteration of 300.

3 Experimentation and Results

3.1 Robotic Platform

To generate the database used to validate the proposed approach, an experimental

platform consisting of three turtlebot 3 Burger (Fig. 2b) under Robotic Operating Sys-
tem (R O S) is established. These robots are equipped with wheel encoders, embedded
OpenCR IMU, A1 RPLIDAR and marvelmind system (indoor localization system).
The system is observed by an optitrack system, which is an accurate localization
system used for ground truth and performance study (see Fig. 2a).

(a) The optitrack configuration (b) The robot used and its sensor placement

Fig. 2 The experimental platform

268 Z. E. Mawas et al.

3.2 Results

The training was done over 5 different sensor fault scenarios for 3 robots evolving
in each of the 6 different trajectories.
To compare the performance of the system, the accuracy is calculated for detection
and isolation MLP, and for both training and testing. It is the set of labels predicted
for a sample must exactly match the corresponding set of labels from the testing
data set.
n samples −1

Accuracy(y, ŷ) = 1/n samples × 1( ŷi = yi ) (10)
i=1

The faults occurring in this system have 2 types: injected and natural faults. Mul-
tiple sensor fault scenarios have been generated for 6 differently shaped trajectories,
by generating the number of repetitions, duration and module for each fault.
Given the nature of the used sensors, three types of errors are implemented: bias
errors on marvelmind data, drift error on the accumulating sensor gyroscope, drift
error on the accumulating sensor encoder, and RPLIDAR faults that are either absence
of information caused by the environment or the technology or bias errors on the
intensity of the sensor beam.
The Learning is done in centralized way. The data from the 3 robots are transfered
to a central processing unit that optimizes the architecture of the MLP for detection
and isolation and trains it then diffuses it to the robots. The testing is done on a sensor
fault scenario for each of the trajectories, and the results are shown in Table 1.
The accuracy changes given the shape of the trajectory. For the detection MLP,
the training resulted an accuracy of 73.15%. As for the isolation MLP, the training
resulted an accuracy of 66.524%. This means that not all the detected fault are iso-
lated. For circular trajectories {1, 2} it performs better than transversal trajectories
{4, 5, 6}. The third trajectory has a transversal then a circular trajectory (carrefour
scenario), and it has the best overall accuracy for detection and isolation. In transver-
sal cases, there are some robots that are not able to detect nor isolate a big majority
of the faults occurring.
Training with the PCA coordinates outperforms the one with all coordinates in
most cases except in trajectory 5 where training with the 8 coordinates gives a better
accuracy.
One trajectory (tra j2 ) is visualized in Fig. 3. This figure shows the evolution of
trajectory for each robot and the corresponding generalized residual. The red color
indicates the detection of a fault in the system. Those faults are also marked on the
generalized residual by a vertical line given their occurrence for each robots, where
some were detected, and others were missed. The accuracy of detection for this
trajectory as shown in Table 1 is 83.52% for Robot 1, with 12.9% of true negative
(miss detection (md)) and 3.52% false positive (false alarm (fa)), that could be coming
from the real fault that occurred during the acquisition of the data. Those two values
do not exceed 1/5 of the whole trajectory. In this table, the case “tra ji detect” denotes
Hybrid Model/Data-Driven Fault Detection … 269

Table 1 Accuracy of the MLP for detection and isolation

Robot 1 Robot 2 Robot 3
tra j1 detect 0.6829 0.5975 0.6951
tra j1 isolpca 0.5853 0.5365 0.5731
¯
tra j1 isol pca 0.5975 0.5121 0.6097
tra j2 detect 0.8352 0.6882 0.7529
tra j2 isolpca 0.776 0.6411 0.7
¯
tra j2 isol pca 0.7823 0.6764 0.7529
tra j3 detect 0.8078 0.8374 0.7635
tra j3 isolpca 0.8029 0.8128 0.7241
¯
tra j3 isol pca 0.7487 0.8128 0.7635
tra j4 detect 0.5212 0.6063 0.7553
tra j4 isolpca 0.5106 0.7127 0.6808
¯
tra j4 isol pca 0.4893 0.7340 0.6914
tra j5 detect 0.6857 0.4857 0.6857
tra j5 isolpca 0.2 0.2857 0.2285
¯
tra j5 isol pca 0.3142 0.4571 0.2571
tra j6 detect 0.7319 0.7010 0.6082
tra j6 isolpca 0.4948 0.6804 0.53608
¯
tra j6 isol pca 0.4329 0.6288 0.5154

Fig. 3 Detected faults from the detection MLP

270 Z. E. Mawas et al.

the detection MLP for the “tra ji ”. As for “tra ji isolpca”, the isolation MLP with
¯ with full input, without
input’s dimension reduced using pca, and “tra ji isol pca”
applying pca.
The detection MLP was able to detect the majority of the faults that were injected.
As for isolation, for the marvel sensor, there were 0.58% (md), 8.23% (fa) and an
accuracy of 91.17%. For the Lidar 2.941% (md), no (fa) and an accuracy of 97.05%.
For the gyroscope, 7.05% (md) 1.17% (fa) and an accuracy of 91.76%. Finally for
the odometry, 3.5% (md) no (fa) and an accuracy of 96.47%.

4 Conclusion and Perspectives

This paper presented a study of the fusion of the model-based and data-driven meth-
ods for the thresholding phase of the diagnosis of a multi-vehicle cooperative locali-
sation system. This approach is based on the observation of the system using multiple
sensor and fusing the data that has validated the diagnosis step without implying the
presence of an error.
The main benefit of this approach is the fact that it doesn’t need prior knowledge
on the system. A non-faulty case for the trajectory cannot always be available in
order to base the study on, because the occurrence of real errors cannot be prevented
during acquisition.
This method can be further developed to take into account the evolution factor of a
residual in a trajectory. This can be done by establishing a memory in the learning of
the model. Another way to look at it would be to study the evolution of the value of the
residual instead of just its value, and apply regression to create a signature function
that describes and predicts the occurrence of a fault given the current behavior of the
system.

Acknowledgements This work benefits from the financial support of the ANR french Research
Agency within the project LOCSP No 2019-CE22-0011.

References

1. Kurazume, R., Nagata, S., Hirose, S.: Cooperative positioning with multiple robots. In: Pro-
ceedings of the 1994 IEEE International Conference on Robotics and Automation, vol. 2, pp.
1250—1257 (1994)
2. Roumeliotis, S.I., Bekey, G.A.: Distributed multirobot localization. IEEE Trans. Robot. Autom.
18(5), 781–795 (2002)
3. Al Hage, J., El Najjar, M.E., Pomorski, D.: Multi-sensor fusion approach with fault detection
and exclusion based on the Kullback-Leibler divergence: application on collaborative multi-
robot system. Inf. Fusion 37, 61–76 (2017)
4. Abci, B., El Najjar, M.E.B., Cocquempot, V.: Sensor and actuator fault diagnosis for a multi-
robot system based on the Kullback-Leibler divergence. In: 2019 4th Conference on Control
and Fault Tolerant Systems (SysTol), pp. 68–73 (2019)
Hybrid Model/Data-Driven Fault Detection … 271

5. El Mawas, Z., Cappelle, C., El Najjar, M.E.B.: Fault tolerant cooperative localization using
diagnosis based on Jensen Shannon divergence. In: 2022 25th International Conference on
Information Fusion (FUSION), pp. 1–8 (2022)
6. Durrant-Whyte, H., Henderson, T.C.: Multisensor Data Fusion, pp. 585–610. Springer Berlin
Heidelberg, Berlin, Heidelberg (2008)
7. Goel, P., Dedeoglu, G., Roumeliotis, S.I., Sukhatme, G.S.: Fault detection and identification
in a mobile robot using multiple model estimation and neural network. In: Proceedings 2000
ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation.
Symposia Proceedings (Cat. No. 00CH37065), vol. 3, pp. 2302–2309 (2000). https://doi.org/
10.1109/ROBOT.2000.846370
8. Wang, C., Shang, W., Sun, D.: Monitoring malfunction in multirobot formation with a neural
network detector. Proc. Inst. Mech. Eng., Part I: J. Syst. Control. Eng. 225(8), 1163–1172
(2011)
9. Christensen, A.L., O’Grady, R., Birattari, M., et al.: Fault detection in autonomous robots based
on fault injection and learning. Auton. Robot. 24, 49–67 (2008)
10. Frank, P.M.: Fault diagnosis in dynamic systems via state estimation—a survey. In: Tzafestas,
S., Singh, M., Schmidt, G. (eds.) System Fault Diagnostics, Reliability and Related Knowledge-
Based Approaches, pp. 35–98. Springer, Netherlands, Dordrecht (1987)
11. Julier, S., Uhlmann, J.K.: General decentralized data fusion with covariance intersection (CI).
In: Hall, D., Llinas, J. (eds.) Multisensor Data Fusion, vol. 3, pp. 319–344. CRC Press (2009)
12. Basseville, M.: Information: entropies, divergences et moyennes (1996)
13. Do, M.N.: Fast approximation of Kullback-Leibler distance for dependence trees and hidden
Markov models. IEEE Signal Process. Lett. 10(4), 115–118 (2003)
14. Goodfellow, I., Bengio, Y., Courville, A.: Deep Learning. The MIT Press (2016)
Development of a Hybrid Safety System
Based on a Machine Learning Approach
Using Thermal and RGB Data

Nicolas Jathe, Hendrik Stern, and Michael Freitag

Abstract Using liquefied natural gas (LNG) as an alternative to conventional marine

fuels can contribute to reaching carbon neutrality. Handling LNG requires tight safety
measures due to the risk of explosion and frostbite. The present paper introduces a
live monitoring approach to detect safety violations (persons in dangerous areas)
and LNG leakage. The system utilizes a dual-camera system with regular RGB, and
thermal vision and a machine learning approach suitable for the naval environment
and movements like roll, pitch, and yaw. In lab tests, we showed a reliable detection
of gas leakage and improved person detection.

1 Introduction and Motivation

Liquified natural gas (LNG) is refrigerated natural gas that can be used for trans-
portation due to the significantly higher density in the cryogenic state compared to
its natural state. LNG can be transported to or from regions efficiently that cannot
be connected to pipeline systems [1]. This transport and supply of LNG requires
adequate infrastructure, and vessels must be equipped for this purpose [2]. Here,
safety requirements must be observed, mainly because of the risk of explosion and
the extreme danger posed by cryogenic gases to living beings (freezing) [1].
Several requirements must be fulfilled at facilities and vessels to obtain ade-
quate LNG safety. First, preventive safety measures can be executed, e.g., adequate

N. Jathe (B) · M. Freitag

BIBA - Bremer Institut für Produktion und Logistik GmbH at the University of Bremen,
Hochschulring 20, 28359 Bremen, Germany
e-mail: [email protected]
M. Freitag
e-mail: [email protected]
H. Stern
Faculty of Production Engineering, University of Bremen,
Badgasteiner Str. 1, 28359 Bremen, Germany
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 273
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_24
274 N. Jathe et al.

maintenance of pipe systems or vessels, to lower the risk of failures. Here, digital
assistance systems can be used [3]. Second, further safety measures such as primary
containment, secondary containment, safeguard systems, and separation distance are
used. They apply throughout the whole LNG value chain, from production, liquefac-
tion, and shipping to storage and re-gasification [4, 5]. This paper aims to contribute
to LNG safety in the field of safeguard systems. It describes a novel monitoring
approach for ensuring safety during LNG transfer operations (e.g., bunkering pro-
cess). It is based on a combined RGB and thermal (IR) camera, which enables auto-
matic hazard detection within the use cases (detection of persons in the hazardous
area as well as detection of gas leaks during the bunkering process) in combina-
tion with Deep Learning algorithms. The monitoring approach was subsequently
evaluated in laboratory tests.

2 State of the Art: Safety Systems for Leakage and Person

Detection at LNG Vessels

The handling of LNG can potentially lead to hazardous events that adequate safety
measures must prevent. Potential hazard events may be attributed to the specific prop-
erties of LNG. Such events can be explosions, gas venting, frostbite, or vaporization
[4, 5]. As described above, various safety requirements for LNG systems are designed
to prevent these hazardous events. In addition to structural design measures that serve
to store and move the LNG in safe and stable tanks and pipe systems (primary and
secondary containment), safeguard and safety systems are of great importance. In
this case, the aim is to rapidly detect a possible gas leak utilizing detectors (e.g., gas,
liquid, or fire sensors) and thus keep its effects as minor as possible. If a gas leak is
detected, such a system would initiate an emergency shutdown and warn the crew
[4, 5]. In addition, a fourth safety measure is to maintain safe distances [4, 5].
Our focus is on the consideration of safety systems. Various research efforts exist
to provide additional protection against hazardous events through more advanced
sensor technology (e.g., infrared sensors or cameras). In research work by Lin et al.
[6], piezoelectric acoustic sensors were used to detect gas leaks using characteristic
acoustic signals that differ from acoustic signals in the non-leaking state. In addition
infrared sensors or cameras have been used in various works to detect gas leaks
in conjunction with image processing by detecting concentration changes of the
leaking gas [7, 8]. Use cases have been leakage from pipelines [7] or gas-powered
household appliances in conjunction with an audible alarm [8]. In studies using
machine learning approaches, gas detection was implemented using RGB images
[9] or IR images [10] for smoke detection. Here, the focus was on leaks at pipelines
[9] or inside gas production power plants [10].
Development of a Hybrid Safety System Based on a Machine Learning Approach … 275

3 Results

In this paper, the two main tasks are a monitoring approach for covering the bunkering
process and fast and reliable detection of gas leakage (Sect. 3.1) and a dual camera
approach to object detection used for detecting persons (Sect. 3.2), with an additional
focus on safety violation prediction approach for moving workers via object tracking
(Sect. 3.2.4).

3.1 Leakage Detection

Fast leakage detection is crucial in safely handling LNG. This section describes
several algorithms for detecting fast and reliable potential leakages in maritime envi-
ronments. Wong et al. [11] have introduced a threshold-based approach to detect
temperature events. If the temperature reaches the threshold, the detector is triggered.
The approach takes only temperature values converted to color pixels as input, which
could be problematic if the conversion is done in a normalizing way. Further, since
a leak of LNG would take time to cool down the surrounding area, relying only on
a threshold approach for the temperature appears not to be fast enough to prevent
hazardous events. Other approaches are based on leakage detection on IR images.
Here, Bin et al. [10] used a ResNet [12] approach to detect leaking gases visible in the
infrared band. A similar approach was performed by Xu et al. [13] with a Fast-RCNN
[14] architecture. However, in maritime environments strong winds and fog could
make detecting leakage only via detecting the gas itself unreliable. Consequently,
the following section illustrates a novel approach to detecting the emersion point of
the leakage.

3.1.1 Basic Approach: Leakage Detection on Thermal Images

To detect leakage via IR imaging, the following definition is used: Let It ∈ Rw×h
be an IR image of size w × h at timestamp t ∈ N and It,i, j with i ∈ [1, . . . , w] and
j ∈ [1, . . . , h] its corresponding values, then the Boolean leak-indicator L t,i, j at the
location i, j and time t is defined as
⎧
⎨ 1, if It,i, j < rmin
L t,i, j = & (It,i, j − It−1,i, j )/t < cmin (1)
⎩
0, else.

with rmin as IR threshold, the time t between t and t − 1 and the IR change rate
threshold cmin . Leaking gas is cooling the leaking spot rapidly and is detected if
the temperature change is greater than the IR change rate threshold. The value cmin
is always a trade-off between sensitivity and specificity and should be chosen for
276 N. Jathe et al.

Fig. 1 IR image of the demonstrator with detected leakage points illustrated in red. Different
scenarios were tested by varying the air pressure inside the pipe system and by controlling a valve.
The system had leaking points between the linked parts detected by the monitoring approach.
The value for rmin was adjusted accordingly. (All the used IR data in this paper have a resolution
of 0.04 mk and are stored on a 16-bit basis. With the conversion formula of T ◦ C = rawvalue ∗
resolution −273, 15 ◦ C this results in a temperature range between −273, 15◦ and 2348, 29 ◦ C.
Therefore, every displayed temperature image here is color adjusted to its full color range)

each combination of gas and surrounding conditions. The value for rmin should be
higher than the temperature of the monitored gas. For LNG the value for rmin has to
be greater than −161 ◦ C to apply. Also rmin should remain at a reasonable level to
reduce the false positives rate.
A demonstrator was created to evaluate the monitoring approach’s functionality
for detecting leakage. It is designed as a system of pipe connections, which can
be cooled down significantly compared to the environment by adding solid carbon
dioxide. At an ambient temperature of 20 ◦ C, up to −75 ◦ C was reached in the pipe
system. In prepared areas, the demonstrator has leaks where cold gas could leak out.
Thus, the demonstrator intends to replicate a (faulty) LNG bunkering system without
posing a risk to the operators involved in the tests. Figure 1 shows the demonstrator
and the detected leakage as described in Eq. 1.

3.1.2 Extended Approach: Improved Leakage Detection on Thermal

Images

The experiments performed with the demonstrator indicated that the monitoring
approach described in Eq. 1 could detect leakage in a closed lab experiment. However,
for the leakage detection approach to work reliably, the location covered by It,i, j needs
to be the same as in It+1,i, j , which is not given for moving parts or camera movement.
Also, the values in It,i, j and It+1,i, j need to be related to the same object. Otherwise
false positive detections could occur when objects are temporally occluded in one
frame and visible in another.
The following section describes a neural network architecture to overcome both
mentioned problems. The used architecture is an enhanced U-Net [15] with a Spatial
Transformer Network (STN) [16], with τθ as a 2D affine transformation. The STN
concatenated with the original first skip-connection (SC) used in [15] formed the
Development of a Hybrid Safety System Based on a Machine Learning Approach … 277

used connection to link the first stages of the contracting and the expansive path.
Further SCs are equal to the original. The input size is 2 × 252 × 336, representing
the frames It and It+1 . The output size is 1 × 252 × 336. Hence the convolution
layers use padding to preserve the resolution. The U-Net is mainly used for detecting
occlusions, and the STN is used to reverse potential movement between the frames.
The neural network was trained on captured data (69500 IR images) of the leakage
detection at the demonstrator with a 5-fold cross-validation split. Here, we applied
data augmentation consisting of random vertical and horizontal flipping and random
affine transformations with a fill value of the minimum of the frame. Additional noise
was not used since the detection via Eq. 1 is highly sensitive to temperature changes.
The training procedure described above to detect the leakage caused training
instability due to the extreme imbalance of leakage and non-leakage. Thus, to get
the training to be more stable, the network’s main goal was not to detect the leakage
but to transform the data so that the output frame was congruent to the first frame
and the pixels occluded were equal to the first image. That way, no false detection
via Eq. 1 is made. Figure 3 shows a chart of the approach. The loss function was a
weighted mix between the MeanSquaredError (MSE) for the complete image and
the MSE for the center part in a 1:10 ratio. In Fig. 2 an example of the augmented
input, the target, and the prediction is shown. Reliable leakage detection using only
the approach from Eq. 1 is not possible in actual use cases with moving parts. The
data was split into five equally sized parts with added artificial movement to test
the approach utilizing the neural network described in this section. Then a five-fold
cross-validation was used to evaluate the network’s performance: The data was split
into five equally sized parts, with four parts used for training and the remaining part
used for validation. This was repeated five times, i.e., until each part was used for
validation. That approach achieved over all five folds a mean F1 score for leakage
detection of 0.78 with a minimum of 0.73. Here, the detection via Eq. 1 was used as
ground truth without the added movement.

3.2 Object Detection

For security reasons, some areas are prohibited to people during an ongoing bunkering
process. The following section describes a surveillance solution to effective monitor
these zones under challenging illumination and different environmental conditions
based on RGB and IR cameras.

3.2.1 Approach: Object Detection on RGB Data

For person detection on RGB images, the yolox [17] algorithm is a good choice. The
authors provide pretrained weights for 300 epochs on COCO [18] on their Github
page, which achieves state-of-the-art results. The performance on the chosen test
set is provided in Sect. 3.2.3 and covers the RGB part of the proposed dual camera
278 N. Jathe et al.

Fig. 2 Example images for the network: a First input frame It augmented with random flip and
an affine transform. b The target frame It+1 with the same flip and transform as for the first input
frame. The target frame is unknown to the network. c Second input frame resulting from the target
frame with an additional affine transform. d Output frame of the network. The network solved the
first problem by correctly reverting the second affine transform. Further, the network extracted the
occluding arm on the right side of the second input frame from the image, reducing the possibility
of false positive detections

Fig. 3 The enhanced U-Net detects possible movement from frame It to frame It+1 and adjusts
frame It+1 accordingly. The adjusted frame It+1 and the normal frame It should now be congruent,
and Eq. 1 is used to detect possible leakage

surveillance system. Akula and Sardana [19] indicate that the high variability of IR
signatures is problematic for handcrafted feature detectors and propose to use CNN-
based methods. Thus, this paper proposes an approach based on an adapted yolox
architecture for the IR part.
Development of a Hybrid Safety System Based on a Machine Learning Approach … 279

3.2.2 Approach: Object Detection on Thermal Data

The algorithm uses the raw data of an IR camera and is based on the yolox [17]
approach for object detection in computer vision. The following changes and addi-
tions were made to the initial algorithm:
Architecture adjustments The architecture uses the yolox backbone and has only
one input channel since no color information is given. The head consists of the
decoupled one proposed in yolox [17]. Since the scenario only requires the class
person, the output size was lowered accordingly. Anchors were not adapted to the
training data.
Training and preprocessing details The architecture was trained on the raw IR
data of the training part of the Teledyne FLIR Free ADAS Thermal Dataset v2.1
The dataset contains annotated RGB, thermal raw, and thermal to greyscale images
captured in various locations and under different conditions. For this paper only
annotated objects of the category persons of the 15 provided categories are relevant.
The preprocessing of the raw data consists of a normalizing step to keep the gradients
low and improve learning speed. The unnormalized input of raw data stored as 16-
bit greyscale images resulted in unstable learning at later training stages and weaker
performance. Mapping the image extrema to 0 and 255 and linear interpolating in
between made the network uncertain in cases where variance was high. Artificially
changing the extrema but not altering an object and its surroundings in the same image
resulted in different output confidence values for the network for that object, which
introduces uncertainty due to the normalizing method. The used normalization of the
raw data was the subtraction by the training data’s mean and division by its standard
deviation. The loss function was the same composition of weighted loss functions
provided by yolox [17]. The weights got randomly initialized. The training consisted
of 300 epochs using cosine annealing for learning rate reduction.
Insights into data augmentation techniques Despite having 10742 training images,
the following data augmentation methods were used in the training step for better
results: Applied noise consisting of Gaussian noise with σ = 1 ◦ C and salt and pepper
noise with values from within image range. Further, randomly applied perspective
transforms ranging between 0◦ and 10◦ were performed. Additionally, we used simple
augmentation as random cropping and vertical flipping. Out of [20], Mosaic, and
MixUp were considered. An adapted Mosaic, originally a method to combine four
images into one new training image by stitching cropped versions of the original
images together, was used for training. The adaption randomly combined two, three,
or four images in one or two rows and up to three columns. MixUp is an advanced
augmentation method mainly introduced for visual object detection classification.
It combines two images of different objects by (weighted) cross-fading them into
each other so that both objects are still visible and changes the target score from one
hot encoded to one representing the weights of the object classes to enable more
efficient learning. However, blending two IR images would result in losing the IR
spectrum characteristic of the object, and only the shape characteristics would remain.

1 https://www.flir.com/oem/adas/adas-dataset-agree/.
280 N. Jathe et al.

Table 1 Average precision at 50% intersection over union for the FLIR V2 dataset on the corre-
sponding validation sets for detecting persons. The architectures used the weights provided by Ge
et al. [17] on the colored and the thermal to greyscale images. The architectures handling the raw
IR data used the weights after the training described in Sect. 3.2.2
Used architectures AP50 on FLIR V2
RGB Greyscale Raw
Yolox_nano 11.6 8.8 17.8
Yolox_x 39.6 30.9 60.1

The training process showed that convergence was only possible without applying
MixUp. However, it might be possible to use this technique on trained networks to
boost robustness.

3.2.3 Application and Validation

For comparability, the performance of the proposed algorithms is measured on the

validation part of the Teledyne FLIR Free ADAS Thermal Dataset v2. Table 1 shows
the performance of the used detections. The time needed for inferencing and non-
max suppression on a Nvidia GTX 1080ti for the yolox_nano is 5 ms (∼200 fps) and
for the yolox_x 42 ms (∼24 fps).

3.2.4 Combined Approach: Movement Monitoring to Predict Safety

Violations

To further improve person detection and tracking under various conditions, fusing
the detection on the RGB and the IR data together by using all bounding boxes
of the RGB and the bounding boxes of the IR detection, and applying non-max-
suppression, taking the offset of the cameras into account, makes the overall detection
more robust. This combined approach achieved a slightly better performance than the
two individual approaches on a small intern dataset with known offset. The dataset
is not big enough to quantify the results but indicates that the system can perform a
domain shift, since the data source is different and the training data does not contain
any part of this dataset. With the knowledge of where persons are, the system links on
each timestamp the detected persons with the nearest bounding boxes of the previous
frame. With this information of movement, the system predicts via Kalman filtering
over the midpoints of the bounding boxes the most likely path of the person and can
trigger a warning if the path violates predefined prohibited zones. Also, the system
performs an alarm if the person is located in the prohibited zone. This approach is
simple and was successfully tested in a lab environment with different persons as
test subjects. In Fig. 4 the proposed system is illustrated.
Development of a Hybrid Safety System Based on a Machine Learning Approach … 281

Fig. 4 Bounding boxes marking detected person and the arrows indicating the predicted path,
calculated by a Kalman filter of the previous movement

4 Summary and Conclusion

This research work developed and evaluated a monitoring solution for ensuring safety
during LNG transfer operations. The monitoring solution is based on a combined
RGB and IR camera, which, combined with Deep Learning algorithms, enables
automatic hazard detection within the use cases (detection of persons in the hazardous
area and detection of gas leaks during the bunkering process).
Extensive laboratory tests showed that the selected approach enables reliable
detection of the use cases based on IR images and thus represents a promising alter-
native to a monitoring approach based on RGB images due to its robustness to
maritime environmental conditions.

Acknowledgements The authors would like to thank the German Federal Ministry of Economic
Affairs and Energy (BMWi) for their support within the project “LNG Transfer—LNG Safety”
(grant number 16KN062731).

References

1. Mokhatab, S., Economides, M.: Onshore LNG Production Process Selection (2006). https://
doi.org/10.2118/102160-MS
2. Mokhatab, S., Mak, J.Y., Valappil, J., Wood, D.A.: Handbook of Liquefied Natural Gas. Gulf
Professional Publishing (2013)
3. Stern, H., Leder, R., Lütjen, M., Freitag, M.: Human-centered development and evaluation of an
AR-assistance system to support maintenance and service operations at LNG ship valves, vol.
09, pp. 272–294 (2021). ISBN 978-3-95545-396-1. https://doi.org/10.30844/wgab_2021_17
4. Alderman, J.A.: Introduction to LNG safety. Process. Saf. Prog. 24(3), 144–151 (2005). https://
doi.org/10.1002/prs.10085
5. Foss M.M., Delano, F., Gülen, G., Makaryan, R.: LNG safety and security, Center for Energy
Economics (CEE) (2003)
6. Lin, W., Jiang, L., Haiyan, W.: Study of non-intrusive gas pipeline leak detection with acoustic
sensor. IFAC Proc. Vol. 46(20), 27–32 (2013). https://doi.org/10.3182/20130902-3-CN-3020.
00039
7. Liu, B., Ma, H., Zheng, X., Peng, L., Xiao, A.: Monitoring and detection of combustible
gas leakage by using infrared imaging. In: 2018 IEEE International Conference on Imaging
Systems and Techniques (IST), pp. 1–6 (2018). https://doi.org/10.1109/IST.2018.8577102
282 N. Jathe et al.

8. Fraiwan, L., Lweesy, K., Bani-Salma, A., Mani, N.: A wireless home safety gas leakage detec-
tion system. In: 1st Middle East Conference on Biomedical Engineering (2011). https://doi.
org/10.1109/MECBME.2011.5752053
9. Marshall, Park J.-S., Song, J.-K.: FCN based gas leakage segmentation and improvement
using transfer learning. In: 2019 IEEE Student Conference on Electric Machines and Systems
(SCEMS 2019), pp. 1–4 (2019). https://doi.org/10.1109/SCEMS201947376.2019.8972635
10. Bin, J., Rahman, C.A., Rogers, S., Liu, Z.: Tensor-based approach for liquefied natural gas
leakage detection from surveillance thermal cameras: a feasibility study in rural areas. IEEE
Trans. Ind. Inform. 17(12), 8122–8130 (2021). https://doi.org/10.1109/TII.2021.3064845
11. Wong, W.K., Tan, P.N., Loo, C.K., Lim, W.S.: Machine condition monitoring using omnidirec-
tional thermal imaging system. In: 2009 IEEE International Conference on Signal and Image
Processing Applications, pp. 299–304 (2009). https://doi.org/10.1109/ICSIPA.2009.5478665
12. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: 2016
IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778 (2016). https://
doi.org/10.1109/CVPR.2016.90
13. Xu, K., Yuan, Z., Zhang, J., Ji, Y., He, X., Yang, H.: SF6 gas infrared thermal imaging leak-
age detection based on faster-RCNN. In: 2019 International Conference on Smart Grid and
Electrical Automation (ICSGEA), pp. 36–40 (2019). https://doi.org/10.1109/ICSGEA.2019.
00017
14. Girshick, R.B.: Fast R-CNN. CoRR, abs/1504.08083 (2015)
15. Ronneberger, O., Fischer, P., Brox, T.: U-net: convolutional networks for biomedical image
segmentation. CoRR, abs/1505.04597 (2015)
16. Jaderberg, M., Simonyan, K., Zisserman, A., Kavukcuoglu, K.: Spatial transformer networks.
CoRR, abs/1506.02025 (2015)
17. Ge, Z., Liu, S., Wang, F., Li, Z., Sun, J.: YOLOX: exceeding YOLO series in 2021. CoRR,
abs/2107.08430 (2021)
18. Lin, T.-Y., Maire, M., Belongie, S.J., Bourdev, L.D., Girshick, R.B., Hays, J., Perona, P.,
Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: common objects in context. CoRR,
abs/1405.0312 (2014)
19. Akula, A., Sardana, H.K.: Deep CNN-based feature extractor for target recognition in thermal
images. In: TENCON 2019 - 2019 IEEE Region 10 Conference, pp. 2370–2375 (2019). https://
doi.org/10.1109/TENCON.2019.8929697
20. Bochkovskiy, A., Wang, C.-Y., Liao, H.-Y.M.: Yolov4: optimal speed and accuracy of object
detection. CoRR, abs/2004.10934 (2020)
A Comparative Study on Damage
Detection in the Delta Mooring System
of Spar Floating Offshore Wind Turbines

Christos S. Sakaris, Anja Schnepf, Rune Schlanbusch, and Muk Chen Ong

Abstract The most common type of Floating Offshore Wind Turbine (FOWT)
installed in Norway, is the spar FOWT in which a delta mooring system (DMS)
is used. The mooring system of a FOWT is an essential part for its station-keeping,
whose loss can lead to the collapse of the FOWT and the endangerment of the
human safety. Thus, early detection of damages in the mooring system is vital.
In this study, damage detection in the DMS of a spar FOWT under varying envi-
ronmental conditions (ECs) is investigated through a comparison of the Multiple
Model-AutoRegressive (MM-AR) method, the Multiple Model-Power Spectral Den-
sity (MM-PSD) method and the Functional Model Based Method (FMBM). The
MM-AR and the MM-PSD methods are based on multiple PSD based or AR models
and the FMBM on a single Functional Model (FM) for the description of the healthy
FOWT’s dynamics under varying ECs. The results show that successful and precise
damage detection in a spar FOWT’s DMS can be achieved through the employed
statistical methods as the MM-AR method and the FMBM detect all the examined 7
healthy and 16 damage cases whereas the MM-PSD method misses only one damage
case. The results also show that the parametric AR models and FM describe more pre-
cisely the FOWT dynamics under varying ECs in comparison to the non-parametric
PSDs.

C. S. Sakaris (B) · R. Schlanbusch

Technology Department, Norwegian Research Centre, Jon Lilletuns vei 9 H, 3. et,
4879 Grimstad, Norway
e-mail: [email protected]
R. Schlanbusch
e-mail: [email protected]
A. Schnepf · M. C. Ong
Department of Mechanical and Structural Engineering and Materials Science,
University of Stavanger/CoreMarine, Stavanger, Norway
e-mail: [email protected]
M. C. Ong
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 283
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_25
284 C. S. Sakaris et al.

1 Introduction

Floating Offshore Wind Turbines (FOWTs) supported by a spar type platform, are
most commonly used in Norway. Currently, the only operational spar FOWT in
Norway is the Zephyros FOWT, formerly Hywind Demo, which has been developed
by Equinor and it is connected to the onshore grid [1]. However, a floating wind farm
consisted of spar FOWTs, named Hywind Tampen, is under construction and it will
be connected to power offshore oil and gas facilities [2]. The station-keeping system
of the aforementioned spar FOWTs is a delta mooring system (DMS), also called
crowfoot mooring system, consisting of three chain mooring lines.
The mooring lines constitute a critical part in maintaining the station-keeping of
a FOWT and thus early detection of damages in this part is vital. Otherwise, failure
to detect a damage may lead to an increase in the mooring lines’ tension, loss of
stability, high maintenance cost, and possibly to the FOWT’s collapse. However,
damage detection in the mooring lines of FOWTs has been investigated in very few
studies under constant [3] and varying [4–6] environmental conditions (ECs) and
through methods based on data-based models (developed with acquired signals from
the structure) such as Power Spectral Density (PSD) [3, 5] and Neural Network [4,
6]. In the baseline (training) phase of some of these methods, the use of a large
number of data records under different ECs is required [6]. Additionally, there are
methods in which it is not clear how the different ECs are handled in the baseline
phase [5]. Also only damage cases are examined in the inspection (real time) phase of
some methods, thus leaving the methods’ effectiveness in damage detection unclear
[4, 5]. These methods treat damage detection as a classification problem in which
an unknown state is classified as a healthy or a damage state.
Recently, damage detection in the tendons of the multibody TELWIND platform
of a 10 MW FOWT under varying mean wind velocity (MWV) and significant wave
height (SWH), has been investigated through the Functional Model Based Method
(FMBM) [7].
The aim of the present study is the investigation of damage detection in a spar
FOWT’s DMS through a comparison of data-based statistical methods. These are
the Multiple Model-AutoRegressive (MM-AR) method, Multiple Model-PSD (MM-
PSD) method [8] and the FMBM [7] which have been applied for damage detection
on different structures successfully and they are part of the Statistical Time Series
(STS) methods. The STS methods constitute an effective tool for robust damage
detection and they present a number of advantages such as needing no physics-
based models, achieving damage detection in a statistical decision context and with
partial data-based models developed with response signals from a limited number of
sensors, taking into account of various types of uncertainties through statistical tools
and getting applied without interrupting the inspected structure’s normal operation
[9].
The spar FOWT used in the present study constitutes a modification of the artifi-
cial 5 MW OC3-Hywind FOWT [10]. Various cases with the healthy and damaged
A Comparative Study on Damage Detection in the Delta Mooring System … 285

FOWT under varying MWV and SWH are examined and the simulated damages are
considered at different locations in a single delta line and of different magnitudes.

2 The Spar FOWT and the Simulations

The spar FOWT. In the current study, the examined spar FOWT is a modified 5
MW OC3-Hywind FOWT which is based on the Hywind concept [10]. The artificial
OC3-Hywind FOWT has been studied extensively in the literature [11, 12] with its
original mooring system being a plain catenary mooring system and consisting of
three quasi-static catenary lines and an artificial yaw stiffness.
The modification consists of the replacement of the original mooring system with
the DMS which is part of the design of the spar FOWTs Zephyros and Hywind
Tampen. In the DMS, each of the three mooring lines consists of two delta lines
and a catenary line that are connected through a delta plate. The delta lines are also
connected to two fairleads on the spar platform whereas the catenary line is anchored
at a water depth of 320 m, as shown in Fig. 1a. The properties of the catenary lines
of the modified spar FOWT are the same as for the original OC3-Hywind FOWT’s
catenary lines [10], whereas the properties of the delta lines are the half of the catenary
lines’ properties.
Simulations and damages. MWV and SWH are considered as varying ECs in
this study and various healthy and damage cases are examined for seven pairs of
MWVs/SWHs shown in Table 1. Each considered damage case is simulated by a
stiffness reduction of a specific magnitude (%) at a specific location in a delta line
of a mooring line. The considered damage magnitudes correspond to 30 and 70%
stiffness reduction, whereas the considered damage locations correspond to 3 and
16% of the delta line’s length (see Fig. 1b). Due to the dependence of SWH on MWV,
each healthy case is represented as F w with w denoting the WV and each examined
w
damage case is represented as Fm,q with m being the damage magnitude and q the
damage location.
The numerical model of the modified OC3-Hywind FOWT is set up in the numer-
ical software OrcaFlex according to [13]. Based on this model, heave acceleration
signals are generated under the considered healthy and damage cases. In relation
to the varying MWV and SWH considered during the simulations, the MWV is
computed based on the power law velocity profile in [14]. The Pierson-Moskowitz
spectrum is used for generating fully developed sea states in deep water and the

Table 1 Values of the varying mean wind velocity (MWV) and significant wave height (SWH)
MWV 8 10.3 11.4 13.8 15 17.3 18
(m/s)
SWH (m) 0.85 1.4 1.72 2.52 2.98 3.96 4.29
286 C. S. Sakaris et al.

Fig. 1 a The Spar FOWT and b the position of the sensor and the considered damages

values of the spectrum’s SWH and the peak period are obtained with the empirical
formulations in [15, pp. 150–155]. The surface current velocity is 0.5 m/s and the
mean current velocity throughout the entire water depth is calculated based on the
power law profile in [16].
During each simulation, an acceleration signal is generated based on the location
of one sensor placed on the delta plate, as shown in Fig. 1b. Table 2 presents the
details and the numbers of the conducted simulations. It must be noted that the
signals corresponding to the 7 baseline phase simulations under the healthy FOWT
are used for the training of each method and they are different from their counterparts
in the inspection phase. Each signal is mean corrected and normalized based on the
sample’s standard deviation.
Effects of varying MWV, SWH and damages on FOWT’s dynamics. The MWV
variability leads to a change in the FOWT’s dynamics. Moreover, a comparison of
11.4 11.4
PSDs in Fig. 2 shows that damages of different magnitude F30,3 , F70,3 and the healthy
state under the same MWV have almost the same effects on the dynamics. This shows
that some damages have small effects and their detection is very difficult. The PSDs
are estimated based on the Welch estimator [7] (Window length 400 samples, 0%
overlap, Hamming window, frequency resolution of 0.025 Hz).
A Comparative Study on Damage Detection in the Delta Mooring System … 287

Table 2 Number of simulations per healthy/damage state and details of vibration signals
Structural state No. of simulations—baseline phase No. of simulations—inspection
phase
Healthy 7 (one per F 8 , F 10.3 , F 11.4 , F 13.8 ,
7 (one per F 8 , F 10.3 , F 11.4 , F 13.8 ,
F 15 , F 17.3 , F 18 ) F 15 , F 17.3 , F 18 )
Damaged – 10.3 , F 10.3 , F 10.3 ,
9 (one per F30,16 30,3 70,16
10.3 8 8 , F 11.4 , F 11.4 ,
F70,3 , F30,3 , F70,3 30,3 70,3
17.3 )
F30,3
Sampling frequency: f s = 10 Hz, Signal bandwidth: [0–1.4] Hz
Signal length: N = 20000 samples (≈1000 s)

Fig. 2 Comparison of Welch-based PSD estimates using heave acceleration signals for healthy &
damage states under MWV 11.4 m/s

3 Damage Detection Methods

In the present study, unsupervised damage detection in the DMS of a spar FOWT
under varying MWV and SWH is investigated through a comparison of three data-
based methods. In these methods, the healthy FOWT’s dynamics under varying ECs
is represented by features in a proper subspace and damage detection is achieved by
checking if the current dynamics resides in the subspace (see Fig. 3). In the MM-
AR and MM-PSD methods, the subspace is described by multiple models such as
Auto-Regressive (AR) models and PSDs correspondingly and the features are the AR
parameters or PSD magnitude values [8]. In the FMBM, the subspace is described
by a single Functional Model (FM) and the features are the FM’s parameters [7].
The methods consist of two phases, the baseline phase which is performed based on
data from known structural states and the inspection phase which is based on current
data while the structure is under an unknown health state.

3.1 MM-AR and MM-PSD Methods [8]

Baseline phase. Initially in this phase, M response signals corresponding to a sample

of the considered MWVs, are acquired from the healthy FOWT. Due to the depen-
dence between MWV and SWH, only MWV is used. The sampled MWVs cover the
288 C. S. Sakaris et al.

Fig. 3 General concept of damage detection through the methods MM-PSD, MM-AR and FMBM

range [wmin , wmax ] via the discretization wr ∈ w1 , w2 , . . . , w M . Then each signal

characterized by a specific MWV wr , is used for estimating a PSD for the MM-PSD
method [7] or an AR model for the MM-AR method via standard identification pro-
cedures [7]. A feature vector mo,wr is obtained from each model and it contains the
PSD magnitude values of the frequencies in the considered bandwidth or the AR
model parameters.
The effectiveness of the MM-AR method may be improved if the AR parameters
with variance over a user-defined threshold are excluded from the method’s feature
vector, as the increased variances indicate that these parameters are more sensitive
to the varying MWV. The variance of each parameter is based on the AR models
corresponding to the considered MWVs. The same procedure can be followed for
the improvement of the MM-PSD method where the frequencies with PSD variance
over a user-defined threshold are excluded.
Inspection phase. A new response signal is acquired under the FOWT’s current
(unknown) health state and it is used for estimating a new model (PSD or AR model)
and obtaining a new feature vector mu . Then the distance fr between mu and each
feature vector mo,wr from the baseline phase, is calculated and it is used for obtaining
the distance metric Q with Q= min fr [8]. Feature vectors mu and mo,wr must have
r
the same length. Finally, a damage in the FOWT is detected when Q is greater than
a user-defined limit that is selected based on the values of Q in the baseline phase. In
the current study, two distances fr are used, the Euclidean distance in the MM-PSD
method and the Mahalanobis distance in the MM-AR method.

3.2 FMBM [7]

Baseline phase. In this initial phase, a Functionally Pooled-AutoRegressive (FP-

AR) model is estimated based on M response signals from the healthy FOWT
corresponding to a sample of the considered MWVs [7] (the same M signals used in
A Comparative Study on Damage Detection in the Delta Mooring System … 289

the baseline phase in Sect. 3.1). The structure of a FP-AR(na) p model is the following
[7]:


na
 
ywr [t] + ai (wr ) · ywr [t − i] = ewr [t], ewr [t] ∼ iid N 0, σe2 (wr ) (1)
i=1

with na designating the AR order, respectively, ywr [t] (t = 1, . . . , N ) the response

signal with length of N samples under WV wr and ewr [t] the disturbance signal that is
a white (serially uncorrelated), normal distributed
 p (N ), zero-mean signal with vari-
ance σe2 (wr ). The AR parameters ai (wr ) = j=1 ai, j · G j (wr ), are explicit func-
tions of wr by belonging to a p-dimensional functional subspace spanned by the
(mutually independent) functions G 1 (wr ), G 2 (wr ), . . . , G p (wr ). These functions
are polynomials of one variable (like Chebyshev, Legendre, Jacobi etc.). The ai, j
designates the AR coefficients of projection. The model is estimated through iden-
tification procedures for the selection of the model orders, the determination of the
model’s functional subspace dimensionality p and the model validation [7]. In the
context of the p determination, the ai, j are estimated based on linear regression tech-
niques such as Ordinary Least Squares (OLS) [7] or Weighted Least Squares (WLS)
[17].
Inspection phase. A response yu [t] signal corresponding to a known/measured
MWV w, is acquired under the FOWT’s current (unknown) state. Then yu [t] and
w are used in the re-parametrized FP-AR(na) p model (from the baseline phase) in
terms of w and the residual signal e[t, w] is obtained [7]:


na
yu [t] + ai (w) · yu [t − i] = e[t, w] (2)
i=1

Damage detection is achieved by checking the uncorrelatedness (whiteness) of

e[t, w] through a statistical hypothesis test. If e[t, w] is white then the FOWT is
in a healthy state, otherwise a damage is detected. In the current study, the Pena-
Rodriguez hypothesis test is employed for the whiteness check and a damage is
detected when the test’s normal distributed quantity D exceeds the distribution’s
critical limits, −Z 1−α >D and/or Z 1−α <D (−Z 1−α , Z 1−α the critical limits, α the
risk level) [7].

4 Damage Detection Results

MM-AR and MM-PSD methods. In the MM-PSD method’s baseline phase, a

Welch-PSD estimate (Matlab function: pwelch.m) is obtained based on each of the
M = 7 acceleration signals from the healthy FOWT and under a different MWV
(see Table 2). The considered MWVs cover the range wr ∈ [8, 18] m/s through the
290 C. S. Sakaris et al.

Fig. 4 a PSD variance per frequency and b AR parameter variance per parameter. Damage detection
results based on the c MM-PSD & d MM-AR methods and e the FMBM with limits (—) based
on risk level α = 4.25 · 10−1

discretization wr ∈ [8, 10.3, 11.4, 13.8, 15, 17.3, 18] m/s. Based on these PSDs, the
PSD variance per frequency is determined and the feature vectors of reduced dimen-
sion are obtained. In the MM-AR method’s baseline phase, an AR(260) model is
estimated (Matlab function: arx.m) based on each of the M = 7 signals, the model
parameters and their covariance matrices are extracted and the reduced feature vec-
tors are obtained to be used in the Mahalanobis distance. Figure 4a, b depict the
PSD variance per frequency and the variance per AR parameter with the excluded
frequencies and parameters corresponding to variances above the designated limits
(horizontal lines). The details of the employed models and reduced feature vectors
are shown in Table 3.
In the inspection phase of both of the methods, damage detection is examined
based on 7 healthy cases and 9 damage cases (see Table 2). The damage detection
A Comparative Study on Damage Detection in the Delta Mooring System … 291

Table 3 Baseline phase details of the damage detection methods

Detection Estimated Estimation No.of No.of Feature vector dimen- Distance
method model method signals models sion
Initial Reduced
MM-PSD PSD Welch M =7 7 [57 × 1] [50 × 1] Euclidean
MM-AR AR(260) OLS M =7 7 [260 × 1] [255 × 1] Mahalanobis
FMBM FP- WLS M =7 1 – – –
AR(140)6
PSD Welch-based estimation: segment length 400 samples, overlap 0%, window type
Hamming, frequency resolution 0.025 Hz
Condition number: 6.54 · 106 –2.19 · 109 (AR models), 2.97 · 1013 (FP-AR model)
FP-AR(140)6 model’s functional subspace: p = 6 univariate Shifted Legendre polynomials

results are presented in terms of Q in Fig. 4c, d with no false alarms occurring in both
of the methods. As for missed damages, the MM-PSD method fails to detect only
one damage case whereas the MM-AR method is able to detect all of the damage
cases.
FMBM. In the baseline phase, an FP-AR(140) model is estimated based on the
M = 7 acceleration signals (see MM-AR and MM-PSD methods of Sect. 4). The
selected functional subspace consists of p = 6 univariate Shifted Legendre polyno-
mials. Thus, the identified model may represent the healthy FOWT under any MWV
in the considered range [8–18] m/s. Details of the FP-AR(140)6 model are presented
in Table 3.
In the inspection phase of the FMBM, damage detection is examined based on
the same 16 cases used in the MM-AR and MM-PSD methods of Sect. 4. Figure 4e
depicts the damage detection results in terms of D and no false alarms or missed
damages occur.

5 Conclusions

Damage detection in the DMS of a spar FOWT has been investigated in the present
study. This investigation has been conducted through a comparison of three data-
based statistical methods: the MM-AR method, the MM-PSD method and the
FMBM. Based on the examined healthy and damage cases under varying MWV
and SWH, the MM-AR method and FMBM have been able to achieve a perfect dam-
age detection whereas the MM-PSD method has missed only one damage. This has
happened because the AR models (MM-AR) and FMs (FMBM) have been able to
describe the FOWT dynamics more precisely in comparison to the PSDs (MM-PSD)
and that some damages have had small effects on the structural dynamics. Future
steps include the examination of damage detection and quantification in the DMS
292 C. S. Sakaris et al.

through MWVs considered only in the inspection phase (not in the baseline phase)
and the use of a larger variety of damage magnitudes and data-based methods.

Acknowledgements This research work has been funded by the Research Council of Norway
through the projects: (i) Analytics for asset Integrity Management of Windfarms (AIMWind), a
collaboration between University of Agder, Norwegian Research Center AS and Technical Univer-
sity Delft, grant no. 312486. Origo Solutions is included as advisory partner. (ii) Design Optimisation
of Power Cable, Shared Electrical Line and Mooring configurations for Floating Offshore Wind
Turbines, grant no. 320902.

References

1. Driscoll, F., Jonkman, J., Robertson, A., Sirnivas, S., Skaare, B., Nielsen, F.: Validation of
a FAST model of the Statoil-hywind demo floating wind turbine. Energy Procedia 94, 3–19
(2016)
2. Equinor: Hywind Tampen, PL050 - PL057 - PL089, PUD del II - Konsekvensutredning. Tech.
rep., Statoil (2019)
3. Jamalkia, A., Ettefagh, M., Mojtahedi, A.: Damage detection of TLP and spar floating wind
turbine using dynamic response of the structure. Ocean Eng. 125, 191–202 (2016)
4. Dehkharghani, P., Ettefagh, M., Hassannejad, R.: Mooring damage identification of floating
wind turbine using a non-probabilistic approach under different environmental conditions. J.
Mar. Sci. Appl. 20, 156–169 (2021)
5. Liu, Y., Ferrari, R., Wu, P., Jiang, X., Li, S., Wingerden, J.: Fault diagnosis of the 10MW
floating offshore wind turbine benchmark: a mixed model and signal-based approach. Renew.
Energy 164, 391–406 (2021)
6. Gorostidi, N., Nava, V., Aristondos, A., Pardo, D.: Predictive maintenance of floating offshore
wind turbine mooring lines using deep neural networks. J. Phys.: Conf. Ser. 2257 (2022)
7. Sakaris, C., Yang, Y., Bashir, M., Michailides, C., J., W., Sakellariou, J., Li, C.: Structural
health monitoring of tendons in a multibody floating offshore wind turbine under varying
environmental and operating conditions. Renew. Energy 179, 1897–1914 (2021)
8. Kaliorakis, N., Iliopoulos, I., G., V., Sakellariou, J., Fassois, S., Deloukas, A., Leoutsakos, G.,
Chronopoulos, E., Mamaloukakis, C., Katsiana, K.: On the on-board random vibration-based
detection of hollow worn wheels in operating railway vehicles. Lect. Notes Civ. Eng. 128,
480–489 (2020)
9. Fassois, S.D., Sakellariou, J.: Statistical time series methods for structural health monitoring.
In: Boller, C., Chang, F.K., Fujino, Y. (eds.) Encyclopedia of Structural Health Monitoring, pp.
443–472. Wiley, Chichester (2009)
10. Jonkman, J.: Definition of the floating system for phase IV of OC3. Tech. Rep. NREL/TP-500-
47535, National Renewable Energy Laboratory, U.S. Department of Energy Office of Energy
Efficiency & Renewable Energy (2010)
11. El Beshbichi, O., Xing, Y., Ong, M.: An object-oriented method for fully coupled analysis of
floating offshore wind turbines through mapping of aerodynamic coefficients. Mar. Struct. 78,
102979 (2021)
12. Xu, X., Day, S.: Experimental investigation on dynamic responses of a spar-type offshore
floating wind turbine and its mooring system behaviour. Ocean Eng. 236 (2021)
13. Schnepf, A., Lopez-Pavon, C., Devulder, A., Johnsen, Ø., Ong, M.: Suspended power cable
configurations for floating offshore wind turbines in deep water powering an FPSO. In: Proceed-
ings of the 41st International Conference on Ocean, Offshore & Arctic Engineering (OMAE).
Hamburg, Germany (2022)
A Comparative Study on Damage Detection in the Delta Mooring System … 293

14. Li, H., Ong, M., Leira, B., Myrhaug, D.: Effects of soil profile variation and scour on structural
response of an offshore monopile wind turbine. J. Offshore Mech. Arct. Eng. 140(4), 042001
(2018)
15. Holthuijsen, L.: Waves in Oceanic and Coastal Waters. Cambridge University Press (2007)
16. Orcina: OrcaFlex. https://www.orcina.com/webhelp/OrcaFlex/ (2022)
17. Sakellariou, J., Fassois, S.: Functionally pooled models for the global identification of stochastic
systems under different pseudo-static operating conditions. Mech. Syst. Signal Process. 72–73,
785–807 (2016)
Robust and Fault Tolerant Control
A Cooperative Centralized MPC
for Collision Avoidance of a Fleet of AVs

Manel Ammour, Rodolfo Orjuela, and Michel Basset

Abstract An original centralized model predictive controller (CMPC) with time

varying Sigmoid-based safety constraints is developed for the control of a fleet of
autonomous vehicles (AVs) and presented in this paper. The defined CMPC frame-
work generates the appropriate control inputs for each vehicle in the network. This
formulation allows each vehicle to attain its objective (reference lane tracking and
driving at the desired speed) while maintaining a safety inter vehicular distance.
The results of an illustrative scenario are presented to show the consistency of the
proposed algorithm.

1 Introduction

Multi-agent systems have gained great interest in recent years due to their potential
in enhancing traffic safety and efficiency in executing their missions such as area
exploration [12], surveillance [1], rescue [6], formation control on unmanned ground
vehicles (UGVs) [2], etc.
Cooperative control requires the consideration of the behavior of each agent in
the network. To this aim, many approaches have been investigated and classified into
three main schemes: centralized, distributed, and decentralized. For the centralized
scheme, a single controller receives the information from all the agents and computes
the necessary control inputs for each of them [7]. In the distributed scheme, each
agent has its own controller which uses its state information and that of its neighbors
to compute its appropriate control inputs [7]. Finally, in the decentralized scheme,

M. Ammour (B) · R. Orjuela · M. Basset

Institut de Recherche en Informatique, Mathématique, Automatique et Signal (IRIMAS UR
7499), 68093 Mulhouse, France
e-mail: [email protected]
R. Orjuela
e-mail: [email protected]
M. Basset
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 297
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_26
298 M. Ammour et al.

each agent has its own controller and measures only its own states [11]. In this paper,
a centralized architecture is adopted to take into account the states of all the agents
and to avoid conflicts in the generated control actions.
The presence of more than one agent in a small area to accomplish a specific
mission invokes the consideration of collision avoidance to ensure the successful
completion of the agents’ missions. Consequently, collision avoidance in crowded
and dynamic environments is one of the important topics addressed by researchers.
A quadratic programming is formulated in [15] to handle collision avoidance of a
multi-robot system, in which centralized safety barrier certificates are established to
dictate how the nominal controller can be changed to avoid pairwise robot collisions.
Then, a decentralization of the safety barriers is performed by making individual
robots consider only nearby robots while guaranteeing conservativeness of solution.
The problem of formation control is investigated in [9] based on a distributed MPC
for multiple linear second order agent systems with collision avoidance and safety
constraints while ensuring the closed loop stability. Decentralized algorithms are
also employed to develop collision avoidance strategies where each agent computes
its avoid set with respect to the others and takes action when one or more agent
approaches its boundary [10]. A group of autonomous flying vehicles is controlled
through a distributed cooperative MPC control framework in [8] where the cost func-
tion of the MPC is designed to integrate the mission objectives (collision avoidance,
reaching the targets, traveling as a flock). A different strategy is adopted in [13],
where a message passing algorithm for cooperative collision avoidance is devel-
oped. Each vehicle in the network is controlled by a MPC controller so that it can
achieve its objective and avoid colliding with a nearby vehicle. Convoy control of
autonomous vehicle is presented in [14] using a two level MPC architecture where
the high level MPC is devoted to the convoy formation and the generation of the
reference trajectory for each vehicle and the low level MPC is for tracking purposes.
This study is an extension of our prior work [5], where a single vehicle is controlled
in the presence of neighboring vehicles. In this work, a networked model predictive
control (Net-MPC) strategy is proposed to control a fleet of autonomous vehicles.
Typically, a cooperative centralized architecture is developed. The aim here is to
control the vehicles so that each vehicle follows its reference without colliding with
other vehicles. The use of Net-MPC is motivated by its predictive nature and its
ability to forecast the positions and velocities of all vehicles. Hence, networked-
based trajectory planning is reactive since the provided control inputs are based on
the current and the predicted behaviors of the agents. The main contribution of this
paper is the proposal of an online cooperative collision avoidance trajectory planning
based on a centralized networked model predictive control architecture, using a set of
Sigmoid-based safety constraints to drive each vehicle to a collision free zone when
the safety inter vehicular distance is not kept. The suggested formulation guarantees
the safe navigation and the online generation of a collision free trajectory for each
vehicle.
The paper is organized as follows: the description of the problem treated in this
paper is presented in Sect. 2. The plant modelling is described in Sect. 3 while the
A Cooperative Centralized MPC for Collision Avoidance of a Fleet of AVs 299

centralized networked MPC formulation is introduced in Sect. 4. The simulation

results are discussed in Sect. 5 and the conclusion is stated in Sect. 6.

2 Problem Formulation

his paper deals with the autonomous driving of a fleet of autonomous vehicles. The
aim is to ensure the cooperative motion while guaranteeing a safety inter-vehicular
distance for all vehicles. This study considers a network of N autonomous vehicles
driving on a two-lane one-way road. An illustration of a network composed of four
autonomous vehicles is presented in Fig. 1.
Each vehicle is desired to drive at its own target speed on its reference lane while
respecting the safety distance. For this purpose, the vehicles must communicate
their states to the center. The latter then computes the appropriate control actions to
control each vehicle in the network and drive them to a collision free zone, when
necessary. The consistency of the solutions provided by the center depends highly
on the cooperation of the vehicles. This feature is implicitly taken into account
by the controller adopted in this work. It is handled by the fact that the predicted
positions and reference trajectories, as well as the current states of all the agents, are
transmitted to the Net-MPC controller. Moreover, these predictions are taken into
account by the Net-MPC controller when computing the control actions to guarantee
the safe navigation of the agents. In this work, the problem is addressed through a
centralized architecture. The modeling of this architecture can be made using graph
theory.

Definition 1 A graph G = (V, E) is a pair that consists of a set of vertices V

and a set of edges E ∈ V × V. The number of vertices of a graph is denoted by
N = |V| ∈ N and the number of edges by E = |E| ∈ N. A vertex is denoted by
vi ∈ V = v1 , v2 , . . . , v N and an edge by ei j = (vi , v j ) ∈ E, where vi , v j ∈ V [3].

Fig. 1 A network of autonomous vehicles on a two-lane one-way road

300 M. Ammour et al.

Fig. 2 Example of the

CMPC graph of the
considered vehicle network

Each vehicle in the network represents a vertex vi in the graph. The communication
between the vehicles and the center is represented by edges. The interaction between
agents is also represented by edges (coupling), as depicted in Fig. 2. In this paper,
Sigmoid based safety constraints [5] form the coupling between the agents.
These safety constraints are integrated into the Centralized MPC (CMPC) frame-
work to generate the adequate solution to the optimization problem. These safety
constraints allow the convexification of the driving areas of each agent by prohibit-
ing it from driving in the vicinity of another agent occupying the same lane. The
detailed description of this formulation will be presented in the next section.
It is worth to mention that the following assumptions are considered: (1) all the
vehicle states are measurable, (2) the communication between the agents is well
established, (3) the study focuses on the case where a lane change maneuver is
possible, i.e., no lane is blocked, (4) each vehicle must take action with respect to
its leading vehicle. Hence, the leader of each lane is supposed to be greedy and does
not change its behavior.

3 AV Fleet Modelling

The purpose of the proposed algorithm is to control the motion of a fleet of

autonomous vehicles and to ensure their navigation in a collision free area. To this
end, the kinematic of each vehicle in the network is described by a point mass
model. Consider a set of N agents. Let xveh i =[xi , yi , vxi v yi ]T denotes the state vector
of each agent i ∈ N , where xi , yi are the longitudinal and lateral positions respec-
tively, and vxi , v yi are the longitudinal and the lateral velocities, respectively. Let
u veh i = [axi , a yi ] is the control input vector of the ith vehicle in the network, where
axi , a yi are the longitudinal and lateral
 accelerations,
 respectively. Initially, each agent
has its desired center-line yr e fi ∈ yl1 , yl2 . This reference may be changed depend-
ing on the lateral gap between the yi position and the closest lane of the ith vehicle
during a collision avoidance maneuver.
A Cooperative Centralized MPC for Collision Avoidance of a Fleet of AVs 301

In a state space form, the longitudinal and lateral motions of an agent vi ∈ V at

each time step T are described as follows:

xveh i (k + 1) = Ai xveh i (k) + Bi u veh i (k), i = 1, . . . , N (1a)

Yi (k) = Ci xveh i (k) (1b)

where:
⎡ ⎤ ⎡ ⎤
1 0 T 0 0 0
⎢0 1 0 T ⎥ ⎢ 0 0 ⎥
Ai = ⎢ ⎥, Bi = ⎢ ⎥ , Ci = 0 1 0 0
⎣0 0 1 0 ⎦ ⎣T 0 ⎦ 0010
0 0 0 1 0 T

The centralized augmented model used to predict the future states of the vehicles is:

Ac 0 X veh (k) B
X veh tot (k + 1) = + c U (k) (2)
0 I Yr e f (k) 0

with:
⎡ ⎤ ⎡ ⎤
A1 B1
⎢ .. ⎥ ⎢ .. ⎥
Ac = ⎣ . ⎦ , Bc = ⎣ . ⎦ (3)
AN BN
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
xveh 1 yr e f1 u veh 1
⎢ ⎥ ⎢ . ⎥ ⎢ . ⎥
X veh = ⎣ ... ⎦ , Yr e f = ⎣ .. ⎦ , U = ⎣ .. ⎦ (4)
xveh N yr e f N u veh N

4 Networked Model Predictive Control Path Planning

The objective of this work is to ensure the safe driving of each vehicle in the network.
In fact, each vehicle in the network aims to: (1) follow its reference center-line yr e fi ,
(2) drive at the desired speed vdesi , (3) navigate in a collision free area and avoid
colliding with the neighboring vehicles, (4) satisfy vehicle dynamic constraints and
road limits.
The objectives 1 and 2 are formulated as a minimization of a cost function within
the MPC framework, described as follows:

N N p −1 Nc −1
min J = ( αi (yik − yrke fi )2 + βi (vxki − vxkdes )2 + γi (u veh i )2 )
u(.) i
i=1 k=0 k=0
302 M. Ammour et al.

where αi , βi , γi are positive weighting coefficients, N p and Nc are the prediction

and control horizon, respectively. The control input u i = [axi , a yi ]T denotes
the increment in the longitudinal and lateral acceleration of the ith agent in the
network. The first term αi (yik − yrke fi )2 is used to penalize the deviation from the
lateral coordinate of the reference center-line and achieve objective 1, while the term
βi (vxki − vxkdes )2 satisfies objective 2. The objectives 3 and 4 are satisfied by the con-
i
straints explained in the following section.

4.1 Dynamic Constraints

Many limitations should be handled to have realistic and feasible control actions
provided by the CMPC. The considered constraints are presented hereafter. ∀vi ∈ V:
Actuator constraints. The following conditions are taken into account to ensure
that the control inputs are within the admissible range and to attain objective 4:

aximin < axi < aximax (5a)

aximin < axi < aximax (5b)

a yimin < a yi < a yimax (5c)

a yimin < a yi < a yimax (5d)

Velocity limits. Each vehicle should drive with a speed within the allowed range
specified by the traffic rules and vehicle capabilities. To this end, the following
constraint is established for each agent:

vximin < vxi < vximax (6)

4.2 Safety Constraints

The safety constraints aim to ensure that each vehicle can navigate the road without
colliding with its boundaries and nearby agents.Therefore, to achieve objective 3,
the following constraints are imposed. ∀vi ∈ V:
Road boundaries. Each vehicle should stay within the road limits and avoid hitting
the road edges by respecting the following constraint:
wi wi
yimin + < yi < yimax − (7)
2 2
A Cooperative Centralized MPC for Collision Avoidance of a Fleet of AVs 303

Fig. 3 The delimited navigable as well as non navigable zones for each vehicle in the network
using Sigmoid barriers

where the term w2i refers to the half width of vehicle vi to take into account the vehicle
size.
Sigmoid-based safety constraints. The proposed collision avoidance strategy is
based on Sigmoid safety barriers. These constraints encounter a non navigable zone
around an agent as depicted in Fig. 3.
More precisely, a lane change maneuver and a navigation to a collision-free zone
are required when a vehicle vi ∈ V approaches its lead vehicle v j ∈ V and tends
not to maintain the required safe distance. Each pair of vehicles within a network
lane must meet the Sigmoid-based safety constraint. Therefore, ∀vi , v j ∈ V:

ylat
yi ≥ ± −ζi (−xi +S fi )
(8)
1+e

with : yi = yi − yr e fi and xi = x j − xi the longitudinal distance between vehicle

vi and its leading vehicle v j . The sign of the Sigmoid-based barrier depends on which
lane vehicle vi is located. It is positive (+) when the vehicle is in the right lane and
it is desirable to execute a left lane change maneuver and negative (−) when it is in
the left lane and wants to perform a right lane change maneuver.
The desired lateral gap ylat represents the distance that should be kept between
two agents in the case where the longitudinal safety distance is not maintained and
a collision avoidance maneuver is initiated. The shape of the lane change maneuver
executed by an agent vi to avoid colliding with its leading vehicle v j is adjusted by the
geometric parameter ζi in (8). This parameter depends on vehicle dynamics and road
state (icy, wet and dry) and the longitudinal velocity to guarantee the feasibility of the
lane change maneuver. It is modified in real-time according to the driving situation.
The tuning of this parameter is detailed in the previous work [4]. The reader is
encouraged to consult the mentioned work to have more details about the derivation
of this parameter. The element S fi in (8) is time varying longitudinal safety distance
that should be kept between each pair of vehicles. This safety distance depends on
the speed of the considered agent and the inter vehicular time (TIV) that must be
maintained. It is stated here, as follows:
304 M. Ammour et al.

S fi = vxi TIV (9)

When a vehicle executes a lane change maneuver, its reference lane is changed to
the nearest lane.

5 Simulation and Results

Numerous scenarios with different initial positions and speeds, as well as various
reference lanes and target speeds, were tested to assess the proposed algorithm. An
illustrative scenario of the proposed CMPC collision free strategy for a fleet of four
autonomous vehicles is selected and presented in this section. Each vehicle is driving
in its reference center-line and is assigned a desired speed. The initial positions and
velocities of each vehicle are summarized in Table 1 and the design parameters of
the centralized MPC are reported in Table 2. In this scenario, vehicles 1 and 2 are
driving in the left lane and are following yl1 = 5.25 m, whereas vehicles 3 and 4
are in the right lane and have yl2 = 1.75 m as starting center-line reference. Initially,
vehicle 1 is moving at higher speed than vehicle 2. They are separated by 30 m,
hence the required safety distance S f1 = T I V vx1 = 60 m is not maintained. In order
to avoid collision, vehicle 1 performs a right lane change and modifies its reference
lane to reach the collision free zone delimited by its associated Sigmoid barrier.
Additionally, it decelerates to reach its desired speed of 25 m/s as depicted in Figs. 4
and 5.
Vehicle 3 is 40 m behind vehicle 4, the safety distance S f3 = TIVvx3 = 50 m is
thus not kept. Therefore, vehicle 3 changes the lane while accelerating (see Fig. 5) to
reach its desired velocity vxdes3 = 27 m/s and becomes the leader of vehicle 2 in the
left lane. But since vehicle 2 drives faster than vehicle 3 (vx2 = 30 > vx3 = 27) m/s
and gets closer to its leader, it changes the lane to avoid collision. The leading

Fig. 4 The corresponding

trajectory and longitudinal
velocity of each vehicle 5

0
200 400 600 800

30

25

20
0 5 10 15 20 25
A Cooperative Centralized MPC for Collision Avoidance of a Fleet of AVs 305

0.5
0 1
-0.5 0.5
-1 0
-1.5
-0.5
0 10 20 0 10 20

0.8 0
0.6
0.4 -0.5
0.2
0 -1
-0.2
-0.4
0 10 20 0 10 20

Fig. 5 The longitudinal and lateral accelerations of each vehicle

vehicle 4 of the left lane maintains its reference lane and decelerate to reach its target
speed. As shown in the presented results, the generated accelerations are within the
limits imposed during the design of the centralized MPC (see Table 2). Furthermore,
the executed lane change maneuvers are comfortable since the lateral acceleration
is a yi < 0.2 g for all the vehicles. The proposed CMPC-based collision avoidance
approach manages successfully in defining the drivable zone of each agent to prevent
collisions.

6 Conclusion

In this paper, a new collision avoidance strategy based on the Sigmoid function barrier
is presented. Each agent in the network receives the proper control action from the
centralized MPC control center in order to avoid collision with the vehicle in front
of it. Each vehicle manages to drive in a collision free area and keep the appropriate
safety distance with its neighbors. The proposed algorithm considers the case when
it is possible to perform a lane change but this is not always the case. In fact, it is
possible to have near agents in the target lane that prevent the lane change. In this
case, it is necessary to perform some braking actions. This matter will be addressed
in future works.
306 M. Ammour et al.

Appendix

Table 1 Initial conditions and desired velocities

Vehicle xi [m] yi [m] vxi [m/s] v yi [m/s] vxdesi [m/s]
1 20 5.25 30 0 25
2 50 5.25 20 0 30
3 120 1.75 25 0 27
4 160 1.75 30 0 27

Table 2 Design parameters

ymin = 0 m ymax = 7 m TIV= 2 s T = 0.2 s
vxmin = 0 m/s vxmax = 40 m/s Nc = 3 N p = 20
wi = 1 m ylat = 3.5 m α=1 β = 10
acxmin = −4 m/s2 acxmax = 2 m/s2 γ1 = 1 γ2 = 5
acymin = −2 m/s2 acymax = 2 m/s2
acxmin = −1.5 m/s2 acxmax = 1.5 m/s2
acymin = −0.5 m/s2 acymax = 0.5 m/s2

References

1. Acevedo, J.J., Arrue, B.C., Maza, I., Ollero, A.: A decentralized algorithm for area surveillance
missions using a team of aerial robots with different sensing capabilities. In: IEEE International
Conference on Robotics and Automation (ICRA), pp. 4735–4740. China, Hong Kong (2014)
2. Ailon, A., Zohar, I.: Control strategies for driving a group of nonholonomic kinematic mobile
robots in formation along a time-parameterized path. IEEE/ASME Trans. Mechatron. 17(2),
326–336 (2012)
3. Alrifaee, B.: Networked model predictive control for vehicle collision avoidance. Ph.D. thesis,
RWTH Aachen University, Germany (2017)
4. Ammour, M., Orjuela, R., Basset, M.: Trajectory reference generation and guidance control
for autonomous vehicle lane change maneuver. In: 28th Mediterranean Conference on Control
and Automation (MED), pp. 13–18. Saint-Raphaël, France (2020)
5. Ammour, M., Orjuela, R., Basset, M.: Collision avoidance for autonomous vehicle using MPC
and time varying sigmoid safety constraints. IFAC-PapersOnLine 54(10), 39–44 (2021)
6. Beck, Z., Teacy, W., Jennings, N., Rogers, A.: Online planning for collaborative search and
rescue by heterogeneous robot teams. In: 15th International Conference on Autonomous Agents
and Multi-Agent Systems (2016)
7. Belkadi, A.: Conception de commande tolérante aux défauts pour les systèmes multi-agents.
Theses, Université de Lorraine, France (2017)
8. Bertrand, S., Marzat, J., Piet-Lahanier, H., Kahn, A., Rochefort, Y.: MPC strategies for coop-
erative guidance of autonomous vehicles. AerospaceLab J. 8, 18 (2014)
9. Dai, L., Cao, Q., Xia, Y., Gao, Y.: Distributed MPC for formation of multi-agent systems with
collision avoidance and obstacle avoidance. J. Frankl. Inst. 354(4), 2068–2085 (2017)
A Cooperative Centralized MPC for Collision Avoidance of a Fleet of AVs 307

10. Hoffmann, G.M., Tomlin, C.J.: Decentralized cooperative collision avoidance for acceleration
constrained vehicles. In: 2008 47th IEEE Conference on Decision and Control, pp. 4357–4363.
Cancun, Mexico (2008)
11. Lindqvist, B., Sopasakis, P., Nikolakopoulos, G.: A scalable distributed collision avoidance
scheme for multi-agent UAV systems. In: IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS). Prague, Czech Republic (2021)
12. Nieto-Granda, C., Rogers, J.G., Christensen, H.I.: Coordination strategies for multi-robot
exploration and mapping. Int. J. Robot. Res. 33(4), 519–533 (2014)
13. Ong, H.Y., Gerdes, J.C.: Cooperative collision avoidance via proximal message passing. In:
2015 American Control Conference (ACC), pp. 4124–4130. Chicago, IL, USA (2015)
14. Qian, X., de La Fortelle, A., Moutarde, F.: A hierarchical model predictive control framework
for on-road formation control of autonomous vehicles. In: 2016 IEEE Intelligent Vehicles
Symposium (IV), pp. 376–381 (2016)
15. Wang, L., Ames, A.D., Egerstedt, M.: Safety barrier certificates for collisions-free multirobot
systems. IEEE Trans. Robot. 33(3), 661–674 (2017)
Fault Tolerant Control of Markov Jump
Systems Using an Asynchronous Virtual
Actuator

Damiano Rotondo

Abstract This paper investigates a fault tolerant control (FTC) strategy using vir-
tual actuators for a class of discrete-time Markov jump systems (MJSs). The main
idea behind this FTC method is to add a block that masks the actuator fault from
the controller/observer point of view, so that the nominal gains can be kept in the
loop without need for retuning them. In line with recent developments in the field of
MJSs, it is assumed that state-feedback controller, state observer, and virtual actuator
operate asynchronously, which means that their active mode might differ from the
controlled plant’s active mode. Based on a stochastic Lyapunov function approach,
design conditions to ensure the fault tolerant stabilization are proposed in the form
of linear matrix inequalities (LMIs). The efficacy of the proposed method is demon-
strated using a numerical example that shows the achieved tolerance against the
complete loss of an actuator due to a fault.

1 Introduction

Markov jump systems (MJS) is the name given to a particular class of stochastic
switching systems for which the underlying switching mechanism corresponds to a
Markov process [1]. These systems are suitable for describing plants that experience
sudden changes in their model, and have found applications in many fields, such as
biology [2] and power grids [3], among many others.
Recent years have seen a surge in the development of theoretical results for asyn-
chronous MJSs that go beyond the more classical synchronous design, as performed
by, e.g., [4]. In a synchronous design, the control system elements (typically the
controller and, possibly, the state observer) switch along with the plant. However, in
the real world, due to uncertainties about the current mode of operation, the presence
of communication delays, and the possibility of loss of data packets, the switching of

D. Rotondo (B)
Department of Electrical Engineering and Computer Science (IDE), University of Stavanger,
4009 Stavanger, Norway
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 309
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_27
310 D. Rotondo

the control system elements might be only partially related to the current operational
mode of the plant. Hence, there is the need to account for this asynchronicity during
the design phase, to avoid the overall stability and performance to be affected in
unexpected and undesired ways [5, 6].
In many control applications, faults can occur and deteriorate the system perfor-
mance up to the loss of stability. For this reason, research on fault detection and
isolation (FDI), fault estimation (FE) and fault tolerant control (FTC) have been a
hot topic during the past few decades. The ongoing research considers many classes
of systems, among which MJSs [7, 8]. The FTC methods developed so far can be
classified into different families, based on the underlying design philosophy. The
fault-hiding paradigm attempts at hiding the fault from the control system perspec-
tive by activating a dedicated block. When the activated block deals with actuator
faults, then it is called virtual actuator, a concept that was first formalized by Jan
Lunze in [10]. Initially developed for LTI systems, virtual actuators have been applied
successfully to many other systems, such as Lure [11], interval [12] and descriptor
[13] systems. However, to our best knowledge, FTC for MJSs via virtual actuator
has not been studied yet.
Inspired by the research mentioned above, the main contribution of this paper is
to describe how asynchronous virtual actuators can be used to achieve fault tolerance
in MJSs. It is shown that the overall system comprising the plant to be controlled, the
state-feedback controller, the state observer and the virtual actuator can be brought
into an equivalent representation using a change of coordinate frame. This repre-
sentation is suitable for obtaining analysis conditions that can be used as a first step
towards design conditions in the form of LMIs, which can be solved to calculate all
the required gains. The overall approach relies on a candidate Lyapunov function
that allows deriving sufficient conditions for mean square exponential stability of the
reconfigured system.

2 Problem Formulation

Let us consider the following discrete-time Markov Jump System (MJS):


x(k + 1) = Ar (k) x(k) + Br (k) u(k)
(1)
y(k) = Cr (k) x(k)

where x(k) ∈ Rn represents the state vector, u(k) ∈ Rm denotes the control input
and y(k) ∈ R p are the sensor outputs. The state-space matrices appearing in (1)
are governed by a Markov jump process through the stochastic variable r (k) ∈ L =
{1, 2, . . . , L}. The jump probability for different values of r (k) is given by:

Pr {r (k + 1) = j|r (k) = i} = πi j (2)

L
with πi j ∈ [0, 1] and j=1 πi j = 1 for all combinations of i, j ∈ L.
Fault Tolerant Control of Markov Jump Systems … 311

In this paper, we will consider actuator faults that affect the system (1), so that
f
under fault occurrence the input matrices become Br (k) , thus transforming (1) into:
 f
x(k + 1) = Ar (k) x(k) + Br (k) u(k)
(3)
y(k) = Cr (k) x(k)

We will consider severe faults, in the sense that given the nominal and faulty matrices
corresponding to a specific mode r (k), the following holds:
   
f
rank Br (k) < rank Br (k) (4)

At the same time, we assume that the system (3) does not lose its stabilizability
properties, so that it can still be controlled satisfactorily, albeit the necessary changes
in the overall control loop are implemented. It is worth stating that the physical
interpretation of the rank condition (4) is that the faults cannot be tolerated by means
of a simple redistribution of the control inputs. The main objective of this paper
is to design a fault tolerant control scheme which employs virtual actuators. The
main idea of this FTC method is to reconfigure the faulty plant so that the nominal
controller and state observer can be kept in the loop without retuning them. This is
achieved by adding a dynamic system whose purpose is to mask the fault. When an
actuator fault appears, the virtual actuator computes an appropriate action u(k) using
the nominal controller’s output u c (k), taking into account the available information
about the fault.
The expression for the virtual actuator is given as follows:
⎧    
⎪
⎪ xva (k + 1) = Aτ (k) + Bτ(k) Mτ (k) xva (k) + Bτ (k) − Bτ(k) u c (k)
⎪
⎪  †
⎪
⎨ f  
u(k) = Bτ (k) Bτ (k) u c (k) − Mτ (k) xva (k)
(5)
⎪ yc (k) = y(k) + Cτ (k) xva (k)
⎪
⎪
⎪  †
⎪
⎩ B = B f f
τ (k) B
τ (k) B
τ (k) τ (k)

where xva ∈ Rn is the virtual actuator state and Mτ (k) is the virtual actuator gain to
be designed. The stochastic variable τ (k) is a probabilistic mapping of r (k) from
space L to space S = {1, . . . , S} with transition probabilities:

Pr {τ (k) = q|r (k) = i} = θiq (6)

S
with θiq ∈ [0, 1] and q=1 θiq = 1 for all combinations of i ∈ L and q ∈ S.
The overall fault tolerant control system is completed by the following nominal
state observer and estimate-feedback control law:
  
x̂(k + 1) = Aτ (k) x̂(k) + Bτ (k) u c (k) + L τ (k) Cτ (k) x̂(k) − yc (k)
(7)
u c (k) = K τ (k) x̂(k)
312 D. Rotondo

where x̂ ∈ Rn is the estimated state, L τ (k) is the observer gain and K τ (k) is the
feedback gain.
Let us define the following mapping between the state variables in a new coordi-
nate frame x1 (k), x2 (k), x3 (k) and the state variables in the original coordinate frame
x(k), x̂(k), xva (k):
⎧ ⎧
⎨ x1 (k) = x̂(k) − x(k) − xva (k) ⎨ x(k) = x2 (k) − x3 (k)
x2 (k) = x(k) + xva (k) ⇔ x̂(k) = x1 (k) + x2 (k) (8)
⎩ ⎩
x3 (k) = xva (k) xva (k) = x3 (k)

Then, the overall system obtained as the interconnection of (3), (5) and (7) is
described by (the dependence of r (k) and τ (k) on the current sample k is omit-
ted to ease the notation):
⎧
⎪
⎪ x (k + 1) = Âr11τ x1 (k) + Âr12τ x2 (k) + Âr13τ x3 (k)
⎨ 1
x2 (k + 1) = Âr21τ x1 (k) + Âr22τ x2 (k) + Âr23τ x3 (k) (9)
⎪
⎪
⎩
x3 (k + 1) = Âr τ x1 (k) + Âr τ x2 (k) + Âr τ x3 (k)
31 32 33

where:
  † 
Âr11τ = Aτ + L τ Cτ + Bτ − Brf Bτf Bτ K τ
  † 
Âr12τ = Aτ − Ar + L τ (Cτ − Cr ) + Bτ − Brf Bτf Bτ K τ
  † 
Âr13τ = Ar − Aτ + L τ (Cr − Cτ ) + Brf Bτf Bτ − Bτ Mτ
  † 
Âr21τ = Bτ + Brf Bτf Bτ − Bτ K τ
  † 
Âr22τ = Ar + Bτ + Brf Bτf Bτ − Bτ K τ
  † 
Âr23τ = Aτ − Ar + Bτ − Brf Bτf Bτ Mτ
 
Âr31τ = Âr32τ = Bτ − Bτ K τ
Âr33τ = Aτ + Bτ Mτ

It is worth highlighting that, due to the assumption that the control system blocks
are asynchronous with respect to the plant (r (k) = τ (k)), the augmented system (9)
does not appear in a block-triangular structure.
Fault Tolerant Control of Markov Jump Systems … 313

3 Analysis Conditions

In this section, we obtain analysis conditions that allow checking if an MJS reconfig-
ured via virtual actuators is mean square exponentially stable or not. These conditions
will serve later the purpose of deriving design conditions that can be used to com-
pute the gains Mq , L q and K q , q ∈ S. In the following, we will denote r (k) = i and
τ (k) = j, respectively.

Theorem 1 The system (9) is mean square exponentially stable if there exist positive
definite matrices Q l,iq ∈ Sn and Pl,i ∈ Sn , l = 1, 2, 3, such that:

S
 
θiq diag Q 1,iq , Q 2,iq , Q 3,iq ≺ diag P1,i , P2,i , P3,i (10)
q=1

⎡ ⎤
−Q 1,iq     
⎢ 0 −Q 2,iq     ⎥
⎢ ⎥
⎢ 0 0 −Q 3,iq    ⎥
⎢ ⎥
⎢ Â11 12 13 −1
− P̌1,i   ⎥ ≺0 (11)
⎢ iq Âiq Âiq ⎥
⎢ 21 −1 ⎥
⎣ Âiq 22
Âiq 23
Âiq 0 − P̌2,i  ⎦
−1
31
Âiq 32
Âiq 33
Âiq 0 0 − P̌3,i

for any i ∈ L, q ∈ S.

Proof Let us consider the following candidate Lyapunov function:

3 3
V (k) = Vl (k) = xl (k)T Pl,i xl (k) (12)
l=1 l=1

and let us define the matrices P̌l,i , l = 1, 2, 3, i = 1, . . . , L as follows:

L
P̌l,i = πi j Pl, j (13)
j=1

Then, by taking into account (9), the following holds:

⎛ ⎞
3 S 3
E {V (k)} = x̄(k)T ⎝ θiq Âl,iq
T
P̌l,i Âl,iq ⎠ x̄(k) − xl (k)T Pl,i xl (k) (14)
l=1 q=1 l=1

 T  
where x̄(k) = x1 (k)T , x2 (k)T , x3 (k)T and Âl,iq = Âl1 l2 l3
iq Âiq Âiq . Then, using
(10), one gets that:
314 D. Rotondo

S
 3

E{V (k)} < θiq x̄(k) T
Âl,iq
T
P̌l,i Âl,iq x̄(k)
q=1 l=1
S

− θiq x̄(k)T diag Q 1,iq , Q 2,iq , Q 3,iq x̄(k) (15)
q=1

By ensuring that:

3

Âl,iq
T
P̌l,i Âl,iq − diag Q 1,iq , Q 2,iq , Q 3,iq ≺ 0 (16)
l=1

one gets that E {V (k)} < 0, which ensures mean square exponential stability.
Finally, (11) is obtained from (16) using Schur complements, which completes the
proof. 

4 Design Conditions

Let us make the assumption that the output-related matrices are all constant, that is,
Ci = C. A consequence of this fact is that the matrices Âr12τ and Âr13τ that appear in
(9) get simplified, as follows:
  † 
Âr12τ = Aτ − Ar + Bτ − Brf Bτf Bτ K τ
  † 
Âr13τ = Ar − Aτ + Brf Bτf Bτ − Bτ Mτ

Under these conditions, the following theorem provides conditions for designing the
unknown gains Mq , L q and K q , q ∈ S.
 
Theorem 2 Let C = V1 V2 0 V3T be the singular value decomposition (SVD) of
the output matrix C, with V1 V1T = I and V3 V3T = I . The system (9) is mean square
exponentially stabilizable if there exist positive definite matrices Q̄ l,iq ∈ Sn and P̄l,i ∈
Sn , l = 1, 2, 3, and matrices q ∈ Rn× p , q ∈ Rm×n , Wq ∈ Rm×n , Z 1 ∈ R p× p , Z 2 ∈
R(n− p)× p , Z 3 ∈ R(n− p)×(n− p) such that:

S
 
θiq diag Q̄ 1,iq , Q̄ 2,iq , Q̄ 3,iq ≺ diag P̄1,i , P̄2,i , P̄3,i (17)
q=1
Fault Tolerant Control of Markov Jump Systems … 315

⎡ ⎤
− Q̄ 1,iq     
⎢ 0 − Q̄     ⎥
⎢ 2,iq ⎥
⎢ 0 0 − Q̄    ⎥
⎢ 11 3,iq ⎥≺0 (18)
⎢ Āiq Ā 12
Ā 13
ϒ1,i   ⎥
⎢ iq iq ⎥
⎣ Ā21 22
Āiq 23
Āiq 0 ϒ2,i  ⎦
iq
31
Āiq 32
Āiq 33
Āiq 0 0 ϒ3,i

for any i ∈ L, q ∈ S, where:

L
ϒl,i = πi j P̄l, j − Z − Z T , l = 1, 2, 3 (19)
j=1


V2−1 V1T Z 1 0
Z = V3 VT (20)
Z2 Z3 3

   † 
f 
11
Āiq = Aq Z + q 0 V3T + Bq − Bi Bqf Bq q (21)
 † 
f 
12
Āiq = Aq Z − Ai Z + Bq − Bi Bqf Bq q (22)
  † 
= Ai Z − Aq Z + Bi Bqf Bq − Bq Wq
13 f
Āiq (23)
 
f  †
21
Āiq = Bq + Bi Bqf Bq − Bq q (24)
 
f  †
22
Āiq = Ai Z + Bq + Bi Bqf Bq − Bq q (25)
 † 
f 
23
Āiq = Aq Z − Ai Z + Bq − Bi Bqf Bq Wq (26)
 
31
Āiq = Āiq
32
= Bq − Bq q (27)
33
Āiq = Aq Z + Bq Wq (28)

Then, the stabilizing controller, observer, and virtual actuator gains can be obtained
as:
K q = q Z −1 , Mq = Wq Z −1 , L q = q Z 1−1 (29)

Proof Let us define the matrices Q̄ l,iq and P̄l,i , l = 1, 2, 3, as follows:

Q̄ l,iq = Z T Q l,iq Z (30)

P̄l,i = Z T Pl,i Z (31)

316 D. Rotondo

Then, by pre-and post-multiplying (10) by Z T and Z , respectively, it follows that

(10) and (17) are equivalent. By applying appropriately the Young’s relation, the
following holds:
⎛ ⎞ ⎛ ⎞−1
L L L
−1
πi j P̄l, j − Z − Z T = Z T ⎝ πi j Pl, j ⎠ Z − Z − Z T  − ⎝ πi j Pl, j ⎠ = − P̌l,i (32)
j=1 j=1 j=1


Let us pre-and post-multiply (11) by diag Z T , Z T , Z T , I, I, I and its transpose
which, after taking into account (30)–(31) and the inequality (32), leads to:
⎡ ⎤
− Q̄ 1,iq     
⎢ ⎥
⎢ 0 − Q̄ 2,iq   
 ⎥
⎢ ⎥
⎢ 0 − Q̄ 3,iq    ⎥
⎢ 0 ⎥
⎢ ⎥≺0 (33)
⎢ Â11 Z Â12 Z Â13 Z ϒ1,i   ⎥
⎢ iq iq iq ⎥
⎢ ⎥
⎢ Â21 Z Â22 Z Â23 Z 0 ϒ2,i  ⎥
⎣ iq iq iq ⎦
31
Âiq Z 32
Âiq Z Âiq Z 0 0 ϒ3,i
33

where the matrices ϒ1,i , ϒ2,i and ϒ3,i are defined as in (19). Let us structure the matrix
Z as in (20). Then, by introducing the auxiliary matrices q  K q Z , Wq  Mq Z
and q  L q Z 1 , the following is true:

  V2−1 V1T Z 1 0    
L q C Z =L q V1 V2 0 V3T V3 V T = L q Z 1 0 V3T = q 0 V3T
Z2 Z3 3
(34)

jk jk
Then, one has Aiq Z = Āiq for j, k = 1, 2, 3, which completes the proof. 

5 Numerical Example

Let us consider an MJS as in (1) with two different modes (L = 2) described by matri-
ces A1 , B1 , C1 and A2 , B2 , C2 , respectively, with C1 = C2 to satisfy the assumption
introduced in Sect. 4:
 
1 0.3 3.7 2.8  
A1 = B1 = C1 = 1 1
0.1 1.1 1 0.5
 
1 0.1 3.5 2.1  
A2 = B2 = C2 = 1 1
0.1 0.5 0.8 1.8
Fault Tolerant Control of Markov Jump Systems … 317

Note that both modes are open-loop unstable due to eigenvalues of the state matrix A
in 1.23 and 1.02, respectively. The probabilistic mappings (2) and (6) are described
by the following matrices:
 
0.8 0.2 0.7 0.3
π= θ=
0.2 0.8 0.1 0.9

The considered fault scenario is a complete loss of the second actuator, which results
in faulty input matrices:
 
f 3.7 0 f 3.5 0
B1 = B2 =
1 0 0.8 0

At first, a virtual actuator-less design has been considered, which means using a
reduced version of Theorem 2 without decision variables Q̄ 3,iq , P̄3,i and Wq , where
the third row/column of the matrices
 in (17),
 the third
 and sixth rows/columns of the
 f f
matrix in (18), and the terms Bq − Bi Bq Bq q appearing in (21)–(22) and
(24)–(25) have been removed. Note that such design is akin to the one considered in
[6], albeit the difference in that a H∞ performance metric for disturbance rejection
is not considered. This design will be later referred to as without FTC. On the other
hand, the straight application of Theorem 2 has returned controller/observer/virtual
actuator gains that will be referred to in the following as with FTC. The simulation
results have been obtained with initial conditions x(0) = [1, 1]T , x̂(0) = [0, 0]T
and xva (0) = [0, 0]T and with asynchronicity between the jump modes of the plant
and those of the control system. The simulated fault scenario is a sudden switch
f f
from matrices B1 , B2 to faulty matrices B1 , B2 occurring at sample k = 500. The
comparison between simulation results without and with the proposed FTC strategy
is shown in Fig. 1. For visualization purposes, a logarithmic scale is used on the
y-axis as it enables comparing convergent and divergent trends of the state variables.
It can be clearly seen that the proposed FTC strategy is able to maintain closed-loop
mean square exponential stability in spite of the actuator fault, whereas the more
traditional design without FTC fails in achieving so.

6 Conclusions

This paper has investigated the FTC of discrete-time MJSs by designing an asyn-
chronous virtual actuator. The proposed FTC system is able to preserve mean square
exponential stability in spite of severe actuator faults, such as the complete loss of
an actuator. This has been demonstrated using simulations of a numerical example.
Future research will focus on developing an FDI algorithm that can be integrated
with this FTC strategy.
318 D. Rotondo

log10(|x1 (k)|)
-50

without FTC
-100 with FTC

0 100 200 300 400 500 600 700 800 900 1000

0
log10(|x2 (k)|)

-50

-100

0 100 200 300 400 500 600 700 800 900 1000
Sample k

Fig. 1 System states x1 and x2

References

1. Zhao, P., Kang, Y., Zhao, Y.-B.: A brief tutorial and survey on Markovian jump systems:
stability and control. IEEE Systems, Man, and Cybernetics Magazine. 5(2), 37-C3 (2019)
2. Wilkinson, D.J.: Stochastic modelling for quantitative description of heterogeneous biological
systems. Nat. Rev. Genetics. 10(2), 122–133 (2009)
3. Lai, J., Lu, X., Yu, X., Monti, A.: Stochastic distributed secondary control for AC microgrids
via event-triggered communication. IEEE Trans. Smart Grid. 11(4), 2746–2759 (2020)
4. Hou, T., Ma, H.: Exponential stability for discrete-time infinite Markov jump systems. IEEE
Trans. Autom. Control. 61(12), 4241–4246 (2015)
5. Wang, L., Wu, Z.-G., Shen, Y.: Asynchronous mean stabilization of positive jump systems with
piecewise-homogeneous Markov chain. IEEE Trans. Circuits Syst. II: Express Briefs. 68(10),
3266–3270 (2021)
6. Zhang, M., Shen, C., Wu, Z.-G.: Asynchronous observer-based control for exponential sta-
bilization of Markov jump systems. IEEE Trans. Circuits Syst. II: Express Briefs. 67(10),
2039–2043 (2019)
7. Li, X., Ahn, C.K., Lu, D., Guo, S.: Robust simultaneous fault estimation and nonfragile output
feedback fault-tolerant control for Markovian jump systems. IEEE Trans. Syst. Man Cybern.
Systems. 49(9), 1769–1776 (2018)
8. Yang, H., Yin, S., Kaynak, O.: Neural network-based adaptive fault-tolerant control for Marko-
vian jump systems with nonlinearity and actuator faults. IEEE Trans. Syst. Man Cybern. Sys-
tems. 51(6), 3687–3698 (2020)
9. Xue, M., Yan, H., Wang, M., Shen, H., Shi, K.: Lstm-based intelligent fault detection for fuzzy
Markov jump systems and its application to tunnel diode circuits. IEEE Trans. Circuits Syst.
II: Express Briefs. 69(3), 1099–1103 (2021)
10. Lunze, J., Steffen, T.: Control reconfiguration by means of a virtual actuator. IFAC Proc.
Volumes. 36(5), 131–136 (2003)
Fault Tolerant Control of Markov Jump Systems … 319

11. Pedersen, A.S., Richter, J.H., Tabatabaeipour, M., Jóhannsson, H., Blanke, M.: Fault tolerant
emergency control to preserve power system stability. Control. Eng. Practice. 53, 151–159
(2016)
12. Rotondo, D., Cristofaro, A., Johansen, T.A.: Fault tolerant control of uncertain dynamical
systems using interval virtual actuators. Int. J. Robust Nonlinear Control. 28(2), 611–624
(2018)
13. Wang, Y., Rotondo, D., Puig, V., Cembrano, G.: Fault-tolerant control based on virtual actuator
and sensor for discrete-time descriptor systems. IEEE Trans. Circuits Syst. I: Regul. Papers.
67(12), 5316–5325 (2020)
14. Zhou, K., Khargonekar, P.P.: Robust stabilization of linear systems with norm-bounded time-
varying uncertainty. Syst. & Control. Letters. 10(1), 17–20 (1988)
Robust Safety Control for Automated
Driving Systems with Perception
Uncertainties

Yan Feng Yu, Kaveh Nazem Tahmasebi, Peng Su, and Dejiu Chen

Abstract Safety assurance and trustworthiness guarantees represent some of the

most challsenging problems in the development of next generation automated driving
and driving assistance systems. A systematic approach, with measures ranging from
development-time modeling and simulation support to operation-time mechanisms
for situation-awareness and adaptation, becomes necessary for tackling the problems.
This paper presents a novel approach to safety control for automated driving under
the condition of uncertain perception due to emergent properties in the environment
or sensor faults. Based on the theory of optimal control, the safety controller is
built upon a control barrier function and a model predictive control function. The
effectiveness of the proposed strategy is evaluated on a simulation scenario created
in the open-source autonomous driving simulator CARLA.

Keywords Autonomous driving system · Control barrier function · Model

predictive control · Measurement uncertainties · Kalman filter

1 Introduction

Currently, increasingly novel technologies are being integrated into automotive vehi-
cles to enable Advanced Driver Assistance Systems (ADAS) or fully Automated
Driving Systems (ADS). In the development of such systems, safety assurance and

Y. F. Yu · K. Nazem Tahmasebi (B) · P. Su · D. Chen

Unit of Mechatronics and Embedded Control System, KTH Royal Institute of Technology,
Brinellävgen 83, 10044 Stockholm, Sweden
e-mail: [email protected]
Y. F. Yu
e-mail: [email protected]
P. Su
e-mail: [email protected]
D. Chen
e-mail: [email protected]
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 321
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_28
322 Y. F. Yu et al.

trustworthiness guarantees represent some of the most challenging tasks. The goal
is to at least assure that, as long as other road users act according to certain assumed
behaviors, an automated driving (AD) vehicle will not crash or be able to mitigate
collisions by unpredictable behaviors of other road users. Therefore, one fundamental
requirement in the development of ADAS/ADS is to provide fail-safe/fail-operational
capabilities so that the system will attain minimal risk conditions in all operational
situations. To this end, one key issue is related to the support for qualified perception
of complex operational conditions, where the presence of measurement uncertain-
ties and unexpected system anomalies could result in hazardous situations, resulting
potentially in serious injuries or fatality to people involved. Nevertheless, challenges
have been indicated by recent traffic accident statistics. In the latest report concern-
ing the period July 2021 to May 2022 by the US National Highway Traffic Safety
Administration, it appears that 392 collisions with ADAS have been reported. These
have resulted in six people being killed and five seriously injured [1].
This paper presents an approach to safety control by integrating Model Predictive
Control (ä) and Control Barrier Function (CBF) for effectively attaining the minimal
risk conditions. The result of the proposed safety control method is evaluated by
a case study of Adaptive Cruise Control (ACC) in a highway condition simulated
with CARLA, an open-source AD simulator [2]. The reminder of this paper is orga-
nized as follows: Sect. 2 introduces the related concepts and technologies. Section 3
elaborates the design concepts and theoretical foundations for the proposed control
approach. Section 4 presents the implementation of the proposed safety control sys-
tem in different cases regarding state estimation method based on CARLA. Finally,
the conclusion is given in Sect. 5.

2 Related Work

MPC [3, 4] as one advanced control framework is widely used in robotics and
autonomous systems. One special strength of MPC is its capability of predicting
future behaviors of dynamic process based on linear and nonlinear time-varying
models given a set of inputs and constraints. For AD vehicles, MPC can be used for a
range of control functions, such as ACC [5] and Lane-Keeping Control [6]. The com-
mon strategy is to treat these control functions as the tasks of solving model-based
optimization problems over certain predefined prediction horizon based on parame-
ters of vehicle dynamics (i.e., positions of throttle, brake, steering) and environmental
constraints (i.e., minimal safe distances to other vehicles).
The motivation behind the choice of CBF algorithm has been its support for stip-
ulating the properties required for system safety assurance [7]. By safety, we refer
to the system’s capability of being protected from hazardous operational conditions,
i.e. posing unacceptable risks to the AD vehicle and its environment. The control
objective is related to risk mitigation by actively avoiding unsafe regions of the
state space [8]. To apply CBF to real-time control problems, it is necessary to trans-
form continuous-time CBF to discrete-time CBF [9]. This also makes it possible to
Robust Safety Control for Automated Driving Systems with Perception Uncertainties 323

integrate CBF into the MPC framework [10, 11] as a useful method for specifying
additional constraints. One benefit by such an integration is that the control policy
can become less greedy as future state information can be more effectively taken
into account. For ADS, through the unification of MPC and CBF, the vehicle can
plan necessary actions for risk mitigation at earlier stages without requiring a large
horizon as typically expected. This helps to reduce the load of computation, which
is considered to be one of the main challenges for successful MPC implementation
For AD vehicles, the perception of environmental conditions by sensors suffers
from uncertainties from external disturbances and measurement perturbations [12].
Such uncertainties, if improperly treated, could cause biased control decisions with
negative impact on system performance and safety. To avoid the system robustness
loss under measurement disruptions, Kalman Filter (KF) [13], as one conventional
method for coping with noisy measurements, has been adopted for providing oper-
ation condition (e.g. vehicle state) estimation for the control actions. KF produces
predictions of system states based on a joint probability distribution over the state
variables for each timeframe with a series of measurements observed over time. One
further method to promote control robustness against external disturbances on sys-
tem dynamics is the Zeroing Control Barrier Function (ZCBF) method [14]. By this
paper, we investigate how the ZCBF method can be integrated in MPC by considering
the uncertainty caused by common faults in the perception system.

3 Key Design Concepts

In this section, we elaborate on necessary preliminaries to the safety optimal control

method integrated with an estimation algorithm for disrupted systems. The notations
employed in this paper are explained below.
Notation Given an arbitrary set S, the boundary and interior of this set are denoted by
∂S and Int(S), respectively. A transpose matrix is denoted by A , and the expectation
of a random variable X is denoted by E(X). An identity matrix is shown by I. A
set S is called (forward) invariant for a dynamic system ẋ = f (x) if for every
x(0) ∈ S ⇒ x(t) ∈ S for all t ∈ R+ . The class K function is defined as a strictly
increasing continuous function α : [0, a) → R+ , with a > 0 and α(0) = 0. This
class K function α belongs to class K∞ if a = ∞ and lim α(r ) = ∞.
r →∞
The discrete-time linear uncertain control system is defined as follows:

xt+1 = Axt + Bu t + wt (1)

yt = C xt + vt (2)

where xt ∈ Rn , u t ∈ Rm , and yt ∈ Rk denote the state, control input, and output

of the system at time step t, respectively. wt ∈ Rn is the exogenous process noise

with zero mean that generates the nominal covariance matrix Q  = E(wt wt ) as
324 Y. F. Yu et al.

a process noise uncertainty. vt ∈ Rk is the exogenous measurement/output noise


with zero mean that generates the nominal covariance matrix R  = E(vt vt ), as a
measurement uncertainty. The matrices A ∈ R , B ∈ R
n×n n×m
, and C ∈ R k×n
are the
state transition, control input, and system output matrices, respectively.
Control Barrier Function: Considering [9] and [10], we formulate a ZCBF as a
robust CBF approach. A continuously differentiable function h : Rn → R is defined
such that the corresponding superlevel set C of this function would be our safe set:
 
C = x ∈ X ⊂ Rn : h(x) ≥ 0
 
∂C = x ∈ X ⊂ Rn : h(x) = 0 (3)
 
Int(C) = x ∈ X ⊂ Rn : h(x) > 0 .

Function h is a CBF if it satisfies (4), where there exists a control input u ∈ U ⊂ Rm

and an extended class K∞ function α [15, 16].

ḣ(x, u) ≥ −α(h(x)) ⇔ C is invariant. (4)

In discrete time domain, by using h(xk , u k ) := h(xk+1 ) − h(xk ) and the simplified
form of extended class K function α(h(x)) = γ h(x), we reformulate (4) as:

h(xk , u k ) ≥ −γ h(xk ) ⇔ C is invariant. (5)

where scalar γ > 0 as a class K function, defined as a relaxation constant in this

paper. A large γ increases the optimization problem feasibility but may result in the
CBF constraint not being activated. A small γ imposes stricter conditions for safety
based on CBF.
Kalman Filter: According to the uncertain control system model (1) and (2), which
has the process and measurement noises, a simple linear KF algorithm is formulated
to deal with these disturbances by estimating the measured and predicted variables.
Besides, the process noise uncertainty, Q  , and the measurement uncertainty, R  , we
also assume the state estimation uncertainties of the current state with a covariance
 
matrix Pt|t = E((xt − x̂t|t )(xt − x̂t|t ) ). Additionally, Pt|t−1 denote the estimation
uncertainties calculated by the previous filter estimation. Once KF [13, 17] is ini-

tialized with x̂0|0 and P0|0 , it operates in a loop containing a time update (6) and a
measurement update (7), where K t is known as Kalman Gain.

x̂t+1|t = A x̂t|t + Bu t
 
(6)
Pt|t−1 = A Pt−1|t−1 A T + Q t−1 .


 
−1
K t = Pt|t−1 C T C Pt|t−1 C T + Rt
 
x̂t|t = x̂t|t−1 + K t z t − C x̂t|t−1 (7)
  
Pt|t = Pt|t−1 − K t C Pt|t−1 .
Robust Safety Control for Automated Driving Systems with Perception Uncertainties 325

Safety Optimal Control Algorithm (MPC-CBF) Formulation: The mathematical

model of the proposed safety optimal control method, integration of MPC and CBF,
subjected to the discrete-time linear control system (1) should be formulated in (8).
Its solution is similar to MPC [18] except for an additional constraint of CBF, h,
added to guarantee the associated safe set C to be forward invariant.

N −1

∗   
J (xt ) = minimize [xt+k Qxt+k + u t+k Ru t+k ] + xt+N P xt+N (8a)
u t:t+N −1
k=0
xt+k ∈ X , u t+k ∈ U k = 0, . . . , N − 1 (8b)
 
h (xt+k , u t+k ) ≥ −γ h xt+k|t k = 0, . . . , N − 1
 
h(xt+k+1|t ) − h(xt+k|t ) ≥ −γ h xt+k|t (8c)
xt+N ∈ X f (8d)
x t ∈ X0 (8e)

where xt , xt+k , xt+N , and u t+k are the initial state, the state at step k, the end of
state at the terminal horizon step, and the predicted control action at step k obtained
from time-step t, respectively. Q, R, and P are the positive semi-definite weight
matrices. The value function J ∗ (xt ), formulated as a quadratic optimization problem,
is minimized subject to the control system model (1), and the constraints on; (i) state
and input control (8a), (ii) terminal and initial constraints (8c) and (8d), and (iii) CBF
constraint, (8b), as a safe constraint added to the conventional MPC constraints.

4 Implementation and Case Study

To evaluate the proposed safety control method with KF algorithm, a case study of
adaptive cruise control (ACC) is designed based on the concepts described in Sect. 3.
The overall control structure shown in Fig. 1.
Fault Modelling of Perception Components: The system uses multi-sensors for
perceiving the environmental conditions. Learning-Enable Components (LEC) are
used for the operational conditions estimation as describied in [19]. In particular,
a RGB-camera is used to measure the relative distance between the ego and lead-
ing vehicles. To investigate the LECs’ ability to compute the relative distance in

Fig. 1 Overall control structure for ACC. The Environment is implemented by CARLA, an open-
source AD simulator. Noises/faults are injected into the Perception System block. The Control
Strategy block provides an integration of MPC and CBF
326 Y. F. Yu et al.

Fig. 2 Different intensity salt-and-pepper noise injected into the RGB-camera

Fig. 3 The predicted

difference between faulty
and true relative distance by
SVR-the sampling rate is set
to 0.04 s

faulty situations, salt-and-pepper noise is injected with 80% failure probability with
different intensities into the RGB-camera, as shown in Fig. 2. The effect of the corre-
sponding faults are modeled using Support Vector Regression (SVR), as a predictor
algorithm, based on the absolute value of the discrepancy between the ground truth
and the measured distances. This fault model is used as an exogenous measurement
noise in the control system model. The prediction result is shown in Fig. 3.
Safety Optimal Control Design:  The vehicle
 dynamics is given by the following
state variables for ACC, xk = v f vl z ,where v f and vl are the velocity of the
AD and leading vehicles, respectively; and z is the relative distance between AD
and leading vehicles. The corresponding control input variable u k = a F is the AD
vehicle acceleration. According to the vehicle kinematics behavior, matrices A and
B in (1) are derived as follows:
⎡ ⎤ ⎡ ⎤
1 0 0 Ts
A = ⎣ 0 1 0⎦ , B = ⎣ 0 ⎦ . (9)
−Ts Ts 1 1 2
2
Ts

where Ts = 0.2s refers to the sampling time. The required parameters for the safety
optimal control algorithm include: N = 8, referring to the control horizon; γ =
5, referring to the class K function; Q = P = I3 , and R = 2 · I2 , referring to the
Robust Safety Control for Automated Driving Systems with Perception Uncertainties 327

weights matrices. For the basic vehicle dynamics model in Eq. (1), the following
constraints on the states and inputs are defined :
 
X = xt+k ∈ Rn : xt+k ≤ [20, 14, ∞ },
 
X f = xt+k ∈ Rn : [0.2, 0, 7] ≤ xt+k ≤ [20, 14, ∞] , (10)
 
U = u t+k ∈ R : −g ≤ u t+k ≤ a f g, g = 9.8m/s .
m 2

The maximum vehicle acceleration is delimited to a fraction of g by a f = 0.3. The

lower bound on the terminal constraint refers to a slack value to ensure the AD
vehicle stops before the leading vehicle. The AD vehicle desired velocity is set to
20 m/s, and the leading vehicle velocity varies between 0 and 14 m/s as indicated
in the two first elements of the stage and terminal state upper-level constraints. The
third element, maximum relative distance, is considered as infinity. In addition, the
lower-level of state constraint is managed based on a CBF constraint, specifying a
safe distance for ACC:
z ≥ Th v f . (11)

According to a general rule for keeping a minimum relative distance [20], the Th is
set to 1.8 s. It should be noticed this constraint is not activated during braking when
both z and v f become very small. Consequently, the intuitive choice of CBF would
be h(x) = z − Th v f which yields the admissible set C (3). Furthermore, to ensure
that the CBF constraint does not conflict with other constraints, the CBF constraint
is reformulated as in Equation (12) by means of a definiton of h F (x) [21], which
yields a set C F = {x|h F (x) ≥ 0} for all x ∈ C F .

1 v f al − vl a f
2 2
h F (xk ) = z − Th v f − . (12)
2 a f al g

Besides the CBF constraint mentioned above, h F (x) is exclusively employed in the
situation where the leading vehicle start braking. The adopted constant al = 0.3 has
the same role as the a f to restrict the leading vehicle acceleration. During sudden
braking situation, the terminal constraints X f must ensure the collision avoidance.
Kalman Filter Design: By only considering the measurement noise added directly
to z as the system output yt , the KF parameters are set as follows: R  = 25, which

decides a choice of nominal covariance matrix for measurement noise; P0|0 = I3 ,

which refers to the covariance matrix for estimation uncertainties; and Q = 0, which
refers to nominal covariance matrix for process noise.
Simulation Results: The evaluation of the proposed safety optimal control approach
is supported by a software-in-the-loop simulation using Python and CARLA1 The
driving scenario is given by a situation where an AD vehicle follows another pre-
ceding vehicle on highway according to a pre-existing scenario model given by the

1 The simulation model and the codes are available in https://github.com/PiggyMouth/MPC-

CBF_for_ADS.
328 Y. F. Yu et al.

Fig. 4 Velocity of the

preceding vehicle

ScenarioRunner of CARLA. The initial state is chosen to be x0 = [0, 0, 100] . The

preceding vehicle velocity behaves with a pattern shown in Fig. 4. The computed
control request, a F , is fed to the longitudinal controller in CARLA as throttle/brake
request after a normalization with the range between 0 and 1. According to the
data derived from simulations of behaviors with different throttle/braking levels, the
following throttle/braking function is defined:
aF
throttle = + 0.5, 0 ≤ a F , brake = 0.
10
aF (13)
brake = , −a f g < a F ≤ 0, throttle = 0.
−10
full brake = 1, −g < a F ≤ −a f g, throttle = 0.

Case 1: Figure 5 shows the simulation results of the situation without any perception
noise. The sudden drop of control input and velocity shows the ACC actions taken
to ensure a safe distance to the leading vehicle by applying the full braking. The
fluctuations of control input and vehicle velocity at the end of the simulation phase
is due to the fact that the AD vehicle tries keeping 7 m safe distance as defined in the
terminal constraint. As shown in Fig. 5a, initially AD vehicle reaches to its desired
velocity, 20 m/s, and then starting reducing its velocity to follow the leading vehicle
velocity at 14 m/s as their relative distance decreases.

Velocity of following vehicle Control input Relative distance

Fig. 5 Simulation results without any faults

Robust Safety Control for Automated Driving Systems with Perception Uncertainties 329

Velocity of following vehicle Control input Relative distance

Fig. 6 Simulation results with perception noise w/o KF-c Salt-and-pepper noise is injected into
measured relative distance

Velocity of following vehicle Control input Relative distance

Fig. 7 Simulation results with perception noise with KF-c Salt-and-pepper noise is injected into
measured relative distance

Case 2: Here, the perception noise is added to the AD vehicle and its simulation
results without KF are shown in Fig. 6. Since the perception noise directly influences
the control performance, the control input varies more largely in order to follow the
required acceleration for keeping the corresponding relative distance. Based on the
outcomes, integration of MPC and ZCBF is relatively resilient to noises without KF.
Case 3: Case 3 covers the same situation in Case 2 by adding KF. The results are
shown in Fig. 7. As shown in Fig. 7c, the effect of perception noise on the measured
relative distance is diminished by applying KF. Consequently, the control input and
the velocity become smoother. In addition, the AD vehicle velocity initially can
reach its desired value and then remains about the leading vehicle velocity with
slight variation compared to the previous case.

5 Conclusion and Future Work

In this paper, we have presented an approach to safety control of ADS under the
uncertainties of perception sensors by exploiting an integration of ZCBF and MPC.
A comparison of different experimental results showed the overall performance of
the proposed control strategy, on the basis of fault injection and software-in-the-loop
simulation. We have also described how the perception noise were injected into the
system. Further work will enrich the operational situations by adding model-based
330 Y. F. Yu et al.

specification of the environment and system uncertainties, and the system model
by using nonlinear state-space to allow further investigations for controller design
optimisation and refinement.

Acknowledgements This work is supported by 1. KTH Royal Institute of Technology with the
industrial research project ADinSOS (2019065006); and 2. the Swedish government agency for
innovation systems (VINNOVA) with the cooperative research project Trust-E (Ref: 2020-05117)
within the programme EUREKA EURIPIDES.

References

1. NHTSA, Summary report: Standing general order on crash reporting for automated driv-
ing systems (2022). https://www.nhtsa.gov/sites/nhtsa.gov/files/2022-06/ADS-SGO-Report-
June-2022.pdf
2. Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: CARLA: an open urban driving
simulator. In: Proceeding of the 1st Annual Conference on Robot Learning, pp. 1–16 (2017)
3. Rawlings, J., Mayne, D., Diehl, M.: Model Predictive Control: Theory, Computation, and
Design. Nob Hill Publishing (2017)
4. Son, T.D., Nguyen, Q.: Safety-critical control for non-affine nonlinear systems with application
on autonomous vehicle. In: 2019 IEEE 58th Conference on Decision and Control (CDC), pp.
7623–7628 (2019)
5. Moser, D., Schmied, R., Waschl, H., del Re, L.: Flexible spacing adaptive cruise control using
stochastic model predictive control. IEEE Trans. Control. Syst. Technol. 26(1), 114–127 (2018)
6. Xu,Y., Chen, B., Shan, X., Jia, W., Lu, Z., Xu, G.: Model predictive control for lane keeping
system in autonomous vehicle. In: 2017 7th International Conference on Power Electronics
Systems and Applications - Smart Mobility, Power Transfer Security (PESA), pp. 1–5 (2017)
7. Wieland, P., Allgöwer, F.: Constructive safety using control barrier functions. IFAC Proc. Vol.
40(12), 462–467 (2007)
8. Ames, A.D., Coogan, S., Egerstedt, M., Notomista, G., Sreenath, K., Tabuada, P.: Control
barrier functions: theory and applications. In: 2019 18th European Control Conference (ECC).
pp. 3420–3431 (2019)
9. Agrawal, A., Sreenath, K.: Discrete control barrier functions for safety-critical control of dis-
crete systems with application to bipedal robot navigation. In: Robotics: Science and System
(2017)
10. Zeng, J., Zhang, B., Sreenath, K.: “Safety-critical model predictive control with discrete-time
control barrier function. In: American Control Conference (ACC) 2021, pp. 3882–3889 (2021)
11. Zeng, v., Li, Z., Sreenath, K.: Enhancing feasibility and safety of nonlinear model predictive
control with discrete-time control barrier functions. In: 2021 60th IEEE Conference on Decision
and Control (CDC), pp. 6137–6144 (2021)
12. Jha, S., Banerjee, S.S., Cyriac, J., Kalbarczyk, Z.T., Iyer, R.K.: Avfi: Fault injection for
autonomous vehicles. In: 48th Annual IEEE/IFIP International Conference on Dependable
Systems and Networks Workshops (DSN-W), pp. 55–56 (2018)
13. Kordic, V.: Kalman Filter. InTech (2010)
14. Xu, X., Tabuada, P., Grizzle, J.W., Ames, A.D., Robustness of control barrier functions for
safety critical control**this work is partially supported by the national science foundation
grants 1239055, 1239037 and 1239085. IFAC-PapersOnLine. 48(27), 54–61, (2015) Analysis
and Design of Hybrid Systems ADHS
15. Grandia, R., Taylor, A., Singletary, A., Hutter, M., Ames, A.: Nonlinear model predictive
control of robotic systems with control lyapunov functions. In: Robotics: Science and System
XVI
Robust Safety Control for Automated Driving Systems with Perception Uncertainties 331

16. Khalil, H.K.: Nonlinear Systems, 3rd edn. Prentice Hall (2002)
17. Huang, Y., Zhang, Y., Li, N., Wu, Z., Chambers, J.A.: A novel robust student’s t-based kalman
filter. IEEE Trans. Aerosp. Electron. Syst. 53(3), 1545–1554 (2017)
18. Borrelli, F., Bemporad, A., Morari, M.: Predictive Control for Linear and Hybrid Systems.
Cambridge University Press (2017)
19. Su, P., Chen, D.: Using fault injection for the training of functions to detect soft errors of dnns
in automotive vehicles. In: Dependability and Complex Systems, pp. 308–318 (2022)
20. A comparison of headway and time to collision as safety indicators. Accid. Anal & Prev, 35(3),
427–433 (2003)
21. Ames, A.D., Xu, X., Grizzle, J.W., Tabuada, P.: Control barrier function based quadratic pro-
grams for safety critical systems. IEEE Trans. Autom. Control. 62(8), 3861–3876 (2017)
Algorithms and Methods for
Fault-tolerant Control and Design
of a Self-balanced Scooter

Ralf Stetter, Markus Till, and Marcin Witczak

Abstract Due to complex operation and surrounding conditions, faults are inevitable
during the use of complex technical systems such as two-wheel scooters. This paper
explains algorithms and methods which allow the accommodation of faults through
the design and control of such systems on the functional level of product concretiza-
tion. A main component of the approach is a fuzzy decision engine. This engine
combines the evaluation of residuals and the current state and is based on specialist’s
expertise.

1 Introduction

The central topics of this paper are fault-tolerant design (FTD) and the concepts
and algorithms which allow to realize FTD [11]. In the last decades, an enormous
amount of research covered the topic fault-tolerant control (FTC), i.e. the capabil-
ity of a control system to accommodate faults which inevitably occur in complex
technical systems [1, 2, 9, 14]. Powerful algorithms, methods and systems were
created that realize FTC—current research aspects are tracking control with ellip-
soidal bounding [5], reliability aware model predictive control [4] and parameter
identifiability [8]. However, this kind of control can be supported by certain design
aspects—the design aspects referred to as fault-tolerant design [9]. In general, it is
sensible to distinguish certain layers of fault-tolerant design based on the level of

R. Stetter (B) · M. Till

Department of Mechanical Engineering, Ravensburg-Weingarten University (RWU), 88250
Weingarten, Germany
e-mail: [email protected]
M. Till
e-mail: [email protected]
M. Witczak
Institute of Control and Computation Engineering, University of Zielona Góra, 65-246 Zielona,
Góra, Poland
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 333
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_29
334 R. Stetter et al.

Fig. 1 Levels for characteristics of fault-tolerant design

abstraction of the involved models of the involved technical system (Fig. 1—adapted
from [9]).
This paper discusses FTD on the functional level using the example of a two-wheel
self-balanced scooter with a steering lever, which is one of the use cases of a current
research project. In this large scale research project the digital product lifecycle
(DiP) is explored. Several variants of the self-balanced scooter were developed and
are used for the exploration of all elements of a product-lifecycle ranging from design
over production to operation and recycling. Figure 2 shows two realized variants of
the scooter (compare [7])—the variant shown on the left side was focused on cost
efficiency while the variant shown on the right side disposes of a weight-optimised
frame.

2 Fuzzy Decision Engine

The functional level describes, amongst others, the processes within a technical
system which allow to fulfil the central requirements. Functions can be connected
with the direct purpose of the system, e.g. for a self-balanced scooter, the central
function is to transport a person. However, much more functions are realised in such
a system such as a steering function, control functions and diagnosis functions. Fault-
tolerant design on the function level can be achieved, amongst others, by means of
Algorithms and Methods for Fault-tolerant Control and Design … 335

Fig. 2 Realised variants of a

self-balanced scooter

Fig. 3 Central parameters of

a self-balanced scooter

virtual redundancy. Based on the given example of the scooter, a fuzzy decision
engine, which is based on an analytical redundancy—a virtual sensor, is elucidated
in this section [12].
A self-balanced scooter basically consists of a platform which disposes of two
wheels that are independently driven by electric motors. For the further elaborations,
it is concluded that the scooter disposes of two sensor arrays. A first sensor array
is able to measure the inclination angle of the main platform of the scooter and,
consequently, the inclination angle of the person (α). The individual velocity of both
wheels of the scooter is measured by a second sensor array. From this information,
the velocity of the complete scooter can be calculated ( ż). For the elaborations in this
section it is assumed that both wheels drive with identical speed and that the scooter
is driving on a straight line. It is further assumed that both wheels are continuously in
contact with the floor, that both wheels are not objected to slipping as well as that the
driver is not behaving unusually or unexpectedly and can consequently be modelled
as a rigid body. These two central parameters are represented in Fig. 3.
336 R. Stetter et al.

If one seeks to analyze the dynamics of self-balanced scooters, the basis can be
a dynamic analysis of an inverted pendulum, because the main characteristics are
similar. This kind of dynamic analysis can lead to a state space model building upon
the proposal of [3] and employing the simplifications proposed by [15]; compare
also [13]. From an initial continuous form, the discrete form can be created employing
the Euler methods. It is possible to formulate a discrete state space model that includes
disturbances and faults in the following form:

x k+1 = Ak x k + B k uk + B k f a,k + W 1 w1,k , (1)

yk = C x k + C f f s,k + W 2 w 2,k , (2)

with

Ak = I + Ts · A, B k = Ts · B. (3)

In these equations x stands for the state vector, y denotes the measurement vector
and u stands for the input vector. The vector f a,k describes an actuator fault whereas
the vector f s,k stands for a sensor fault. w1,k denotes an exogenous disturbance vector
(including the discretization error) with a distribution matrix W 1 which is assumed
as known. w 2,k stands for the measurement uncertainties with a distribution matrix
W 2 which is assumed as known. The state consists of the distance covered by the
self-balanced scooter (z), the velocity of this scooter ( ż), the inclination angle of
this scooter (α) and the angular velocity of this angle (α̇). The input u denotes the
voltage applied to the electrical drive motors on both sides. The system matrices of
this state and the parameters in these state matrices are described in [12].
From the measurement of the state an input estimation can be achieved using
unknown input estimation; compare for instance [10]. In the case of the self-balanced
scooter, one has to accept that this system has an unknown input dk that must be
estimated for enabling the detection and identification of faults. Earlier research
activities concerning a product platform lead to the proposal of an adaptive estimator
for a comparable task [10]. The proposed estimator disposes of the capability to
estimate the unknown input dk and it does not need to rely on an estimation of the
state of the system xk . In the case of the self-balanced scooter, the measurement may
be described using the subsequent equation:

yk = C k x k + V v k . (4)

In this equation, v k denotes the measurement noise and V denotes its distribution
matrix which is assumed as known. It is possible to employ an unknown input
estimator [10]. The basis can be formulated for the given system in the subsequent
form:
Algorithms and Methods for Fault-tolerant Control and Design … 337

d̂k−1 = M k ( yk − C k x̂k−1|k−1 ). (5)

In this equation, x̂k−1|k−1 stands for an estimate of x k−1 . It is possible, to define the
innovation ỹk by
ỹk  yk − C k x̂k−1|k−1 . (6)

In the proposed estimation process, it is possible to estimate the unknown input

d̂k−1 from the measurement yk employing the matrix M k . Furthermore, it is possible
to compare this estimation with the voltage actually applied and to synthesize a
residual that may be applied for fault identification purposes:

z u,k = uk − û k . (7)

The next step in this process is the application of a fuzzy decision engine. In
this kind of intelligent system, the residual information resulting from the process
steps described above may be combined with certain elements of the state, in the
given case the inclination angle (α). The fuzzy decision engine is able to combine
information that results from residuals with information concerning the state and
specialist’s expertise in order to support a well-founded decision whether a certain
fault is present in the system. It is possible to expand a fuzzy decision engine into a
fuzzy virtual actuator [10]. As explained above, the fuzzy decision engine can decide
whether a certain fault is present.
It is possible to derive membership functions μz and μα for both the residual z and
elements of the state (in the given case the inclination angle α). These membership
functions μz and μα allow an primary evaluation of the available input informa-
tion. Research with similar intentions identified trapezoidal membership functions
as appropriate [6]. Their main strength is to represent an available input information
in an efficient manner. Two possibilities to find the width of the membership func-
tions exist. The first possibility is to analyse the maximum and minimum values of
the residuals. The second possibility is to analyse the maximum and minimum values
of the state variables. The advantages behind this is that it is possible to determine
the values of the parameters which define the width and inclination experimentally.
In order to be able to accommodate process noises, disturbances and mismatches
between the plant and the analytical models, it is sensible that the core is realised
as an interval around zero; it is possible in this kind of problem to find a sensible
size of this interval by means of the analysis of experimental data [6]. During the
development process of the proposed fuzzy decision engine, bot design experts and
control experts generally agreed that residuals in general can be a very powerful
fault indication, but they also agreed that a combination with other measured state
variables in a fuzzy logic system can be extremely fruitful. For the scooter, the
specialist’s expertise concerning two probable faults (first a sensor fault; second an
actuator fault) could be represented employing three membership functions for the
residual z i and the state variable α—which both serve as inputs of the fuzzy system.
In this example, the specialist’s expertise concerning a certain fault (e.g. a sensor
fault) could be formulated employing three membership functions for the residual
338 R. Stetter et al.

z i , which is an input of the fuzzy inference system. The first input membership
function μz1 was formulated as follows:
⎧
⎨ 1, z < c1
d1 −z
μz 1 = d1 −c1
, c1 ≤ z ≤ d1 (8)
⎩
0, z > d1 .

In these equations c1 and d1 stand for parameters which can be found based on
the specialist’s expertise or based on experimental data. The main intention of this
membership function is to characterize the situation that the residual indicates a fault
and that the value of the residual is negative. The next membership function μz2
indicates the case that the residual is close to zero, thus not indicating a fault. This
membership function was formulated as follows:
⎧
⎪
⎪ 0, (z < a2 ) or (z > d2 )
⎨ z−a2 , a ≤ z ≤ b
2 2
μz2 = b2 −a2 (9)
⎪
⎪ 1, b2 < z < c2
⎩ d2 −z
d2 −c2
, c2 ≤ z ≤ d2

In these equations, a2 , b2 , c2 and d2 stand for parameters which can be determined

based on specialist’s expertise or based on experimental data. The main intention of
this membership function is to characterize the situation that the residual indicates
no fault. The third membership function μz3 may be formulated in the subsequent
manner:
⎧
⎨ 0, z < a3
μz3 = bz−a 3
, a 3 ≤ z ≤ b3 (10)
⎩ 3 −a3
1, z > b3

In these equations, a3 and b3 stand for parameters which can be determined based on
specialist’s expertise or based experimental data. The main intention of this mem-
bership function is to characterize the situation that the residual indicates a fault and
that the value of the residual is positive. Similarly, also three membership functions
may be defined for each output variable. The first one μo11 may be formulated in the
subsequent form:
⎧
⎨ 1, a f o1 < co11
do11 −a f o1
μo11 = do11 −co11
, co11 ≤ a f o1 ≤ do11 (11)
⎩
0, a f o1 > do11

In these equations, a f o1 denotes the respective aggregated fuzzy output and co11
and do11 denote parameters which can be determined based on specialist’s expertise
or based on experimental data. A value in the given range indicates that another fault
(e.g., an actuator fault) has occurred, but not the fault associated with this fuzzy
Algorithms and Methods for Fault-tolerant Control and Design … 339

output. The second output membership function μo12 could be formulated in the
subsequent form:
⎧
⎪ 0, a f o1 < ao12 or a f o1 > do12
⎪
⎨ a f o1 −ao12 , a ≤ a f o ≤ b
o12 1 o12
μo12 = bo12 −ao12 (12)
⎪ 1,
⎪ bo12 < a f o1 < co12
⎩ do12 −a f o1
do12 −co12
, co12 ≤ a f o1 ≤ do12

In these equations, a f o1 denotes the respective aggregated fuzzy output and ao12 ,
bo12 , co12 and do12 stand for parameters which can be determined based on specialist’s
expertise or experimental data. A value in this range indicates that no fault has
occurred. The third output membership function μo13 may be formulated as follows:
⎧
⎨ 0, a f o1 < ao13
a f o1 −ao13
μo3 = bo13 −ao13
, ao13 ≤ a f o1 ≤ bo13 (13)
⎩
1, a f o> bo13

In these equations, a f o1 denotes the respective aggregated fuzzy output and ao13
and bo13 are parameters which can be determined based on specialist’s expertise or
experimental data. A value in the given range indicates that the associated fault (the
sensor fault) has occurred.

3 Evaluation

Two prominent fault scenarios can be distinguished in the given case of the self-
balanced scooter:
• The first one consists of a sensor fault which results in an incorrect voltage reading.
• The second one consists of an actuator fault infesting one or even two drive
motors—this fault will directly lead to an effect on the inclination angle.
Both faulty cases have been simulated. The residual z and the state variable α for a
sensor fault are represented in Fig. 4 ([12] Reprint with permission from Springer).
The residual z and the state variable α for an actuator fault are represented in
Fig. 5.
In Figs. 4 and 5 the blue line represents the residual z and the red line represents the
inclination angle α for a fault happening at k = 4000. It is obvious that the residual
might be similar in both cases, but that the state variable will present additional
helpful information. Both are input information for the fuzzy decision engine. The
results of this fuzzy decision engine are shown in Fig. 6 ([12] Reprint with permission
from Springer) respectively Fig. 7.
In Figs. 6 and 7 the blue line represents the result of one output membership
function—the one which aims to detect the specific sensor fault. For this output
340 R. Stetter et al.

Fig. 4 Residual z and state variable α in the case of a sensor fault

Fig. 5 Residual z and state variable α in the case of an actuator fault

Fig. 6 Output of the fuzzy interference system for the sensor fault scenario
Algorithms and Methods for Fault-tolerant Control and Design … 341

Fig. 7 Output of the fuzzy interference system for the actuator fault scenario

membership function, a positive value represents the information that—taking into

consideration both the residual AND the state variable—a conclusion is achieved that
the specific sensor fault is present (see Fig. 6). For this output membership function,
a negative value represents the information that—taking into consideration both the
residual AND the state variable—a conclusion is achieved that the specific sensor
fault is not present and that another fault is present (see Fig. 7). In Figs. 6 and 7
the red line represents the result of the other output membership function—the one
which aims to detect the specific actuator fault. For this output membership function,
a positive value represents the information that—taking into consideration both the
residual AND the state variable—a conclusion is achieved that the specific actuator
fault is present (see Fig. 7). For this output membership function, a negative value
represents the information that—taking into consideration both the residual AND
the state variable—a conclusion is achieved that the specific actuator fault is not
present, but that another fault is present (see Fig. 6). It is clearly visible that in Fig. 7
the opposite information is indicated than in Fig. 6; this result corresponds clearly
to the simulated input scenario. It is also obvious that for the sensor fault and the
actuator fault the presence of the fault is immediately detected; thus accentuating
the effectiveness of the proposed fuzzy decision engine. The implementation of the
specific realization of an analytical redundancy presents consequently an appropri-
ate high level measure intended to increase the fault-tolerance of the self-balanced
scooter.
342 R. Stetter et al.

4 Conclusions

Today’s products in the industrial and consumer market enable an enormous func-
tionality and performance. This is possible by means of more and more complex
systems. The drawback of this complexity is the growing possibility for faults; one
might say that faults are unpreventable in such systems. In the field of fault-tolerant
control, powerful algorithms were developed in recent years; since some years they
are supported by fault-tolerant design. Current investigations underline the immense
importance of the connection of the algorithm development with the system devel-
opment [11]. This paper presents a fuzzy decision engine which enables the identi-
fication and accommodation on the functional level of product concretization.

References

1. Blanke, M., Kinnaert, M., Lunze, J., Staroswiecki, M.: Diagnosis and Fault-Tolerant Control.
Springer-Verlag, New York (2016)
2. Ding, S.X.: Model-based Fault Diagnosis Techniques: Design Schemes, Algorithms, and Tools.
Springer-Verlag, Berlin/Heidelberg (2008)
3. Grasser, F., D’Arrigo, A., Colombi, S., Rufer, A.: Joe: a mobile, inverted pendulum. IEEE
Trans. Ind. Electron. 40(1), 107–114 (2002)
4. Khoury, B., Nejjari, F., Puig, V.: Reliability-aware zonotopic tube-based model predictive
control of a drinking water network. Int. J. Appl. Math. Comput. Sci. 32(2), 197–211 (2022)
5. Kukurowski, N., Mrugalski, M., Pazera, M., Witczak, M.: Fault-tolerant tracking control for
a non-linear twin-rotor system under ellipsoidal bounding. Int. J. Appl. Math. Comput. Sci.
32(2), 171–183 (2022)
6. Mendonca, L.F., Sousa, J.M.C., da Costa, J.M.G.: Fault isolation using fuzzy model-based
observers. IFAC Proc. Volumes. 39(13), 735–740 (2006)
7. Schuster, J., Pahn, F.: Entwicklung und bau zweier konzeptionell unterschiedlicher segways.
Bachelor-Theses Ravensburg-Weingarten University (RWU) (2018)
8. Srinivasarengan, K., Ragot, J., Aubrun, C., Maquin, D.: Parameter identifiability for nonlinear
lpv models. Int. J. Appl. Math. Comput. Sci. 32(2), 255–269 (2022)
9. Stetter, R.: Fault-Tolerant Design and Control of Automated Vehicles and Processes. Insights
for the Synthesis of Intelligent Systems. Springer-Verlag, Cham (2020)
10. Stetter, R.: A fuzzy virtual actuator for automated guided vehicles. Sensors. 20(15), (2020)
11. Stetter, R.: Algorithms and methods for the fault-tolerant design of an automated guided vehicle.
Sensors. 22(12), (2022)
12. Stetter, R., Witczak, M., Till, M.: Fault-tolerant design of a balanced two-wheel scooter. In:
Bartoszewicz, A., Kabziński, J., Kacprzyk, J. (eds.) Advanced Contemporary Control, pp.
1399–1410. Springer, Berlin (2020)
13. van der Veen, J.: Stabilization and trajectory tracking of a segway. University of Groningen,
Faculty of Science and Engineering (2018)
14. Witczak, M.: Fault diagnosis and fault-tolerant control strategies for non-linear systems. Lec-
ture Notes in Electrical Engineering, vol. 266. Springer International Publishing, Heidelberg,
Germany (2014)
15. Younis, W., Abdelati, M.: Design and implementation of an experimental segway model. In:
Proceedings of the 2nd Mediterranean Conference on Intelligent Systems and Automation,
(2009)
Robust Control of a Customized Lane
Change Maneuver

Benoit Vigne, Rodolfo Orjuela, Jean-Philippe Lauffenburger,

and Michel Basset

Abstract A full acceptation of Autonomous Vehicles (AV) could be achieved by

the customization of different driving services. One of these could consist of choos-
ing the driving style and consequently its impact on the generated trajectory shapes,
particularly in overtaking maneuvers. A previous work introduced a driving style
adaptive path planning method able to reproduce different behaviors during over-
taking maneuvers. This paper focuses on the integration of the aforementioned path
planning algorithm in the control scheme of an autonomous vehicle. The generated
adapted path is the reference of a robust feedback feed-forward robust controller.
The overall path generation and control are validated in simulation as well as in real-
world conditions. The obtained results show a good correlation between simulation
and experimental results.

1 Introduction

Although road deaths have decreased significantly over the last 50 years, 3219 people
still died in crashes in France in 2021 [8]. The main causes are alcohol and speed,
but overtaking and lane-changing maneuvers also represent potential dangerous sit-
uations (833 crashes with 14 deaths and 60 hospitalized people in 2020). Advanced
Driver Assistance Systems (ADAS) and automated driving vehicles offer promising
solutions to break this scourge by providing more security. However, this interesting
perspective is not sufficient to allow AV to be accepted by the population. Surveys of

B. Vigne (B) · R. Orjuela · J.-P. Lauffenburger · M. Basset

Université de Haute-Alsace, IRIMAS UR7499, Mulhouse, France
e-mail: [email protected]
R. Orjuela
e-mail: [email protected]
J.-P. Lauffenburger
e-mail: [email protected]
M. Basset
e-mail: [email protected]
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 343
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_30
344 B. Vigne et al.

Lateral
generation Lateral control
Behavioral
Global planning
planning
Longitudinal Longitudinal
generation control
Perception
Localization Navigation Control

Fig. 1 Global architecture of autonomous driving control

public opinion reveal concerns about the security of vehicles without driver control
[13]. To enhance trust in this new technology, one way consists on the development
of human-like interfaces between the user and the system [11]. Another aims to adapt
the system to the user behavior, i.e. by personalizing the vehicle dynamic to the driv-
ing style [4]. In this context of adaptation, a previous study proposed a method to
personalize the path shapes with one driving style parameter in an overtaking maneu-
ver [14]. This method showed good path planning results in simulation in different
conditions (varying driving styles, car and truck overtaking). This paper assumed an
efficient control loop able to track the generated adapted path. The main purpose of
this article is to present the control architecture of the dedicated motion planning and
highlight the performance of the planning-control loop either in simulation and in
real experiments.
Motion planning refers to the Navigation layer in the global architecture of the
AV [6]. As shown in Fig. 1, the Navigation integrates a global planner establishing a
route from a start to a destination point dealing with mission constraints (road type,
time spent, particular points of interest,. . .). This global path has to be adapted to the
local driving situation by a behavioral planner which generates dedicated maneuvers
(overtaking, road crossing, merging on highway,. . .). Finally, a local planner is used
for trajectory generation, i.e. lateral deviation and speed profile generation according
to safety and vehicle dynamic constraints. The Navigation layer is preceded by a
Perception/Localization layer carrying out the positioning and obstacle detection
challenges and followed by a Control layer aiming to provide the vehicle control
signals required to follow the defined path, i.e. throttle/braking and steering controls.
This paper focuses on the lateral path planning and control layers described in
Fig. 1 with a particular interest on a driving style parametrized lane change maneuver.
The outline of the paper is as follows: Sect. 2 presents the parametrized path planner.
The lateral controller description is done in Sect. 3. The experimental validation
process is detailed in Sect. 4 with results and discussion. Finally, Sect. 5 concludes
the paper.
Robust Control of a Customized Lane Change Maneuver 345

Fig. 2 Lane change path function of driving style parameter α

2 Personalized Path Generation

In robotics and generally speaking in vehicle control applications, parametric func-

tion interpolation methods are commonly used to generate lateral deviation paths.
The mathematical models are typically 3 to 5 order polynomial functions [12], con-
sidering that 5th order functions are optimal with respect to jerk minimization [15].
Clothoid curves have the property to provide a continuous curvature [3]. Bézier
curves are known for their low computational cost [7]. This list is not exhaustive and
more details could be find in [6]. The sigmoid curve has the benefit to be limited to
4 parameters for its computation. Its definition is:
dlat
y= dlong (1)
−a(x− −b)
1 + exp 2

The lateral deviation y depends on the longitudinal vehicle position x and four
parameters (Fig. 2): dlat corresponds to the entire lateral deviation from the initial
lane to the adjacent lane and depends on the road width. dlong is the global length of
the lane change maneuver and depends on the time to execute the entire maneuver.
So, these two parameters are predefined. a is a form factor representing the curvature.
Finally, b is a longitudinal translation allowing a spatial delay. a and b are used to
adjust the shape of the lane change path to the driving style. Previous studies [1] and
[14] have already shown that for security, dynamic and continuity constraints, both
parameters a and b have to be bounded, i.e. a ∈ [amin , amax ] and b ∈ [bmin , bmax ].
Independently and intuitively, the form factor a can be associated to a relaxed driving
style because the lower a is, the smoother the path is, with low jerk and acceleration
rates. So, a cost criterion relative to a and amin is defined in (2) with an optimization
function to be minimized:
⎧  2
⎨ a − amin
Jr elaxed = min
⎩
a amax − amin (2)
amin ≤ a ≤ amax
346 B. Vigne et al.

Fig. 3 Sigmoid parameters function of driving style parameter α

The sporty driving style is associated to the b parameter. The higher b is, the tighter
the path is, reducing the driving time (and distance) on the adjacent lane which gives
a sporty feeling. So, a cost criterion relative to b and bmax is defined in (3) with an
optimization function to be minimized:
⎧  2
⎨ bmax − b
Jtight = min
⎩
b bmax − bmin (3)
bmin ≤ b ≤ bmax

As a conclusion, the cost criterion (2) defines a relaxed style on the one hand and
the cost criterion (3) defines a sporty style on the second one. To manage a continuous
transition from one to the other, a weighting factor α varying from 0 to 1 links both
as follows: 
J = α Jr elaxed + (1 − α)Jtight
(4)
0≤α≤1

Finally, the optimization process finds the best couple [a, b] matching with the
predefined driving style parameter α as seen in Fig. 3. Both parameters a and b
increase with a driving style evolving from relaxed to sporty involving increasingly
sharp path shape as seen in Fig. 2. After this trajectory generation, a lateral controller
described in the coming section is used to track the reference path.

3 Lateral Guidance

The control design is based on the lateral dynamic bicycle model (Fig. 4) described
by the following state space form:

ξ̇ = Aξ + Bδ f (5)

with the state vector ξ = [y ψ]T where y is the lateral deviation in the vehicle local
frame, ψ its heading angle and δ f the front wheel steering angle. The matrices A
and B depend on vehicle parameters and their definition can be found in [9].
Robust Control of a Customized Lane Change Maneuver 347

Fig. 4 lateral and orientation error for path following

The controller shown in Fig. 5 provides the appropriate steering angle δ f in order
to follow a reference path. For that purpose, it has to minimize the lateral error e y
and the orientation error eψ between the vehicle and the reference path as shown in
Fig. 4 and defined by:

e˙y = v y + vx eψ eψ = ψ − ψr e f (6)

Here, the center of percussion (CoP) is preferred to the CoG to define the lateral
error (Fig. 4) due to its position in front of the vehicle allowing a better anticipation
of the trajectory similarly to a look-ahead control method [2]. The lateral CoP error
is defined by:
eCoP = e y + xCoP eψ (7)

Iz
with the CoP position defined by xCoP = where Iz represents the vehicle inertia
mL f
moment. From the bicycle model (5), both errors (6) and (7), the tracking error model
used to design the lateral guidance controller is:

ξ̇CoP = ACoP ξCoP + BCoP δ f + DCoP ωr e f (8)

where ξ = [eCoP ėCoP eψ ėψ ]T is the error state vector, ωr e f = [ψ̇r e f ψ̈r e f ] is the
disturbance vector with the desired yaw rate and yaw acceleration. The controller
used to track the parametrized path is based on [2] using both a feed-forward and a
robust state-feedback control:

δ f = uFB + uFF (9)

where u F F represents the feed-forward action and u F B is the feedback action.

348 B. Vigne et al.

Fig. 5 controller
architecture

The feed-forward aim is to reduce the impact of the disturbance ωr e f (t) on the
lateral error. From the definition of the matrix DCoP (see [2]), the feed-forward control
is obtained in order to remove the action of the ψ̇r e f in the disturbance vector:
 
m 2Rl C f L f m
uFF = + Vx ψ̇r e f + xCoP ψ̈r e f (10)
2Rl C f mVx 2Rl C f

L +L
with Rl = fL r r .
The state feedback action ensures the lateral stability by guarantying the error
vector exponential convergence towards zero as well as it attenuates the impact of
wr e f on the state variables. For this purpose, the following control law is proposed:

u F B (t) = −K ξ(t) (11)

where K represents the gain of the feedback. The Linear Matrix Inequality (LMI)
method is used to formulate this robust control problem and its resolution is done with
the vehicle model parameters given in [2]. The obtained gains for K are: [0.0417 −
−0.0352 0.9971 0.0408].

4 Experimental Validation Process

Experimental platform
The experimental validation process requires at least a vehicle and an appropriate
zone (test track) to carry out trials in secured conditions. The driving platform is
usually a commercial vehicle customized with sensors, actuators and embedded
Robust Control of a Customized Lane Change Maneuver 349

computing systems allowing technical skills and time in development. The test zone is
generally a private driving track due to the security conditions and the law agreements
to be respected.
The employed test vehicle is based on a Renault Grand Scenic (Fig. 6a) with
an architecture for autonomous driving shown in Fig. 6c. In the case of this study,
the Perception layer provides to the Navigation layer the current vehicle position
in the North East Down (NED) frame by using two real time kinematic (RTK)
global positioning systems (GPS). Proprioceptive metrics (vx , v y , ψ, ψ̇) required
for the error state vector computation are obtained from an inertial navigation system
which provides also the lateral acceleration a y . The lateral control is performed by a
servomotor coupled to the steering column and the constant longitudinal speed by the
vehicle cruise control module. All Algorithms for localization, trajectory planning
and control are implemented in the middleware RTMAPS (Real Time, Multisensor
applications) from Intempora [10]. RTMaps runs on an embedded system integrating
an industrial-grade GPU computing platform with an Intel core i7-9700 and 2×16
GB of RAM. It allows also to timestamp, record, synchronize and replay the collected
data. Trials are performed on a private track at the National Automobile Museum
of Mulhouse [5]. Its is formed by two rings links together and a peripheral track as
shown in Fig. 6b. It is about 100 m long and 40 m wide.
Experimental protocol
Experiments are conducted on the straight line of the museum track. First, the vehi-
cle drives on the track from a starting point during 80 m and records position points
with a sampling time Ts of 100 ms (Fig. 6b). The point set defines the global path.
Independently, a local path is generated for a driving style parameter α and a longi-
tudinal speed vx . Afterwards, the path is sampled with a spatial period x = vx Ts .
Finally, the local path is merged with the global one to obtain the reference path for
the lateral controller. This process is carried out off-line.
From the starting point, the autonomous vehicle performs the trajectory accord-
ing to the reference path and records metrics relative to the vehicle dynamics and
the controller. The experimentation is performed with three different driving style
parameter (α = 0: relaxed style, α = 0.5: daily style and α = 1: sporty style) at a
longitudinal velocity vx of 30 km/h. The road width dlat is fixed to 3 m and the total
length maneuver dlong to 80 m.
Experimental results
Reference trajectories for each driving profile are shown in Fig. 7a. The sportier the
driving style, the steeper the path shape, resulting in progressive lateral acceleration.
Due to the experimental conditions, the reference and vehicle starting positions are
slightly different generating undesired initial lateral and heading errors. But after
around 30 meters, the lateral controller corrects the trajectories and the vehicle lane
changes match with the expecting references (Fig. 7a). By removing the initial tran-
sient trajectories, i.e. the first 30 m (Fig. 7b), lateral errors reveal a maximal deviation
of 9 cm between the reference and the trajectory corresponding to 2.5% of the entire
lane change deviation (3 m). Heading errors show a maximal deviation of –0.002 rad
350 B. Vigne et al.

IMU + CAN Ser

Steering column
GPS RTK
Embedded
Eth
system Ser
LIDARs RTMaps Brake

Eth CAN
Camera Acceleration

Fig. 6 Artemips platform a, test environment b and architecture c

proving the controller efficiency. After a median filter processing in order to reduce
the signal noise, the maximum value of the vehicle lateral acceleration appears to
be depending on the driving style parameter but without exceeding the pre-set limits
(amax = ±2 ms−2 ) parametrized in the optimization proces .
Finally, the experimentation proves the feasibility of the lane change customiza-
tion, as well as the global to local path transform and the efficiency of the lateral
controller for low speeds. However, the test track is too short to validate the algo-
rithms at high speeds.
Simulation results
In order to extend the validation with more varying driving scenario, the algorithms
are tested by a simulation process requiring to replace the real test vehicle by a
mathematical model.
In our case, a 2 Degree Of Freedom (DOF), bicyle model, is adopted. It returns the
different state variables allowing localization and correction. The lateral controller is
implemented following the functional diagram in Fig. 5. Numerical testing conditions
are similar to the experimental process: vx = 30 km/h, α ∈ [0; 0.5; 1]. The time step
is 10 ms. Curves are shown in the Fig. 7c and present the same characteristics as the
experimental results. At low speed, the different non-linearities of the autonomous
vehicle (i.e ground/tire interaction forces, body deformation, ...) are irrelevant and the
2 DOF model is efficient: For the different driving styles, the correlation coefficient
r between simulated and real trajectories is near 1 as shown in Fig. 7d.
Robust Control of a Customized Lane Change Maneuver 351

Relaxed Relaxed
Daily Daily
Sporty Sporty

Relaxed Relaxed
Daily Daily
Sporty Sporty

Relaxed Relaxed
Daily Daily
Sporty Sporty

Relaxed Relaxed
Daily Daily
Sporty Sporty

Relaxed Relaxed
Daily Daily
Sporty Sporty

Complete lane change experimentation Zoom on [30-80] m

Relaxed
Daily
Sporty

Real
Simulated r=0.96
Relaxed
Daily
Sporty

Relaxed
Daily
Sporty

Real
Simulated r=0.95

Relaxed
Daily
Sporty

Real
Simulated r=0.99

Complete lane change simulation Trajectories comparison

Fig. 7 Lane change maneuver results

352 B. Vigne et al.

5 Conclusion and Perspectives

The shape of a lane change maneuver trajectory is personalized with one sole param-
eter allowing to provide from a comfortable to a sporty driving path. The theoret-
ical concept has been validated on one side by an experimental process on a real
autonomous vehicle and on another side by simulation tests. However, to satisfy a
full validation, the reference path generation has to be embedded in the real-time
process. Moreover, due to the length of the test track, the maneuver is performed
at low speed. To overcome these drawbacks and accelerate the transition between
simulations and experimentations, it is aimed in future works to develop a full digital
twin vehicle on an open source simulator.

Acknowledgements The authors gratefully acknowledge the help of our team members, Thomas
Josso-Laurain and Jonathan Ledy, whose contributions were essential for the experimental process.
The authors would also like to thanks the National Automobile Museum of Mulhouse for the access
to its test track.

References

1. Ammour, M., Orjuela, R., Basset, M.: Trajectory Reference Generation and Guidance Control
for Autonomous Vehicle Lane Change Maneuver. In: 2020 28th Mediterranean Conference on
Control and Automation (MED). IEEE, Saint-Raphaël, France (2020)
2. Boudali, M., Orjuela, R., Basset, M.: A comparison of two guidance strategies for autonomous
vehicles. In: 20th IFAC World Congress. IFAC, Toulouse, France (2017)
3. Alia, C., Gilles, T., Reine, T., Ali, C.: Local trajectory planning and tracking for autonomous
vehicle navigation using clothoid tentacles method. In: 2015 IEEE Intelligent Vehicles Sym-
posium (IV). IEEE, Seoul, Korea (South) (2015)
4. Chu, D., Deng, Z., He, Y., Wu, C., Sun, C., Lu, Z.: Curve speed model for driver assistance
based on driving style classification. IET Intell. Transp. Syst. 11(8), (2017)
5. Musée National de l’Automobile. https://www.musee-automobile.fr (2022)
6. González, D., Pérez, J., Milanés, V., Nashashibi, F.: A review of motion planning techniques
for automated vehicles. IEEE Trans. Intell. 17(4), 1135–1145 (2016)
7. Moreau, J., Melchior, P., Victor, S., Moze, M., Aioun, F., Guillemard, F.: Reactive path planning
for autonomous vehicle using bézier curve optimization. In: 2019 IEEE Intelligent Vehicles
Symposium (IV). IEEE, Paris, France (2019)
8. ONSIR. Road accidents: final data. Technical Report, ONSIR (2021)
9. Rajamani, Rajesh: Vehicle Dynamics and Control. Springer, US, Bosto, MA, USA (2012)
10. INTEMPORA Rtmaps. https://intempora.com (2022)
11. Ruijten, P.A.M., Terken, J.M.B., Chandramouli, S.N.: Enhancing trust in autonomous vehicles
through intelligent user interfaces that mimic human behavior. Multimodal Technol. Interact.
2(4), (2018)
12. Said, A., Talj, R., Francis, C., Shraim, H.: Local trajectory planning for autonomous vehicle with
static and dynamic obstacles avoidance. In: 2021 IEEE International Intelligent Transportation
Systems Conference (ITSC). IEEE, Indianapolis, IN, USA (2021)
13. Schoettle, B., Sivak, M.: A survey of public opinion about autonomous and self-driving vehicles
in the US, the UK, and Australia. Technical Report (2014)
Robust Control of a Customized Lane Change Maneuver 353

14. Vigne, B., Orjuela, R., Lauffenburger, J-P., Basset, M.: A personalized path generation for an
autonomous vehicle overtaking maneuver. In: 11th IFAC Symposium on intelligent autonomous
vehicles, IAV 2022. IFAC, Prague, Czech Republic (2022)
15. Ziegler, J., Stiller, C.: Spatiotemporal state lattices for fast trajectory planning in dynamic on-
road driving scenarios. In: 2009 IEEE/RSJ International Conference on Intelligent Robots and
Systems. IEEE, St Louis, MO, USA (2009)
Robust L2 Proportional Integral
Observer Based Controller Design
with Unmeasurable Premise variables
and Input Saturation

Ines Righi, Sabrina Aouaouda, Mohammed Chadli, and Said Mammar

Abstract In this paper, a descriptor Takagi-Sugeno (T-S) fuzzy system is considered,

subject to disturbances, sensor and actuator faults, in the presence of unmeasurable
premise variables and input saturation, which the later are considered as uncertain-
ties of the system. A robust L2 Proportional-Integral (PI) observer based controller
is designed. This observer based controller, can estimate both the system states, the
faults, and stabilize closed-loop states trajectories. An integrated robust controller
strategy is adopted by a novel structure of non Parallel Distributed Compensation
(non-PDC) control law. Moreover, the stability conditions of controller design are
expressed in term of Linear Matrix Inequalities (LMIs) constrained optimization
problem, by the proposed one-step design procedure. With the help of LMI TOOL-
BOX in MATLAB, we can easily design the type PI observer for efficient robust
state/fault estimation and solve the L2 observer based controller design problem of
discrete-time T-S fuzzy systems. The applicability of the proposed scheme is illus-
trated via a numerical example.

I. Righi · S. Aouaouda
University of Souk Ahras, LEER- BP 1553, Souk, Ahras 41000, Algeria
e-mail: [email protected]
S. Aouaouda
e-mail: [email protected]
M. Chadli (B) · S. Mammar
Université Paris-Saclay Evry, IBISC, 91020 Evry, France
e-mail: [email protected]
S. Mammar
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 355
D. Theilliol et al. (eds.), Recent Developments in Model-Based and Data-Driven Methods
for Advanced Control and Diagnosis, Studies in Systems, Decision and Control 467,
https://doi.org/10.1007/978-3-031-27540-1_31
356 I. Righi et al.

1 Introduction

Among the systems studied in control theory, systems with unknown inputs (UI) are
of particular interest. The challenge is thus to obtain a good estimation of the state
without any a priori knowledge on the unknown input. Fortunately, a T-S and LPV
fuzzy model provide an effective way to express the complicated nonlinear systems,
via a set of local linear models interpolated by membership functions . As a result, the
nonlinear control systems theory can be widely exploited to analyze and synthesize
the nonlinear systems expressed as multiple model approach [3]. Therefore, there is
a rapidly growing interest in FE and FTC problems for nonlinear system based on
the T-S fuzzy method and many important results have been reported for the topic in
the literature [5–7, 11].This paper focuses on discrete-time T-S fuzzy systems with
UI and disturbances, subject to states and input constraints. However, in many real
systems, the frequent occurrence of unknown faults often lead to performance degra-
dation and even instability of the system [1, 2]. In order to strengthen the system
reliability and guarantee system stability, the FE and FTC have received considerable
attention during the past few decades, and plenty of results of these research fields,
have been reported in the literature [9, 12]. Using FE/FTC results to design faults
observer and FEs for nonlinear systems directly is a challenging issue. For exogenous
disturbances, several minimum variance observers have been proposed for such T-S
systems [15]. However, in the present paper the disturbances have been assumed to
be with finite energy. Hence, the use of L2 observer is particularly suited, as this kind
of observer aims to minimize the transfer of energy between the disturbances and
the estimation error [17]. For that, the main contribution of this paper corresponds to
the design of an observer in charge of the sensor and actuator fault estimation simul-
taneously along with the system state, with fewer disturbances than the one presents
at the system dynamics.This paper is dedicated to the study of the observer design
based on the L2 approach, for T-S system with unmeasurable premise variables; the
main objective is to address the FE problem for a class of constrained T-S fuzzy
models subject to sensor and actuator faults, this controller can also be designed in
order to maintain stability, acceptable dynamic performance and steady state of the
overall system, also reduce the conservatism compared to previous work, such as [6,
17, 21], despite the presence of faults. This paper is organized as follows: Sect. 2
formulates the brief description of some notations and problem statements, for mod-
eling of the T-S structure of studied system. The description of the proposed observer,
the modified constrained robust controller, the control problem definitions and the
presentation of main results with the LMI based design conditions for robust L2 PI
observer based controller, are presented in Sect. 3. The effectiveness of the proposed
methods is clearly demonstrated by means of example in Sect. 4, and finally section
V provides some conclusions (Fig. 2).
Robust L2 Proportional Integral Observer Based Controller Design … 357

2 Notations and Problem Statements

2.1 Notations

The following notations are adopted to represent conveniently the different expres-
sions, given a set of nonlinear function: h i (.), vk (.), i ∈ Ir , k ∈ Ire are the nonlinear
scalar functions. This work focuses on unmeasurable premise variables, whose mea-
surements can be obtained from the observer design, which depends on the state
vector, and can be equivalently represented by a vector of states, expressed as h i (z k )
and vk (z k ), satisfying the convex sum property. For a vector x and z, x(k), z(k)
defined by xk , z k and x(k + 1) defined by xk+1 , the same for the other vector Ir
denotes the set (1, 2, · · · , r ), re denotes the set (1, 2, · · · , re ), R + represents the set
of positive real integer. I denote the identity matrix. An asterix * symbolizes the
symmetric block matrices. Nn denotes the set (1, 2, · · · , n).

2.2 Problem Statements

In the following section, the controller is derived using the descriptor form Suffi-
cient LMI constraints are derived from Lyapunov theory. Compared to [21], in the
following section a constrained controller is proposed, for this we consider the fol-
lowing class of T-S fuzzy model, subject to input saturation, external disturbances,
sensor/actuator fault:
⎧ re
⎨  v (z )E x r
k k k k+1 = h i (z k )(Ai xk + Bi sat (u k )) + Fa f k + Bω ωk
k=1 i=1 (1a)
⎩
yk = C xk + Fs f k

The state-space matrices E k , Ai , Bi , C, Bω are of the appropriate dimensions,Fa ,Fs ,

and the matrix C involved in (1a) is assumed to be a full row rank, k is a current sam-
ples, where i ∈ Ir represent the ith linear right hand-side submodel, k ∈ Ire represents
the kth linear left hand-side submodel of T-S fuzzy model (1a). Besides E, A,
B corresponding to the kth/ith subsystem contains the bounded uncertain terms,
which can be rewritten as: E = Ha De,k Ne , A = Ha Da,k Na , B = Hb Db,k Nb ,
with Ha , Hb , Na , Nb , He , Ne are known constant matrices, with Da,k , Db,k and De,k
are unknown matrices functions bounded, for all index  = a, b and e, i ∈ Ir , k ∈ Ire ,
T
a one has D,k D,k ≤ I .
System (1a) can be represented by a polytopic form:
⎧ re
⎨  v (zˆ )E x r
k k k k+1 = h i (zˆk )(Ai xk + Bi sat (u k )) + Fa f k + Bω ωk
k=1 i=1 (1b)
⎩
yk = C xk + Fs f k
358 I. Righi et al.

where the membership functions are denote h i (zˆk ) and vk (zˆk ) vary within the convex
sets 1 and 2 :

1 = (h i (zˆk ) ∈ Ir ; h i (zˆk ) = (h 1 (zˆk ), ...., h r (zˆk ))T (2a)

2 = (vk (zˆk ) ∈ Ire ; vk (zˆk ) = (v1 (zˆk ), ...., vre (zˆk ))T (2b)

Note that h i (zˆk ) and vk (zˆk ) depend on the variable zˆk verifying the convex sum
 T
property. The augmented form is adopted, where x ∗ = xk xk+ . The T-S system
(1a) can be equivalently rewritten in the following compact augmented form:

∗
E ∗ xk+1 = (A∗ xk∗ + B∗ sat (u k ) + Fa∗ f k + Bω∗ ωk
(3a)
yk = C ∗ xk∗ + Fs f k

0 I 0 I 0
A∗ = ; B∗ = ; E∗ = (3b)
(Aĥ + A) (E v̂ + E) (Bĥ + B) 00

3 Control Problem

3.1 Control Law

In order to satisfy the desired constrained controller performance, the augmented

controller design, is defined for system (3a):
⎧ −1∗ ∗
∗
⎪
⎨u k = Fĥ v̂ Hĥ v̂ xk
ψ(u k(l) ) = u k(l) − sat (u k(l) ) (4a)
⎪
⎩
ψ(0) = 0
  Hĥ v̂ 0
Fĥ∗v̂ = Fĥ v̂ 0 ; Hĥ∗v̂ = . (4b)
0 0

The gains Fĥ∗v̂ and Hĥ∗v̂ are matrices controllers to be determined, xk∗ is the estimated
augmented state variable of xk∗ . The architecture of the proposed constrained robust
controller design is based on the scheme depicted in Fig. 1. System (1a) or (3a)
achieves observability conditions, as detailed in [23]. To derive the controller laws,
a PI observer based controller is synthesized to estimate both faults and states for
system (3a) and has the augmented form:
Robust L2 Proportional Integral Observer Based Controller Design … 359

Fig. 1 The proposed robust controller design scheme

⎧ ∗
⎪
⎪ ˆ ∗ = A∗ xˆk ∗ + B ∗ (u k ) + Fa∗ fˆk + L ∗P ĥ v̂ (yk − yˆk )
E xk+1
⎪
⎨ y = C ∗ xˆ ∗ + F fˆ
k k s k
(5a)
ˆ = fˆk − L
⎪ f k+
⎪
⎪ I a ĥ v̂ (yk − yˆk )
⎩
u k = Fĥ∗v̂ Hĥ−1∗
v̂
xk∗

0
L ∗P ĥ v̂ = . (5b)
L P ĥ v̂

L ∗P ĥ v̂ is the augmented proportional gain for estimating the augmented variable state.
L I a ĥ v̂ is the integral gain for estimating the fault (Fig. 3).
The estimation error between the system (4a) and the observer (5a) is given by:
∗
e0,k = xk∗ − xˆk∗ (6)

The fault error is defined by:

e f,k = f k − fˆk (7)

In the following, substituting u k of (4a) into (3a) we consider the augmented form
as follows:

∗
E ∗ xk+1 = (A∗ + B ∗ F ∗ H −1∗ )xk∗ + Fa∗ f k + Bω∗ ωk − B ∗ (Fĥ v̂ H −1∗ e0,k
∗ − ψ(u ))
k
ĥ v̂ ĥ v̂ ĥ v̂ (8)
yk = C ∗ xk∗ + Fs f k
360 I. Righi et al.

Fig. 2 The control inputs 0.3

u
satu
0.25 umax
umin
0.2

0.15

0.1

0.05

−0.05
0 1 2 3 4 5 6 7 8
t(s)

Fig. 3 States variables and 0.5

x x x
their estimate TS1 NL1 es1

0
x1

−0.5
0 1 2 3 4 5 6 7 8
t(s)

1
x x x
TS2 NL2 es2

0.5
x2

−0.5
0 1 2 3 4 5 6 7 8
t(s)

Now, using the augmented state vector is defined by

 T
X k = xk∗ f k (9)

(9), and system (8) becomes:

⎧ −1∗ ∗
∗ ∗ ∗ ∗ ∗
⎨ X k+1 = Aĥ ĥ v̂ X k − B ψ(u k ) + Bω ωk − B Fĥ v̂ Hĥ v̂ e0,k
⎪
yk = C X k (10a)
⎪
⎩
u k = Fĥ∗v̂ Hĥ−1∗
v̂
X k − Fĥ∗v̂ Hĥ−1∗
v̂
∗
e0,k
Robust L2 Proportional Integral Observer Based Controller Design … 361
⎧
⎪ (A∗ + B∗ F ∗ H −1∗ ) Fa∗   H∗ 0
⎪
⎪ A = ĥ v̂ ĥ v̂ ; ∗ =
E∗ 0
; F ∗ = F ∗ 0 ; H∗ = ĥ v̂
⎪
⎪
⎨ ĥ ĥ v̂ 0 If 0 If ĥ v̂ ĥ v̂ ĥ v̂ 0 0
⎪
⎪  
⎪
⎪ ∗ Bω∗ B∗
⎪
⎩Bω = ; B∗ = ; C = C∗ 0
0 0
(10b)

The dynamic error is defined by:

E ∗ e0,k+
∗ ∗ − E ∗ x ∗ˆ = (A − L
= E ∗ xk+ ∗ + (F ∗ − L
C)e0,k F ) − B ∗ ψ(u k ) + Bω ω + 1
k+ ĥ v̂ P ĥ v̂ a I ĥ v̂ s ĥ
(11a)
with 1 = (A∗ + B ∗ Fĥ∗v̂ Hĥ−1∗
v̂
)xk∗ − B ∗ Fĥ∗v̂ Hĥ−1∗
v̂
∗
e0,k − B ∗ ψ(u k )

The dynamics of the fault estimation, with considering the fault constant: f k+ = f k
[16]. Considering the dynamics (11a) and (11b), the augmented error dynamics is
obtained as follows:

e f,k+ = f k − fˆk = L I ĥ v̂ Ce0,k
∗
+ (L I ĥ v̂ Fs + Ie, f )e f,k (11b)

Now, we consider the augmented system, by combining the system (8) and the error
dynamics (11), the closed-loop T-S system is obtained as follows:
⎧
⎨ek+ = M ek − B ψ(u k ) + Bω ωk
ĥ ĥ v̂ ĥ
(12a)
⎩u k = F ∗ H−1∗ X k − F ∗ H−1∗ ek
ĥ v̂ ĥ v̂ ĥ v̂ ĥ v̂

(Aĥ v̂ − L P ĥ v̂ C)− (Fa∗ − L I ĥ v̂ Fs ) E∗ 0

Mĥ ĥ v̂ = ; = . (12b)
L I ĥ v̂ C (L Iĥv̂ Fs + Ie f ) 0 Ie f

Now, we consider the augmented system x k = [X k ek ]T , by combining the system

(8) and the error dynamics (11), the closed-loop T-S system is obtained as follows:
⎧
⎪
⎨ E x k+1 = Aĥ ĥ v̂ x k − Bψ(u k ) + Bω ωk
yk = Cx k (13a)
⎪
⎩u = K∗ L−1∗ x
k ĥ v̂ ĥ v̂ k

  Hĥ v̂ 0
Kĥ∗v̂ = Fĥ v̂ −F ĥ v̂ L∗ĥ v̂ = (13b)
0 0

Moreover, to handle the non-linearity sat (u k ) , the dead-zone function ψu k will be

employed [15], where:
ψ(u k(l) ) = u k(l) − sat (u k(l) ) (14)
362 I. Righi et al.

we use the generalized sector condition proposed by [15] to deal with the dead zone
function. Also, the set: Du = (x k(l) ∈ R n x ; |(u k(l) − vk(l) )| < u 2max(l) ; l ∈ Inl ), with
the auxiliary signal: vk = Wĥ∗v̂ L−1∗
ĥ v̂
x k used as a degree of freedom in the design con-
ditions, and the condition: ψ(u k(l) )Sĥ−1 v̂
(ψ(u k ) − vk(l) ) holds. One way to construct
the estimate DoA is to employ level sets taken from the Lyapunov function associ-
ated with the closed-loop system. To this end, a non quadratic Lyapunov function is
considered
T −1
V (x k ) = x kT E P h E x k (15)

a level set associated with the Lyapunov function can be defined as in the following
lemma:
Lemma 3.1 Suppose thatV (x k ) given in (15) is a Lyapunov function for system
(13a). Then, a possible level set is given by:
 r
−T −1 T −1
Lv = ε(E P h E, 1) = i=1 ε(E P i E, 1) ⊂ Dx (16)

T −1 T −1
ρ > 0, ε(E P h E, ρ) = (x k ∈ R n x : x kT E P h E x k < ρ) (17)

Definition 1 The state trajectories of T-S descriptor system (13a) are contained
within the following polyhedral set (validity domain):

Dx = (|NmT x k | < 1; m ∈ In m ) (18)

Where the given matrix Nm represents the state constraints of system (13a), for
all x k ∈ R n x . 

3.2 LMI-Based Design Conditions of Constrained Descriptor

System

Property 1 [local stability]. Given a scalar α  , the initial condition x(0) belong to
T −1
a specific set in the state-space, which ε(E P h E, ρ) is a region of asymptotic
stability (RAS) for the saturated system. In the presence of disturbances, the controller
T
guarantees that the trajectories of (13a) are bounded, there exist a matrix P i > 0 and
T −1
a positive scalar ρ > 0 such that, for any x(0) ∈ ε(E P h E, ρ) and in the presence
of ω, the trajectories of the saturated system remains inside the polyhedral set Dx k ,
T −1
and do not leave the ellipsoid ε(E P h E, ρ) and converges exponentially to the
equilibrium point with a decay rate less than α  , satisfying the following property.
Property 2 [L2 gain-performance]. Given vector Nm defined in assumption 1, and a
positive scalar δ depending in the type of disturbances involved in the dynamics of
system (13a), see [15] we distinguish the following control problem.
Robust L2 Proportional Integral Observer Based Controller Design … 363

Control problem. In the presence of ωk . There exist positive scalar ρ and γ such that
for all x k ∈ Lv , the corresponding closed-loop trajectory (13a) remains inside the
validity domain Dx k in (18). Moreover the L2 -gain of the state vector x k follows:

||x k ||22 ≤ γ 2 ||ωk ||22 + ρ, k > 0 (19)

Where the objective is to attenuate the effects of exogenous input ωk on the

augmented state space by minimizing γ and ρ.
The objective now is to compute the gains of observer based controller (5a),
to ensure the stability of the closed-loop system (13a), guarantying the trajectories
performance for all ωk not nul, sufficient conditions to achieve this objective are
given through the following theorem:

Theorem 1 For a given the discrete-time T-S descriptor system (3a) with a non-
linearities parameter z k ∈ 1 and 2 , under input saturation with the proposed
observer based controller (5a), whose validity domain is defined by Dx ,is
locally exponentially stable if there exist a positive definite matrices Pi11 = Pi11 T
>
n x ×n x
0, Pi12 = Pi12 > 0, Pi2 = Pi2 > 0, Pi4 = Pi4 > 0, (H11 , H12 , H21 , H22 ) ∈ R
T T T
,
(Pi11 , Pi12 ) ∈ R n x ×n x , Pi2 ∈ R n f ×n f , Pi4 ∈ R n e f ×n e f , an observer gains (L I 11 , L I 12 ,
L I 21 , L I 22 ) ∈ R n u ×n x , (L P11 , L P12 , L P21 , L P22 ) ∈ R n y ×n x , apositive diagonal matri-
ces S jk ∈ R n u ×n u a positive scalars τ11 ,τ12 ,τ31 ,τ32 ,τb ,γ = γ 2 ,ρ, η where (i, j) ∈
(Ir × Ir ), k ∈ Ire , such that:
⎧
⎨min γ , ρ, η
such that
⎩ E T P −1 E ≥ 0
h

ρ + γ 2δ < 1 (20)
⎡ ⎤
 (1,1) ∗ ∗ ∗ ∗
⎢ 0 Pi2 ∗ ∗ ∗ ⎥
⎢ ⎥
⎢ 0 0  (1,1)
∗ ∗ ⎥
⎢ ⎥ ≥ 0, m ∈ Nm , i, j ∈ Nr (21)
⎣ 0 0 0 Pi4 ∗ ⎦
T jk∗ −W2∗jk T jk∗ −W4∗jk u 2max /ρ
⎡ ⎤
 (1,1) ∗ ∗ ∗ ∗
⎢ 0 Pi2 ∗ ∗ ∗ ⎥
⎢ ⎥
⎢ 0 0  (1,1) ∗ ∗ ⎥
⎢ ⎥ ≥ 0, m ∈ Nm , i, j ∈ Nr (22)
⎣ 0 0 0 Pi4 ∗ ⎦
∗
Nm∗ H jk 0 0 0 1/ρ

Tikj < 0
(23)
2
Tk
r −1 ii
+ Tikj + T jiT < 0,
364 I. Righi et al.
⎡ ⎤
 + XT GX ∗ ∗
Tikj = ⎣ ZT −G I ∗ ⎦ (24)
θ12
T
0 −θ22

4 Illustrative Example

This section gives design examples for the following nonlinear model:
⎧
⎪ 2 −1
⎨2x1,k+ − (1 + x1,k ) x2,k+ = x2,k x1,k − 0.5βx2,k + 0.1u k
2

2 −1
(1 + x1,k ) x1,k+ + 2x2,k+ = 0.7x1,k + sin(x1,k )x2,k (25)
⎪
⎩
yk = x1,k

Then, the nonlinearities are chosen as:


2 −1
z k = [z 1 , z 2 , z 3 ]T ; z 1 = (1 + x1,k ) , z 2 = x2,k
2
, z 3 = sin(x1,k ) (26)

Solving the LMIs of Theorem 1, the unknown gains of the PI observer based
controller are obtained. The constants α  , δ and ε are selected as 1, 0.2 and 1, and
since the main objective is to estimate the state variables and the faults.

5 Conclusion

It was considered that the descriptor T-S fuzzy system was affected by external
disturbance, sensor and actuator faults, with unmeasurable premise variables and
input saturation. The used strategy was based on the L2 performance criteria to
be robust against disturbance and faults. Furthermore, it was demonstrated that the
proposed approach is suitable to estimate system states and sensor/actuator faults
by a Proportional-Integral PI observer based controller, and stabilize the states and
controller into equilibrium point. Finally, a numerical example was presented to show
the effectiveness and applicability of the proposed approach.

References

1. Edwards, C., Spurgeon, S.K., Patton, R.J.: Sliding mode observers for fault detection and
isolation. Autom. 36(4), 541–553 (2000)
2. Su, X.J., Liu, X.X., Song, Y.D.: Event-triggered sliding mode control for multi-area power
systems. IEEE Trans. Ind. Electron. 64(8), 6732–6741 (2017)
3. Chadli, M.: «Stabilité et Commande de Systemes décrits par des Structures Multimodeles:
Approche LMI »Thèse INPL-CRAN, Nancy, France (2002)
Robust L2 Proportional Integral Observer Based Controller Design … 365

4. Li, X.J., Yang, G.H.: Dynamic observer-based robust control and fault detection for linear
systems. IET Control. Theory Appl. 6(17), 2657–2666 (2012)
5. Zhang, K., Jiang, B., Staroswiecki, M.: Dynamic output feedback fault tolerant controller
design for Takagi-Sugeno fuzzy systems with actuator faults. IEEE Trans. Fuzzy Syst. 18(1),
194–201 (2010)
6. Zhang, K., Jiang, B., Shi, P.: Analysis and design of robust H fault estimation observer with
finite-frequency specifications for discrete-time fuzzy systems. IEEE Trans. Cybern. 45(7),
1225–1235 (2015)
7. Dong, J.X., Hou, J.T.: Output feedback fault-tolerant control by a set theoretic description of
T-S fuzzy systems. Appl. Math. Comput. 301, 117–134 (2017)
8. Jia, Q.X., Chen, W., Wang, P.: Design of a PD-type learning observer for reconstruction of
actuator faults in descriptor systems. IET Control. Theory Appl. 11(1), 17–24 (2017)
9. Huang, S.J., Yang, G.H.: Fault tolerant controller design for T-S fuzzy systems with time-
varying delay and actuator faults: a K-step fault-estimation approach. IEEE Trans. Fuzzy Syst.
22(6), 1526–1540 (2014)
10. Chadli, M., Abdo, A., Ding, S.X.: H2/H fault detection filter design for discrete-time Takagi
Sugeno fuzzy system. Autom. 49, 1996–2005 (2013)
11. Liu, M., Cao, X.B., Shi, P.: Fault estimation and tolerant control for fuzzy stochastic systems.
IEEE Trans. Fuzzy Syst. 21(2), 221–229 (2013)
12. Liu, M., Cao, X.B., Shi, P.: Fuzzy-model-based fault tolerant design for nonlinear stochastic
systems against simultaneous sensor and actuator faults. IEEE Trans. Fuzzy Syst. 21(5), 789–
799 (2013)
13. Aouaouda, S., Chadli, M., Boukhnifer, M., Karimi, H.R.: Robust fault tolerant tracking con-
troller design for vehicle dynamics: a descriptor approach. mechatronics. 30, 316–326 (2015)
14. Aouaouda, S., Chadli, M.: Model reference tracking control for uncertain takagi-sugeno sys-
tems subject to sensor faults. IFAC-PapersOnLine 48(21), 1250–1255 (2015)
15. Tarbouriech, S., Garcia, G., da Silva Jr, J.M.G., Queinnec, I.: Stability and Stabilization of
Linear Systems With Saturating Actuators. Springer Science & Business Media (2011)
16. Escobet, T., Bregon, A., Pulido, B., Puig, V.: Fault Diagnosis of Dynamic Systems. Springer
International Publishing (2019)
17. Do, M.H., Koenig, D., Theilliol, D.: Robust H proportional-integral observer-based controller
for uncertain LPV system. J. Frankl. Inst. 357(4), 2099–2130 (2020)
18. Alwi, H., Edwards, C., Marcos, A.: Fault reconstruction using a LPV sliding mode observer
for a class of LPV systems. J. Frankl. Inst. 349(2), 510–530 (2012)
19. Zhen-Hua, W.A.N.G., Rodrigues, M., Theilliol, D., Yi, S.H.E.N.: Sensor fault estimation filter
design for discrete-time linear time-varying systems. Acta Autom. Sin. 40(10), 2364–2369
(2014)
20. Tanaka, K., Wang, H.O.: Fuzzy Control Systems Design and Analysis: a Linear Matrix Inequal-
ity Approach. John Wiley & Sons (2004)
21. Guzman, J., López-Estrada, F.R., Estrada-Manzo, V., Valencia-Palomo, G.: Actuator fault esti-
mation based on a proportional integral observer with nonquadratic Lyapunov functions. Int.
J. Syst. Sci. 1–14 (2021)
22. Boyd, S., Ghaoui, L., Feron, E.: Linear Matrix Inequalities in Systems and Control Theory.
SIAM, Philadelphia, PA, USA (1994)
23. Zhang, K., Jiang, B., Shi, P., Cocquempot, V.: Observer-Based Fault Estimation Techniques.
Springer International Publishing (2018)
24. Righi, I., Aouaouda, S., Chadli, M., Khelil, K.: Robust controllers design for constrained
nonlinear parameter varying descriptor systems. Int. J. Robust Nonlinear Control. 31(17),
8295–8328 (2021)
25. Righi, I., Aouaouda, S., Chadli, M., Khelil, K.: Robust L2 control for uncertain TS descriptor
model with input saturation. In: 2020 28th Mediterranean Conference on Control and Automa-
tion (MED), pp. 428–433. IEEE (2020)