2014 Sanchez ReviewOfForceReconstructionTechniques (Edited)

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Journal of Sound and Vibration ] (]]]]) ]]]–]]]

Contents lists available at ScienceDirect

Journal of Sound and Vibration


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

Review

Review of force reconstruction techniques


J. Sanchez, H. Benaroya n
Department of Mechanical and Aerospace Engineering, Rutgers University, 98 Brett Road, Piscataway, NJ 08854, United States

a r t i c l e i n f o abstract

Article history: An important engineering problem is the recovery of the input of a system given its output.
Received 9 September 2013 This is a difficult problem to solve in that it is often an ill-defined problem. Such ill-posedness
Received in revised form is problematic since noise becomes very influential and results in inaccurate or non-unique
22 January 2014
solutions. To combat this ill-posedness, additional constraints are typically applied to redefine
Accepted 19 February 2014
the problem, leading to a well-defined problem with a unique solution. Current input
Handling Editor: M.P. Cartmell
reconstruction methods span the spectrum of analysis and computation, and we have
grouped them into three categories: Direct, Regularization, and Probabilistic/Statistical. Each
of these groups is divided into several subsets that offer different perspectives in which to
view the reconstruction problem. Our primary interests lie in the behavior of mechanical
systems and, as such, we have focused on the literature in these fields. However, applicability
includes other fields with the same and similar governing equations.
& 2014 Elsevier Ltd. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2. Direct methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1. Mathematical analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1.1. Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.2. Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3. Regularization methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1. Tikhonov regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2. Frequency range truncation methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2.1. Mode selection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.2.2. Windowing function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.2.3. Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.3. Optimization methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.3.1. Optimal filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.3.2. Conjugate gradient method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.3.3. Levenberg–Marquardt iterative regularization method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3.4. Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4. Weighted basis functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4.1. Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

n
Corresponding author.
E-mail addresses: josanche@scarletmail.rutgers.edu (J. Sanchez), benaroya@rci.rutgers.edu (H. Benaroya).

http://dx.doi.org/10.1016/j.jsv.2014.02.025
0022-460X & 2014 Elsevier Ltd. All rights reserved.

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
2 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

4. Probabilistic/Statistical methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.1. Random excitation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.1.1. Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2. Recurrence plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2.1. Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.3. Bayesian methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.3.1. Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.4. Adaptive estimation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.4.1. Sequential realization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.4.2. Kalman Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.4.3. Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

1. Introduction

In engineering, a mathematical model of a system is developed to analyze its behavior. Often times, a system model is
formulated using an approach involving the forces or the energy of the system and produces a model of the form

! !
Lð x ðtÞÞ ¼ F ðtÞ;
! !
where L is an operation on a vector x and F ðtÞ is the force acting on the system. The model can be linear or nonlinear.
Problems such as these, where the forces are known but the system behavior is unknown, are typically referred to as the
forward problem.
!
Such models are developed with the understanding that input force, F ðtÞ, is known and we seek to establish the
!
behavior, x ðtÞ, of the system. Sometimes this may not be the case, where we can measure the response of the system but
are unable to measure the forces acting on the system. This presents an interesting problem where we know the system
behavior but seek to characterize the forces creating that behavior. This is the opposite of the forward problem and is thus
referred to as the inverse problem. Unfortunately, the inverse problem has received little attention in relation to the forward
problem and as a result, the number of appropriate references is quite small. The discipline is relatively young and much of
the focus has been placed on system identification rather than force reconstruction. As a result, our group of references is
essentially complete.
The inverse problem is very different from the forward problem. The analysis for the forward problem is straightforward,
typically involving the solution of a system of differential equations. These are methods with which we are quite familiar
and are capable of implementing with ease. But the inverse problem can be troublesome. The influence of noise is
significantly more apparent on the inverse problem than the forward problem. Thus, the solution of the inverse problem is
more susceptible to indeterminancies such as singularities or non-uniqueness of solution and this requires a higher level of
analytical rigor than the forward problem. Several classes of methods have been developed and fall into one of the following
categories:

 Direct methods.
 Regularized methods.
 Probabilistic/Statistical methods.

Each of these methods has its strengths and drawbacks and each will be further discussed in the following sections.

2. Direct methods

This section summarizes the methods that make direct use of the physical or mathematical model to formulate the
inverse problem and does not include any manipulation due to a filter or any additional constraints on the given
mathematical model. These methods are carried out in the time domain or the frequency domain.

2.1. Mathematical analysis

Mathematical models are often developed to model physical structures and thus the solutions of the models provide the
physical description. Due to nonlinearities inherent in the behavior of a structure, the mathematical analysis is often
restricted to the linear regime of the structural behavior to provide an estimate of the action taking place. The use of these

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]] 3

linear models has provided a general solution of the form


1
uðx; tÞ ¼ ∑ Φn ðxÞΓ n ðtÞ; (1)
n¼1

where uðx; tÞ is typically the displacement. The use of Eq. (1) is often referred to as modal superposition analysis.

2.1.1. Methods
This is shown in the work performed by Jacquelin and Hamelin [1]. They sought to recover the force from the recordings
of three strain measurements applied to the split Hopkinson pressure bar. Using the solution of d'Alembert for the one-
dimensional wave equation and concepts for strain from elasticity theory, the force is recovered by first determining the
strain at the end of the bar, which is given by
ϵ0 ðt Þ ¼ 12 fϵA ðt  T A Þ þ ϵA ðt þ T A Þ
þϵB ðt  T B Þ þ ϵB ðt þT B Þ
 ϵC ðt þ T B  T A Þ  ϵC ðt  T B þ T A Þg; (2)
and the force would, then, be given by
F 0 ðtÞ ¼ ESϵ0 ðtÞ; (3)
where ϵA , ϵB , and ϵC are the strains measured at locations A, B, and C along the bar respectively and ϵ0 is the strain at the end
of the bar. E is Young's modulus and S is the cross sectional area of the bar. The sensors are located at points xA, xB, and xC
along the bar, and
xA
TA ¼
c0
xB
TB ¼
c0
xC
TC ¼ ¼ T A þ T B;
c0
where c0 is the speed of the elastic longitudinal wave and TA, TB, and TC represent the time it takes for the longitudinal wave
to reach locations A, B, and C respectively.
Law et al. [2] apply the modal superposition model to a Bernoulli–Euler beam with a moving force applied, given by
∂2 vðx; tÞ ∂vðx; tÞ ∂4 vðx; tÞ
ρ þC þ EI ¼ δðx ct Þf ðt Þ; (4)
∂t 2 ∂t ∂x4
where vðx; tÞ is the beam deflection at point x at time t, ρ is the beam density, C is the viscous damping, E is Young's modulus,
I is the second moment of inertia, f(t) is the force, l is the length of the beam, c is the speed of the force motion, and δðtÞ is the
Dirac delta function. The force is reconstructed using equations derived from knowledge of either the bending moments or
accelerations, or both. Making use of the orthogonality properties in the modal superposition model, and discretizing the
problem, leads to a solution of the form
Bf ¼ m; (5)
where f is the force vector and m is the moment vector. B is an ðN 1Þ  ðN B 1Þ matrix, f is an ðNB  1Þ  1 vector, and m is
an ðN  1Þ  1 vector. NB is the number of segments into which the system is discretized and equal to l=ðcΔtÞ. If N ¼ NB , B is a
lower triangular matrix and f can be found by directly solving the system of equations. Otherwise, the use of the least
squares method is required. Force identification from acceleration measurements and combined measurements produces an
equation of the same form as Eq. (5). This procedure can, also, be performed in the frequency domain using a Fourier
Transform, as shown by Chan et al. [3].
Karlsson [4] sought to reconstruct the force in a more general way. The assumed solution was of the form
uðx; tÞ ¼ Re½uðxÞeiωt  and the force taking the form Fðx; tÞ ¼ Re½FðxÞeiωt . The relation between the displacement and the force
can be rewritten as L½uðÞðxÞ ¼ FðxÞ, where L½ is a linear differential operator. The force is assumed to be of the form
FðxÞ ¼ ∑N j ¼ 1 F j f j ðxÞ. The shape functions, fj(x), are known and the objective is to find the amplitudes, Fj. Measurements are
taken at M different points and the measurement quantity (i.e., acceleration, strain, etc.) is related to the displacement by
m
r i ¼ R½uðÞðxmi Þ, where ri is the measured response at point xi . Applying Green's function, G, and the measurement response
relation, R, a transfer function matrix is obtained. This produces a solution of the form rp ¼ HF, where rp is the predicted
response vector, H is the M  N transfer function matrix, and F is the force amplitude vector. The force amplitude vector is
found by using the least squares method. If rankðHÞ ¼ N, then the solution can be solved simply by F i ¼ H þ r m , where H þ is
the pseudoinverse of H. If rankðHÞ oN, then there does not exist a single unique solution and the problem is ill-posed.
Avitabile et al. [5] went about reconstructing the force in the frequency domain of a linear system. This would produce a
linear algebraic system of the form Sy ¼ HSf , where Sy is the Fourier spectrum of the response, H is the transfer function, and
Sf is the Fourier spectrum of the input force. Using some simple matrix analysis, we are able to produce an equation of the
form Gff ¼ H  1 Gyy H  h , where Gff ¼ Sf Snf , Gyy ¼ Sy Sny . The superscripts,  1 and h, correspond to the matrix inverse and
hermetian, respectively. Using this equation, the applied force is reconstructed from the measurement data.

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
4 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

Vyas and Winks [6] used modal analysis to reconstruct the forces acting on turbine blades. The blade displacements are
given by
m
xðz; tÞ ¼ ∑ f 1i ðZÞq1i ðtÞ (6a)
i¼1

m
yðz; tÞ ¼ ∑ f 2i ðZÞq2i ðtÞ (6b)
i¼1

m
θðz; tÞ ¼ ∑ f 3i ðZÞq3i ðtÞ (6c)
i¼1

where Z¼ z/l and the functions, f 1i ðZÞ, f 2i ðZÞ, f 3i ðZÞ, are the shape functions. The displacements, x, y, θ, are taken at m points
along the blade length. Defining the forcing functions as F x ðZ; tÞ, F y ðZ; tÞ, F θ ðZ; tÞ, and making use of the orthogonality
properties of the shape functions, a system of linear differential equations is produced. Transformation of the forces and
displacements into generalized coordinates, then discretizing the system with respect to time, produces the relation, η ¼ NS.
In this relation, η and S correspond to the generalized coordinates and generalized forces, respectively. The force is recovered
by performing the inverse of the generalizing process.

2.1.2. Results and discussion


The methods presented here produce varying levels of accuracy, which suggests a profound sensitivity to noise in the
system. This ill-posedness problem is an important topic of discussion and methods developed to remedy the noise
situation are discussed in detail in subsequent sections of this paper. The sensitivity of a system to noise is typically observed
by analyzing the singular values of the transfer function. This is achieved by observing the nature of the singular values. If
the singular values gradually decay to zero with no particular gap, then the problem is ill-posed. If they gradually decay to
zero with a well-defined gap between two singular values, then the problem is not only ill-posed, but it is additionally rank
deficient [7].

3. Regularization methods

To overcome the ill-posedness of a problem, a method of regularization must be utilized. This involves the addition of
conditions to produce a well-posed problem and these conditions can be physical or mathematical.

3.1. Tikhonov regularization

Tikhonov regularization is a commonly used method. Jacquelin et al. [7] presented its use in deconvolution problems.
This form of regularization involves defining a norm that provides a smoothing condition on the least squares problem. For a
general deconvolution problem, the least squares approach with Tikhonov regularization is of the form
min J ½GfFg ½S J 2 subjected to min ΩðfFgÞ; (7)
F F

where ½G is a transfer matrix generated by sampling the convolution integral to discretize the problem, fFg is the load
vector, ½S is the measured response, and ‖‖2 stands for the Euclidean norm. The function Ωð½FÞ is typically of the form
Ωð½FÞ ¼ J ½LfFg J 2 ; (8)
where ½L is a parameter to be chosen. It has been shown that Eq. (7) is equivalent to the following formulation:
min½ J ½GfFg  ½S J 2 þαΩðfFgÞ and minΩðfFgÞ; (9)
F F

where α is the regularization parameter. It is calculated by plotting J ½LfFg J 2 and J ½GfFg ½S J on a graph using log–log
scaling. This produces an L-shaped curve. This curve represents the relationship between the goodness of fit and the
smoothness of the solution. On one side of the L curve, the fit is good but under-smoothed. On the other side of the L curve,
the fit is over-smoothed at the expense of the quality of the fit. Thus, the optimal value for α is determined by choosing the
value that produces the solution closest to the corner of the L-curve.
Problems involving Tikhonov regularization are presented in the next section.

3.2. Frequency range truncation methods

As stated before, the singular values are closely related to the ill-conditioning of a problem. If the singular values
gradually decay to zero then the problem is ill-posed, and additionally, the existence of gap between singular values implies
rank-deficiency. One way to recondition the problem would be to truncate the problematic singular values, thus eliminating
the ill-posedness of the problem. The same concept is applied in other ways as well to combat ill-conditioning.

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]] 5

3.2.1. Mode selection


Jiang and Hu [8] reconstructed the dynamic load acting on an undamped Euler beam. In dimensionless form, the Euler
beam is governed by

ðκ 1 LÞ4 ∂2 wðx; t Þ ∂4 wðx; t Þ 


þ ¼ f x; t Þ; (10)
4π 2 ∂t
2
∂x 4

where x ¼ x=L is the standardized


pffiffiffiffiffiffiffiffiffiffiffiffi spatial variable scaled by the length L, t ¼ t=T 1 is the standardized time scaled by
T 1 ¼ 2π=ω1 where ω1 ¼ κ 21 EI=ρA, w ¼ w=L is the standardized transverse displacement, f ¼ L3 f =ðEIÞ is the standardized
force, κ1 is the wavenumber of the first modal shape, ρ is the material density, E is Young's modulus of the material, A is the
cross sectional area, and I is the second moment of the area. Using modal analysis (Eq. (1)) and assuming the standardized
form of the force can be represented as f ðx; t Þ ¼ CW j ðxÞeiωt , the displacements and accelerations can be represented as
  4π 2 1
qm t ¼ c δjm 2 eiωt (11a)
ðκ 1 LÞ4 ωm  ω2

  4π 2 1
q€ m t ¼ c δjm eiωt : (11b)
ðκ1 LÞ4 1 ω 2m =ω 2

From this, the following scale factors are defined as


1
SF w ¼ (12a)
ω 2m ω 2

1
SF a ¼ (12b)
1  ω 2m =ω 2 ;

where w is the standardized displacement and a is the standardized acceleration. The objective in defining the scale factors
is to reconstruct the distributed force over a specified frequency range and spatial modes. The frequencies are obtained from
the response data, and from that, the scale factors are calculated. Only those modes whose scale factors are larger than a
certain threshold will be used to reconstruct the load.
Jiang and Hu [9], also, applied this method to a thin plate. In standardized form, the equation of motion is given by

π 2 ð1 þλ2 Þ2 ∂2 Zðξ; η; τÞ
þ Λ½Z ðξ; η; τÞ ¼ pðξ; η; τÞ
4 ∂τ2
∂4 ∂4 ∂4
Λ  4 þ2λ2 2 þλ4 4 ; (13)
∂ξ ∂ξ ∂η2 ∂η

where ξ  x=a and η  y=b are the standardized spatial variables, τ ¼ t=T is the standardized time with T 
pffiffiffiffiffiffiffiffiffiffiffi
2a2 ρh=D=ðπð1 þλ2 ÞÞ, Z  w=a is the standardized transverse displacement, p  a3 f =D is the standardized distributed force,
3
λ  a=b is the aspect ratio, D ¼ Eh =12ð1  ν2 Þ, ρ is the density, E is Young's modulus, ν is Poisson's ratio, and h is the plate
thickness. Using modal analysis in the standardized form, the scaling factor produced for the thin plate is
1
SF  : (14)
Ω2k  Ωþ i2ζ k Ωk Ω

3.2.2. Windowing function


Djamaa et al. [10] used a frequency range truncation method to reconstruct the force applied to a thin cylindrical shell.
The governing equation is given by
      
Eh ν ∂u 1 ∂v w
   2
1  ν2 a ∂z a2 ∂θ a
2  !
h ∂4 w 2 ∂4 w 1 ∂4 w
þ þ 2 þ þρhω2 w ¼  F w : (15)
12 ∂z4 a ∂z2 ∂θ2 a2 ∂θ4

Using the finite difference method, the force is reconstructed numerically. Each term from Eq. (15) can be approximated as
follows:

∂4 w 1  
2δ4ij ðzÞ ¼ 4 wi þ 2;j  4wi þ 1;j þ 6wi;j 4wi  1;j þ wi  2;j (16a)
∂z4 Δz

∂4 w 1  
4
2δ4ij ðθÞ ¼ 4 wi;j þ 2  4wi;j þ 1 þ6wi;j 4wi;j  1 þ wi;j  2 (16b)
∂θ Δθ

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
6 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

∂4 w 1 
2δ4ij ðz; θÞ ¼ 2 2 wi þ 1;j þ 1 2wi þ 1;j þ wi þ 1;j  1 2wi;j þ 1
∂z2 ∂θ2 Δz Δθ
þ 4wi;j  2wi;j  1 þ wi  1;j þ 1  2wi  1;j þ wi  1;j  1 Þ (16c)

∂u 1  
2 ui þ 1;j ui  1;j (16d)
∂z 2ΔZ

∂v 1  
2 v  vi;j  1 : (16e)
∂θ 2Δθ i;j þ 1
By substitution into the governing equation, these finite difference relations give us the following force relation:

Eð1 þ jηÞh ν   1  
F ij ¼ u  ui  1;j þ 2 vi;j þ 1  vi;j  1 :
1  ν2 2aΔz i þ 1;j 2a Δθ
2  !
h 4 2 4 1 4
þwij þ δ ðzÞ þ 2 δij ðz; θÞ þ 4 δij ðθÞ ρhω2 wij : (17)
12 ij a a

To combat noise, a technique that filters allowable frequencies in the frequency domain is utilized. The function in the
frequency domain is defined as
b ¼ 1;
hðkÞ k A ½ kc ; kc 

b ¼0
hðkÞ otherwise: (18)

The regularized force would, then, be defined as


Fðz; θÞ ¼ Fðz; θÞ hðz; θÞ: (19)
|fflfflffl{zfflfflffl} |fflfflffl{zfflfflffl}
Filtered Original

In the spatial domain, the filter is given by


    
kcz z kcθ aθ sin ðkcz zÞ sin ðkcθ aθÞ
hðz; θÞ ¼ 1 þ cos 1 þ cos ; (20)
2f z 2f θ 4π 2 zaθ

where kcz ; kcθ are the cutoff frequencies in the axial and circumferential directions and f z ; f θ are the “Form Factors of the
Filter” along the axial and circumferential directions, respectively. The “Form Factors of the Filter” are the ratio of the half-
length of the Hanning window and the cutoff wavelength.

3.2.3. Results and discussion


The results for each of the methods were promising. The windowing function was able to clarify a noisy force distribution
field and provided a clear picture of the location of the impact acting on the structure.
The mode selection technique experiments were significantly more involved. For the Euler beam, several cases
showcased the potential of the method. And while it showed promise, the variation of the mode selection criteria produced
varied results. Thus, a major concern with the use of these methods is the ability to determine a suitable point of truncation.
This could potentially be remedied by a thorough analysis of the noise present in the system. By analyzing the noise, an
expression for an optimal point of truncation could potentially be derived, providing the necessary closure to the above
analysis.

3.3. Optimization methods

In addition to truncation methods, techniques that seek to minimize error estimates are used to obtain the best estimate
for the system input. These methods employ improvements to the least-squares method to provide a single optimal solution
for a given problem.

3.3.1. Optimal filtering


One such method involves the inclusion of terms that are meant to act as filters for noise inherent in the system. This is
exhibited by Huang et al. [11] in their reconstruction of the traction forces acting on a cell. The displacement model was
based on the linear elastic theory of the Boussinesq solution. This solution is achieved through convolution integration in
the frequency domain:
! !
u ðkx ; ky Þ ¼ ½Gðkx ; ky Þ T ðkx ; ky Þ; (21)

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]] 7

where u is the displacement, ½G is the matrix Green's function, T is the traction vector, and ðkx ; ky Þ are the wavenumbers.
Green's function from the Boussinesq solution is given by
2 3
2 2
  2π 4 ð1  sÞk þsky  skx ky
5;
G kx ; ky ¼ A 3 2 2
(22)
k  skx ky ð1  sÞk þ skx
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2
where k ¼ kx þky and A ¼ ð1 þsÞ=πE. The objective is to find the inverse of this problem, such that
! !
T ðkx ; ky Þ ¼ ½Gðkx ; ky Þ  1 u ðkx ; ky Þ; (23)

where
" #
R11 ðkÞ R12 ðkÞ
½Gðkx ; ky Þ  1 ¼ ; (24)
R21 ðkÞ R22 ðkÞ
!
and traction vector, T ðkx ; ky Þ, can be written in component form as
( )
! T x ðkx ; ky Þ
T ðkx ; ky Þ ¼ : (25)
T y ðkx ; ky Þ

In the absence of noise, Eq. (23) can be expanded as


T x ðkx ; ky Þ ¼ R11 ðkx ; ky Þux ðkx ; ky Þ þR12 ðkx ; ky Þuy ðkx ; ky Þ (26a)

T y ðkx ; ky Þ ¼ R21 ðkx ; ky Þux ðkx ; ky Þ þ R22 ðkx ; ky Þuy ðkx ; ky Þ: (26b)

In the presence of noise, the deformation field can be approximated by u b ðkÞ ¼ uðkÞ þ nðkÞ, where the noise, n(k), is
assumed to be zero mean and with a standard deviation, s. The deconvolution problem is an ill-posed problem and requires
regularization to combat noise. This is achieved by the use of four filtering parameters, Φxx ðkx ; ky Þ, Φxy ðkx ; ky Þ, Φyx ðkx ; ky Þ, and
Φyy ðkx ; ky Þ. The traction force would then be estimated as

Tb x ðkx ; ky Þ ¼ R11 ðkx ; ky Þu


b x ðkx ; ky ÞΦxx ðkx ; ky Þ
b y ðkx ; ky ÞΦxy ðkx ; ky Þ
þR12 ðkx ; ky Þu (27a)

Tb y ðkx ; ky Þ ¼ R21 ðkx ; ky Þu


b x ðkx ; ky ÞΦyx ðkx ; ky Þ
b y ðkx ; ky ÞΦyy ðkx ; ky Þ;
þR22 ðkx ; ky Þu (27b)

where Tb x and Tb y are the estimated traction forces. To ensure the accuracy of the result, the following relations must be
minimized:
Z 1 Z 1 
min jTb x ðkx ; ky Þ T kx ;ky ðkÞj2 dkx dky (28a)
1 1

Z 1 Z 1 
min jTb y ðkx ; ky Þ  T y ðkx ; ky Þj2 dkx dky : (28b)
1 1

The minimization of Eqs. (28) produces


8
>
> G6 G2 G5 G4
>
> ; kx ; ky a0
>
> G3 G2 G1 G4
  <
Φxx kx ; ky ¼ 0; kx ; ky ¼ 0
>
>
>
> jux j2
>
> otherwise
: ju j2 þ jn j2
x x

8
G G  G1 G6
  < 5 3 ; kx ; ky a 0
Φxy kx ; ky ¼ G3 G2  G1 G4
:
0 otherwise
8
> Q 5Q 3  Q 1Q 6
>
> ; kx ; ky a 0
>
> Q 3Q 2  Q 1Q 4
>
  < 0; kx ; ky ¼ 0
Φyy kx ; ky ¼
>
>
>
> juy j2
>
> otherwise
: juy j2 þ jny j2

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
8 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

8
  < Q 6Q 2  Q 5Q 4 ;
>
kx ; ky a0
Φyx kx ; ky ¼ Q 3Q 2  Q 1Q 4 (29)
>
:0 otherwise;

where
G1 ¼ 2R211 ðjux j2 þjnx j2 Þ (30a)

G2 ¼ G3 ¼ R11 R12 ðuny ux þuy unx Þ (30b)

G4 ¼ 2R212 ðjuy j2 þjny j2 Þ (30c)

G5 ¼ 2R211 jux j2 þ R11 R12 ðuny ux þuy unx Þ (30d)

G6 ¼ 2R212 juy j2 þR11 R12 ðunx uy þ ux uny Þ (30e)

Q 1 ¼ 2R221 ðjux j2 þjnx j2 Þ (30f)

Q 2 ¼ Q 3 ¼ R21 R22 ðuny ux þuy unx Þ (30g)

Q 4 ¼ 2R222 ðjuy j2 þjny j2 Þ (30h)

Q 5 ¼ 2R221 jux j2 þ R21 R22 ðuny ux þuy unx Þ (30i)

Q 6 ¼ 2R222 juy j2 þR21 R22 ðunx uy þ ux uny Þ: (30j)

To perform this analysis, the noise components need to be estimated and substituted into the filtering parameters. Due to
the direct relationship between noise levels and the components of the matrix, ½Gðkx ; ky Þ  1 , the amplification of noise
typical in the inversion process is repressed by the filtering parameters.

3.3.2. Conjugate gradient method


The conjugate gradient method is also known as the iterative regularization method, as the regularization is applied
during iteration. It involves the solution of three problems: the direct problem, the sensitivity problem, and the adjoint
problem. The procedure is discussed by Huang [12], who applied the conjugate gradient method to a spring–mass–damper
system with displacement-dependent spring and damper:
_ ¼ yðtÞ;
xðtÞ t 4 0; xð0Þ ¼ x0 (31a)

CðxÞ KðxÞ f ðtÞ


y_ ðt Þ ¼  y xþ ; yð0Þ ¼ y0 : (31b)
M M M
The solution to the inverse problem would be determined by the minimization of the functional:
Z tf
J½f ðtÞ ¼ ðxðt; f ðtÞÞ XðtÞÞ2 dt; (32)
0

where x(t) is the estimated displacement and X(t) is the measured displacement.
The estimation of the force, bf ðtÞ, is given by the recursive formulas:
nþ1 n
b
f ðtÞ ¼ b
f ðtÞ βn P n ðtÞ (33a)

P n ðtÞ ¼ J 0n ðtÞ þ γ n P n  1 ðtÞ (33b)


R t ¼ tf
ð J 0n ðtÞÞ2 dt
γ n ¼ R t t¼¼tf0 0n  1 ; γ 0 ¼ 0; (33c)
t¼0 ðJ ðtÞÞ2 dt
where βn is the search step size at iteration n, P n ðtÞ are the search directions, γn is the conjugate coefficient, and J 0n is the
gradient direction.
To calculate βn, a perturbation problem must be presented and solved. When f(t) undergoes a variation, Δf ðtÞ, we assume
that x(t), y(t), C(x), and K(x) are perturbed by ΔxðtÞ, ΔyðtÞ, ΔCðxÞ and ΔKðxÞ, respectively. Substitution into and subtraction of
the original direct problem produce a differential equation in terms of the perturbed variables, which is to be solved.
Substitution of Eq. (33a) into Eq. (32) and minimizing the resulting equation with respect to βn yield
R t ¼ tf n
Δxðt; P n Þ½xðt; b
f Þ  XðtÞ dt
βn ¼ t ¼ 0 R t ¼ t f n
: (34)
t ¼ 0 ½Δx ðt; P Þ dt
2

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]] 9

The term, J 0n ðtÞ, is calculated by solving Eq. (32), using Lagrange multipliers, λ1 ; λ2 , and perturbing Eq. (32). This produces
differential equations for λ1 ; λ2 :
dλ1 KðxÞ ðdK=dxÞx ðdC=dxÞy
 ¼ þ þ λ2 ðt Þ  2ðx X Þ
dt M M M

t 4 0; λ1 ðt f Þ ¼ 0 (35a)

dλ2 CðxÞ
 ¼ λ2 ðt Þ þ λ1 ðt Þ
dt M

t 4 0; λ2 ðt f Þ ¼ 0: (35b)
R t ¼ tf 0 0
Using the definition, ΔJ ¼ t¼0 ðJ Δf Þ dt, the expression for J becomes
 λ2 ðtÞ
J 0 f ðt Þ ¼ : (36)
M
The criterion for the termination of the algorithm, given by ε, is calculated by
ε ¼ s2 t f ; (37)

where s is the statistical variance of the displacement measurements.


Huang [13], also, successfully applied this procedure to a multi-degree of freedom system with displacement-dependent
parameters. The system of equations is given by
2
d x1 ðtÞ ½C 1 ðx1 Þ þC 2 ðx2 Þ dx1 ðtÞ C 2 ðx2 Þ dx2 ðtÞ
¼ þ
dt 2 M1 dt M2 dt
½K 1 ðx1 Þ þK 2 ðx2 Þ K 2 ðx2 Þ f 1 ðtÞ
 x1 ðt Þ þ x 2 ðt Þ þ ; t 40 (38a)
M1 M2 M1

2
d xi ðtÞ C ðx Þ dxi  1 ðtÞ ½C i ðxi Þ þC i þ 1 ðxi þ 1 Þ dxi ðtÞ
¼ i i 
dt 2 Mi dt Mi dt
C i þ 1 ðxi þ 1 Þ dxi þ 1 ðtÞ K i ðxi Þ
þ þ x ðt Þ
Mi dt Mi i  1
½K ðx Þ þ K i þ 1 ðxi þ 1 Þ
 i i x i ðt Þ
Mi
K ðx Þ f ðtÞ
þ i þ 1 i þ 1 xi þ 1 ðt Þ þ i ; t 40; i ¼ 2 to I  1 (38b)
Mi Mi

2
d xI ðtÞ C I ðxI Þ dxI  1 ðtÞ C I ðxI Þ dxI ðtÞ
¼ 
dt 2 MI dt MI dt
K I ðxI Þ K I ðxI Þ f I ðtÞ
xI  1 ðt Þ  xI ðt Þ þ ; t 40 (38c)
MI MI M1
subject to the initial conditions:
xi ð0Þ ¼ xi;0 (39a)

dxi ð0Þ
¼ yi ð0Þ ¼ yi;0 ; i ¼ 1 to I: (39b)
dt
The basic procedure remains unchanged, but due to the increase in the number of degrees of freedom, the level of
computation is increased.
Huang et al. [14] then applied this procedure to external forces acting on cutting tools. The tool was modeled as an
Euler–Bernoulli beam subjected to a transverse load, f ðx; tÞ, applied to a beam of length L,
∂4 yðx; tÞ ∂2 yðx; tÞ
EI þρA ¼ f ðx; t Þuðx  xc Þ; (40)
∂x4 ∂t 2
where x is a point along the length of the beam from 0 to L, uðx xc Þ is the unit step function which restricts the transverse
load to the region on the beam from x ¼ xc to x ¼L.
Yen and Wu [15,18] used the conjugate gradient method to reconstruct the impact acting on rectangular plates. The plate
was modeled using the Reissner–Mindlin plate theory and the Rayleigh–Ritz method. The displacements of the plate are
assumed to have the form of Eq. (1). Using variational techniques, the equations of motion are recast as
" #" # " #" # " #
0 0 q€1 K11 K12 q1 0
þ ¼ ; (41)
0 M22 q€2 K21 K22 q2 P2

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
10 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

where q1 is the vector containing the in-plane modal amplitude functions, q2 is the vector of transverse modal amplitude
functions, M22 is the mass matrix, Kij are the stiffness matrices. P2 is found by multiplying and integrating the forcing
function, f ðx; tÞuðx  xc Þ, by the mode shapes in the x and y directions to produce the time-dependent terms of P2. This is
solved using eigenmode expansions and Green's functions. Using N points of strain measurements on the plate, the
following relation is obtained:
RðkÞ ¼ GðkÞ P; k ¼ 1; 2; …; N; (42)

where RðkÞ is the vector of n strain measurements for the kth sensor, GðkÞ is the matrix of Green's functions, and P is the force
to be determined. Using Eq. (42), we are able to find the location of impact and reconstruct the force as follows. To find the
location of the impact force on the plate, using the fact that the convolution integral is commutative, the relationship
GðiÞ RðjÞ ¼ GðjÞ RðiÞ is deduced. To find the location and the force profile, two error functions are defined
N N
Fðx; yÞ ¼ ∑ ∑ f ij ‖GðiÞ RðjÞ  GðjÞ R ðiÞ ‖2 (43a)
i ¼ 1 j ¼ 1;j a i

N
E ¼ ∑ ‖GðjÞP RðjÞ‖2 ; (43b)
j¼1

where J J is the Euclidean norm, fij is the scaling factor given by f ij ¼ 1=J GðiÞ RðjÞ J , and E is the error. Using the conjugate
gradient method to minimize these functionals, the force and its location were identified.

3.3.3. Levenberg–Marquardt iterative regularization method


This method follows from the Gauss–Newton method, which is the application of Newton's method to the least-squares
problem. This procedure iteratively calculates the parameters that minimize the sum of the squares of the error. The solution
is of the form xk þ 1 ¼ xk þ sðxk Þ where xk and xk þ 1 are the solutions of the given problem at iteration k and k þ 1, respectively,
and sðxk Þ is the step function. The objective of this method is to define a step function, sðxk Þ, to ensure the convergence of xk
to its optimal solution.
This method can be applied to force reconstruction problems, as exhibited by Gunawan [16]. He sought to reconstruct
pulse-type impact forces with the system modeled using convolution integration. Discretizing the convolution integral
produces an equation of the form, Ax¼b. To solve for the reconstructed force, the residue function must be defined as
eðxÞ ¼ Ax  b. The least-squares function to be minimized is then defined as f ðxÞ ¼ 12 eðxÞT eðxÞ. To calculate the step at iteration
k, given by sðxk Þ, the gradient of the least-squares function is required

gðxÞ ¼ ∇f ðxÞ ¼ AT eðxÞ; (44)


where g(x) is the gradient of f(x). The step function is, thus, given by

sðxÞ ¼  ðAT A þ νIÞ  1 gðxÞ; (45)


such that xk þ 1 ¼ xk þsðxk Þ for iteration k and where ν is the Levenberg–Marquardt regularization parameter.
An approach is used to determine an appropriate value for ν. In this method, the parameter, ν, is adjusted based on its
performance. It is measured by defining a variable, r, as the ratio of an estimate of the reduction of the objective function,
Δf theo , to the actual reduction, Δf act . This is given by
Δf act
r¼ ; (46)
Δf theo
where
Δf theo ¼  gðxk ÞT ðxk  x þ Þ; (47)
and
Δf act ¼ f ðxk Þ  f ðx þ Þ; (48)
where x þ here is first determined by using ν ¼ 0 in the iteration relation, x þ ¼ xk þ sðxk Þ. The parameter, ν, is then adjusted
by a specified value, either ωup or ωdown, depending on its relation to the parameters: μ0, μlow, and μhigh. If r o μ0 or μhigh o r,
ν is increased by ωup or decreased by ωdown, respectively, and x þ is recomputed by

Ac ¼ AT A þ νI (49)

sðxk Þ ¼  Ac 1 gðxk Þ (50)

x þ ¼ xk þsðxk Þ: (51)
This procedure is repeated until μlow o r oμhigh . When μlow o r oμhigh is true, x þ is accepted as the next solution, xk þ 1 .
Gasparo et al. [17] considered nonlinear inverse problems of the form FðxÞ ¼ y when employing this technique.
The method derived is analogous to Runga–Kutta methods for the solution of ordinary differential equations and is aimed at

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]] 11

regularizing ill-posed problems. The data, yδ , are assumed to have a noise level δ Z0 such that J yδ  y J r δ. The Levenberg–
Marquardt method was applied with two fixed parameters, α1 4 0 and α2 40. The method is as follows:
δ
b ¼ F 0 ðxδ Þn ðyδ  Fðxδ ÞÞ
ðF 0 ðxδn Þn F 0 ðxδn Þ þ α1 IÞh (52a)
n n n

δ
b
zδn ¼ xδn þ 12 h (52b)
n

δ
ðF 0 ðzδn Þn F 0 ðzδn Þ þ α2 IÞhn ¼ F 0 ðzδn Þn ðyδ Fðxδn ÞÞ (52c)

δ
xδn þ 1 ¼ xδn þ hn : (52d)
The superscript, δ, indicates the dependency of the iterations on δ.

3.3.4. Results and discussion


Though these procedures seek various methods of regularization on an optimization problem, the results for each were
very promising. All these methods produced very acceptable results and were fairly insensitive to noise present in the
system, but still certain concerns remain. The techniques that use the Levenburg–Marquart method do not always have a
defined method for selecting the regularization parameters. Gunawan [16] used a trust region approach to determine the
parameter, but Gasparao et al. [17] did not have a method for parameter determination. There does not appear to be a
specific method for regularization parameter selection, and this could potentially be problematic in its present state.
Perhaps the introduction of an additional analysis on the regularization parameters analogous to Lagrange multipliers could
provide the necessary information to produce a more fully defined regularization technique.

3.4. Weighted basis functions

Liu and Shepard [19] presented an alternative to modal analysis for spatially varying dynamic forces. The force and the
response are constructed using a set of tailored basis functions. The force and the response would, then, be given by
N
FðxÞ ¼ ∑ W i χ i ðxÞ (53a)
i¼1

N
νðxÞ ¼ ∑ W i ψ i ðxÞ; (53b)
i¼1

where χ i ðxÞ are the basis functions that span the forcing space, Wi is the set of weighting coefficients, and ψ i ðxÞ are the basis
functions of the response. This analysis is based on the idea that the force described by χi of unit amplitude will produce the
response, ψi, of unit amplitude. Using the measured response, νðxÞ, the weighting coefficients can be determined by
GW ¼ B; (54)
where G, known as the Gram matrix, and B consist of the terms:
Z
Gij ¼ ψ i ðxÞψ j ðxÞ dΩ (55a)
Ω
Z
Bij ¼ ψ i ðxÞνj ðxÞ dΩ: (55b)
Ω

Once G and B are known, W is determined by W ¼ G  1 B. Using the terms in W, the force is reconstructed by using Eq. (53a).
It should be noted that the basis functions, χi, are not required to be modal functions, but for a solution to be found, each ψi
corresponding to χi must be linearly independent of all other ψj corresponding to χj.

3.4.1. Results and discussion


When applied to simple problems without noise, the proposed method provides solutions more akin to the actual force
input than solutions computed with modal basis functions. Additionally, the proposed method required fewer basis
functions, thus providing a more efficient algorithm than modal functions, and appeared to refrain from succumbing to the
Gibbs phenomenon for discontinuous distributions. While this method alone may be useful for simple force reconstruction
problems, the potential for ill-conditioning remains present. To determine the posedness of the problem, the singular values
of the Gram matrix, G, must be analyzed. If the singular values of G gradually decay to zero and the ratio between the largest
and smallest singular values is small, then the problem is ill-posed. To overcome this, the authors propose the use of
Tikhonov regularization, and the problem is now given by
min J GW α  B J 22 þ α2 J W α J 22 ; (56)
with the regularization parameter, α, determined by the L-curve method. The addition of regularization for a problem with
noise produces satisfactory results despite the noise inclusion.

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
12 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

4. Probabilistic/Statistical methods

In this section, methods employing a probabilistic or statistical perspective to gain insight into the driving force of the
system are reviewed. From this perspective, multiple techniques have been developed with significantly different
procedures and objectives.

4.1. Random excitation

Granger and Perotin [20] sought to identify a distributed random excitation, f ðx; tÞ, applied to a vibrating structure. Using
a modal model, the displacement is given by
uðx; tÞ ¼ ∑ϕi ðxÞqi ðtÞ: (57)
i

The generalized equations of motion, then, become

mi ðq€ i þ 2ζ i ωi q_ i þ ω2i qi Þ ¼ Q i ðtÞ; (58)

where ωi is the natural frequency of the ith mode, ζ i is the damping coefficient of the ith mode, and Qi is the generalized
RL
excitation, given by Q i ðtÞ ¼ 0 f ðx; tÞϕi ðxÞ dx. It is assumed that f ðx; tÞ is a stationary Gaussian process, and thus, Q i ðtÞ, qi ðtÞ,
and uðx; tÞ are stationary Gaussian. This means that they are completely defined by their cross-spectral densities, SQ i Q j ðωÞ,
Sqi qj ðωÞ, and Suu ðx1 ; x2 ; ωÞ. The spectral density of the force, Sff, is found by first identifying the modal parameters. From there,
modal coordinates, Sqi qj , and thus, the generalized excitations, SQ i Q j , can then be estimated. This would enable us to solve for
the force spectral density, Sff.
The problem concerning the identification of the modal parameters is essentially a least squares problem. As such, the
potential for ill-posedness is apparent, and thus, the addition of constraints to regularize the problem is necessary. This is
achieved using a Tikhonov regularized least squares problem and is given by

min J ½ϕqα  u J 2 þ αJ Bqα J 2 (59a)


subject to ρ2 ðαÞ  J ½ϕqα  u J 2 ¼ δ J u J 2 : (59b)

The solution of Eqs. (59) is in the frequency domain by using the expression of the norm of a process, y(t), in terms of its
truncated Fourier transform using the Weiner–Kinchin relationship:
Z 1
1 1  n 
J yJ 2 ¼ lim E y ðω; T Þyðω; T Þ dω; (60)
2π T-1  1 T

where yn is the Hermitian transpose and EðÞ denotes the mathematical expectation. By way of changes in variables and
orthogonal transformations, and using the singular value decomposition, the solution for qα is, thus, given by
bα ¼ ½TðωÞu;
q (61)
bα is the estimate for qα . The cross-spectral density matrix of the modal coordinates is, then
where q

½Sb
q αb

ðωÞ ¼ ½TðωÞ½Suu ðωÞ½TðωÞT ; (62)

where
si ðωÞ
½T ðωÞ ¼ 1=jBi ðωÞj½V ðωÞ ½Un ðωÞT : (63)
si ðωÞ þ α

BðωÞ is an n  n matrix with diagonal elements:


"  2  #
ω ω
Bj ðωÞ ¼ mj ω2j 1  þ 2iζj (64)
ωj ωj
pffiffiffiffiffiffiffiffi
for i ¼  1 and j ¼ 1; …; n. The terms, si, are the singular values of matrix ½Φ ¼ ½ϕ½1=jBi ðωÞ, and the singular value
decomposition is given by ½Φ ¼ ½UðωÞ½rðωÞ½VðωÞT . The matrix ½Un  contains the first n columns of ½U.
Additionally, the analysis performed provides the following relation for the Tikhonov regularization parameter relation,
ρðαÞ, given by
Z 1 !2
1 n α
ρ2 ðαÞ ¼ ∑ Sgi gi ðωÞdω
2π i ¼ 1  1 s2i ðωÞ þα
m
Z 1
1
þ ∑ Sg g ðωÞdω: (65)
2π i ¼ n þ 1  1 i i

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]] 13

Defining the following terms:


m
Z 1
1
e20  lim ρ2 ðαÞ ¼ ∑ Sgi gi ðωÞ dω (66a)
α-0 2π i ¼ n þ 1 1

Z 1
1 m
e21  lim ρ2 ðαÞ ¼ ∑ Sgi gi ðωÞ dω; (66b)
α-1 2π i ¼ 1 1

the solution to (b) of Eq. (59) can be given by


Z 1
1 n Sgi gi ðωÞ
∑ dω e2δ ¼ 0; (67)
2π i ¼ 1 1 ð1 þ s2i ðωÞγÞ2

where γ ¼ 1=α, e2δ ¼ maxðδðe21  e20 Þ; 0Þ. The term, Sgi gi ðωÞ, is the cross-spectral density of gi, where g ¼ ½UðωÞT u is a vector. The
solution to Eq. (67) provides the optimal Tikhonov regularization parameter for the given regularized least-squares problem.

4.1.1. Results and discussion


The solution procedure was verified by comparing the results produced by the use of a reference force acting on three
structures of variable characteristics, and the results of which were very agreeable. The comparison between experimental
results and numerical simulations produced comparable results in the linear domain with variances being mainly due to
measurement noise. When nonlinearities are introduced, the largest relative discrepancy was approximately 25 percent,
which was only for a single observed variable whereas all other variables produced results with a better level of agreement
[21].

4.2. Recurrence plots

The previous section assumed that the force excitation was a stationary process. In this section, a method is proposed
using recurrence plots, which are often useful in the visualization of a nonstationary time series.
Tanio et al. [22] utilized recurrence plots to develop a method to reconstruct the driving force acting on a dynamical
system, which is defined in discrete form as
si þ 1 ¼ f ðsi ; γ i Þ; xi ¼ hðsi Þ; (68)
where si is the state of the system, γi is a slowly varying parameter, and xi is the observation at time i.
To construct a recurrence plot, first define a vector, v, such that
vi ¼ ðxi ; xi þ τ ; …; xi þ ðm  1Þτ Þ; i ¼ 1; …; N m ; (69)

where τ is the time delay, m is the embedding dimension, and Nm ¼ N  ðm 1Þτ. Then, consider a two-dimensional plane
where both axes represent time. The distribution of the points on this plane is given by
(
1 when J vi  vj J r r
Rði; jÞ ¼ (70)
0 otherwise;

where J  J is the maximum norm and r is a ceiling criterion, which determines the distribution of points in the
recurrence plot.
It has been shown that the recurrence plot of the driving force can be approximated from the recurrence plot of the
driven system. Once the recurrence plot of the driving force has been determined, the time series can be approximated. This
is achieved by constructing a rank order, ρ, of the time series fxi gi ¼ 1;…;N such that the elements of x are organized into
nondecreasing order. The rank order, ρ, is a permutation on f1; …; Ng, and the objective is to find a permutation, denoted by
s, that would satisfy this condition. In the recurrence plot, the rank order permutes the rows so that the points in the
columns are consecutive, and it is a measure of this consecutiveness that is required to retrieve the time series.
To determine the consecutiveness of a column in the permuted recurrence plot, the mean deviation from the mean value
is defined to provide a metric of consecutiveness. For a recurrence plot, Rði; jÞ, the ith column has M points at the indices
K i ¼ fj1 ; …; jM g. When permuted by s, the ith column becomes K si ¼ fsðj1 Þ; …; sðjM Þg. The mean deviation from the mean value
is, then, defined as
  1 M   
MD K si ¼ ∑ s 2ptjm 〈sðjÞ〉j; (71)
Mm¼1

m ¼ 1 sðjm Þ denotes the mean value. Thus, the following optimization problem is presented:
where 〈sðjÞ〉 ¼ ð1=MÞ∑M
Nm
minimize ∑ MDðK si Þ subject to s A SNm ; (72)
i¼1

which represents the minimization of the mean deviation of all columns of the recurrence plot, and SNm represents the
permutation group on 1; …; N m .

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
14 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

In using Eq. (71), the solution may not be an accurate representation of the driving force. Thus, the objective function
must be modified, and subjected to some form of regularization. This is done by assuming that the driving force is
continuous. The addition of this assumption provides two techniques: the penalty method and the thick diagonal method.
In the penalty method, the optimization problem is modified:
Nm   a k Nm  j
minimize ∑ MD K si þ ∑ ∑ jsði þjÞ sðiÞj;
i¼1 kj¼1 i¼1

subject to s A SNm ; (73)


where a is a positive constant and k is a natural number. The penalty method has similarities to the Tikhonov method
discussed earlier as the addition of the second term adds a measure of continuity to the signal being reconstructed. In the
thick diagonal method, the recurrence plot is modified:
(
1; Rði; jÞ ¼ 1 or ji jj rw;
b jÞ ¼
Rði; (74)
0 otherwise;

where w is a parameter for the width of points around the diagonal line. This method is meant to act as a type of
compensator for the original recurrence plot. It seeks to add points that may be desirable but were lost in the original
recurrence plot. The optimization problem, then, becomes
Nm s
b Þ
minimize ∑ MDðK subject to s A SNm ; (75)
i
i¼1

where Kb s corresponds to the amended recurrence plot, Rði;


b jÞ, permuted by s. The approximate solution to the combinatorial
i
optimization problem is found using a search algorithm that compares the values of the objective function produced by
different permutations.

4.2.1. Results and discussion


The method using the recurrence map produced promising reconstruction of the driving force when applied to a tent
map with a driving force using both the penalty method and the thick diagonal method. But like the other methods
observed previously, there are no present methods to determine optimal values for the parameters involved in the
modification of the objective function. As the parameters vary, the reconstructed driving force in the presented solution is
found to be either correct, or diagonal. This means the solution is either obviously right or obviously wrong. This is due to
the addition of the continuity constraint. If it is applied too strongly, an inaccurate reconstruction results.

4.3. Bayesian methods

Zhang et al. [23] used Bayesian techniques to reconstruct forces for two situations: force reconstruction with uncertain
frequency response functions, and with uncertain modal parameters. The benefit of the Bayesian approach is that it provides
a thorough accounting of noise by the use of a posterior probability distribution, given by
pðxjDÞ ¼ pð½YjxÞpðxjI Þ=pðDÞ; (76)
where x is the vector of unknown parameters, ½Y is a matrix of the measured data, I is the prior information, and D is the
union of the sets of I and ½Y. The conditional probability density function, pð½YjxÞ, is the likelihood function, which
represents the probability of observing ½Y given a set of parameters, x. pðxjI Þ is the prior probability density function, and it
is representative of our knowledge of x prior to experimentation. pðDÞ is the evidence function. It is used for model class
R
selection, and is given by pðDÞ ¼ pð½YjxÞpðxjI Þdx.
For the problem of force reconstruction with uncertain frequency response functions (FRF), the measured response is
given by
YðωÞ ¼ HðωÞFðωÞ þδH ðωÞFðωÞ þ NðωÞ; (77)
where δH ðωÞ represents the model uncertainty, δH ðωÞFðωÞ is the uncertainty of the response due to the uncertain FRF, HðωÞ is
the transfer function, NðωÞ is the measurement noise, YðωÞ is the column vector of the ns measured responses at nω discrete
frequencies, and FðωÞ is the force. To apply the Bayesian approach, the following prior distributions are used:
F  N c ðF0 ; ½CF0 Þ (78a)

sN 2  ΓðkN ; βN Þ (78b)

F0  N c ðU 0 ; s2U 0 Þ; 8 i ¼ 1; …; nω (78c)

sF0 2 ðωi Þ  ΓðkF ; βF Þ; 8 i ¼ 1; …; nω ; (78d)

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]] 15

2
for the unknowns, F, F0 , sN, and ½CF0 1  which represent the vector of force frequency components, the vector of mean values,
measurement noise variance and the diagonal covariance matrix, respectively. N c and Γ represent normally distributed or
gamma distributed data, and the terms in parentheses represent the mean and variance, respectively. Substitution of
Eqs. (78) into Eq. (76), and maximizing the posterior probability density function with respect to the force, F, produce the
following estimate:
!1 !
ns ns
b n 2
F ¼ ∑ ½H k  ½H k  þ s ½C 1 n 2 1
∑ ½H k  Y k þs ½C F0 : (79)
N F0 N F0
k¼1 k¼1

The solution of this problem is accomplished by using a Monte Carlo Markov chain method. This provides the solution to the
joint probability density function:
pðF; F0 ; ½CF0 1 ; sN 2 jDÞ;

but to fully reconstruct the force, the marginal posterior probability density function, pðFjDÞ, is required. This is asympto-
tically approximated by
 1 n h i 
p FjDÞ ∑ pðFjF0 ðiÞ; CF0 1 ; sN 2 ; D ; (80)
ni¼1
where n is the number of samples returned by the Monte Carlo Markov chain method that was implemented.
This approach was, also, applied to a force reconstruction problem with uncertain modal parameters. The unknown force
is modeled by
F ¼ ½BX þ η; (81)
where ½B is the matrix of basis functions, X is the vector of coefficients, and η is the vector of residual errors, which is
zero-mean and normally distributed with unknown variance, s2η .

4.3.1. Results and discussion


Using Bayesian techniques, the force was able to be reconstructed using the measured data independent of additional
constraints to stabilize the results. Additionally, the authors provided a comparison between Bayesian methods and the
Tikhonov regularization method. They showed that these were equivalent but that the Bayesian method inherently regular-
ized the problem, whereas Tikhonov regularization requires additional effort to determine a suitable regularization
parameter. The Bayesian approach yielded more accurate reconstructions than Tikhonov regularization. This is especially
true involving estimates of the maximum value of the peak of a force.

4.4. Adaptive estimation

The term, adaptive estimation, refers to estimation that is akin to that of adaptive systems used in control theory. There is
a measured signal with unknown parameters, and a signal that estimates the state of the system using estimates for the
unknown parameters. A control law is employed to vary the unknown parameter until the signals converge. The adaptive
estimators employ iterative control laws based on probabilistic or statistical methods to update the unknown parameter.

4.4.1. Sequential realization


Güntürkün [24] used an adaptive estimator to estimate the input signal for nonlinear systems, given by
xðnÞ ¼ gðxðn 1Þ; uðnÞÞ þ ωðnÞ (82a)

yðnÞ ¼ hðxðnÞÞ þ νðnÞ: (82b)


An estimator for the point, y(n), based on the previous point, yðn  1Þ, is used, and defined as
zðnÞ ¼ f~ ðyðn  1ÞÞ: (83)
This is used in an online prediction error computation, eðnÞ ¼ yðnÞ  zðnÞ, estimated by
∂f ðyðn  1Þ; uðnÞÞ
eðnÞ uðnÞ þ ϵðnÞ; (84)
∂uðnÞ
where ϵðnÞ is the overall approximation error. Excess error is reduced first by an adaptive filter, whose output is described as
∂f ðy; uÞ
qðnÞ ¼ uðnÞ þ ϵ0 ðnÞ; (85)
∂uðnÞ
where ϵ0 represents the reduced noise. Then, using Wold's decomposition, which states that any discrete-time stochastic
process can be represented as the sum of a predictable part and an unpredictable part, Eq. (85) can be rewritten as
qðnÞ ¼ ηðnÞuðnÞ þ θðnÞ; (86)
where η is the predictable part and θ is the stochastic part. A relation of the form, bðnÞ ¼ ∑1
q k ¼ 1 ak qðn  kÞ ¼ an qðnÞ,
T
is used to

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
16 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

estimate q(n). The inversion is performed by using a least mean square technique with Tikhonov regularization, given by
bðnÞÞ2  þλ‖q
J ¼ E½ðqðnÞ  q bðnÞ  q
bðn 1Þ‖2 : (87)
Minimization of Eq. (87) with respect to the vector of the weights, an , and the step size, μn, produces
an þ 1 ¼ an þμn ½αðnÞ λγðnÞqðnÞ (88a)

μn þ 1 ¼ μn þ ζ½αðnÞ  λγðnÞΨ Tn qðnÞ (88b)

Ψ n þ 1 ¼ ½I ð1 þ λÞμn qT ðnÞqðnÞΨ n þ½αðnÞ  λγðnÞqðnÞ; (88c)


b ðnÞ, and γðnÞ ¼ q
where ψ n ¼ ∇μn an , αðnÞ ¼ qðnÞ  q b ðnÞ  q
b ðn  1Þ. It should be noted that no recursive relation is given for u(n)
due to the generality of the analysis. To implement this method, a model for u(n) is required, and thus, no recursive relation
can be derived until it is actually applied to a particular problem.
An error lower bound was provided by the use of the Posterior Cramer–Rao Lower Bound. For a sample of measured data,
y, a k þ1 dimensional estimated random parameter, θ, the joint probability density function for the pair ðy; θÞ, given by
pðY; ΘÞ, and the estimate of θ, θ,b the error estimation is given by

P ¼ E½ðb
θ  θÞðθb  θÞT  Z I  1 ; (89)
where I is the ðk þ 1Þ  ðk þ1Þ Fisher information matrix with the components given by
∂2 logðpðY; ΘÞÞ
Iij ¼ E  ; i; j ¼ 1; …; k þ 1: (90)
∂Θi ∂Θj
The lower bound is modeled for the general nonlinear filtering problem given by
x1n þ 1 ¼ g 1 ðxn Þ þ ω1n

x2n þ 1 ¼ g 2 ðxn Þ þ ω2n

yn ¼ hðxn Þ þ νn ; (91)
where the state x has been split into two parts, and x1n x2n,
with the respective nonlinear mappings, g1 and g2 and additive
noise, ω1n and ω2n. The function, h, is the measurement mapping.

4.4.2. Kalman Filtering


Another common type of adaptive estimation is known as the Kalman filter and has been used in multiple ways in force
reconstruction techniques.
Ma et al. [25] applied a Kalman filter and a recursive least squares algorithm to find the force acting on a beam. Using the
finite element method, the equations of motion are given by
€ þCYðtÞ
MYðtÞ _ þKYðtÞ ¼ FðtÞ; (92)
where M denotes the mass matrix, C denotes the damping matrix, K denotes the stiffness matrix, F is the input force vector,
€ Y_ and Y denote the vectors of acceleration, velocity and position, respectively. To rewrite in state space form, the
and Y,
_ T . This allows for the model to be rewritten as
vector, X, must be defined as XðtÞ ¼ ½YðtÞ Y
_
XðtÞ ¼ AXðtÞ þ BFðtÞ (93a)

ZðtÞ ¼ HXðtÞ; (93b)


where
0nn Inn 0nn
A¼ ; B¼ : (94)
M  1K M  1 C M1
and XðtÞ ¼ ½X 1 ðtÞ X 2 ðtÞ … X 2n  1 ðtÞ X 2n ðtÞT . H is the identity matrix, and Z is the observation vector. In discretized form and
associated with noise, Eqs. (93) become
Xðk þ 1Þ ¼ ΦXðkÞ þ Γ½FðkÞ þ wðkÞ (95a)

XðkÞ ¼ ½X 1 ðkÞ X 2 ðkÞ … X 2n  1 ðkÞ X 2n ðkÞT (95b)

Φ ¼ expðAΔtÞ (95c)
Z ðk þ 1ÞΔt
Γ¼ expfA½ðk þ1ÞΔt  τgB dτ (95d)
kΔt

wðkÞ ¼ ½w1 ðkÞ w2 ðkÞ … wn  1 ðkÞ wn ðkÞT (95e)

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]] 17

FðkÞ ¼ ½F 1 ðkÞ F 2 ðkÞ…F n  1 ðkÞ FnðkÞT (95f)

ZðkÞ ¼ HXðkÞ þ vðkÞ (95g)

ZðkÞ ¼ ½Z 1 ðkÞ Z 2 ðkÞ … Z 2n  1 ðkÞ Z 2n ðkÞT (95h)

vðkÞ ¼ ½v1 ðkÞ v2 ðkÞ … vn  1 ðkÞ vn ðkÞT ; (95i)


where X(k) is the state vector, Φ is the state transition matrix, Γ is the input matrix, Δt is the sampling interval, and F(k) is
the deterministic input. The term, wðkÞ, is the noise vector which is assumed to be zero mean and white noise with variance
E½ðwðkÞwT ð jÞ ¼ Q δkj , where Q is the process noise covariance matrix and δkj is the Kronecker delta. The term, vðkÞ, is the
measurement noise vector which is assumed to be zero mean and white noise with variance E½vðkÞvT ðjÞ ¼ Rδkj , where R is
the measurement noise covariance matrix. The equations of the Kalman filter are
Xðkjk  1Þ ¼ ΦXðk  1jk  1Þ (96a)

Pðkjk 1Þ ¼ ΦPðk  1jk  1ÞΦT þ ΓQ Γ T (96b)

SðkÞ ¼ HPðkjk 1ÞHT þ R (96c)

Ka ðkÞ ¼ Pðkjk  1ÞHT S  1 ðkÞ (96d)

PðkjkÞ ¼ ½I Ka ðkÞHPðkjk  1Þ (96e)

ZðkÞ ¼ ZðkÞ  HXðkjk  1Þ (96f)

XðkjkÞ ¼ Xðkjk  1Þ þKa ZðkÞ: (96g)


The equations of the recursive least-squares algorithm are
Bs ðkÞ ¼ H½ΦMs ðk 1Þ þ IΓ (97a)

Ms ðkÞ ¼ ½I Ka H½ΦMs ðk  1Þ þI (97b)

Kb ðkÞ ¼ γ  1 Pb ðk 1ÞBTs ðkÞ½Bs γ  1 Pb ðk  1ÞBTs Þk þ SðkÞ  1 (97c)

Pb ðkÞ ¼ ½I Kb ðkÞBs ðkÞγ  1 Pb ðk 1Þ (97d)

b ¼ Fðk
FðkÞ b  1Þ;
b  1Þ þKb ðkÞ½ZðkÞ Bs ðkÞFðk (97e)
where P is the filter's error covariance matrix, SðkÞ is the innovation covariance, Ka ðkÞ is the Kalman gain, Bs ðkÞ and Ms ðkÞ are
sensitivity matrices, ZðkÞ is the innovation, Kb ðkÞ is the correction gain for updating Fb ðkÞ, and Pb is the error covariance of the
estimated input vector, Fb ðkÞ. The parameter, γ, is used as a compromise between fast adaptive capability and loss of accuracy.
The terms, Xðkjk  1Þ, are representative of the conditional expected value. It indicates that the expected value of X expected
at time k is based on the observations made at time k  1.
Xu et al. [26] used an extended Kalman filter to identify the pressure load on a structure. The filter is meant to be applied
to a general model and is expressed as
zt ¼ hðxt Þ þ vt ; (98)
where hðxt Þ represents the structural response corresponding to the loading parameters, xt, and is considered to be
nonlinear, zt is the measurement observation, and vt is the noise vector at time, t, with covariance matrix, Rt. The algorithm
for reconstructing the force, xt, is as follows:
xt ¼ xt  1 þ Kt ðzt  hðxt  1 ÞÞ (99a)

Kt ¼ Pt HTt R t 1 (99b)

Pt ¼ Pt  1 Pt  1 HTt ðHt Pt  1 HTt þ Rt Þ  1 Ht Pt  1 (99c)

∂hðxÞ
Ht ¼ j ; (99d)
∂x x ¼ xt  1
where Kt is the Kalman gain matrix, Pt is the covariance matrix of the state estimate, and Ht is the sensitivity matrix.
To reconstruct the force, xt, this algorithm is repeated until the ith iteration, xit, is sufficiently close to the observed
measurement, i.e., J hðxit Þ  zt J oδ where δ is the stopping criterion, hðxit Þ is the structural response function evaluated at
iteration, i, of the reconstructed force estimate at time t, xit, and zt is the measured structural response.

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
18 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

Lourens et al. [27] used an augmented Kalman filter to identify forces on structures. The state space model and force
model in discrete time are given by
xk þ 1 ¼ Axk þ Bpk þwk (100a)

pk þ 1 ¼ pk þ ηk ; (100b)
where A and B are system matrices, wk is the system noise, pk is the force input, and ηk is the noise related to the force.
Combining Eqs. (100) and defining a new state vector, xa, the augmented model is
xak þ 1 ¼ Aa xak þζk ; (101)
where
" # " #
A B xk wk
Aa ¼ ; xak ¼ ; ζk ¼ :
0 I pk ηk

The time observation equation with unknown noise is


dk ¼ Ga xak þvk ; (102)
where dk is the measured data vector and the matrix, Ga, is augmented from output influence and direct transmission
matrices, G and J, and is given by Ga ¼ ½G J.
The Kalman filter equations to solve for the measurement update of the augmented system are
Lk ¼ P kjk  1 GTa ðGa P kjk  1 GTa þRÞ  1 (103a)

a a a
x kjk ¼ b
b x kjk  1 þ Lk ðdk  Ga b
x kjk  1 Þ (103b)

P ðkjkÞ ¼ P kjk  1  Lk Ga P kjk  1 ; (103c)

where Lk is the Kalman gain, Pkjl is the error covariance matrix, and R is the covariance matrix of the measurement noise, vk.
The discrete time-update equations are, then, given by
a a
x k þ 1jk ¼ Aa b
b x kjk (104a)

P k þ 1jk ¼ Aa P kjk ATa þ Q a ; (104b)


where Qa is the augmented covariance matrix of the system noise, wk, and noise related to the force, ηk. This is given by
Q 0
Qa ¼ ;
0 S
bak þ 1jk ,
where Q is the covariance matrix associated with wk and S is the covariance matrix associated with ηk. The term, x
indicates the estimate at time k þ 1 based on the observation made at time k.

4.4.3. Results and discussion


The methods proposed in this section each offer specific pros and cons to the method of reconstruction.
The method of sequential realization was tested by the authors on slowly varying and rapidly varying input forces. While
slowly varying forces can be reconstructed with a high degree of agreement under varying noise levels, rapidly varying
forces have been shown to be more sensitive to higher levels of noise, and thus, can be subjected to improvement.
The Kalman filter presented by Ma et al. [25] was able to track the unknown input forces, but signal quality appears
to degrade with the increase in system noise. Additionally, a time lag between the estimate and the exact signal is
approximately equal to three time steps.
The method of the Kalman filter proposed by Xu et al. [26] was tested under various loading conditions with the use of
either a strain or displacement sensor. For quasi-static loads, the results for either sensor were quite accurate, producing
errors of no more than 7 percent for displacement measurements and 2 percent for strain measurements. It should be noted
that the displacement measurements required fewer iterations than the strain measurements when estimates were made
based on measurements with only a single sensor.
Dynamic loading conditions were, also, tested under 2 cases: with 1 sensor used, and with 2 sensors used. The results of
reconstructions with a single sensor varied significantly, and produced errors in excess of 50 percent. The authors postulated
that data from a sensor at a single point produced an ill-conditioned problem. To test this, they conducted an experiment
using sensors at two points. The results from this case were significantly improved over the single point case, providing
estimates with errors of no more than 8.1 percent.
This was similarly addressed by Lourens et al. [27] with the augmented Kalman filter. The example of a force identifi-
cation of a hammer impact exhibited some levels of ill-posedness when a single acceleration sensor was used. When the
number of sensors used was increased to 6 collocated sensors, the force was almost identically identified by the augmented
Kalman filter. When the sensors were non-collocated, the quality of the results diminished and the system became ill-posed.

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]] 19

5. Conclusions

A number of methods of force reconstruction have been presented. A common problem among these methods is the
absence of a properly defined method for the determination of the regularization parameter, as exhibited, for example, in
the mode selection analysis. But while these methods may be impractical in application due to this open-endedness, they
could very well prove invaluable to the advancement and development of reconstruction techniques as they can reveal the
connections between certain properties and the ill-posedness of the given problem.
To overcome this impracticality in the application of techniques with an unoptimized regularization parameter, the
analysis should be amended to include the parameter in the analysis. This inclusion of the regularization parameter in the
technique is exhibited by methods such as Lagrange multipliers in the variational calculus or in the optimal filtering
technique presented in Huang et al. [11]. In these methods, the regularization parameters are among the outputs of the
reconstruction technique employed, and thus, the technique procedure is more robust.
In addition to the regularization parameter, the inclusion of noise in the analysis could provide significant improvements to the
advancement of reconstruction techniques. The observation of the noise behavior leads to more complete and realistic
reconstruction methods. This is observed in Zhang et al. [23], Huang et al. [11] and the Kalman filter methods. Reconstruction
techniques that account for noise led to more concise and self-contained results, and in a sense, were self-regularizing. Thus,
methods that seek to better reign in the system noise with a more rigorous analysis would be exceedingly beneficial to the subject
area of force reconstruction.
For those who wish to use this reference as a basis to engage in research in the subject of force reconstruction, we
encourage the effort. Our objective in writing this paper is to gather and organize the major efforts to date. But since the
subject of force reconstruction is a relatively young field, the number of references used to write this paper is relatively
small. Thus, the opportunity to advance the field of force reconstruction is abundant. We believe that several opportunities
are present in the sub-groupings presented herein. These may be of either a theoretical or numerical nature, and we believe
that this is especially true for recursion-based methods. The potential for advancement is not limited to the topics covered
within this paper. Probability theory provides ample opportunity for the advancement of this field in addition to what was
reviewed in this paper, and the application of new fields, such as control theory, can offer new insights into the field of force
reconstruction.

Acknowledgments

The authors would like to thank the U.S. Department of Homeland Security, Transportation Security Laboratory, in
particular Charles Chih-Tsai Chen and Mark W. Nansteel for their interest, involvement, and support.

References

[1] E. Jacquelin, P. Hamelin, Force recovered from three recorded strains, International Journal of Solids and Structures 40 (2003) 73–88.
[2] S.S. Law, T.H.T. Chan, Q.H. Zeng, Moving force identification: a time domain method, Journal of Sound and Vibration 201 (1997) 1–22.
[3] T.H.T. Chan, S.S. Law, L. Yu, T.H. Yung, Moving force identification studies. I: theory, Journal of Sound and Vibration 247 (2001) 59–76.
[4] S.E.S. Karlsson, Identification of external structural loads from measured harmonic responses, Journal of Sound and Vibration 196 (1996) 59–74.
[5] Peter Avitabile, Fabio Piergentili, Ken Lown, Identifying dynamic loadings from measured responses, Journal of Sound and Vibration 33 (1999) 24–28.
[6] N.S. Vyas, A.L. Wicks, Reconstruction of turbine blade forces from response data, Mechanism and Machine Theory 36 (2001) 177–188.
[7] E. Jacquelin, A. Bennani, P. Hamelin, Force reconstruction: analysis and regularization of a deconvolution problem, Journal of Sound and Vibration 265
(2003) 81–107.
[8] X.Q. Jiang, H.Y. Hu, Reconstruction of distributed dynamic loads on an Euler beam via mode-selection and consistent spatial expression, Journal of
Sound and Vibration 316 (2008) 122–136.
[9] X.Q. Jiang, H.Y. Hu, Reconstruction of distributed dynamic loads on a thin plate via mode-selection and consistent spatial expression, Journal of Sound
and Vibration 323 (2009) 626–644.
[10] M.C. Djamma, N. Ouelaa, C. Pezerat, J.L. Guyader, Reconstruction of a distributed force applied on a thin cylindrical shell by an inverse method and
spatial filtering, Journal of Sound and Vibration 301 (2007) 560–575.
[11] Jianyong Huang, Lei Qin, Xiaoling Peng, Tao Zhu, Chunyang Xiong, Youyi Zhang, Jing Fang, Cellular traction force recovery: an optimal filtering approach
in two-dimensional Fourier space, Journal of Theoretical Biology 259 (2009) 811–819.
[12] C.-H. Huang, A non-linear inverse vibration problem of estimating the external forces for a system with displacement-dependent parameters, Journal
of Sound and Vibration 248 (2001) 789–807.
[13] Cheng-Hung Huang, A nonlinear inverse problem in estimating simultaneously the external forces for a vibration system with displacement-
dependent parameters, Journal of the Franklin Institute 342 (2005) 793–813.
[14] Cheng-Hung Huang, Chih-Chun Shih, Sin Kim, An inverse vibration problem in estimating the spatial and temporal-dependent external forces for
cutting tools, Applied Mathematical Modelling 33 (2009) 2683–2698.
[15] Ching-Shih Yen, Enboa Wu, On the inverse problem of rectangular plates subjected to elastic impact. Part 1: method development and numerical
verification, ASME Journal of Applied Mechanics 62 (1995) 692–698.
[16] Fergyanto E. Gunawan, Levenberg–Marquardt iterative regularization for the pulse type impact-force reconstruction, Journal of Sound and Vibration
331 (2012) 5424–5434.
[17] Maria Grazia Gasparo, Alessandra Papini, Aldo Pasquali, A two-stage method for nonlinear inverse problems, Journal of Computational and Applied
Mathematics 198 (2007) 471–482.
[18] Ching-Shih Yen, Enboa Wu, On the inverse problem of rectangular plates subjected to elastic impact. Part 2: experimental verification and further
applications, ASME Journal of Applied Mechanics 62 (1995) 699–705.
[19] Yu Liu, W. Steve Shepard Jr., An improved method for the reconstruction of a distributed force acting on a vibrating structure, Journal of Sound and
Vibration 291 (2006) 369–387.

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i
20 J. Sanchez, H. Benaroya / Journal of Sound and Vibration ] (]]]]) ]]]–]]]

[20] S. Granger, L. Perotin, An inverse method for the identification of a distributed random excitation acting on a vibrating structure. Part 1: theory,
Mechanical Systems and Signal Processing 13 (1999) 53–65.
[21] S. Granger, L. Perotin, An inverse method for the identification of a distributed random excitation acting on a vibrating structure. Part 1: theory,
Mechanical Systems and Signal Processing 13 (1999) 67–81.
[22] Masaaki Tanio, Yoshito Hirata, Hideyuki Suzuki, Reconstruction of driving forces through recurrence plots, Physics Letters A 373 (2009) 2031–2040.
[23] E. Zhang, J. Antoni, P. Feissel, Bayesian force reconstruction with an uncertain model, Journal of Sound and Vibration 331 (2012) 798–814.
[24] Ulas Gunturkin, Sequential reconstruction of driving-forces from nonlinear nonstationary dynamics, Physica D 239 (2010) 1095–1107.
[25] C.-K. Ma, J.-M. Chang, D.-C. Lin, Input forces estimation of beam structures by an inverse method, Journal of Sound and Vibration 259 (2003) 387–407.
[26] Shaowan Xu, Xiaomin Deng, Vikrant Tiwari, Michael A. Sutton, William L. Fourney, Damien Bretall, An inverse approach for pressure load
identification, International Journal of Impact Engineering 37 (2010) 865–877.
[27] E. Lourens, E. Reynders, G. De Roeck, G. Degrande, G. Lombaert, An augmented Kalman filter for force identification in structural dynamics, Mechanical
Systems and Signal Processing 27 (2012) 446–460.

Please cite this article as: J. Sanchez, & H. Benaroya, Review of force reconstruction techniques, Journal of Sound and
Vibration (2014), http://dx.doi.org/10.1016/j.jsv.2014.02.025i

You might also like