International Journal of Mechanical Sciences 140 (2018) 133–150

Contents lists available at ScienceDirect

International Journal of Mechanical Sciences

journal homepage: www.elsevier.com/locate/ijmecsci

A two-stage approach to updating of mass, stiffness and damping matrices

Sharad Pradhan a,∗, S.V. Modak b
a
Department of Mechanical Engineering, NITTTR, Bhopal 462002, India
b
Department of Mechanical Engineering, Indian Institute of Technology, Delhi 110016, India

a r t i c l e i n f o a b s t r a c t

Keywords: Model updating techniques are used to update a dynamic FE model of a structure so as to obtain its accurate
Model updating representation in terms of mass, stiffness and damping matrices. Some of the existing updating methods update
Complex FRFs all the three matrices simultaneously using either complex FRFs or modal data. These methods, however, are
Normal FRFs
faced with numerical problems in practical implementation due to large difference in the magnitudes of the
Frequency response function
elements of the stiffness and mass matrices on one end and of the damping matrix on the other.
Damping matrix
This paper proposes a two-stage approach for updating mass, stiffness and damping matrices and performs
numerical and experimental investigations to assess the effectiveness of such an approach. The first stage of
updating is based on the concept of normal FRFs, which represent the FRFs of a structure if the structure were
undamped, to update the mass and stiffness matrices. In the second stage, the damping matrix is updated based
on the difference of complex and normal FRFs, which represents the effect of damping in the structure on its
frequency response.
The numerical example of a fixed-fixed beam structure is first considered that allows investigating the impact
of first stage of updating on the second stage. This is followed by an experimental example of an F shape structure.
A new method called ‘hybrid’ method to deal with data incompleteness in the context of two-stage updating is
suggested. Effectiveness of the two-stage method when updating parameters cannot be chosen correctly is also
investigated. The two stage approach presented and the results of the investigations carried out would be helpful
in proper implementation and application of the proposed updating method in practice.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction sented by mass and stiffness matrices, while others also consider updat-
ing of damping matrix. Since the focus of the current paper is on updat-
Accurate mathematical models of engineering structures are needed ing a damped FE model of a structure, further review here is restricted
in order to predict their dynamic characteristics accurately. In this con- to contributions made in this direction.
text finite element model updating is considered as an invaluable tool The approaches for identification of mass, stiffness and damping ma-
for obtaining an updated dynamic finite element (FE) model of a struc- trices of a structure can be broadly classified as either the approaches
ture that accurately represents its dynamics. that consider simultaneous updating of these matrices or the approaches
Updating of the structural dynamic FE models has been an active that consider updating of these matrices in two stages. Lin and Ewins
area of research for the last three decades and several approaches have [5] presented an iterative frequency response function (FRF) based
been proposed as shown in the surveys by Imregun and Visser [1], Mot- method for simultaneous updating of mass, stiffness and damping matri-
tershead and Friswell [2] and in the text by Friswell and Mottershead ces, in which the difference between the measured and analytical FRFs
[3]. Model updating is essentially an inverse problem as it attempts is linearized with respect to the parameters to be updated. Lin et al.
to identify certain unknown or uncertain model parameters from the [6] presented a method for updating of mass, stiffness and damping ma-
knowledge of the test data [4]. Model updating methods can be broadly trices based on sensitivity of eigendata. Ibrahim et al. [7] presented a
classified into direct methods, which are essentially non-iterative ones, technique to directly update mass, stiffness and damping matrices using
and the iterative methods. The direct methods are based on experimen- harmonic input-output. Only the measured input-output responses at a
tal modal data, while the iterative methods are based on either exper- limited number of time instants are required to update the complete FE
imental FRF data or experimental modal data. Many of these updating model. The response function based updating method developed by Lin
methods are aimed at updating only an undamped FE model, repre- and Ewins [5,8] can update mass and stiffness matrices only. The same
method is further extended using complex FRF data to update a pro-

∗
Corresponding author.
E-mail addresses: [email protected] (S. Pradhan), [email protected] (S.V. Modak).

https://doi.org/10.1016/j.ijmecsci.2018.02.033
Received 27 February 2017; Received in revised form 23 January 2018; Accepted 15 February 2018
Available online 16 February 2018
0020-7403/© 2018 Elsevier Ltd. All rights reserved.
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

updating of the damping matrix so as to satisfy the eigenvalue equation.

Nomenclature Lu and Tu [21] proposed a two-level neural network scheme for updat-
ing including damping matrix for the case of a lightly damped structure.
MA Analytical mass matrix In the first stage the effect of damping is neglected and the structural pa-
MX Experimental mass matrix rameters are updated using natural frequencies and mode shapes, while
KA Analytical Stiffness matrix in the second stage, updating of damping matrix is carried out using FRF
KX Experimental Stiffness matrix data. Datta et al. [22] proposed a two-stage direct updating procedure
D Structural damping matrix for mass and damping matrices updating using quadratic orthogonality
ZA Analytical DSM criterion.
ZX Experimental DSM It is seen that many of these approaches are based on the assump-
S Sensitivity matrix tion of light damping and hence are not applicable to a general class
TDOF Translational Degree of freedom of structures. Some other methods are based on accurate knowledge of
RDOF Rotational Degree of freedom mass and stiffness matrices, an assumption which may not always be
u, 𝛽 Vectors of physical parameters to be updated for stage-I satisfied in practice. Yet another strategy which has also been employed
and II respectively is to utilize the real part of the FRFs in the first stage to update mass
x(𝜔) Displacement vector and stiffness matrices and utilize the imaginary part of FRFs or the com-
f(𝜔) Force vector plex FRFs themselves for updating damping in the second stage. This
𝛼N Normal frequency response functions approach is also not strictly correct since real part of the FRFs is also
𝛼C Complex frequency response functions affected by damping.
𝛼𝑅𝐶 Real part of complex FRFs This paper proposes a two-stage framework for updating of all the
𝛼𝐼𝐶 Imaginary part of complex FRFs three matrices. The first stage of updating is based on the concept of
ZC Complex DSM normal FRFs, which represent the FRFs of the structure if the structure
ZN Normal DSM was undamped, to update mass and stiffness matrices. In the second
ΔZC Change in Complex DSM stage, the damping matrix is updated based on the difference of the
ΔZN Change in Normal DSM complex and normal FRFs, which represents the effect of damping in
𝑍̄ DSM portion corresponding to damping the structure on its frequency response.
Δ𝑍̄ Change in damping DSM portion Updating in a sequential manner couples the performance of the
damping matrix updating in the second stage to mass/stiffness matrix
updating in the first stage. Inaccuracies in the updated matrices in first
portional damping matrix [9]. Friswell et al. [10] extended the direct stage may have an adverse impact on the performance of the second
updating approach to updating of stiffness and damping matrices. Kuo stage and this coupling of performance needs to be investigated. The
et al. [11] developed a method for updating mass, damping, and stiff- incompleteness of the reference or measured data and any noise in it
ness matrices simultaneously by minimizing an objective function that also plays a role in making the second stage performance dependent on
defines changes in these matrices. Esfandiari et al. [12] proposed an FRF the first stage. The performance of the two-stage updating framework
based updating approach with the objective of damage detection by cor- in the presence of actual experimental data also needs to be assessed.
relating through sensitivity equations the changes in FRFs of a structure These are the issues, related to two-stage updating of mass/stiffness and
and the changes in the mass, stiffness and damping matrices. Lin and damping matrices, which form the subject matter of this paper. It un-
Zhu [13] further investigated the response function method proposed in dertakes numerical and experimental studies with the objective of ana-
[8] for updating damping matrix alongwith mass and stiffness matrices. lyzing above issues. A numerical study of a fixed-fixed beam structure is
Hu and Li [14] formulated a procedure that updates the mass, damping, presented followed by an experimental study of an F shape structure. A
and stiffness matrices of linear dynamic systems simultaneously from new method called ‘hybrid’ method to deal with data incompleteness in
modal data. Khanmirza et al. [15] proposed simultaneous identification the context of two-stage updating is suggested. Finally, a case study to
of mass, stiffness and damping matrices directly from the measured re- test effectiveness of the two-stage approach when updating parameters
sponses using neural networks. Recently, Hernandez et al. [16] proposed are not chosen correctly is considered.
an iterative time domain formulation for finite element model updating
that updates stiffness and damping parameters. It is observed that simul- 2. Two-stage model updating procedure
taneous updating of mass, stiffness and damping matrices may present
difficulty in the practical implementation since quite often the magni- In the two-stage FE model updating procedure, the updating of mass
tude of the stiffness and mass matrix elements are much more than that and stiffness matrices is performed in the first stage using the Normal
of the damping matrix elements [17,10]. In such a situation, updating Response Function Method (NRFM) [23]. Updating of damping matrix
iterations are more influenced by the stiffness and mass matrices than is performed in the second stage using the complex Response Function
the damping matrix. This may lead to an inaccurate identification of the Method [24]. The formulations for these two stages of updating are
damping matrix. briefly described below.
The approaches based on two stages of updating typically involve
updating of the mass and stiffness matrices in the first stage followed by 2.1. Stage I: mass and stiffness matrix updating
updating of damping matrix in the second stage. For lightly damped
structures, Brown et al. [18] developed a two-stage approach based The updating of mass and stiffness matrices carried out in stage-I is
on complex eigendata sensitivity. Reix et al. [19] considered updating done through an approach that is based on normal FRFs of a structure.
of damping matrix assuming that the mass and stiffness matrices are The governing equations of motion of a multi degree of freedom
perfectly known through a prior updating of them using experimental (DOF) structure with structural damping can be written in matrix form
modal data. The coefficients of a hysteretic damping matrix are then as,
updated using FRF data through least squares estimation. Halevi and ∙∙
𝑀 𝑥 (𝑡)𝑐 + 𝑖𝐷𝑥(𝑡)𝑐 + 𝐾 𝑥(𝑡)𝑐 = 𝑓 (𝑡) (1)
Kenigsbuch [20] performed updating of the damping matrix using com-
plex eigenvectors by assuming that an accurate knowledge of the stiff- Where M, K, and D are the real n × n mass, stiffness and structural
ness and mass matrices is available. Measured complex eigenvectors are damping matrices respectively. x(t)c and f(t) are n × 1 vectors of time-
corrected first so as to satisfy orthogonality relations. This is followed by varying displacements and forces and ‘n’ is the total number of degrees

134
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

of freedom in the finite element model. The superscript ‘c’ indicates the Where SN is the normal sensitivity matrix and Δ𝛼 N is the vector of dif-
quantity corresponding to the damped system. For harmonic excitation, ferences in the FRFs on the right hand side of Eq. (7).
The vector ‘u’ found by solution of Eq. (10) is used to update the
𝑓 (𝑡) = 𝑓 (𝑡)𝑒𝑗𝜔𝑡 and 𝑥(𝑡)𝑐 = 𝑥(𝑡)𝑐 𝑒𝑗𝜔𝑡 (2) vector of physical variables ‘p’ and then an updated version of the FE
model is built using the new set of physical variables. This process is
Let 𝛼𝐴𝑁 represent the normal FRF matrix corresponding to the FE repeated in an iterative way until convergence is obtained.
model of the structure. Normal FRFs are the FRFs of a structure without It should be noted that the uj found in jth iteration represents the
damping. These can be obtained by substituting Eq. (2) into (1), taking vector of fractional correction factors for the current ‘p’, i.e. pj , and does
damping matrix [D] = [0] and then solving the resulting equation for x. not represent a cumulative correction with respect to the ‘p’ existing
The coefficient matrix on the RHS is nothing but the matrix of normal before updating, the p0 . In view of this, at the end of jth iteration the
FRFs 𝛼(𝜔)N given by, ith cumulative fractional correction factor, is given by,
( ) ( ) ( )
𝛼(𝜔)𝑁
𝐴 = (−𝜔 𝑀𝐴 + 𝐾𝐴 )
2 −1
(3)
𝑢𝑖 = 1 + 𝑢1𝑖 ⋅ 1 + 𝑢2𝑖 ...... 1 + 𝑢𝑗𝑖 − 1 (11)
Let the normal FRF matrix for the actual structure be represented by
𝑁 . 𝛼 𝑁 represents the ‘measured’ normal FRF matrix of the structure in In this way, after the convergence, the final set of updating param-
𝛼𝑋 𝑋 eters, stored in vector 𝑢̄ 𝑖 , can be used to obtain the updated mass and
the hypothetical situation when the structure is undamped. Since it is
stiffness matrices.
not possible to remove damping from the actual structure, it is not possi-
ble to measure its normal FRF matrix. The FRFs measured on the actual
𝐶 . Chen 2.2. Stage II: damping matrix updating
structure are complex, represented by complex FRF matrix 𝛼𝑋
et al [17] has shown that there exists a direct relationship between the
𝐶 and the normal FRF matrix𝛼 𝑁 . The Once mass and stiffness matrices have been updated the updating of
measured complex FRF matrix 𝛼𝑋 𝑋
𝑁 𝐶 using following relationship. damping matrix is carried out in stage-II using complex FRFs. The basic
𝛼𝑋 can be directly computed from the 𝛼𝑋
updating equations involved in this stage are described below.
The dynamic stiffness matrix (DSM) corresponding to the FE model
𝛼(𝜔)𝑁 𝐶 𝐶 𝐶 −1 𝐶
𝑋 = 𝛼(𝜔)𝑅 + 𝛼(𝜔)𝐼 (𝛼(𝜔)𝑅 ) 𝛼(𝜔)𝐼 (4)
can be written as,
Where 𝛼(𝜔)𝐶 𝑅
and 𝛼(𝜔)𝐶
𝐼
are the real and imaginary parts of 𝛼(𝜔)𝐶
𝑋
.The 𝑍𝐴𝐶 = 𝑍𝐴𝑁 + 𝑗 𝑍̄ 𝐴 (12)
analytical normal FRF matrix of the structure𝛼(𝜔)𝑁 𝐴
, measured normal
FRF matrix 𝛼𝑋 𝑁 , and change in the normal dynamic stiffness matrix of where 𝑍𝐴𝑁 is the normal DSM, corresponding to the undamped FE model
𝑁
the analytical model Δ𝑍𝐴 are related by, [24], and𝑍̄ 𝐴 is the DSM due to damping in the FE model. For a structure with
structural damping the 𝑍 𝐴 can be written as,
𝛼(𝜔)𝑁 𝑁 𝑁 𝑁 𝑁
𝐴 Δ𝑍𝐴 𝛼(𝜔)𝑋 = 𝛼(𝜔)𝐴 − 𝛼(𝜔)𝑋 (5)
𝑍 𝐴 = 𝐷𝐴 (13)
Where Δ𝑍𝐴𝑁 is given by,
Where ‘D’ is the structural damping matrix. Let ΔDA represent a change
Δ𝑍𝐴𝑁 = Δ𝐾𝐴 − 𝜔2 Δ𝑀𝐴 (6) in DA . Hence,

Thus, it is seen that the updating Eq. (5) involves only those prop- Δ𝑍 𝐴 = Δ𝐷𝐴 (14)
erties that are dependent on only stiffness and mass properties of the The measured complex FRF matrix of the structure measured 𝐶,
𝛼𝑋
structure and don’t depend on its damping property. In view of this, up- 𝑁 , the damping matrix D and the change in the
normal FRF matrix 𝛼𝑋 A
dating of mass and stiffness matrices based on Eq. (5) is a theoretically damping matrixΔDA are related by, [24],
consistent formulation which isolates the process of mass and stiffness 𝐶 𝑁 𝑁 𝐶
matrix updating from the damping matrix. 𝛼𝑋 − 𝛼𝑋 = −𝑗𝛼𝑋 (𝑍 𝐴 + Δ𝑍 𝐴 )𝛼𝑋 (15)
Eq. (5) written for the jth column of 𝛼(𝜔)𝑁 𝑋
and 𝛼(𝜔)𝑁𝐴
yields the The left hand side of Eq. (15) is nothing but difference between the
following equation. The symbol ‘(𝜔)’ showing the dependence on fre- ‘measured’ complex FRFs and the corresponding normal FRFs. Since the
quency has been dropped for the sake of brevity. effect of mass and stiffness matrices of the structure on the frequency
response is completely quantified or captured by 𝛼𝑋 𝑁 , the difference ap-
𝛼𝐴𝑁 Δ𝑍𝐴𝑁 𝛼𝑋,𝑗
𝑁 𝑁
= 𝛼𝐴,𝑗 𝑁
− 𝛼𝑋,𝑗 (7)
pearing on the left hand side of Eq. (15) quantifies the influence of the
Let p = {p1 ,p2 ,......, pnu }be the vector of ‘nu’ number of stiffness and damping of the structure on the frequency response. This means that if
mass related physical parameters of the structure which are to be up- 𝛼𝑋𝐶 and 𝛼 𝑁 are completely known then the left hand side is completely
𝑋
dated during updating. Linearizing Δ𝑍𝐴𝑁 with respect to ‘p’ gives, independent of the mass and stiffness characteristics of the structure.
( ) On the right hand side of Eq. (15) we again have 𝛼𝑋 𝐶 and 𝛼 𝑁 that re-
𝑛𝑢
𝑁
∑ 𝜕𝑍𝐴𝑁 𝑋
Δ𝑍𝐴 = .Δ𝑝𝑖 (8) lates to ‘measured’ FRFs. Thus, it’s only the damping matrix of the FE
𝑖=1
𝜕 𝑝𝑖 model which appears in Eq. (15) and therefore the updating equation
for damping is unaffected by the mass and stiffness characteristics of the
𝑍𝐴𝑁 can be replaced by KA − 𝜔2 MA . Dividing and multiplying each
FE model. Thus, the updating Eq. (15) allows to completely isolate up-
term of the above equation by ‘pi ’ and then writing ‘ui ’ in place of Δpi /pi ,
dating of damping matrix in stage-II from updating of mass and stiffness
the above equation becomes,
matrices in stage-I. This is a very useful characteristic of the proposed
∑𝑛𝑢 ( )
𝜕(𝐾𝐴 − 𝜔2 𝑀𝐴 ) two stage updating procedure.
Δ𝑍𝐴𝑁 = .𝑝𝑖 .𝑢𝑖 (9)
𝑖=1
𝜕 𝑝𝑖 Eq. (15) written using the rth column of LHS is,
𝐶 𝑁 𝑁 𝐶
u = {u1 ,u2 ,......, unu } is the vector of unknown updating parameters. 𝛼𝑋,𝑟 − 𝛼𝑋,𝑟 = −𝑗𝛼𝑋 (𝑍 𝐴 + Δ𝑍 𝐴 )𝛼𝑋,𝑟 (16)
The individual parameters ui s represent fractional correction in the cho-
𝑁 ̄
sen updating parameters pi s . 𝐶
𝑗(𝛼𝑋,𝑟 𝑁
− 𝛼𝑋,𝑟 ) − 𝛼𝑋 𝐶
𝑍𝐴 𝛼𝑋,𝑟 𝑁
= 𝛼𝑋 Δ𝑍̄ 𝐴 𝛼𝑋,𝑟
𝐶
(17)
Eq. (7) after making substitution for Δ𝑍𝐴𝑁 from Eq. (9) can be written
at ‘nf’ number of frequency points chosen from the frequency range of Representing LHS of Eq. (17) by vector ‘br ’ we get,
𝑁 𝐶
interest. The resulting equations can be framed in the following matrix 𝛼𝑋 Δ𝑍 𝐴 𝛼𝑋,𝑟 = 𝑏𝑟 (18)
form,
The updating parameters ‘p’ are chosen as the correction factors,
𝑆 𝑁 (𝑛×𝑛𝑓 )×(𝑛𝑢) 𝑢𝑛𝑢×1 = Δ𝛼 𝑁 (𝑛×𝑛𝑓 )×1 (10) 𝛽 q , to the element damping matrices of the FE model. Therefore, the

135
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

change in the damping matrix can be expressed in terms of the updating

parameters as,
𝑁𝑒
∑ 𝑞
Δ𝐷𝐴 = 𝛽𝑞 𝐷 𝐴 (19)
𝑞=1
𝑞
where 𝐷𝐴 is the qth element damping matrix, 𝛽 q is the corresponding up-
𝑁𝑒
∑
dating parameter and indicates assembly over ‘Ne’ number of finite
𝑞=1
elements or their groups whose coefficients are selected for correction.
Substitution of Eqs. (14) and (19) in Eq. (18) gives,
( 𝑁𝑒 )
∑ 𝑞
𝑁 𝐶
𝛼𝑋 𝛽𝑞 𝐷𝐴 𝛼𝑋,𝑟 = 𝑏𝑟 (20)
𝑞=1
𝑞
After expanding element damping matrix 𝐷𝐴 to the size of the FE
model, Eq. (20) can be written as,
𝑁𝑒
∑
𝑁 𝑞 𝐶
𝛼𝑋 𝐷𝐴 𝛼𝑋,𝑟 𝛽𝑞 = 𝑏𝑟 (21)
𝑞=1

Eq. (21) written down at ‘nf ’ number of frequency points chosen

from the frequency range of interest can be combined and expressed in
the following matrix form,
𝑆 (𝜔)(𝑛×𝑛𝑓 )×(𝑛𝑒) 𝛽𝑛𝑒×1 = 𝑏𝑟 (𝜔)(𝑛×𝑛𝑓 )×1 (22)
Where ‘n’ is the DOF of the FE model and ‘𝛽’ is the vector of ‘ne’
correction factors.
Eq. (22) is implemented iteratively till the convergence is obtained.
Let 𝛽𝑖𝑗 is the ith correction factor found in jth iteration. Its cumulative
value,𝛽 𝑖 , at the end of ‘j’ number of iterations can be calculated as,
Fig. 1. Flow chart of two-stage updating method.
𝛽 𝑖 = (1 + 𝛽𝑖1 ).(1 + 𝛽𝑖2 ).....(1 + 𝛽𝑖𝑗 ) − 1 (23)
The final values of the damping correction factors are used to update
the damping matrix using Eqs. (13), (14) and (19). (3) Choose the correction factors𝛽 q to the elemental FE damping matri-
ces or the groups formed by them as the updating parameters.
(4) Set updating Eq. (18), and then convert it into Eq. (22), using the es-
2.3. Algorithm of the two-stage updating method
timate of normal FRFs, the ‘measured’ complex FRFs and the current
knowledge of the damping matrix.
Based on the mathematical formulations described in
(5) Find ‘𝛽’ by least square solution of Eq. (22). Construct the updated
Sections 2.1 and 2.2, the algorithm of the two-stage updating ap-
damping matrix using Eqs. (13), (14) and (19).
proach is summarized below and the flowchart of the method is
(6) Check for convergence. If the convergence criterion is not satisfied
depicted in Fig. 1.
then iterate by going to step (4) else stop with the damping matrix
Stage-I updating: constructed in step (5) taken as the final updated damping matrix.
(1) Construct an (initial) FE model of the structure using the knowledge
of geometry, material properties and the boundary conditions.
(2) Decide measurement points and measure FRFs at these points. These 3. Issues to be addressed in two-stage updating
are complex FRFs.
(3) Estimate normal FRFs from the measured complex FRFs using This section identifies the issues related to measurement and estima-
Eq. (4). tion of FRFs that need to be addressed in stage I and II updating and
(4) Choose mass and stiffness related updating parameters. discusses the link between the two stages of updating.
(5) Compute analytical FRFs corresponding to the undamped FE model.
(6) Set updating Eq. (7), and then convert it into (10), using estimate of
normal FRFs and the undamped analytical FRFs. 3.1. Issues in stage I updating
(7) Find ‘u’ by least square solution of Eq. (10). Using ‘u’ calculate mod-
ified value of vector of updating parameters ‘p’. The element mass A. Estimation of reference data used in updating:
and stiffness matrices are computed using these parameters and as- The first stage of updating uses Normal FRFs as reference data
sembled to obtain global updated mass and stiffness matrices. in model updating. The expression for estimating this is given in
(8) Check for convergence. If convergence criterion is not satisfied then Section 2.1. This expression requires full measured FRF matrix to be
iterate by going to step (5) else stop with the current updated mass known for estimation. This is a major issue to be addressed in stage-
and stiffness matrices representing the updated undamped FE model. I updating because the measured data is generally incomplete. It has
been shown in reference [23] that one strategy to obtain a reasonably
Stage-II updating:
accurate estimate of the Normal FRFs using incomplete data is to have
(1) Construct an initial estimate of the FE damping matrix of the struc- measurement of sufficiently large number of translational FRFs. This
ture. strategy will be followed in the numerical and experimental investiga-
(2) Use the normal FRFs, corresponding to ‘measured’ complex FRFs, tions carried out in this paper. However, truncation of the measured
estimated in step (3) of stage-I. data does introduce some errors in the estimation.

136
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Fig. 2. Details of beam structure.

-40
Exact Normal FRF
Complex Simulated Experimental FRF
-60 ***** Estimated Normal FRF
Receptance MAG in db (re:1 m/N)

-80

-100

-120

-140

-160

-180
0 100 200 300 400 500 600 700 800 900 1000
Frequency in Hz.

Fig. 3. Overlay of simulated experimental, exact normal and estimated normal

FRF based on 14 TDOFs: case of 25% coordinate incompleteness.

B. Incompleteness of measured FRFs:

Incompleteness of measured FRFs also affects setting up of updating
equations. This is dealt by replacing the Normal FRFs corresponding to
unmeasured DOFs by their FE model counterparts. Once updating equa-
tions have been set the equations corresponding to unmeasured DOFs
are dropped.
C. Noise in measured FRFs:
The measured FRFs generally have some random noise. This ad-
versely affects the estimation of Normal FRFs.
It is noted that all the above three issues may affect accuracy of
mass and stiffness matrices updating in stage-I. This may also affect the
performance of stage-II updating as explained below.

3.2. Issues in stage II updating

A. Estimation of reference data used in updating:

LHS of Eq. (16) is the difference between the measured FRFs and
the corresponding estimated Normal FRFs and can be looked upon as
reference data for damping matrix updating. This clearly indicates that
reference data for second stage updating is also dependent on Normal Fig. 4. Final values of the updating parameters: case of 25% incomplete data.
FRFs and hence is affected by accuracy of their estimation. (a) updated stiffness parameters after stage-I and (b) updated damping param-
B. Incompleteness of measured FRFs: eters after stage-II.
𝑁 needs to be completely known for setting up of up-
In Eq. (16), 𝛼𝑋
dating equations. The strategy followed in this work is to replace the

137
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

D. Noise:
Noise in measured FRFs affects estimation of Normal FRFs and
thereby directly affects the reference data used for stage-II updating.

4. Numerical study

This section investigates the performance of the two stage updating

procedure through a numerical example of a simple but a representative
fixed-fixed beam structure, shown in Fig. 2. The objective of the study
is to test the effectiveness of the two-stage approach for updating of
mass, stiffness and damping matrices and to evaluate the influence of
incompleteness of the experimental data on it. Another important point
of investigation is to study the influence of performance of updating in
the first stage on the second stage. The effect of simulated ‘noise’ on
updating is also studied.

4.1. Details of case study

The size of beam structure is 900 mm × 50 mm × 5 mm. The modulus

of elasticity and density of the structure is taken as 2.0e+11 N/m2 and
Fig. 5. FRF overlay before stage-I updating: case of 25% incomplete data FRF 7800 kg/m3 respectively. The initial FE model of the beam is built using
overlay before stage-I updating: case of 25% incomplete data (‘experimental’ thirty 2D beam elements with nodes at the ends fixed giving a total of 29
complex (‘——’ (red)) and FE model (‘—–’(black)) FRFs). (For interpretation of nodes with two degrees of freedom (translational and rotational) at each
the references to color in this figure legend, the reader is referred to the web node. The damping in the structure is modeled through a proportional
version of this article.) structural damping matrix in the initial FE model.
The simulated experimental data is obtained by introducing certain
known discrepancies in the initial FE model as follows:
Normal FRFs corresponding to unmeasured DOFs by their FE model
counterparts. (a) Young’s moduli of thirty finite elements are perturbed as detailed in
C. Influence of stage-I updating: Table 1.
It is mentioned in part B above that incompleteness is dealt by re- (b) The damping matrix DX of the simulated structure is represented as
placing the Normal FRFs corresponding to unmeasured DOFs by their 𝑁𝑒 ( )
∑ 𝛽𝑖 𝑖
FE model counterparts. This helps to overcome the problem of incom- 𝐷𝑋 = 1+ 𝐷𝐴,𝑒 (24)
𝑖=1
100
pleteness but requires FE mass and stiffness matrices to compute normal
FRFs. It is at this place that the updated mass and stiffness matrices from The summation sign represents assembly of the elemental matrices
stage-I chip in and start influencing stage-II updating. This is how stage- 𝑖
in the above equation. 𝐷𝐴,𝑒 represents the ith elemental damping ma-
II updating gets coupled/linked to stage-I updating. In this way all the trix of the initial FE model. The scalar 𝛽 i in the Eq. (24) represents the
possible sources of errors in stage-I updating, mentioned in Section 3.1, percentage deviation in the ith elemental damping matrix. The 𝛽 i S used
indirectly start affecting performance in stage-II updating since output to obtain DX are also given in Table 1. This gives a non proportional
of stage-I needs to be used in stage-II updating. structural damping matrix.

Fig. 6. FRF overlay after updating: case of 25% incomplete data (a) after stage-I (b) after stage-II.

138
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Fig. 7. Final values of the updating parameters after stage-I updating: case of
50% incomplete data with 1% noise (analytical data approach) (a) updated stiff-
ness parameters (b) % error in updated stiffness parameters. Fig. 8. Final values of the updating parameters after stage-II updating: case of
50% incomplete data with 1% noise (analytical data approach). (a) updated
damping parameters (b) % error in updated damping parameters.
The model obtained in the above manner thus represents the simu-
lated experimental structure. This simulated experimental structure is
solved for obtaining its frequency response. This frequency response is
treated as the ‘measured’ frequency response in the study.
Table 1
For the stage-I updating in which mass and stiffness matrices are up-
Discrepancies between the finite element and the simulated ‘experimental’
dated, the modulus of elasticity of the finite elements are updated. Thirty
model for beam with structural damping.
elements are grouped into ten groups of three finite elements each. The
fractional correction to modulus of elasticity of finite elements of each Group no. Element % deviation in % deviation
group is treated as an updating parameter thus giving a total of 10 updat- numbers Young’s modulus in 𝛽 S
ing parameters. Similarly for stage-II a common fractional correction to
1 1–3 +70% +8000%
damping matrix of the three finite elements in each group is taken as an
2 4–6 +10% +6000%
updating parameter. This also gives 10 number of updating parameters
3 7–9 +50% +9000%
for damping matrix updating. The values shown in Table 1 are, there-
4 10–12 +30% +5000%
fore, nothing but the correct values of the above two sets of updating 5 13–15 +20% +8000%
parameters. 6 16–18 +60% +4000%
7 19–21 +40% +10,000%
4.2. Updating using incomplete ‘experimental’ data 8 22–24 +10% +7000%
9 25–27 +40% +5000%
In practice, it is not realistic that the FRFs corresponding to all the 10 28–30 +70% +9000%
coordinates specified in an FE model have been measured, either due to
physical inaccessibility or due to difficulties faced in the measurement

139
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Fig. 9. FRF overlay: case of 50% incomplete data with 1% noise. (Analytical data approach) (a) after stage-I updating (b) after stage-II updating.

like that for measurement of rotational degrees of freedom. It is there-

fore assumed that the measurements are not available at all the degrees
of freedom of the finite element model. Moreover, it is also generally not
desirable to measure the complete FRF matrix due to the reasons of time
and economy. Two levels of incompleteness are considered by assuming
that translational FRFs are available at all the nodes and then at only
half of the nodes of the FE model. These have been referred as cases of
50% and 25% incomplete data respectively, the percentages represent-
ing the proportion of the total degrees of freedom of an FE model that
are taken as ‘measured’. Measurement of limited number of FRFs poses
two challenges in the implementation of the two-stage updating. The
first challenge relates to the estimation of normal FRFs and the second
challenge relates to implementation of updating equations.
Estimation of normal FRFs: Updating in the two stages requires a prior
estimation of normal FRFs (𝛼𝑋 𝑁 ) for the ‘experimental structure’ from

the knowledge of its complex FRFs (𝛼𝑋 𝐶 ). This estimation is based on

Eq. (4) which requires a square 𝛼𝑋 𝐶 matrix. This estimation is faced with

two types of FRF data incompleteness. The first problem related to in-
completeness is that FRFs at all the DOFs of FE model can not be mea-
sured e.g. for beam and plate like structures generally only translational
FRFs are measured while FRFs corresponding to rotational DOFs (RD-
OFs) are unmeasured. It is shown that if FRFs at a sufficient number of
translational DOFs (TDOFs) are measured then the estimation of nor-
mal FRFs using only the translational FRFs gives a quite close estimate
of the translational normal FRFs [23]. This strategy is adopted to make
up for the unmeasured RDOF-FRFs. However, it should be noted that
rotational normal FRFs still remain unknown.
The second problem related to the incompleteness is that even the
whole FRF matrix corresponding to all the TDOFs, as required above, is
not measured and the general practice has been to measure FRFs corre-
sponding to either one row or one column of TDOF-FRF matrix. In view
of this, the completeness of the 𝛼𝑋𝐶 is doubly hampered as far as estima-

tion of normal FRFs is concerned. This second incompleteness related

to the unmeasured columns/rows of the FRF matrix can be resolved by
regenerating the ‘unmeasured’ TDOF-FRFs in other rows and columns
using the modal model identified based on the ‘measured’ row/column
of the translational FRFs. A normal FRF estimated using above consid-
erations, for the case of 25% incomplete data, is shown overlaid on the
Fig. 10. Percentage error in updated stiffness parameters after stage-I updating: exact normal FRF in Fig. 3. It is seen that the error between the esti-
case of 50% incomplete data with 1% noise. (a) Analytical data approach (b)
mated normal FRF and the exact normal FRF is negligible and hence
hybrid data approach.
the above strategy to deal with FRF data incompleteness is effective for
the estimation of normal FRFs.

140
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Fig. 11. Final values of the updating parameters after stage-II updating: case of Fig. 12. Final values of the updating parameters after stage-II updating: case of
50% incomplete data with 1% noise (hybrid data approach) (a) updated damp- 50% incomplete data with 2% noise (hybrid data approach) (a) updated damp-
ing parameters (b) % error in updated damping parameters. ing parameters (b) % error in updated damping parameters.

Stage-I updating: The incompleteness referred in the above section ing parameters given in Table 1 indicates that the estimated parameters
also influences the setting-up of updating Eq. (7) (and hence Eq. (10)). are nearly exact. It is noted that the computation of the analytical FRFs
This equation requires knowledge of the complete normal FRF matrix, for dealing with incompleteness related to unmeasured RDOF-FRFs re-
which includes both the translational and rotational FRFs (in case of a quire the FE model stiffness and mass matrices obtained at the end of
beam or plate like structure). The translational normal FRFs are already stage-I updating. Since the mass and stiffness matrices could be accu-
estimated and the next difficulty now is associated with the estimation of rately updated in stage-I, there is no unfavorable influence of them on
rotational normal FRFs. In this section, the rotational normal FRFs are the damping matrix updating in stage-II.
approximated by their analytical counterparts. With this replacement FRF comparison after two stages of updating: An overlay of an ‘experi-
the updating Eq. (10) can be framed and updating carried out. Fig. 4(a) mental’ and the corresponding FE model FRF before updating is shown
shows a plot of the stiffness updating parameters estimated during stage- in Fig. 5. The overlay of these two FRFs after stage-I updating is shown in
I updating for the case of 25% incomplete data. A comparison with the Fig. 6(a). It is seen from the overlay that stage-I updating has been able
values of the correct updating parameters given in Table 1 indicate that to remove stiffness/mass matrices modeling errors and hence a good
the estimated parameters are exact. FRF match is seen away from the resonances and the antiresonances.
Stage-II updating: For the updating of damping matrix, again the up- The overlay after stage-II updating is shown in Fig. 6(b). The overlay
dating equations require full FRF matrix. However, since data is incom- shows a perfect match between the ‘experimental’ and FE model FRF
plete, the FRFs are available only at limited number of DOFs. Similar to after stage-II updating.
the previous section, incompleteness of FRF data related to the unmea-
sured RDOF-FRFs is dealt by replacing the ‘unmeasured’ FRFs by their 4.3. Updating using noisy incomplete ‘experimental’ data
analytical counterparts. Fig. 4(b) shows a plot of the damping matrix
updating parameters estimated during stage-II updating for the case of In practice, the measured FRFs are generally contaminated by mea-
25% incomplete data. A comparison with the correct values of the updat- surement noise. To check the robustness of the method under such con-

141
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Fig. 13. Comparison of FRF overlay after stage-II updating: case of 50% incomplete data (hybrid data approach) (a) with 1% noise and (b) with 2% noise (‘experi-
mental’ complex (‘——’ (red)) and updated FE model (‘—–’(black)) FRFs). (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

ditions, the complex FRFs, are polluted with a 2% random noise to ob- therefore have some error left in them. The updated mass and stiff-
tain noisy complex FRFs. ness matrices are subsequently used in stage-II for computation of
The incompleteness of ‘measured’ data is again dealt by replacing FRFs corresponding to unmeasured DOFs to tackle data incomplete-
the ‘unmeasured’ FRFs by their analytical counter parts computed using ness. This develops a link between the updating in stage-I and stage-
latest available estimate of FE mass and stiffness matrices at any stage. II. Thus, the lack of complete data causes the linking of stage-II with
For the case of 50% incompleteness and 1% noise, Fig. 7(a) and (b) stage-I. Due to this, stage-II updating is affected as a result of any
show plots of updated stiffness parameters and corresponding individual inaccuracies that might be left in stage-I updating. Lesser the degree
percentage errors respectively. of incompleteness, lesser would be the influence of stage-I updating
Similarly, Fig. 8(a) and (b) show updated damping parameters and errors on stage-II updating.
corresponding percentage errors respectively. It is seen that the maxi-
mum percentage error in stiffness updating parameters is 5.9%, while An improvement in the method to deal with incompleteness is sug-
the maximum percentage error in damping updating parameters is gested here in the framework of proposed two-stage updating approach
24.79%. Fig. 9(a) and (b) show overlays of an FE model FRF and the to reduce its unfavorable influence on updating. This improvement is
corresponding ‘measured’ complex FRF after stage-I and II updating re- based on using a combination of FE model and ‘experimental’ data. In
spectively. It is seen from the above results that stage-I and II updating the modified method, the ‘unmeasured’ complex FRFs are not directly
have been quite successful in terms of FRF matching, even though there computed using the updated mass and stiffness matrices and the current
is some error in the updated parameters in the two stages. Updating knowledge of damping matrix as is done in previous sections, but, are
for the case of 50% incompleteness and 2% noise is also done and it is regenerated using the modal data. For this purpose the complex ‘mea-
seen that not only the maximum percentage errors in both the stages sured’ FRFs are analyzed to obtain an estimate of the ‘experimental’
of updating have increased but the FRF-match after stage-II is also not natural frequencies and the corresponding damping factors. The mode
so good in the higher frequency range. The situation is further deterio- shapes are obtained by solving the eigenvalue problem using the up-
rated with 25% incompleteness and 1% noise. It is seen from the FRF dated mass and stiffness matrices. In this way, ‘experimental’ natural
comparison that the mismatch of the FRFs is localized around resonant frequencies and the damping factors along with the FE model mode
and anti-resonant frequencies indicating that it is the stage-II updating shapes are used to regenerate the ‘unmeasured’ complex FRFs.
that is being more affected as compared to stage-I updating. In stage-I updating, for the case of 50% incompleteness and 1% noisy
data, the identified parameters are nearly the same using both the ana-
4.4. Treatment of ‘experimental’ data incompleteness using hybrid data lytical and hybrid data based strategies. This is clear from the compari-
son of percentage errors in updating parameters shown in Fig. 10. It is
It is observed from the results given in the previous section that noted that both the strategies involving replacing the ‘unmeasured’ FRF
data with some approximation of them are self-corrective in nature, be-
(a) In both the stages of updating, higher is the level of incompleteness, cause as the FE model improves over the iterations, the approximation
higher is the error in the identified updating parameters. This is due also improves and hence in the limit the error of approximation may
to the fact that the incompleteness of ‘experimental’ data related to get eliminated if the FE model indeed converges to the ‘correct’ model.
the unmeasured rotational FRFs is dealt by substituting the corre- This is the reason why the parameters identified by both the approaches
sponding data from the analytical model. Since the analytical model are not much different. However, in practice, this self-corrective prop-
is not accurate it leads to inaccuracies in the estimation of updating erty will be less effective with analytical data based approach if one is
parameters. not able to choose a ‘correct’ set of updating parameters. With hybrid
(b) The percentage error in stage-II updating is on higher side as com- data approach self-corrective property may still be at least partially ef-
pared to stage-I updating. Due to noise in the ‘measured’ FRFs, the fective. It is however noted that there is a marginal improvement in
stiffness and mass matrices are not updated accurately in stage-I and the convergence characteristics of the updating parameters in stage-I

142
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Fig. 15. FRF overlay after stage-II updating (hybrid data approach): case of
25% incomplete data with 1% noise.

FRFs is much improved when incompleteness is resolved using hybrid

data approach. Hybrid data approach also behaves well when the incom-
pleteness is further increased to 25% as seen from the Figs. 14 and 15
which clearly indicate that the percentage errors in updating parameters
are lower and the FRF overlay is better as compared to corresponding
updating with analytical model based approach.

5. Updating using equivalent updating parameters

In the previous sections the parameters that need to be updated are

assumed to be correctly known. For example, the errors simulated in
the FE model were related to modulus of elasticity and the parameters
updated were also the modulus of elasticity of individual finite elements.
However, in practice, it’s not always possible to correctly identify
the parameters that should be updated and most often it is the engi-
neering judgment, coupled sometimes with sensitivity analysis, which
Fig. 14. Final values of the updating parameters after stage-II updating: case of is generally resorted to in choosing the updating parameters. If param-
25% incomplete data with 1% noise (hybrid data approach) (a) updated damp- eters chosen in this manner don’t match with the ‘correct’ parameters
ing parameters (b) % error in updated damping parameters. then such updated parameters can be viewed as a set of equivalent pa-
rameters yielding correction in the structural matrices that try to re-
duce the error between the ‘experimental’ and the analytical data. This
updating based on hybrid data strategy. Hybrid strategy is found to re- section is devoted to simulate above situation to test effectiveness of
quire fewer numbers of iterations for the convergence. This is because the proposed two-stage updating procedure for its robustness in giving
of the fact that the hybrid data based approach uses a correct estimate an improved FRF correlation when updating parameters chosen don’t
of some of the natural frequencies and corresponding damping factors contain those parameters that are in error. ‘Measured’ data is gener-
right from the first iteration resulting into a faster convergence. With ated through an ‘experimental’ beam structure in which stiffness and
hybrid data approach to deal with incompleteness, the parameters and damping discrepancies are introduced in the form of torsional spring
errors in stage-II updating as shown in Fig. 11(a) and (b) are more accu- stiffnesses (kt1 and kt2) and torsional damping coefficients (dt1, dt2) at
rate than those with analytical data based approach as shown in Fig. 8(a) the two ends. Table 2 gives the value of these parameters. ‘Experimen-
and (b). tal’ complex FRFs are generated through this model from which normal
Similarly, for the case of 50% incompleteness and 2% noisy data, the FRFs are computed as explained before.
errors in stage-II updating are reduced from a maximum error of 263.2% The FE model of the beam structure is same as used in the previous
obtained using analytical data approach to 7.3% (Fig. 12(a) and (b)) sections. The updating parameters for the two stages are also the same
obtained using the hybrid data approach. This indicates that the hybrid i.e. the fractional correction to the modulus of elasticity of the thirty up-
method is a more effective procedure of dealing with data incomplete- dating parameters for stage-I and coefficients of the elemental damping
ness in stage-II because a partial use of ‘experimental’ data to regenerate matrices for stage-II. It is noted that the parameters to be updated don’t
‘unmeasured’ FRFs helps to reduce the influence of any errors in mass exactly contain the parameters that are actually in error. Updating using
and stiffness matrices in stage-I updating flowing down to damping ma- two-stage procedure is performed using complete ‘measured’ FRF data.
trix updating in stage-II. Overlays of FE model and ‘measured’ complex Fig. 16(a) and (b) show bar chart of the identified updating parameters
FRF after stage-II updating using hybrid data approach for the case of in stage-I and II respectively. It is seen that stage-I shows a significant
50% incompleteness with 1% and 2% noisy data are shown in Fig. 13(a) drop in the Young’s modulus of the finite elements at the ends while the
and (b) respectively. These overlays further reflect that the matching of Young’s modulus of other finite elements show only a relatively smaller

143
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Table 2
Torsional spring stiffnesses (kt1 and kt2) and torsional damping coefficients (dt1, dt2) at the two
ends of the ‘experimental’ beam structure.

kt1 (in N-m/radian) kt2 (in N-m/radian) dt1 (in N-m-s/radian) dt2 (in N-m-s/radian)

1.0e+04 1.0e+04 0.7e+04 0.7e+04

Fig. 16. Final values of the updating parameters: case of equivalent updating parameters. (a) Updated stiffness parameters after stage-I and (b) updated damping
parameters after stage-II.

change. The values of the updating parameters therefore indicate that

the stiffness modeling error is situated at the two ends. Similarly, stage-
II updating shows a large change in the damping matrix coefficients of
the finite elements at the ends indicating that the damping modeling
errors are localized near the two ends. In this way, the two stages of
the updating are able to correctly localize the zones of the modeling er-
rors, which in this case are the two ends of the beam, and perform an
equivalent adjustment of the chosen updating parameters.
An overlay of an ‘experimental’ and the corresponding FE model FRF
before updating is shown in Fig. 17. The overlay of these two FRFs after
stage-I updating is shown in Fig. 18(a). It is seen from the overlay that
stage-I updating has been able to make equivalent corrections in the
stiffness/mass matrix and hence a good FRF match is seen away from
the resonances and antiresonances. The overlay of above two FRFs after
stage-II updating is shown in Fig. 18(b). The overlay shows a good match
between the ‘experimental’ and FE model FRF after stage-II updating. In
this way, it is seen that the two-stage updating procedure has success-
fully made an equivalent correction sequentially in the mass/stiffness
matrices and the damping matrix.
The case study in this section demonstrates that even when the up-
dating parameters cannot be chosen correctly, but the chosen parame- Fig. 17. Analytical and ‘experimental’ FRF overlay before stage-I updating: case
ters have the ability to capture the influence of the modeling errors in of equivalent updating parameters.
an equivalent way, then the proposed two-stage updating procedure is
effective in yielding an updated model that has a good correlation with
the measured FRF data.
6.1. Details of the case study

6. Experimental study on an F-shaped structure The F-shape structure has been built by bolting horizontally the two
beam members to a vertical beam member, which in turn has been
This section presents an experimental investigation into two-stage welded at the bottom to a base plate. All the beam members have a
updating of an FE model of an F-shape structure that resembles the skele- square cross-section of side 37.7 mm. Fig. 19 shows the basic dimen-
ton of a drilling or a milling machine tool. sions of the F-structure used in the study.

144
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Fig. 18. Analytical and ‘experimental’ FRF overlay: case of equivalent updating parameters (a) after stage-I updating (b) after stage-II updating.

used to measure the frequency response function using an accelerometer

(B&K model 4368) and an impact hammer (B&K model 8202).
The FRFs are measured over a frequency range of 1 KHz by exciting
the structure at above locations with an impact hammer and measuring
the response with an accelerometer kept fixed at one of the locations.
This gives one row of the translational FRF matrix corresponding to 18
nodes. For the estimation of normal FRFs from the measured complex
FRFs, a full square complex FRF matrix is required. In this work, this is
done by measurement of all the rows of the translational FRF matrix.
This is accomplished by changing the accelerometer position to other
nodes and again measuring all the FRFs as explained above. In this way
the lower triangular part of the FRF is completely measured which effec-
tively gives the full FRF matrix, of order 18 × 18, under the assumption
of symmetricity of the FRF matrix as a result of reciprocity theorem.

6.2. Stage-I updating

This section presents the results of mass and stiffness matrices up-
dating performed in stage-I.
Estimation of normal FRFs: Stage-I updating requires normal FRFs of
the structure which can be computed from Eq. (4). The measured FRF
data is generally incomplete. The strategy followed here is similar to
Fig. 19. F-shape-structure.
that mentioned in section on ‘Updating using incomplete ‘experimental’
data’ of the numerical study, which is based on estimating normal FRFs
using a sufficiently large number of measured translational FRFs. In view
of this the measured translational FRF matrix of order 18 × 18 is used to
compute normal FRFs.
A finite element model of the structure is built using forty eight 2- Fig. 21 shows a plot of an estimated normal FRF corresponding to
D-frame elements with 52 nodes with three degrees of freedom (two the FRF 𝛼 55 × 94 , which indicates response at node 55 due to excitation at
displacements and one rotation) at each of the nodes. The FE mesh is node 94, using above strategy. For comparison the corresponding mea-
shown in Fig. 20. The values of the modulus of elasticity and the density sured complex FRF is also shown. The estimated Normal FRF shows
are taken as 2.0e11 N/m2 and 7800 kg/m3 respectively. In the present quite sharp peaks and valleys at resonance and antiresonance respec-
case due to the presence of three joints the modeling of stiffness at these tively as compared to damped peaks and valleys in the experimental
places is expected to be a dominant source of modeling inaccuracy in the FRF. It is seen that the computed Normal FRFs display an expected char-
FE model. In view of this, the three joints present are modeled by tak- acter of a normal FRF and hence seem to represent well the FRFs of a
ing coincident nodes at each of them. Thus, there are two geometrically hypothetical ‘undamped’ F shape structure.
coincident nodes at a joint instead of just one node. The three corre- With a view to see the effect of number of translational FRFs used in
sponding degrees of freedom at the two nodes are coupled through three the estimation of normal FRFs, estimation is done using complex square
springs respectively. The stiffness of these springs could be the potential Translational FRF matrix corresponding to only 12 nodes. Fig. 22 shows
updating parameters for accounting the flexibilities that might be there a normal FRF computed in this manner with the corresponding mea-
at the joints. Onosokki make dual channel FFT analyzer, CF-7200A, is sured FRFs (𝛼 55 × 94 ) overlaid on them. It is observed that the estimated

145
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Fig. 20. Details of measurement points on F shape structure used in updating.

normal FRF is still accurate though error in estimation can be seen at Table 3
some frequency points. Values of the updating parameters for stage-I updating of F-structure.
Above results validate the strategy of estimation of normal FRFs even
Updating Initial value (in Updated value (in
in the case of experimental data as the results indicate that the estima-
parameters N-m/radian) N-m/radian)
tion is reasonably accurate when the number of translational FRFs is
‘sufficient’. Kt1 3.3E+06 2.9129e+005
Choice of updating parameters for stage-I: One of the important issues Kt2 3.3E+06 2.9956e+005
in FE model updating using iterative methods is the selection of the Kt3 3.3E+06 3.1126e+005
updating parameters. In case of iterative methods, the selection of up-
dating parameters would decide how good a match is obtained between
an FE model data and the experimental data. Choice of updating param- Results of stage-I updating: The undamped FE model of the F shape
eters on the basis of engineering judgment about the possible locations structure is now updated with three torsional spring stiffness parame-
of modeling error in a structure is one of the strategies to ensure that ters. Table 3 shows values of the updating parameters before and after
only the physically meaningful corrections are made. updating. Updating is seen to predict lower torsional stiffnesses at the
In the present structure due to the presence of three joints the mod- three joints. Fig. 23 shows a comparison of undamped FE model point
eling of stiffness at these places is expected to be a dominant source of FRF 𝛼 64 × 64 and the corresponding normal FRF computed from the mea-
localized modeling inaccuracy in the FE model. The stiffness of these sured FRFs. The comparison shows a significant discrepancy between
springs, Kx, Ky and Kt respectively, could be the potential updating pa- the two FRFs. A similar comparison after updating is shown in Fig. 24
rameters allowing to account for the deviation in the stiffness of the re- from which it is seen that the updated undamped model FRF matches
gions covered by the joints. A sensitivity analysis is performed to further very well with the normal FRF of the structure.
reduce the number of unknowns and it is noticed that the eigenvalues The frequency points chosen during updating were restricted to the
are much more sensitive to the torsional stiffness at the three joints than left side of the vertical line shown in Fig. 24 indicating an updating
to other spring-stiffness parameters. In the light of this observation, the frequency range of 0–680 Hz. These figures also show comparison out-
three torsional stiffness parameters are chosen as updating variables for side the updating frequency range (i.e. over a range of 680–1000 Hz).
updating mass and stiffness matrices. The other stiffness values are taken The updated undamped FE model is found to predict the FRF accurately
very large to represent rigid coupling of the corresponding degrees of even in this range from where no frequency points were used in updat-
freedom. ing indicating that the corrections suggested are not only able to give

146
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Overlay of Est. Normal (blue....) & Experimantal (Red___) FRFs

-80

-100 Estimated Normal FRF

Experimental FRF

-120
Receptance MAG in db (re:1 m/N)

-140

-160

-180

-200

-220

-240
0 100 200 300 400 500 600 700 800 900 1000
Frequency in Hz.

Fig. 21. Overlay of experimental (‘——’ (red)) and estimated normal FRFs (‘…….’(blue)) 𝛼 55 × 94 with 18 TDOFs only. (For interpretation of the references to color
in this figure legend, the reader is referred to the web version of this article.)

Table 4
Correlation of measured and FE model natural frequencies for F-structure.

Mode Measured natural Initial FE model natural Initial percentage Updated FE model Percentage error
number frequency (Hz) frequency (Hz) error (%) natural frequency (Hz) after updating (%)

1 34.35 41.5 17.2 34.10 −0.73

2 103.78 120.7 14.0 101.52 −2.22
3 133.56 177.2 24.6 134.72 0.86
4 316.88 370.2 14.4 312.92 −1.26
5 979.48 1013.4 3.34 977.67 −0.18

an improved FRF prediction but also are physically meaningful. A com- Table 5
parison of the measured and the updated model natural frequencies is Groups of finite elements for stage-II updating for F-structure.
given in Table 4. With stage-I updating, the error in natural frequencies
Group no. Finite elements in the group
has been reduced substantially in all the modes.
1 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
6.3. Stage-II updating 2 12, 13, 14, 15, 16, 17
3 18, 19, 20, 21
4 22, 23, 24, 25, 26, 27, 35, 36, 37
At the end of stage-I updating, we have now an updated undamped
5 28, 29, 30, 31
FE model, in the form of updated mass and stiffness matrices, and com- 6 32, 33, 34
plex FRFs corresponding to the experimental structure. At this stage, a 7 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48
comparison of updated undamped FE model FRF 𝛼 64 × 64 and the corre-
sponding measured FRF is shown in Fig. 25. The comparison indicates
that the FRFs match very well in terms of the frequencies of the reso-
nance peaks and antiresonances, indicating that the updated mass and 48 finite elements is divided into 7 groups. These groups are shown
stiffness matrices represent the ‘undamped’ behavior of the experimen- diagrammatically in Fig. 26 and the details of the finite elements in each
tal F-structure accurately. It is however noted that, there is a significant group are given in Table 5. Damping in groups 1, 3, 7 and 5 represent
difference in the FRF magnitude between the FE model and the exper- structural damping in beam members, while damping in groups 2, 4 and
imental FRF at and around resonances and antiresonances. This differ- 6 represent equivalent structural damping in the zones containing joints.
ence in the two FRFs is due to the damping model that is missing from In this way there are seven updating parameters, which are nothing but
the undamped FE model and this difference would form the basis for the correction factors to the coefficients of the damping matrices of each
identifying a structural damping matrix via stage-II updating. group (represented by variables 𝛽 q in Eq. (19)).
Choice of updating parameters: The damping in the structural members Results of stage-II updating: Damping matrix identification is per-
and the joints is the source of energy dissipation in the F shape structure. formed using the algorithm given in ‘stage-II updating theory’ in ‘Two-
To capture the influence of these sources, the complete structure with stage updating method section’. It is noted that normal FRFs need to be

147
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

Overlay of Est. Normal(Blue...) & Experimental(Red___) FRFs

-80

-100 Estimated Normal FRF

Experimental FRF
Receptance MAG in db (re:1 m/N)

-120

-140

-160

-180

-200

-220

-240
0 100 200 300 400 500 600 700 800 900 1000
Frequency in Hz.

Fig. 22. Overlay of experimental (‘——’ (red)) and extracted normal FRFs (‘…….’(blue)) 𝛼 55 × 94 with 12 TDOFs only. (For interpretation of the references to color
in this figure legend, the reader is referred to the web version of this article.)

Overlay of Analytical (black...) & Ext. Normal (red___) FRFs Overlay of Upt. Analytical (black...) & Ext. Normal (red___) FRFs
-80 -80
Updated undamped FE Model FRF
Undamped FE Model FRF Experimental Normal FRF
-100 before updating -100
Receptance MAG in db (re:1 m/N)

Receptance MAG in db (re:1 m/N)

-120 -120

-140 -140
Experimental Normal FRF
-160 -160

-180 -180

-200 -200

-220 -220

-240 -240
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Frequency in Hz. Frequency in Hz.

Fig. 23. Overlay of FRF 𝛼 64 × 64 before stage-I updating: experimental normal Fig. 24. Overlay of FRF 𝛼 64 × 64 after stage-I updating: experimental normal
(‘——’ (red)) and undamped FE model (‘…….’(black)) FRFs. (For interpretation (‘——’ (red)) and updated undamped FE model (‘…….’(black)) FRFs. (For inter-
of the references to color in this figure legend, the reader is referred to the web pretation of the references to color in this figure legend, the reader is referred
version of this article.) to the web version of this article.)

estimated only once as has been done in the previous section. However, a bar chart of the percentage change in the coefficients. The identified
the unmeasured complex FRFs need to be computed at each iteration correction factors for groups 1,3,7 and 5, representing structural beam
using the updated mass and stiffness matrices and the knowledge of members, are not very much different, while the identified correction
damping matrix in the current iteration. factors for groups 2, 4 and 6, representing zones of joints, are higher,
The initial damping matrix before updating is taken as proportional suggesting a higher level of damping in the joints. Overlay of the 𝛼 64 × 64
to the stiffness matrix with the constant of proportionality being 0.0001. updated FE model and the corresponding measured complex FRFs, after
A total of 10 frequency points chosen in the updating frequency range stage-II updating is shown in Fig. 28. The updated FE model FRF is seen
(0–680 Hz) were used in updating. Table 6 shows the values of coeffi- to have a very good match with the corresponding experimental FRF
cients of the damping matrix of the seven groups, while Fig. 27 shows both inside as well as outside the updating frequency range.

148
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

OVERLAY OF ANALYTICAL (Black...) & EXPERIMENTAL (Red___) FRFs 180

G2 G4
-80
Updated undamped FE Model FRF 160
Experimental FRF
-100
140 G6

Frac t ional Correc tion Fac tors

G5
Receptance MAG in db (re:1 m/N)

-120 G3
120
G1
-140 100 G7

-160 80

-180 60

40
-200

20
-220
0
1 2 3 4 5 6 7
-240
0 100 200 300 400 500 600 700 800 900 1000 Parameter no.
Frequency in Hz.
Fig. 27. Correction factors to the elemental matrices: change in updating
Fig. 25. Overlay of FRF 𝛼 64 × 64 after stage-I updating: experimental complex parameters- stage II.
(‘——’ (red)) and updated undamped FE model (‘…….’(black)) FRFs. (For inter-
pretation of the references to color in this figure legend, the reader is referred
to the web version of this article.) OVERLAY OF ANALYTICAL(Black...) & EXPERIMENTAL(Red___) FRFs
-80
Group 2 Updated damped FE Model FRF
o o 13 12 11 10 9 8 7 6 5 4 3 2 1 Experimental FRF
-100
16 Group 1
17
Receptance MAG in db (re:1 m/N)

-120
18
19 -140

20 Group 3
-160
21
22
-180
23 Group 7
-200
Group 4 o o 36 37 38 39 40 41 42 43 44 45 46 47 48

26
-220
0 100 200 300 400 500 600 700 800 900 1000
27
Frequency in Hz.

28
Fig. 28. Overlay of FRF 𝛼 64 × 64 after stage-II updating: experimental complex
29 (‘——’ (red)) and updated FE model (‘…….’(black)) FRFs. (For interpretation of
Group 5 the references to color in this figure legend, the reader is referred to the web
30 version of this article.)
31
32
Y 33 Group 6 7. Conclusions

34 This paper presents a two stage approach to updating of mass, stiff-

X
ness and damping matrices of a structure. It also presents numerical and
Fig. 26. Groups of finite elements for choosing updating parameters for stage-II. experimental investigations using this approach. The two-stage updat-
ing approach has the unique characteristic that the first stage updating
Table 6
is carried out using an estimate of experimental ‘undamped’ FRFs, re-
Values of the updating parameters for stage-II updating for F-structure.
ferred as normal FRFs, instead of the complex measured FRFs used by
Updating parameters Initial value Group number Updated value other existing methods. This theoretically eliminates any possibility of
effect of contribution of damping in the measured complex FRFs on the
𝛽q 0.0001 Group1 0.0102 mass and stiffness matrices updating and hence results into their more
Group 2 0.0169 accurate updating. In the second stage, updating of damping matrix is
Group 3 0.0121 driven by the difference of measured complex FRFs and an estimate of
Group 4 0.0174
the experimental ‘undamped’ FRFs. Since this difference solely repre-
Group 5 0.0123
sents contribution of damping to the measured FRFs, the updating car-
Group 6 0.0135
Group 7 0.009 ried out in stage-II is theoretically insulated from the stiffness and mass
matrices updating in stage-I.

149
S. Pradhan, S.V. Modak International Journal of Mechanical Sciences 140 (2018) 133–150

The two-stage procedure is first investigated on a simulated beam References

structure. Through this study it is found that if complete ‘measured’
data is available then stage-II updating involving damping matrix up- [1] Imregun M, Visser WJ. A review of model updating techniques. Shock Vibr Dig
1991;23:141–62.
dating is completely insulated from the stage-I updating. However, this [2] Mottershead JE, Friswell MI. Model updating in structural dynamics: a survey. J
is found not to be completely true if the ‘measured’ data is incomplete. Sound Vib 1993;167:347–75.
Two methods are suggested to deal with incompleteness in the context [3] Friswell MI, Mottershead JE. Finite element model updating in structural dynamics.
Dordrecht: Kluwer Academic Publishers; 1995.
of the proposed two-stage updating procedure. The first is based on us- [4] Mottershead JE, Datta BN. Special issue on ‘Inverse Problems in MSSP’. Mech Syst
ing the current FE model to compute substitutions for the ‘unmeasured’ Sig Process 2009;23(6):1731–3.
FRFs. The second method suggested makes use of ‘hybrid’ data that is a [5] Lin RM, Ewins DJ. Model updating using FRF data. In: Proceedings of 15th interna-
tional modal analysis seminar, K.U. Leuven, Belgium; 1990. p. 141–63.
combination of using experimentally identified natural frequencies and
[6] Lin RM, Lim MK, Du H. A New Complex inverse eigensensitivity method for struc-
damping factors and analytically obtained mode shapes to regenerate tural damping model identification. Comput Struct 1994;52(5):905–15.
the unmeasured FRFs. It is found that the hybrid data based method to [7] Ibrahim SR, D’ambrogio W, Salvini P, Sestieri A. Direct updating of nonconservative
finite element models using measured input-output. Proc X, IMAC 1992:202–10.
deal with incompleteness results into a more accurate updated model
[8] Lin RM, Ewins DJ. Analytical model improvement using frequency response func-
and helps partially restore the insulation of stage-II from stage-I updat- tions. Mech Syst Sig Process 1994;8(4):437–58.
ing. [9] Imregun M, Visser WJ, Ewins DJ. Finite element model updating using frequency
Simulated study incorporating ‘measurement’ noise and data incom- response function data–I: theory and initial investigation. Mech Syst Sig Process
1995;9(2):187–202.
pleteness shows that the two-stage procedure is able to yield a close [10] Friswell MI, Inman DJ, Pilkey DF. Direct updating of damping and stiffness matrices.
match between the FE and the ‘measured’ FRFs, though the updated AIAA J 1998;36(3):491–3.
parameters are found to have small deviations from their exact values. [11] Kuo Y, Lin W, Xu S. New methods for finite element model updating problems. AIAA
J 2006;44(6):1310–16.
The data incompleteness necessiates the use of updated mass and stiff- [12] Esfandiari A, Bakhtiari-Nejad F, Sanayei M, Rahai A. Structural finite element model
ness matrices in stage-II updating for computation of unmeasured FRFs. updating using transfer function data. Comput Struct 2010;88(1-2):54–64.
This causes linking of stage-II to stage-I updating that otherwise is insu- [13] Lin RM, Zhu J. Model updating of damped structures using FRF data. Mech Syst Sig
Process 2006;20:2200–18.
lated from stage-I in the case of complete data. It is however concluded [14] Hu SJ, Li H, Wang S. Cross-model cross-mode method for model updating. Mech
that if the mass and stiffness matrices are updated accurately in stage-I Syst Sig Process 2007;21(4):1690–703.
then the stage-II updating also leads to a more accurate identification of [15] Khanmirza E, Khaji N, Johari Majd V. Model updating of multistory shear buildings
for simultaneous identification of mass, stiffness and damping matrices using two
the damping matrix.
different soft-computing methods. Expert Syst Appl 2011;38(5):5320–9.
The two-stage procedure is then investigated on an experimental F- [16] Hernandez EM, Bernal D. Iterative finite element model updating in the time domain.
shape structure. This structure being assembled from beam members Mech Syst Sig Process 2013;34(1–2):39–46.
[17] Chen SY, Ju MS, Tsuei YG. Estimation of mass stiffness and damping matrices from
through joints makes it challengeable for stiffness and damping matrix
frequency response function. ASME J Vibr Acoust 1996;118:78–82.
modeling. For the experimental structure the normal FRFs could be esti- [18] Brown GW, Farahat C, Hemez FM. Extending sensitivity to lightly damped structures.
mated well using only the translational FRFs. Stage-I updating is found AIAA J 1997;35(8):1369–77.
to yield accurate stiffness and mass matrices that predicted accurately [19] Reix C, Tombini AGerard, Strobbe J, Dascotte E. Updating the damping matrix using
frequency response data. In: Proceedings of the XIV, IMAC, Dearborn, Michigan;
the FRFs of the hypothetical ‘undamped’ experimental structure. Stage-II 1996. p. 585–90.
updating also yielded an accurate equivalent model of structural damp- [20] Halevi Y, Kenigsbuch R. Model updating of the complex modeshapes and the damp-
ing matrix appropriately capturing the distribution of damping in the ing matrix. Inverse Probl Eng 2000;8:143–62.
[21] Lu Y, Tu Zhenguo. A two-level neural network approach for dynamic FE model up-
structural beam members and the joints. dating including damping. J Sound Vib 2004;275:931–52.
The numerical and experimental investigations into the two-stage [22] Datta BN, Deng SV, Sokolov O, Sarkissian DR. An optimization technique for damped
updating approach thus indicate that the updating of mass and stiffness model updating with measured data satisfying quadratic orthogonality constraint.
Mech Syst Sig Process 2009;23(4):1759–72.
matrices followed by updating of damping matrix is a theoretically con- [23] Pradhan S, Modak SV. Normal response function method for mass and stiffness ma-
sistent, numerically superior and robust approach to mass. stiffness and trix updating using complex FRFs. Mech Syst Sig Process 2012;32:232–50.
damping matrix updating. The results of the investigations carried out [24] Pradhan S, Modak SV. A method for damping matrix identification using frequency
response data. Mech Syst Sig Process 2012;33:69–82.
in this paper would be helpful in proper implementation and application
of the proposed two-stage updating method in practice.

150