Documente Academic
Documente Profesional
Documente Cultură
This article was originally published in the Reference Module in Materials Science and Materials Engineering,
published by Elsevier, and the attached copy is provided by Elsevier for the author’s benefit and for the benefit of the
author’s institution, for non-commercial research and educational use including without limitation use in instruction
at your institution, sending it to specific colleagues who you know, and providing a copy to your institution’s
administrator.
All other uses, reproduction and distribution, including without limitation commercial reprints, selling or licensing
copies or access, or posting on open internet sites, your personal or institution’s website or repository, are prohibited.
For exceptions, permission may be sought for such use through Elsevier’s permissions site at:
http://www.elsevier.com/locate/permissionusematerial
Stavroulakis G.E., Bolzon G., Waszczyszyn Z., and Ziemianski L., Inverse Analysis. In: Saleem Hashmi (editor-in-
chief), Reference Module in Materials Science and Materials Engineering. Oxford: Elsevier; 2016. pp. 1-39.
ISBN: 978-0-12-803581-8
Copyright © 2016 Elsevier Inc. unless otherwise stated. All rights reserved.
Author's personal copy
Inverse Analysis$
GE Stavroulakis, Technical University of Crete, Chania, Greece, and Carolo Wilhelmina Technical University, Braunschweig, Germany
G Bolzon, Politecnico di Milano, Milano, Italy
Z Waszczyszyn, Cracow University of Technology, Kraków, Poland
L Ziemianski, Rzeszow University of Technology, Rzeszów, Poland
r 2016 Elsevier Inc. All rights reserved.
1 Introduction 1
2 Model Calibration versus Diagnostic Identification 2
3 Tools 2
3.1 Mathematical Programming and Numerical Optimization Techniques 2
3.2 Extended Filter Approach 3
3.3 Computing Sensitivity 7
3.4 Soft Computing Techniques and ANNs 9
3.5 Basics of Neural Networks 10
3.5.1 Back-propagation neural network 10
3.5.2 Radial basis function neural network 12
3.5.3 Adaptive neuro-fuzzy inference system 12
3.5.4 Hopfield−Tank neural network 14
3.6 POD-RBF Model Reduction Technique 14
4 Case Studies and Applications 15
4.1 Crack and Flaw Identification Problems 16
4.1.1 Static loadings 16
4.1.2 Dynamic loadings 19
4.1.3 Comparison between algorithms and comments 21
4.2 Inverse Problems in Materials Mechanics 21
4.2.1 Parameter identification in quasi-brittle fracture 22
4.2.2 Parameter identification for ductile materials and fracture 25
4.3 Neural Network Application to Inverse Problems 28
4.3.1 Reliability of elastoplastic structures 28
4.3.2 Identification of real building fundamental periods of vibration 30
4.3.3 Analysis of implicit constitutive relations 32
4.3.4 Updating of structural mathematical models 34
4.3.5 Damage detection using wave propagation 35
References 37
1 Introduction
Techniques and applications of inverse and identification problems are presented in this chapter. A general-purpose formulation
of the inverse problem, based on system identification techniques and least-squares fitting of parameterized models to experi-
mental data, is presented in detail, since this is the most general approach available today. Nevertheless, one must underline here
that no general method is available for the solution of all problems of practical interest. The reasons for this, as it will be discussed
in more detail later in several parts of this text, is that inverse problems lead to difficult optimization problems (ill-posed) whose
solution is not always straightforward with current numerical optimization techniques. Therefore, one should consider semi-
empirical methods and experimental testing techniques as well. All these methods are, usually, of restricted applicability and are
related to one specific application. The compensation for this is that a large amount of experience is incorporated in these
methods, so that their results are, within the limits of their applicability, superior to the ones provided by the general-purpose,
optimization-based method. To be fair one should add here that empirical methods are restricted to the identification of a
significant change in the system's response and, in some cases, try to estimate the extent of the existing damage. Alternatively, the
general error minimization technique is, in principle, able both to detect the existence of a defect and estimate its nature and
extent.
☆
Change History: March 2015. G. Bolzon added Section 3.6 on POD–RBF Model Reduction Technique, improved Section 4.2.2 on Parameter Identification for
Ductile Materials and Fracture, and updated the reference list.
Here the error-minimization-based approach to inverse analysis is called model calibration technique, while all other methods
are collectively characterized as diagnostic identification. The second section of this chapter is devoted to a brief description of
both methods together with their areas of applicability.
The solution of inverse problems is based on classical optimization techniques and more recent data processing techniques like
the ones studied in the area of soft computing. Therefore, the third section is devoted to a brief presentation of the relevant
techniques.
Various applications from material and structural identification are presented in more detail in the last section of this chapter.
Let us assume that one is able to define a suitable mathematical model of the structural system or of the material under
investigation and that this model includes all possible sources of damage and degradation. The latter group may include,
depending on the studied application, cracks and defects of any type, adhesive cracks, joints and delamination mechanisms, or
plasticity and damage. The model contains certain unknown parameters which have to be determined. This is the goal of the
inverse analysis. Using this mathematical model one may calculate, numerically, the response of the mechanical system for every
given excitation. Model calibration means the determination of the unknown parameters of the model, especially of the ones
relevant to defects and damage, by comparing the results of the model with experimental measurements. Data fitting is usually
done by means of an error minimization technique, where the distance between parameterized (by the unknown parameters)
predictions of the mechanical model and measurements of the corresponding experiment is minimized. This formulation is
known as an output error minimization procedure for the inverse problem. Instead of the quadratic norm of the distance between
measurements and parameterized model predictions, one may use other more complicated (even nondifferentiable) norms,
although this is not common in the scientific literature. Furthermore, other affiliated formulations of the inverse problem may be
used, for instance, minimization of the error in the governing equations of the mechanical problem, or in particular parts of it (e.
g., interfaces).
The model calibration technique is a general-purpose method for the solution of inverse and identification problems in
mechanics, provided one has a suitable model and one is able to solve the resulting minimization problem. The first requirement
means that, e.g., the arising defects or other physical mechanisms are included in the mathematical model used for the identi-
fication. In several engineering applications accumulated experience on design and damage evaluation permits us to include the
essentials of a given problem in the model. For new and innovative applications this step is by no means trivial. A remedy of at
least theoretical value is to use several models and let an automatic procedure (e.g., by evolutionary programming) choose among
them the best one.
3 Tools
A number of mathematical tools and data processing techniques which are used in the solution of inverse problems will be briefly
presented in the following sections.
where || || denotes, in general, a weighted norm of the involved error. Often this is the Euclidean norm. In order to fix ideas, let us
consider the crack identification problem outlined in the sequel, where one tries to find the position and the shape of the crack
lying in the interior of an elastic plate from boundary measurements of the elastic deformation under a given static loading. In this
case, the unknown position of a crack and other parameters defining its shape and size are included in the vector z, and y(z) ¼u(z)
is the vector of boundary displacements for every possible value of z. Alternatively, the same specimen with a crack of fixed
position and size subjected to the same loading leads to boundary displacements y0. A successful solution of problem [1] leads to
the determination of the crack parameters corresponding to y0. It should be pointed out here that the structural mapping s(z) is
based on a mathematical model of the mechanical (here, elastostatic) problem, whereas the values of y0 may come from the
laboratory.
In order to understand the difficulties of problem [1], let us first examine the structural mapping s(z). In general, this is a
nonlinear vector-valued function. Therefore, the objective function in this optimization problem, which, for least-squares
Author's personal copy
Inverse Analysis 3
problems, is a composite function of a quadratic one with the structural mapping, is nonconvex. Optimization problems of this
type belong to the class of mathematical programs with equilibrium constraints (see, e.g., the treatise by Luo et al. (1996)). This
observation explains the consequent discussion of more general optimization algorithms with more global properties (cf. for crack
identification problems, Stavroulakis, 2001; Engelhardt et al., 2006b). Of course, local convergence is almost always guaranteed.
Therefore, classical numerical optimization does not lose its importance, provided a good starting point (initial estimate of the
solution) is available.
Classical numerical optimization is a well-known topic, which need not be presented in detail here. For instance, easily
available in widely diffused packages (e.g., Matlab, 2000) are sequential quadratic programming (SQP) algorithms, designed to
solve general constrained optimization problems through the construction of the Lagrangian function which incorporates the
original objective function and its constraints. A sequence of convex quadratic approximations to this Lagrangian is generated by
using information from gradients only, while no second derivatives are evaluated. At each iteration the global optimum point is
computed by linear search along the direction which minimizes the current convex quadratic approximation. The method is quite
effective when gradients can be provided by the user, otherwise they are evaluated by finite difference schemes, but at the price of
significant computing costs and of big efficiency loss.
As an alternative, recourse could be made to evolutionary methods such as neural networks, to be dealt with in the following,
or (direct search) genetic algorithms (GAs) which have been shown to be effective in numerous kinds of hard optimization
problems (Goldberg, 1989; Davis, 1991). The genetic procedure is started with a randomly generated population, each member of
which represents a set of the nz unknown parameters collected by vector x and defined within the nz-dimensional box delimited by
the upper and lower bounds usually provided by a priori knowledge. For each member (‘individual’) of the population, the
structural mapping s(z) is performed by standard, ‘direct’ solving procedures. The fitness of each individual is assessed on the basis
of the corresponding value of the discrepancy function [1]. A subset of ‘better fit’ individuals is then selected from the initial
population and from each one of the subsequent generations, and is used to reproduce a new population by operations like cross-
over and mutation. The process of reproduction is continued until further reduction of the objective function is no longer observed
in the fittest member of the population. The most time-consuming operation within this method is the solution of the direct
problem, which is repeated many times. However, an efficient solver can make GA competitive and a practical alternative to
gradient-based algorithms.
As a common drawback of any nonconvex optimization tool (Arora et al., 1995), it is impossible to ensure their convergence
toward the optimal solution. The aforementioned SQP procedures require an initialization point to be provided by the user, and
different runs can converge toward different local minimum points just by changing this initialization. The best solution, among
the possibly multiple local minimum points, may be singled out simply by comparing the values of the objective function in
them. However, in some situations, multiple sets of optimal values may exist which fit the experimental data to the same degree of
accuracy. Therefore, the reliability of the identified parameters rests on both the confidence level of experimental data and the
appropriate design of the inverse procedure, as it will be shown in the example section.
1. the flow of experimental data yk (ordered by the counter k) collected during laboratory or in situ tests, usually affected by some
measurement noise, vk;
2. the model sk(zk) which describes the process, depending on the unknown parameters collected by vector zk; and
3. an a priori estimate of the parameters, according to previous experiences or to the judgment of an hypothetical expert, which is a
piece of information which can be clearly accommodated as an additional datum at k¼ 0.
All the variables but the model, which is deterministic, are assumed to be stochastic, characterized by Gaussian (or ‘normal’)
probability density distribution.
Author's personal copy
4 Inverse Analysis
The measurement noise vk is therefore assumed to be a (white) Gaussian random process characterized by zero mean and
covariance matrix Rk, in general time dependent, often with zero cross-correlation; in customary notation:
The experimental quantities yk can be related to the parameters by means of the model, as follows, all along the considered test:
yk ¼ sk ðzk Þ þ vk ½4
A Gaussian probability density is then attributed, as a simplifying idealization in this EKF application, to the sequence of
estimates based on the flow of information yk. Equations [3] and [4] can be combined to obtain the conditional probability
density of the experimental data with respect to the parameters:
1
fYjZ ðz; yk Þ ¼ ðð2pÞny detðRk ÞÞ1=2 exp ðyk sk ðzÞÞT ðRk Þ1 ðyk sk ðzÞÞ ½6
2
As is well known (see, e.g., Tarantola, 1994), the two independent sources of stochastic information, labeled (a) and (c) above, can
be combined to give the a posteriori information which is described by a conditional probability density (Bayes’ theorem) in the form
where m¼fY(yk) (yk) does not depend on the unknown parameters z, and plays the role of a normalizing factor. More explicitly, using
eqns [6] and [5], the expression [7] reads
1
fZjY ðz; yk Þ ¼ expfϖk ðzÞg ½8
m
having set
The set of parameters with maximum conditional probability density is then the optimal vector of the following optimization
problem:
max fZjY ðz; yk Þ ¼ minfϖk ðzÞg ½10
z z
In view of eqns [8] and [9], the objective function to minimize consists of the sum of two quadratic forms: the former measures the
discrepancy between experimental information yk and the corresponding computed quantity for the same situation, as in eqn [1], and
the latter reflects the a priori information and represents a regularization term (in Tichonov's sense, see, e.g., Bui, 1994) which fosters
the convergence of the minimization procedure. In fact, the minimum problem [10] can be considered as a particular case of the more
general (regularized) formulation
where r represents a penalty coefficient, and || || the weighted norm ||x||¼(xTW1x)1/2 which coincides with the Euclidean one when W
coincides with the identity matrix.
As a remarkable property, the a posteriori conditional probability density fZ|Y(z, yk) eqn [8] represents a Gaussian (normal)
distribution when sk(z) is a linear operator. This property does not hold for the general nonlinear case, but an iterative inverse
procedure can still be envisaged, which rests on a step-by-step linearization of the operator sk(z) and on the hypothesis of normal
distributions within each step. These assumptions make the EKF technique heuristic. However, although no convergence proof has
been given so far, many computational experiences have shown that this methodology is usually endowed with good convergence
and stability properties (see, e.g., Stavroulakis and Antes, 2000; Engelhardt et al., 2006a).
To linearize the procedure, e.g., around the initial parameter vector x0, one can consider the first-order Taylor expansion
where the tangent operator S0 represents the so-called ‘sensitivity matrix’ to be dealt with in the next section:
∂sk
S0 z ¼ z0 ½13
∂z
Substituting eqn [12] into eqn [9] and then into eqn [8], using known matrix algebra results (see, e.g., Catlin, 1989; Kailath
et al., 2000), and normalizing once more, it is possible to rewrite the normal distribution fZ|Y as follows:
^ 1=2 exp 1 ðz ^zÞT C
fZjY ðz; yk Þ ¼ ðð2pÞnz detðCÞÞ ^ 1 ðz ^zÞg ½14
2
1 1 T 1
1
^z ¼ z0 þ S T0 R1
k S0 þ C0 S 0 Rk ðyk sk ðz0 ÞÞ ¼ z0 þ C 0 S T0 S 0 C 0 S T0 þ Rk ðyk sk ðz0 ÞÞ ½15
^ ¼ E z ^zÞðz^zÞT ¼ S T R1 S 0 þ C 1 1 ¼ C 0 C 0 S T S 0 C 0 S T þ Rk 1 S 0 C 0
C ½16
0 k 0 0 0
Vector ^z collects the resulting estimates of the mean value of the parameters and represents the solution of the minimum
problem [10], while matrix Ĉ contains the relevant variance values.
Equations [15] and [16] can be written in a more compact format as follows:
^ ¼ E z ^zÞðz^zÞT ¼ C 0 B 0 S 0 C 0
C ½18
where
1 T 1 1 T 1
B0 ¼ C 0 S T0 S 0 C 0 S T0 þ Rk ¼ S 0 Rk S 0 þ C 1
0 S 0 Rk ½19
Finally, it can be observed that eqns [15]–[19] implicitly define a recursive scheme for processing each new information, here
denoted by yk, having set
The filter-driven identification iterative procedure can then be summarized by the following algorithmic steps:
• Step 0 Set iteration counter k¼ 0 and initialize the unknown variables by the a priori estimates z0, C0.
• Step 1 Iteration k ¼ k þ 1: the available observation [4] is assumed in the linearized form
where the linearization matrix Sk at step k is calculated by using the estimate of z which is available from the previous step:
S k ∂sk =∂zjz ¼ ^zk 1 .
• Step 2 The new estimation of the state, i.e., the new value of ^z, is given by
where yk is the current observation (measurements) and sð^zk1 Þ is the attained value of the observation (measurements) based on
the estimation of the state values available from the previous iteration step. Matrix Bk is the filter gain, which reads
Bk ¼ C ^ k1 S T þ Rk Þ1
^ k1 S T ðS k C ½23
k k
while the covariance matrix of the estimates is updated according to the rule [18]:
C ^ k1 B k S k C
^k ¼ C ^ k1
In some situations, the following relatively simple projection filter, which disregards the estimate covariance [19], may also
lead to good results (Tosaka et al., 1995):
þ T þ
B k ¼ ðS Tk Rþ
k S k Þ S k Rk ½24
In the above formula, superscript þ denotes the Moore–Penrose generalized inverse of a matrix.
The effectiveness of the identification can be checked by means of some confidence indicators, such as the following; details are
given in Catlin (1989) and Tarantola (1994).
An ellipsoid can be defined in the space of parameters z by the estimates in situation k, centered in the point ^zk :
^ 1 ðz ^zk Þ ¼ 1
ðz ^zk ÞT C ½25
k
It can be easily proven that the square lengths of the principal diameters of this ellipsoid are equal to the eigenvalues of the
current covariance matrix Ĉ k, and that the volume bounded by the ellipsoid is proportional, through factor p, to the product of the
lengths of the principal axes, i.e., to det(Ĉ k). These geometric properties permit to measure and graphically visualize the ‘con-
fidence level’ of the estimates and the effectiveness of the identification procedure during the flow of information up to situation k
(see, e.g., Figure 2). Particularly meaningful are the ellipsoids obtained amplifying homotetically by a factor 3 those described by
eqn [25], which contain the correct value of parameters with a probability of 99%. The diameters reduce progressively while the
ellipse center moves toward the correct value of the parameters.
Author's personal copy
Inverse Analysis 7
Figure 2 Confidence domains for the joint probability of parameters during EKF identification process. Convergence is monotonic for the
variance, oscillating for the mean value.
A ‘relative information index’ (RI) can also be defined to measure the informative content of the probability density fk with
respect to the probability density fk1, at the previous instant:
Z
f k ðzÞ
RIðf k ; f k1 Þ ¼ f k ðzÞlog dz ½26
Rnx f k1 ðzÞ
For Gaussian probability densities, indicator [26] is available in closed form, as the sum of three addends with transparent
meanings:
1 ^ kÞ
detðC h i
^ k1
2RIðf k ; f k1 Þ ¼ ð^zk ^zk1 ÞT C ð^zk ^zk1 Þ log þ Tr C ^ 1 I
^ k 1C ½27
^ k1 Þ
detðC
k
In some applications, e.g., for the calibration of quasi-brittle fracture parameters to be dealt with in the example section, the
response function sk(z) might be known explicitly and the sensitivity analysis reduces to the evaluation of the gradients ∂sk/∂z. In
most applications, however, this functional dependence is not explicit in the argument z. The conceptually simplest way to
evaluate sensitivity is then to make recourse to the finite difference approach, replacing the derivatives ∂sk/∂z with some
approximation Dsk/Dz, being
∂sk Dsk
z ¼ zk ¼ z ¼ zk þ Oð‖Dz‖n Þ
∂z Dz
the last term on the right-hand side indicating the order of the truncation error.
For instance, the first-order forward difference approximation is
Dsk sðzk þ DzÞ sðzk Þ
z ¼ zk ¼
Dz Dz
This technique is easy to implement but usually rather inefficient. In fact, the computing cost is much increased by the need of
evaluating additional system responses at each situation k, in a number proportional to nz, the dimension of the sought parameter
vector z. The computing effort can be eventually reduced if path-independent system response can be hypothesized, but the
difficulty remains of choosing an appropriate size of the perturbation Dz: this should be small enough to reduce the truncation
error but large enough to overcome the numerical round-off errors.
As a still intuitive alternative to this approach, the problem could be tackled as done, e.g., by Aoki et al. (1997) in ductile
fracture application: the whole system response is evaluated first by its mathematical model for a set of the parameter values,
chosen on a grid in the parameter space within the expected output range (see, e.g., Figure 4); these model responses are then
interpolated by suitable functions, e.g., by polynomials of order p. The evaluation of the gradients is therefore straightforward, but
the introduced approximation might be significant for low-order p while the computing cost might still grow beyond any
acceptable level with the increase of p and nz (the number of the sought parameters). In fact, the number of complete analyses to
be performed sums up to pnz .
An alternative technique originating from shape sensitivity analysis (Kleiber, 1993) is gaining increasing popularity within
finite element (FE) or boundary element (BE) approaches. In fact, discretization in space and, eventually, in time, reduces the
Figure 4 Interpolation of the response function within the expected parameter range.
Author's personal copy
Inverse Analysis 9
direct (forward) analysis to the iterative solution of linearized equation systems, namely,
where Kk represents the consistent tangent stiffness matrix at step k, uk is the vector which collects the main unknowns of the
discretized problem (often, incremental displacements), and rk the vector of the residual forces. The dependence of the tangent
matrix Kk on the parameters is usually explicit, while it is implicit for uk(z) and rk(z). Notably, however, the residual rk vanishes at
convergence.
Any quantity related to the system response (e.g., the considered reaction force Q) can be known as a function of the entries of
vector uk; its sensitivity (variation) with respect to the parameters can then be expressed as a function of the derivatives ∂uk/∂z.
Then, Equation system [28] can be formally differentiated with respect to z:
and solved with respect to sought derivatives, for the current value zk of the parameter vector, at the converged solution (rk ¼0).
Therefore,
!
∂uk
1 ∂K k
z ¼ zk ½K k ðzk Þ z ¼ zk uk ðzk Þ
∂z ∂z
being
Z Z
∂K k ∂C k ∂sk
uk BT B dO uk ¼ BT dO ½30
∂z ∂z ∂z
where B is the standard consistency matrix and Ck is the algorithmic tangent constitutive matrix.
So, the basic steps to evaluate sensitivity are the following:
The method requires the above calculations (steps 1 and 2) for each parameter, independently of the other; therefore, it is quite
efficient when path dependence is avoided (e.g., z¼zk for elastic or pseudo-elastic, even nonlinear, problems) while it becomes
time consuming when the system response and the relevant sensitivity depends on previous history. In this latter case, in fact,
sensitivity computation has to be introduced within the iterative procedure which solves the direct problem several times before
getting convergence.
• A. Simulation is related to direct methods of structural analysis, i.e., for known inputs (e.g., excitations of mechanical systems
(MSs)) and characteristics of structures or materials outputs (responses of MS) are searched.
• A0 . Inverse simulation (partial identification) takes place if inputs correspond to known responses of MS and excitations are
searched as outputs of ANNs.
• B. Identification is associated with the inverse analysis of structures and materials, i.e., excitations and responses are known and
MS characteristics are searched.
In this chapter the main attention is focused on the application of ANNs to the identification problems. In the selected case
studies based on the research developed in Poland under authors’ supervision new possibilities of neurocomputing are illustrated
on examples based on analytical or experimental data. The attention is focused on a comparison of numerical efficiency and other
advantages of ANN applications. Improvement of data processing and formulation of ANN architectures, corresponding to the
problems considered, are discussed as well. Further related applications of ANNs for the solution if inverse problems can be found
in Efstathiades et al. (2007), Tsompanakis et al. (2008), and Mroz and Stavroulakis (2005).
where x(p), t(p) are input and output (target) vectors for pth patterns, L and T are number of training and testing patterns,
respectively. The number of training patterns L should correspond to the number of network parameters (NPs). It is desirable to
have L4NP to prevent the overfitting of the trained network, i.e., to have a small error of neural predictions of results with respect
to testing data. The trained network should also have generalization properties when it operates on data not used for training or
testing.
There are many problems related to formulation, training, testing, and checking of neural network properties. The scope of this
chapter does not allow for a more profound analysis of these problems. This can be found in the extensive literature devoted to
neural networks (cf. references in books by Rojas, 1996; Haykin 1999; Waszczyszyn, 1999). Results quoted in the paper were
achieved by means of computer simulators of ANNs (SNNS, 1998; ST Neural Network, 1998; Neural Network Toolbox, 2000;
Fuzzy Logic Toolbox, 2001), where algorithms corresponding to neurocomputing procedures were described.
X
N X
N
vi ¼ wij xj þ bi ¼ wij xj ; yi ¼ Fðvi Þ ½32
j¼1 j¼0
To make it simpler, the layer index l is omitted in the previous relation and the bias (threshold) parameter bi can be treated as a
weight wio for the unit signal xo ¼ 1 (cf. Figure 5(b)). Figures 5(c)–5(e) shows the following, frequently used activation functions:
binary sigmoid
1
FðvÞ ¼ A ½0; 1 for s40 ½33
1 þ expðsvÞ
bipolar sigmoid
1 expðsvÞ
FðvÞ ¼ A ½1; 1 for s40 ½34
1 þ expðsvÞ
Author's personal copy
Inverse Analysis 11
Figure 5 (a) Three-layer BPNN; (b) single neuron i in layer l; and (c)–(e) binary sigmoidal, bipolar sigmoidal, and Gaussian activation functions,
respectively.
Gaussian function
!
ðv cÞ2
FðvÞ ¼ exp ½35
2s2
After all patterns p¼ 1,…, P are introduced, the network errors can be computed, defined according to different measures:
1X P X M
ðpÞ ðpÞ
E¼ ðt yi Þ2
2 p ¼ 1i ¼ 1 i
½36
2E pffiffiffiffiffiffiffiffiffiffi
MSE ¼ ; RMSE ¼ MSE
P
ðpÞ ðpÞ
where ti and yi are target and computed outputs for pth patterns, E is the least-mean-square-error, MSE is the mean-square-error,
and RMSE is the root-mean-square-error. In the case of fixed neuron parameters s and c in the above listed activation functions, the
NPs wij are computed iteratively during the training process using the formula
where s is the number of iteration step and a(s) is the momentum term parameter.
The simplest learning rule corresponds to the gradient method of steepest descent. Assuming a(s) ¼ 0, the weight increments are
∂E
Dwij ðsÞ ¼ ZðsÞgij ðsÞ; gij ðsÞ ¼ s ½38
∂wij
where Z(s) is the adaptive learning rate. In formula [37] the global learning parameters Z(s) and a(s) can be adapted for each epoch
s, (Neural Network Toolbox, 2000). The epoch corresponds to one forward transmission of signals and one error back propa-
gation, associated with all the introduced patterns.
Local adaptive learning rates Zij(s), associated with each network weight, are numerically more efficient. From among many
local learning rules, the resilient-propagation (Rprop) method is worth mentioning
where Zij ðsÞ ¼ fij ½sgnðgij ðsÞgij ðs 1ÞÞ; ck and ck are control parameters for k¼1, …, 4 (SNNS, 1998).
Author's personal copy
12 Inverse Analysis
Figure 6 (a) Scheme of RBFNN and (b) Gaussian activation function as RBF.
X
H
y¼ wk Gk ðxÞ þ b ½40
k¼1
where wk, b are weights and bias, Gk(x) is the RBF as generalized Gaussian function:
!
X
H ðxj ckj Þ2
Gk ðxÞ ¼ exp ½41
j¼1 2ðskj Þ2
RBFs give local approximation in the input variables space. The approximation is related to H centers of position vectors ck for
k¼ 1, …, H and widths sk. Besides these parameters the weights wk and bias b complete the set of RBF network parameters.
The basic problem of RBFNN is the evaluation of the number H of RBFs and their centers ck. There are many methods and
algorithms of different numerical complexity and efficiency (Jang et al., 1997). The simplest approach is associated with location of
centers at selected inputs of H patterns, i.e., ck ¼ x(k) for k¼ 1,…, H. This leads to the so-called regularization neural networks
(Haykin, 1999; Neural Network Toolbox, 2000).
A simple version of RBFNN is related to the assumptions of the same value of the width parameter to all RBFs. After the number
H of RBFs is evaluated, a value of s is fixed and the least-mean-square (LMS) method is used to compute the values of weights wk
for k¼0, 1,…, H (bias is treated as wo ¼ b for xo ¼ 1) (cf. Figure 6(a)). An optimal value sopt can be computed iteratively by means of
criterion
where RMSE (s;T) is the root-mean-square-error for the testing set of patterns.
In the case of variables skj their values are computed by the BN algorithm. This can be combined with computation of weights
wk by the LMS algorithm in the frame of the so-called hybrid method of learning (Jang et al., 1997). The method depends on
learning of weights wk during the forward pass for fixed skj . Using the back pass the width parameters skj are computed for fixed
values of weights wk.
where mA(x) is membership function. In what follows only the Gaussian membership function is considered, which corresponds to
the activation function [35] for F ¼ m and v¼ x
!
ðx cÞ2
mA ðx; c; sÞ ¼ exp ½44
2s2
The main goal of ANFIS is to perform the mapping: f:x0 -y0 , where x0 ARN and y0 AR1 are related to crisp values. After the
fuzzification, i.e., introduction of membership functions mAj for the components xj of the input vector x, internal manipulations are
made on fuzzy sets Aj. ANFIS is based on three main assumptions: (1) partition of input space, (2) Sugeno fuzzy inference rule,
and (3) hybrid learning of the NPs. Figure 7 shows two partitions for 2D input space (x1, x2). As an example, four Sugeno first-
order inference rules are written for the grid partitioning related to Figure 7(a) as
Figure 7 Partitioning of input space: (a) grid partition and (b) scatter partition.
Figure 8 shows a scheme of ANFIS corresponding to grid partitioning and Sugeno rules [45]. Fuzzy neurons corresponding to
the membership functions mki
Aj and weights wk are computed by means of the algebraic product
0 0 0
wk ¼ mk1
A1 ðx1 ÞmA2 ðx2 Þ⋯mAN ðxN Þ
k2 kN
½46
The weight centroid formula is used to compute the crisp output y0 , i.e.,
PH
wk yk
y0 ¼ Pk H¼ 1 ½47
k ¼ 1 wk
Hybrid learning, described previously, is associated with forward computation of consequence parameters ak and bki in eqn
kj kj
[45]. The error back propagation is applied to compute the premise parameters cj and sj of fuzzy neurons shown in Figure 8 (Jang
et al., 1997; Fuzzy Logic Toolbox, 2001).
1. the exponential growth of the number of preliminary analyses to be performed, depending on the number of parameters to be
identified;
2. the condition of box-shaped search domain (Ageno et al., 2009; Bolzon and Talassi, 2013);
3. the fast deterioration of the approximation in the case of noisy interpolated quantities (Bolzon and Buljak, 2011).
An alternative and more effective approach, suggested by Ostrowski and coworkers (Ostrowski et al., 2005; Ostrowski et al.,
2008), rests on the combined use of proper orthogonal decomposition (POD) and radial basis functions (RBFs). Besides dras-
tically reducing computing times and costs, this methodology permits to retain the essential features of the original simulation
problem and to filter most disturbances (Gosh and Joshi, 2013).
The consideration of RBFs, see e.g., Figure 6(b), permits to circumvent the above listed issues 1 and 2. In fact, RBFs can return an
analytical approximation of the investigated system response based on parameter combinations freely distributed in any region of
interest (Buhmann, 2003; Tiago and Leitão, 2006). Robust RBF implementation supports also 3D computer graphics recon-
structions (Carr et al., 2001).
On the other hand, the noise associated to the computations can be reduced, and the accuracy of the overall
procedure improved, by the consideration of data compression schemes based on POD (Ly and Tran, 2001; Liang et al.,
2002).
The operative steps that define the POD-RBF approximation of a numerical simulation model can be listed as follows.
1. An initial set of admissible parameter combinations zi is generated within the search domain. Vectors zi are collected by the
training matrix z.
2. Each vector zi defines one input parameter set for the numerical analysis. The corresponding system response, called snapshot
in the present context, is stored in matrix U in ordered correspondence with the vectors zi in matrix z.
Author's personal copy
Inverse Analysis 15
Figure 9 Indentation curves resulting from finite element simulation (FEM in the legend) of the test and from the corresponding POD–RBF (POD
in the legend): approximation based on 3 (left) and 9 (right) retained modes.
3. Eigen-pairs of the symmetric square matrix D ¼UTU are computed in a sequence, starting from the largest eigenvalue and
truncating the search as the returned value is smaller than some pre-established threshold. In fact, due to the expected
correlation of the information stored in matrix U, a large number of zero or nearly null eigenvalues are associated to matrix D.
The square root of the inverse of the retained eigenvalues define the diagonal matrix K1=2 while the retained eigenvectors
represent response modes stored in matrix V.
4. The retained eigen–pairs define the matrix U ¼ UVK1=2 (such that UT ¼ U1 ), which represents an optimally truncated
orthonormal reference system for reconstructing the system response.
5. Snapshots are approximated by the linear combination U ¼ UA, where matrix A ¼ UT U collects the amplitudes of the new
basis combination.
6. The coefficients stored in the amplitude matrix A are interpolated by RBFs in order to define the vector valued function a(z).
Thus, uðzÞ ¼ UaðzÞ represents the approximated system response for the input parameter set z defined within the region of
interest and not included in the initial selection z.
The accuracy of the reduced model depends on the number of retained modes and on the informative content of the
neglected ones, roughly quantified by the sum of the discarded eigenvalues compared to the larger figures (Gosh and Joshi,
2013). However, truncation has a beneficial filtering effect in the case of noisy data. As an example, the graphs in Figure 9
visualize the response to indentation of an isotropic steel described by the classical Hencky–Huber–von Mises plasticity model
with exponential hardening rule (Bolzon and Talassi, 2012). The output of a finite element simulation of the experiment and
the approximation provided by the POD–RBF interpolation of 534 direct computations based on a random distribution of
input parameter combinations (training points) are compared. The POD–RBF results are produced by 3 or 9 retained modes.
The parameter set that defines the curves in the graphs is situated on the boundary of the training region, a circumstance that
amplifies the numerical imperfections.
The reduction of the large amount of data gathered by the snapshot matrix U to a small number of coefficients stored in the
amplitude matrix A, reflecting almost the same information, is particularly relevant for the exploitation of full-field measurements
in combination with artificial neural networks (ANNs). In fact, ANNs perform well when the number of input data (that represent
the experimental output, in the present context) is comparable to the size of the output vector (the sought constitutive
parameters).
The POD–RBF approach permits to obtain the sought results in almost real time in most optimization approaches. Thus,
inverse analysis procedures become feasible even at industrial scale and can be applied also in the case of on-site experiments,
performed on operated components for structural diagnostic purposes (Bolzon et al., 2012; Bolzon et al., 2015).
The flexibility gathered by RBF interpolation and the filtering capability of POD opens the interesting perspective of using truly
experimental information for recovering an analytical model of the experiment or, at least, for enhancing the reliability of
computationally-based approximations.
The previously outlined methods have been applied by various researchers, including the authors of this chapter, for the solution
of inverse problems in mechanics. Without any claim of completeness, a number of case studies is presented in this last section
and demonstrates the application of various solution methods for inverse problems. More details can be found in the references to
the original publications. The first part of this section, compiled by Stavroulakis, presents crack and flaw identification problems
Author's personal copy
16 Inverse Analysis
using either classical optimization or soft computing and filter algorithms. The next part, compiled by Bolzon, treats inverse
problems arising in material mechanics modeled and solved by mathematical programming techniques or filter algorithms. The
last part, compiled by Waszczyszyn and Ziemianski, includes various applications of neural networks and fuzzy inference tech-
niques to inverse problems.
nondifferentiability of the error measure. Multiple defects require either the introduction of discrete variables or some tricks for
setting very small variables are equal to zero during the solution procedure of the algorithm (see Stavroulakis and Antes, 1997,
1998a,b).
Let us present a few representative results and point out the several algorithms which have been tested for the solution of the
problem. The solution of the mechanical problem has been done by means of the BE method, suitably modified to take into
account unilateral behavior of cracks (if required, see, among others, Antes and Panagiotopoulos, 1992; Stavroulakis et al., 1999).
A typical BE discretization for a plate with a crack with the simplest two-region BE method is shown in Figure 10. More compact
discretization techniques, including specialized crack elements, have also been used (see Alessandri and Mallardo, 1999; Rus and
Gallego, 2002; Stavroulakis et al., 2002), without any other change of the work.
Let us demonstrate the first difficulty with the error minimization approach to inverse analysis by considering the crack
identification problem with a horizontal crack of given size. The coordinates of the crack's center are the unknown variables. The
error function is shown in Figure 11. A suitable (logarithmic) scaling (see Figure 12), demonstrates that the minimization of this
function can be used for the inverse analysis. In fact, this scaling has been used with good results with both classical optimization
(Stavroulakis and Antes, 1998a,b; Engelhardt et al., 2006b) and neural networks (Stavroulakis and Antes, 1997).
Classical numerical optimization, as well as filter-driven optimization techniques (see the discussion of EKF algorithms given
elsewhere in this chapter), have been used for the determination of position, and size of one crack or flaw. Unilaterally working
cracks may cause problems to numerical optimization but can still be treated with filter techniques. The iterations of the various
algorithms for one example are shown in Figure 13. A change of the loading activates the unilateral contact mechanisms along the
Figure 12 Logarithmic scaling of the error function for the crack identification problem.
Author's personal copy
18 Inverse Analysis
Figure 13 Crack size determination with optimization and filter algorithm, for a classical and a unilateral crack.
crack and influences the solution of the inverse problem (for details, the reader is referred to Stavroulakis and Antes (2000), for
classical cracks using filter algorithms see also Tosaka et al. (1995)). Alessandri and Mallardo (1999) report good results for one
unilateral crack obtained with numerical optimization techniques. Numerical optimization has been used for the determination of
classical cracks in elastostatics or related fields (electro- or magnetostatics, for other applications of nondestructive evaluation
techniques) in several publications (see Bezerra and Saigal, 1993; Ingham and Wrobel, 1997; Larson et al., 1999; Rus and Gallego,
2002). Iterative algorithms based on an engineering implementation of gradient optimization techniques belong to this category
as well (see, e.g., Gallego and Suarez, 2000).
A very flexible tool for the solution of defect identification problems, including multiple defects, is the genetic optimization. In
statics, this has been demonstrated by the authors in Stavroulakis and Antes (1998a) and Engelhardt et al. (2006b) (see also
Burczynski and Beluch (2001), for thermoelastic applications and the references given later for dynamical problems). A typical set
of initial–final generations is shown in Figure 14, where the inverse problem concerns the identification of the position of a circular
flaw with a given radius. The iterations of the algorithm are shown in Figure 15, where the fitness functions of the best individual
and the mean value of the whole population are shown. The previously mentioned logarithmic scaling is used in this example as
well. Unfortunately GAs require a large number of function evaluations (i.e., solutions of classical mechanical problems);
therefore, they are time consuming and less suitable for most applications of practical interest. Nevertheless, some combination of
neural networks for the approximation of the modeling results with GAs for the inverse problems, in the spirit of the reliability
analysis discussed in the last set of examples presented in this chapter, may be a viable alternative (cf. Suh et al., 2002). Other
combinations, e.g., connection with fuzzy inference in Kojima et al. (2001), are also possible.
Author's personal copy
Inverse Analysis 19
NNs represent another possibility for the solution of crack identification problems. BPNNs, which have already been outlined
elsewhere in this chapter, have been used extensively by our research team. They approximate directly the inverse structural
mapping y-z, from measurements to unknown variables, using training which is based on a number of examples. The examples
(paradigms) are calculated by repeated application of every available modeling software for the solution of the classical, direct
mechanical problem. A schematic representation is shown in Figure 16.
The position of a classical crack with a given size could be identified, without problems, with this method. The introduction of
unilateral phenomena required the use of two different loading cases (Stavroulakis and Antes, 1997). A set of results concerning
the identification of cracks with various sizes, taken from Engelhardt et al. (2002), is given in Figure 17. Comparable results have
been reported in Xu et al. (2001) and Liang and Hwu (2001).
Figure 17 Finding the position of a crack. Training and testing based on data with several crack sizes.
measurements (waveforms at the accessible boundaries) include a lot of redundant or useless information. For example, the
measured response after the application of a dynamical loading and before any reflection from the sought defect does not carry
information about the defect. Moreover, reflection from known (e.g., other external) boundaries or given interfaces does not help
the solution of the inverse problem. On the contrary, they make the problem more difficult than, say, classical scattering analysis
in an infinite medium. If one decides to use blindly the error minimization formulation, one should keep in mind that every
modeling of a dynamical problem is time consuming. Therefore, methods which require less number of iterations (e.g., filter
algorithms) or a predictable amount of calculations (e.g., training of neural networks) are preferable than classical iterative
optimization techniques. The best solution is to use some characteristic features of the dynamical measurements (like the points of
first reflection of the wave) and, subsequently, use these features as measurements.
Let us first discuss the harmonic analysis problem. A time-harmonic excitation is used. The results are vibration amplitudes for
the given excitation frequency. Although the problem looks similar to the static one, the error function becomes bumpy, i.e., the
optimization problem is nonconvex. Therefore, if no information exists about the solution, an arbitrary choice of the starting point
for the numerical optimization cannot guarantee that the correct solution will be found. In Stavroulakis and Antes (1998a) flaw
identification problems have been solved by genetic optimization, while classical optimization could not give but local con-
vergence. NNs seem to be suitable solution methods as well (cf. Stavroulakis and Antes, 1998b). Analogous results have been
reported in Liu and Chen (2001) and Burczynski and Beluch (2001).
Time-domain analysis is a more general procedure. A simplest example of its application for inverse analysis is the so-called
impact-echo principle which is schematically demonstrated in Figure 18. For technological aspects and applications in civil
engineering the reader may consult the monograph by Sansalone and Streett (1997). Both classical and unilateral delaminations
and cracks have been studied numerically by the impact-echo analysis. The inversion has been done by means of neural networks
(see, e.g., Stavroulakis, 1999). A significant difference between classical and unilateral cracks has been observed. Let us consider an
elastic layer of given thickness and measure the normal displacement at a point of its free boundary in front of the delamination.
Figure 19 plots this waveform for several assumptions (a: no delamination, b: classical delamination, c, d: unilateral delamination
without and with friction, respectively). The identification of the unilateral crack is, in every case, much more difficult. Extended
testing of the method to more complicated 2D problems is an ongoing (early 2000s) research topic of our group. The first results
are in the same direction. Encouraging results in dynamic crack identification have been presented by several authors (see also the
solution of inverse crack identification problems using filter algorithms in Engelhardt et al. (2006a)). Numerical optimization has
been used in Tanaka et al. (1990), Nishimura (1993), and Rus and Gallego (2002); neural networks have been used in Oishi et al.
(2001), Ishak et al. (2002); and GAs in Nag et al. (2002) and Wu et al. (2002).
Author's personal copy
Inverse Analysis 21
The uncertainties implied by measurements and mathematical models are naturally reflected by uncertainties in the parameter
estimates resulting from the identification process. The statistical characterization of the inverse problem solutions can be naturally
pursued in a Kalman filter approach, which attributes a normal (Gaussian) probability density, as a simplifying idealization, to the
computed sequence of estimates.
Consistently with the aim of the present work, some applications of these techniques will be illustrated in the following with
specific reference to failure analyses, involving damage and fracture.
Figure 20 Schematic representation of the three-point bending test and of the wedge-splitting test.
Author's personal copy
Inverse Analysis 23
Figure 21 The reaction force Q versus the imposed displacement q and versus the surface displacements u: results of laboratory test performed
on an epoxy foam (Bolzon et al., 1997).
Progressive pure mode I fracture, without local unloading, results from such experimental setup; this circumstance can be
exploited to reduce the computing cost of the simulations since a holonomic (pseudo-reversible) description of the separation
process can be given, in the spirit of the so-called ‘deformation theory’ of plasticity.
Savings result also from the usually assumed linear elastic behavior of the bulk material outside the a priori known, from
symmetry consideration, separation surface G (see Figure 20). In this case, superposition can be used to relate the distribution of
traction t across G to the external action and to the distribution of the displacement discontinuity (opening displacement) w, as
follows:
Z
tðξÞ ¼ at E ðξÞ þ Zðξ; ZÞwðZÞ dZ; ξ; ZA G ½48
G
where tE(ξ) represents the specimen response to an external action (e.g., an imposed displacement q) of unit amplitude, in the
absence of displacement discontinuities; a is the corresponding, monotonically increasing amplification factor; Z(ξ, Z) denotes the
(Green) influence function which reflects the geometric and elastic properties of the considered specimen and represents the effect,
on t(ξ), of the displacement discontinuity w(Z), considered as an imposed distortion in the otherwise unloaded body.
Tractions are further related to the corresponding opening displacements by a cohesive law (see, e.g., Figure 22) which can be
given by the following holonomic (reversible) analytical description, in the format of a complementarity problem (CP):
where g is a given function, such that g(0) ¼ tc and g(wc) ¼0, where tc and wc represent the material tensile strength and the critical
opening displacement, respectively, while l(ξ) denotes an auxiliary variable field.
Relations [49] describe, at the same time, all the following possible conditions of a material point along ɣ: the still undamaged
state (w ¼ 0; totc); the softening phase (wZ0; t ¼ f(w)); and the complete separation (w4wc; t ¼ 0).
Nonsmooth piecewise-linear softening laws with break-points, like the widely accepted model for mode I fracture in concrete
(CEB-FIP-Model Code, 1993; RILEM Report 5, 1991) represented in Figure 22, can also be described in complementarity format. In
Author's personal copy
24 Inverse Analysis
Figure 22 Nonlinear and piecewise linear interface law in mode I quasi-brittle fracture.
Z
Q ¼ aQE þ RðZÞwðZÞdZ; ZA G ½51
G
This direct problem can be given a formally identical discretized version, consistent with the usual FE or BE modeling (see, e.g.,
Bolzon et al., 1997, 2002a; Tin-Loi and Que, 2001), in terms of nodal displacements and corresponding traction forces, where
Author's personal copy
Inverse Analysis 25
integral operators are simply replaced by algebraic matrices. The corresponding solution for the given amplifier a can be obtained
by single-step numerical procedures, which do not require to recover the actual equilibrium path since local unloading has been
ruled out. The economical gain of this holonomic approach is further emphasized by the reduced number of unknowns entering
the formulation (namely, only variables pertaining to the interface G are involved).
The inverse problem of identifying the unknown fracture parameters contained in the interface law on the basis of laboratory
results can then be cast in the format of a minimization problem under complementarity constraints, within a deterministic or
stochastic context, as introduced in Section 3. Vector z will then collect fracture parameters such as tc, wc, and so on, while the
measurements u and Q will be gathered by vectors yk for each suitable situation k defined by a corresponding a input.
An important part of the computing effort related to the inverse problem in point is represented by the recurrent evaluation of
the sensitivity matrix (Sk) which quantifies the influence of the parameters z on the computed measurable quantities in situation k.
This is a key point of the EKF methodology and of gradient-based solution algorithms for least-squares approaches, whose
computation may become prohibitive as the number of parameters exceeds a few units if finite difference techniques or inter-
polations in the space of the parameters are adopted in order to numerically compute derivatives, as in some of the applications to
be presented in the next section.
Notably, in the present formulation, matrix Sk can give a closed form expression, with significant computational gains, by
exploiting the peculiarities of the forward operator in a way basically similar to that adopted by Bittanti et al. (1985) for elastic–
plastic frames. The nonsmoothness issue deriving from the CP format [49] (earlier noticed in Bittanti et al., 1985) can be
practically circumvented by the selection of the ‘active modes,’ depending on the sign of the activation functions ji and on the
values of the complementary variables w, l at the previous solution step of the iterative procedure converging toward the actual
parameter values. Details can be found in Bolzon et al. (2002a).
The benchmark problems solved show that both sequential and batch procedures are effective for the present identification
purpose (Bolzon et al., 1997, 2002a; Tin-Loi and Que, 2001).
The sequential technique typical of the Kalman filter methodology considers all the available experimental data with the
corresponding measurement noise, but only the information relevant to the current value of the loading factor a is processed at
each computing step, with the advantage of keeping the number of unknowns always small. Possibly linked to the performed
sequence of updating operations, based on step-by-step linearization, a dependence on the initial guess of the final estimates of the
unknown parameters can be observed. This drawback is overcome by the so-called global-iteration step, which consists of repeated
filtering of the same information flow, initialized by the former final estimate of the sought parameters.
In a batch procedure, the number of unknowns is related to that of considered loading situations, which should be carefully
selected to maintain the procedure computationally competitive. Scatter of the identified parameter values is then expected,
depending on the performed choice but devoid of any statistical interpretation. A more severe limitation is linked to the possibility
of detecting local minimum points of the (usually nonconvex) discrepancy function. This occurrence can be singled out by
unexpected high optimal value of the function, and the corresponding parameter estimate can be (hopefully) corrected by a
different initialization of the minimization algorithm.
Extension of these approaches to the identification of parameters governing debonding and delamination of composites, as
considered, e.g., by Corigliano and Mariani (2001), or mixed-mode fracture of jointed (discontinuous) rock masses is at hand.
Geotechnical applications are particularly intriguing since parameters are to be evaluated from in situ measurements, by means of
field instrumentation during the construction stage or the operation time, e.g., in gas/oil exploitation activities (e.g., Gioda and
Maier, 1980; Cividini et al., 1983; Lee and Kim, 1999; Simoni and Schrefler, 2001; Xiang et al., 2002). In fact, the real material state
would be perturbed by the extraction of specimens, and could not be reproduced in the laboratory.
The main difficulty in the application of such inverse procedures to mixed-mode fracture of homogeneous materials is related
to the design of meaningful laboratory experiments (Guo et al., 1994). In fact, as clearly observed in concrete fracture when the
specimen size is large enough with respect to the aggregate dimensions, each elementary crack-growth step is essentially produced
by an opening mechanism along a curvilinear trajectory, and dissipation due to grain friction and interlock is negligible compared
with that associated to material separation.
Figure 23 Plate specimens with a central hole or with double-side round notches under tension used in material identification experiments.
Mahnken (1999) has addressed the problem by specializing to this context the FE-oriented identification technique which
exploits the algorithmic tangent matrix, proposed by Mahnken and Stein (1996) within a least-squares minimization approach.
The quite general underlying methodology has been applied also to different related fields, such as gradient enhanced damage
mechanics (Mahnken and Khul, 1999), thermoelastic damage (Mahnken, 2000), and fluid saturated porous media (Mahnken and
Steinmann, 2001).
Alternative mathematical model and identification procedure for elastoplastic solids has been formulated as an optimi-
zation problem by Yoshida et al. (1998), using noiseless computational information from bending test as main source of
information. The difficulties related to the complex nature of the experimental tests to be simulated have been partially cir-
cumvented by introducing an approximate response function, defined in the space of the sought parameters and obtained through
Lagrangian interpolation of a few direct (forward) analyses carried out for fixed parameter values, chosen in a reasonable (in a
Bayesian sense) rectangular domain. The main drawback of this approach is that computing costs increase exponentially with the
number of the unknown parameters and with the degree of the interpolation which has a strong influence on the accuracy of the
estimation.
The same strategy has been used by Aoki et al. (1997) and by Corigliano et al. (2000), in conjunction with Kalman filter
methodology, to identify the parameters entering the generalized Gurson model. The experimental information processed by the
filter was obtained from plate specimens with a central hole or with double-side round notches under tension (see Figure 23) and
from single notched specimens under three-point bending. Results pointed out the central role in this context of the proper
experiment design and of the ‘expert’ initialization of the identification procedure.
Difficulties in converging toward the optimal solution due to noisy data and inaccurate model equations suggested the use of
robust optimization techniques based on (zeroth order) evolutionary algorithms (GA, NN), even for parameter identification
procedures based on the direct knowledge of stress–strain responses (Furukawa and Yagawa, 1997, 1998).
Significant computational gains can be achieved by the implementation of the model reduction techniques introduced in
Section 3.6. The effectiveness of this approach has been tested by Bolzon and Talassi (2013) on the identification problem of seven
independent parameters entering the constitutive model of anisotropic metals, on the basis of experimental information collected
from an indentation test schematized in Figure 24.
The mechanical response of the considered material class is linear elastic, represented by the relationships:
1 1 1
exy ¼ sxy ; eyx ¼ syz ; ezx ¼ szx
2Gxy 2Gyz 2Gzx
!
1 1 1 1 1
kx ¼ þ ; kxy ¼ ; and so on
2 R2y R2z R2x R2xy
Figure 25 POD-RBF training (dots) and verification (stars) yield limit combinations.
Figure 26 Profiles of the residual imprint obtained by FE simulation with input values Ex ¼ 205 GPa, Ey ¼ 199 GPa, Gxy ¼96.4 GPa, sYx ¼ 556 MPa,
sYy ¼ 382 MPa, sYxy ¼ 1200 MPa, n¼0.049 and the corresponding reconstruction based on the parameter values identified by means of the POD-
RBF approximation.
considered case, in fact, the average number of iterations was 15, 8 computations were required at each iterations to evaluate
the derivatives by a finite difference scheme, and the optimization algorithm was initialized by 10 different trial vector in the
parameter space, in view of the possible convergence toward local minimum points and of the possible existence of multiple
solutions. It is also worth observing that the cubic Lagrange interpolation of this system response would require more than 16 000
preliminary analyses.
numerically inefficient for large-scale problems because of high computational time. That is why neural networks are suggested
for generating samples for the CMC simulation (Papadrakakis et al., 1996; Hurtado, 2001). This idea was developed by Kaliszuk
and Waszczyszyn (2003) and examined on an example of a plane elastoplastic frame with random yield points of frame
members.
The probability of failure is defined by the following relationship:
Z
Pf ¼ ProbfGðXÞ R Sr0g ¼ f ðXÞ dX ½52
GðXÞr0
where pf is the probability of failure, R is the resistance of structure, S is the actions (loads) applied to structure, and X¼ {XR, XS} is
the vector of random variables. The complement of pf is the probability of reliability pr ¼ 1–pf that the structure will not fail.
The Monte Carlo simulation corresponds to computation of integer eqn [52]. Following the law of large numbers the CMC
estimator of the probability of failure is
1X N
pf ¼ IðX i Þ
Ni¼1
( ½53
1 for GðX i Þr0
IðX i Þ ¼
0 for GðX i Þ40
A two-bay four-storey frame (Pulido et al., 1992) was used as an example to present the numerical efficiency of the discussed
approach. The frame data, shown in Figure 27, was analyzed by the FE program ELPLAF-v1 (Waszczyszyn and Pabisek, 2002). A
perfect elastoplastic material, distributed yielding zones, and the second-order geometrically nonlinear theory were assumed. The
ultimate load corresponds to the limit load of the frame global buckling, Pult ¼ PLbuckling.
The computations were carried out for two inputs: X1 ¼ Rc and X2 ¼ Rb, where Rc, Rb are yield points of the columns and beams,
respectively. The normal probability function for both yield points was assumed with the mean value m ¼240 MPa and standard
deviation s¼ 24 MPa. The values of the frame ultimate load were computed as the neural network output yR ¼Pult(Rc, Rb). The
ultimate load for mean yield points was Ro ¼ Pult(m, m) ¼ 1042.3 kN and then the patterns were computed for Rc, RbA[m–3s,
m þ 3s], corresponding to regular mesh of 16 16 points which gave L¼256 patterns for the network training. T ¼100 testing
inputs were randomly selected points from the above defined input domain. A very simple network BPNN þ M: 2 3 1 was
trained by the BP þ momentum learning method and then tested using the MATLAB Neural Network Toolbox (2000). The
following errors are listed:
1Xv
avr ep ¼ ep; max ep ¼ max fepg
Vp¼1 p
for ep ¼ 1 yðpÞ =t ðpÞ j100% ½54
where V¼L, T is the number of patterns for the network training and testing, respectively and t(p), y(p) are output values for known
and neurally computed outputs for the pth pattern. In Table 1 besides the relative errors the standard error St e ¼RMSE the
regression coefficient r was also given, computed for the full set of pairs {(t,y)(p)}, where: p ¼1, …, L þ T.
Three cases of Monte Carlo simulation were analyzed. They are put together in Table 2. In case 1 two random variables were
considered, i.e., X1 ¼ Rc, X2 ¼ Rb. In cases 2 and 3 only one random variable was assumed, i.e., either Rc or Rb are taken into account
as a random variable.
The probability of the frame reliability pr is shown in Figure 28 as a function of the load factor P for the three considered cases. It
is visible that cases 1 and 2 coincide. Another situation takes place in case 3 in which the yield point of beams is fixed, i.e.,
Rc ¼240 MPa. That is why at Pult ¼ 1042 kN, the probability of reliability falls drops to zero, i.e., pr(P ¼ 1042)¼0.
Figure 28 Reliability curve pr(P) of the elastoplastic frame for three cases of the random variable selection.
The reliability curve pr(P) enables us to estimate the value of the load factor P at the assumed probability of the frame
reliability. For instance, in Figure 28 it is shown that for pr ¼ 0.9 the corresponding load equals P¼927 kN for cases 1 and 2, and
P¼ 1023 kN for case 3 of random variable selection, respectively.
The computations were performed on PC AMD A1000 under Windows 98. During the computations, the CPU time was
evaluated. It was stated that the time of computation of a single pattern by the FE program ELPLAF-v1 (Waszczyszyn and Pabisek,
2002) was in the range 10.2–83.2 s, average B32 s (2077 s for generating 67 patterns). The projection of this time for generating
1000 samples for the Monte Carlo simulation gives 3.2 104 s. The CPU time needed for evaluation of one point at the
probability curve (Figure 28), using the neural network for generating 1.0 105 samples, was 1.2 s. This means that the neural
network needs the CPU time of order 10–4 lower than using the FE program for generating 1000 samples per one value of pf.
Table 3 Relative errors and statistical coefficients for the fundamental period of vibration of five-storey buildings
Errors and statistical coefficients BPNN 3–4–1 RBFNN 3–22–1 ANFIS 3–8–1 Empirical formulas
[55] [56]
uniform vertical deflection of soil basement Cz. That is why the following empirical formula was proposed by Ciesielski et al.
(1992):
0:98
T1 ¼ pffiffiffi ½55
3Cz ½s
1:2
T1 ¼ p ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ½56
3
0:003ðkb þ ks Þ=b þ Cz ½s
where kb ¼SiEIi/a, ks ¼ SiGAi/a are equivalent bending and shear stiffness of walls in the segment plan (cf. Figure 29(a)) where a, b
are length and width of the segment plan, and E, G are elasticity moduli.
The neural identification problem was analyzed with respect to 13 types of five-storey buildings (Waszczyszyn et al., 1999;
Kuzniar et al., 2000). On the base of vibration records, the experimental values of the basic period of vibrations Texp(s) were
computed. The set of P¼31 patterns was split into L¼22 training patterns and T ¼9 testing patterns.
The BP network of structure 3–4–1 was used, corresponding to the following input and output vectors:
The network was trained by the computer simulator SNNS (1998) and the Rprop learning rule [39]. After S ¼1000 epochs
the network errors were MSEL, MSETo3.0 104. In Table 3 the relative errors average and maximal are put together,
ðpÞ ðpÞ
according to formulas [54] assuming t ðpÞ ¼ Texp ; yðpÞ ¼ Tcom for the pth pattern. In the table, the relative errors and statistical
coefficients corresponding to neural prediction of a simple regularization neural network RBFNN (Waszczyszyn et al., 1999)
are given.
The most successful was the application of the ANFIS neurofuzzy network discussed in Section 3.5. Using Fuzzy Logic Toolbox
(2001) and scatter partition of the input data (cf. Figure 7(b)), eight membership Gaussian functions were automatically adopted
for each input variable in eqn [57] (cf. Waszczyszyn and Slonski (2000)).
The comparison of results points out that the neural predictions are much better than those by empirical formulas
[55] and [56] (cf. Table 3). The ANFIS estimation is superior to the estimation by the neural networks BPNN
and RNN.
The soil–structure interaction coefficient Cz is of fuzzy character since it is difficult to evaluate its crisp value. That is why this
coefficient can be treated as a linguistic variable. Following the approach suggested by Juang et al. (1999) instead of one input
Author's personal copy
32 Inverse Analysis
variable Cz, three values were introduced (Kuzniar and Waszczyszyn, 2002), associated with values of corresponding triangular
membership functions:
Figure 30 shows an example of variable [58] as CL ¼ {0.5764, 0.4231, 0.0} corresponding to Cz ¼ 110 MPa m–1.
The problem was reformulated taking into account only two control variables, i.e., coefficient Cz and the building segment
width b. Instead of eqn [57] the following input variables were introduced:
x21 ¼ fCz ; bg-x41 ¼ mS; mM ; mL ; b ½59
Variables [59] were used in two BP neural networks: (1) BPNN(c): 2–3–1 with crisp inputs [59], (2) BPNN(l): 4–2–1 with the
linguistic variable [58].
The simulator SNNS and Rprop learning rule were used for the training and testing of the discussed networks. In Figure 31(a)
the training processes with errors MSEL are shown for both networks. It is evident that network 2 with the linguistic variable CL is
learnt much more efficiently than the network with the crisp variable Cz.
In Table 4 the results of neural simulation are put together for the networks BPNN(c) and BPNN(l). It is obvious that the
introduction of the linguistic variable [58] leads to a network superior to the network with the corresponding crisp input variable.
This conclusion is valid also with respect to success rate SR shown in Figure 31(b). Success rate is related to the empirical
cumulative curve
PC
SR ¼ 100% ½60
P
where PC is the number of patterns predicted correctly by the neural networks for the relative error cone RE (%) corresponding to
epA[0, RE], and P is total number of patterns.
Table 4 Training and testing errors for networks BPNN (c) and BPNN (l)
Networks Number of training epochs MSEL ( 104) MSET ( 104) avr ep (%)
Figure 31 (a) Training errors MSEL (s) for the networks BPNN (c) and BPNN (l), and (b) success rate SR in RE cones.
From among many implicit models, the Chaboche model is commonly accepted. In what follows only the uniaxial model is
considered as a special case of the Chaboche model (Chaboche and Rousslier, 1983). The analytical form of this model is related
to the flow rule, kinematic, and isotropic hardening:
js χj R k n
e_ vp ¼ sgnðs χ Þ ½61
K
χ_ ¼ H_evp Dχj_evp j
R _¼ ðh ddRÞj_evp j ½62
where: evp ¼e ee are viscoplastic, total, and elastic strains, respectively, s is total stress, χ is back stress related to kinematic
hardening, R is the drag stress related to isotropic hardening, and k, K, n, H, D, h, d are seven material parameters.
The implicit neural model is related to patterns with the following input and output vectors:
which can be computed for the subsequent time instants t¼ p, associated with a time increment Dt : yt þ 1 ¼ yt þ y_ tDt. Initial values
vp
are e0 ¼e (t¼ 0), s0 ¼ Eðe0 e0 Þ, and other values are derived from experimental data using the algorithm given by Furukawa and
Yagawa (1998).
Below only a simple problem, related to monotonic uniaxial tension of steel specimens, is discussed. The specimens were
tensioned with three constant strain rates, i.e., e_ ¼0.0001, 0.01, 0.5%. After a number of numerical experiments, a BPNN of
structure 4–6–6–3 was formulated. All neurons had the same binary activation functions with s¼ 1 in formula [33]. The SNNS
simulator (1998) was used for the training by means of Rprop learning rule [38]. Figure 32 shows neural stress–strain curves on the
plane (s, e) after 2000 epochs. The patterns were completed to learn only one curve for a fixed value of strain rate e_ ¼ const.
In a paper by Furukawa and Yagawa (1998), no testing and generalization evaluation was made since the main goal was to
show that neural simulation is much better than the results shown in Figure 32 for the best fitting of Chaboche's model parameters.
Another problem is related to the prediction of concrete fatigue durability. An extensive experimental evidence, corresponding
to more than 400 tests performed in 14 laboratories, was compiled by Furtak in 1984 (cf. references in Waszczyszyn, 2000;
Kaliszuk et al., 2001).
The concrete specimens were subjected to cycles of compressive loadings and fatigue durability was defined as the limit number
of cycles N which caused the specimen fatigue damage. The problem of fatigue durability prediction was formulated as the
following mapping:
where χ ¼fcN/fc is the ratio of compressive fatigue strength of concrete and strength fc, R ¼smin/smax is the ratio of minimal and
maximal strengths in compressive cycle of loading, and f(Hz) is the frequency of cycle loading.
An empirical formula was derived as the following implicit relation between variables [65]:
1 1:16Cf
logN ¼ log þ logð1 þ BRlogN ½65
A χ
where A¼0.008–0.118 log (sI/fc), Cf ¼ 1 þ 0.07(1–R) log f, and B¼0.118(sII/sI–1), sI and sII are critical strengths.
Author's personal copy
34 Inverse Analysis
P¼218 patterns was selected from the experimental tests in order to complete a consistent set which was randomly split into a
training set of L¼ 118 patterns and a testing set of T¼ 100 patterns.
Table 5 shows results of the neural simulation by different networks. The most accurate is the approximation by RBFNN
(Waszczyszyn, 2000), then by ANFIS (computations by Slonski in 2001). A comparable accuracy was also achieved by BPNN
(Kaliszuk et al., 2001). The neural approximation is significantly better than the prediction by the empirical formula [66].
Figure 33 Frame models: (a) laboratory model, (b) FEM model FRAME1, and (c) FEM model FRAME2.
o1 o2 o3 o4
The vibrations of the laboratory model were excited by an impulse load. The responses were measured by the accelerometer
applied at different points shown in Figures 33(b) and 33(c). Of course, the masses of gauges were taken into account in the FE
analysis.
Two BPNNs; 4–10–2 were used. The input vector corresponds to four eigenfrequencies; two outputs are related to identified
parameters of the models FRAME1 and FRAME2, respectively:
The sigmoidal hidden neurons were used and the Lavenberg–Marquard learning method was explored (cf. Neural Network
Toolbox, 2000). Using an FE program P ¼ 11 560 patterns were computed from which 50% of patterns were randomly selected as
a training set, and the remaining patterns were used for testing of the trained networks.
Table 6 shows the measured eigenfrequencies. After their introduction as the inputs of the trained networks, the identified
values of structural parameters were computed and used in the FE analysis in order to compute eigenfrequencies of the updated
mathematical models (cf. Miller and Ziemianski, 2001).
It is obvious that the results of parameter identification in the model FRAME2 (maximal absolute error of 4.2% for frequencies
o4) are slightly better than in FRAME1 (error 8.6%).
Figure 34 (a) A rod with notch and triangle microimpulse and (b) propagation of wave velocity at a selected measurement point.
Figure 35 (a) Replicator as an auto-associative BPNN, (b) and (c) splitting of trained replicator into compressor and decompressor.
Figure 36 BPNNs for notch parameters identification: (a) standard BPNNs; (b), (c), and (d) cascade networks BPNNc for c ¼I, II, III.
The replicator is an auto-associative neural network, i.e., the mapping x-x is performed for xARN. In the considered case, the
BPNN of structure N–n–N is formulated. After the training the replicator can be split into the compressor and decompressor (cf.
Figures 35(b) and 35(c)) where noN corresponds to the number of sigmoidal neurons.
The data compression by replicators was successfully used in the analysis of buckling loads of cylindrical shells with manu-
factured imperfections (Waszczyszyn and Bartczak, 2002). Both the data compression and decompression were performed in the
analysis of vibration records (Ghaboussi and Chu-Chin Lin, 1998) and in the simulation of building displacements caused by
paraseismic excitations (Kuzniar and Waszczyszyn, 2001). In the considered problem N ¼200 discrete values, taken from the
velocity record, were compressed into n¼ 12 values at each measurement point. In this way 4 12 ¼ 48 inputs, corresponding to
four measurement points, were formulated for the network to identify three damage parameters related to the notch, i.e., location
l, depth h, and width b.
Two types of BP neural networks were used. Besides standard BPNNs: 48–12–8–3, three cascade BPNNc were formulated for
c ¼I, II, III (cf. Figure 36). The training and testing of i networks shown in Figure 34 were performed by 800 patterns selected from
over 2500 patterns computed by ADINA (1999). The SNNS (1998) computer simulator and Rprop learning rule were used to train
and test both the replicator and BPNNs.
Thanks to the application of BPNNc, the error cone RE (%) of the neural approximation for all the tested patterns was reduced
to 2%, 3%, and 5% for the prediction of depth h, location l, and width b, respectively (Ziemianski and Piatkowski, 2000).
The approach discussed above was also applied to the identification of stiffness changes in rods with failures associated with
jump type changes of Young's modulus E or yield point s0 in a narrow zone of the rod (Piatkowski and Ziemianski, 2001).
Author's personal copy
Inverse Analysis 37
References
Adeli, H., Park, H.S., 1998. Neurocomputing for Design Automation. Boca Raton, FL: CRC Press.
ADINA System Online Manuals, 1999. Theory and Modeling Guide. Watertown: ADINA R&D.
Ageno, M., Bolzon, G., Maier, G., 2009. An inverse analysis procedure for the material parameter identification of elastic−plastic free-standing foils. Struct. Multidisc. Optim.
38, 229–243.
Alessandri, C., Mallardo, V., 1999. Crack identification in two-dimensional unilateral contact mechanics with the boundary element method. Comp. Mech. 24 (2), 100–109.
Antes, H., Panagiotopoulos, P.D., 1992. The Boundary Integral Approach to Static and Dynamic Contact Problems. Equality and Inequality Problems. Basel, Boston, Berlin:
Birkhäuser.
Aoki, S., Amaya, K., Sahashi, M., Nakamura, T., 1997. Identification of Gurson's material constants by using Kalman filter. Comp. Mech. 19, 501–506.
Arora, J.S., Elwakeil, O.A., Chachande, A.I., Hsieh, C.C., 1995. Global optimization methods for engineering applications: A review. Struct. Optim. 9, 137–159.
Bazant, Z.P., 2002. Concrete fracture models: Testing and practice. Eng. Frac. Mech. 69, 165–205.
Bezerra, L.M., Saigal, S., 1993. A boundary element formulation for the inverse elastostatics problem (iesp) of flaw detection. Int. J. Num. Methods Eng. 36, 2189–2202.
Bittanti, S., Maier, G., Nappi, A., 1985. Inverse problems in structural elasto-plasticity: A Kalman filter approach. In: Sawczuk, A., Bianchi, G. (Eds.), Plasticity Today:
Modelling, Methods and Applications. London: Elsevier Applied Science, pp. 311–329.
Bolzon, G., 2014. Advances in experimental mechanics by the synergetic combination of full-field measurement techniques and computational tools. Measurement 54, 159–165.
Bolzon, G., Buljak, V., 2011. An effective computational tool for parametric studies and identification problems in materials mechanics. Comput. Mech. 48, 675–687.
Bolzon, G., Cocchetti, G., 2003. Direct assessment of structural resistance against hydraulic fracture. Int. J. Num. Anal. Methods Geomech. 27, 353–378.
Bolzon, G., Fedele, R., Maier, G., 2002a. Parameter identification of a cohesive crack model by Kalman filter. Comp. Methods Appl. Mech. Eng. 191, 2847–2871.
Bolzon, G., Gabetta, G., Molinas, B., 2015. Integrity assessment of pipeline systems by an enhanced indentation technique. ASCE J. Pipeline Sys. Eng Pract. 6 (1), 04014010.
Bolzon, G., Ghilotti, D., Maier, G., 1997. Parameter identification of the cohesive crack model. In: Sol, H., Oomens, C.W.J. (Eds.), Material Identification Using Mixed Numerical
Experimental Methods. Dordrecht: Kluwer Academic, pp. 213–222.
Bolzon, G., Ghilotti, D., Maier, G., 2002b. Strength of periodic elastic-brittle composites evaluated through homogenization and parameter identification. Euro. J. Mech. A/Solids
21, 355–378.
Bolzon, G., Maier, G., Panico, M., 2004. Material model calibration by indentation, imprint mapping and inverse analysis. Inter. J. Solids Struct. 41, 2957–2975.
Bolzon, G., Molinas, B., Talassi, M., 2012. Mechanical characterisation of metals by indentation tests: An experimental verification study for on-site applications. Strain 48,
517–527.
Bolzon, G., Talassi, M., 2012. Model reduction techniques in computational materials mechanics. In: Zavarise, G., Boso, D.P. (Eds.), Bytes and Science. Barcelona: CIMNE,
pp. 131–141.
Bolzon, G., Talassi, M., 2013. An effective inverse analysis tool for parameter identification of anisotropic material models. Int. J. Mech. Sci. 77, 130–144.
Brühwiler, E., Wittmann, F.H., 1990. The wedge splitting test, a new method of performing stable fracture mechanics tests. Eng. Fract. Mech. 35, 117–125.
Buhmann, M.D., 2003. Radial Basis Functions: Theory and Implementations. Cambridge, UK: Cambridge University Press.
Bui, H.D., 1994. Inverse Problems in the Mechanics of Materials: An Introduction. Boca Raton, FL: CRC Press.
Burczynski, T., Beluch, W., 2001. The identification of cracks using boundary elements and evolutionary algorithms. Eng. Anal. Boundary Elem. 25 (4−5), 313–322.
Carr, J.C., Beatson, R.K., Cherrie, J.B., et al., 2001. Reconstruction and representation of 3D objects with radial basis functions. In: Fiume, E. (Ed.), Computer Graphics
Proceedings (Proceedings of SIGGRAPH), Annual Conference Series. New York: ACM Press, pp. 67–76.
Catlin, D.E., 1989. Estimation, Control and the Discrete Kalman Filter. New York: Springer. (Applied Mathematical Science, 71).
CEB-FIP, 1993. Model Code 1990: Design Code. London: Thomas Telford.
Chaboche, J.L., Rousslier, G., 1983. On the plastic and viscoplastic constitutive equations: P. II. Trans. ASME J. Pres. Ves. Technol. 105, 159–164.
Ciesielski, R., Kuzniar, K., Maciag, E., Tatara, T., 1985. Empirical formulae for fundamental natural periods of buildings with load bearing walls. Arch. Civil Eng. 38, 291–299.
Ciesielski, R., Kuzniar, K., Maciag, E., Tatara, T., 1992. Damping of vibration in precast buildings with bearing concrete walls. Arch. Civil Eng. 41, 329–341.
Cividini, A., Maier, G., Nappi, A., 1983. Parameter estimation of a static geotechnical model using a Bayes’ approach. Int. J. Rock Mech. Mining Sci. 20 (5), 215–226.
Cocchetti, G., Maier, G., Shen, X.P., 2002. Piecewise linear models for interfaces and mixed mode cohesive cracks. Comp. Model. Eng. Sci. 3, 279–298.
Corigliano, A., Mariani, S., 2001. Parameter identification of a time-dependent elastic-damage interface model for the simulation of debonding in composites. Compos. Sci.
Technol. 61, 191–203.
Corigliano, A., Mariani, S., Orsatti, B., 2000. Identification of Gurson−Tvergaard material model parameters via Kalman filtering technique I. Theor. Int. J. Fract. 104, 349–373.
Courage, W.M.G., Schreurs, P.J.G., Janssen, J.D., 1990. Estimation of mechanical parameter values of composites with the use of finite element and system identification
techniques. Comp. Struct. 34, 231–237.
Davis, L.D., 1991. Handbook of Genetic Algorithms. New York: Van Nostrand Reinhold.
Denarié, E., Saouma, V., Iocco, A., Varelas, D., 2001. Concrete fracture process zone characterization by fiber optics. ASCE J. Eng. Mech. 127, 494–502.
Elices, M., Guinea, G.V., Gòmez, J., Planas, J., 2002. The cohesive zone model: Advantages, limitations and challenges. Eng. Fract. Mech. 69, 137–163.
Elices, M., Guinea, G.V., Planas, J., 1992. Measurement of the fracture energy using three-point bend tests: Part 3. Influence of cutting the P-d tail. Mater. Struct. 25,
327–334.
Engelhardt, M., Likas, A., Stavroulakis, G.E., 2002. Neural crack identification. In: Mang, H.A., Rammersorfer, F.G., Eberhardsteiner, J. (Eds.), Proceeding of 5th World Congress
on Computational Mechanics WCCM V, Vienna, Austria, July 7−12, 2002. Austria: Vienna University of Technology. ISBN 3-9501554-0-6, Available at: http://ccm.tuwien.
ac.at
Engelhardt, M., Antes, H., Stavroulakis, G.E., 2006a. Crack and flaw identification in elastodynamics using Kalman filter techniques. Comput. Mech. 37, 249–265.
Engelhardt, M., Schanz, M., Stavroulakis, G.E., Antes, H., 2006b. Defect identification in 3-D elastostatics using a genetic algorithm. Optim. Eng. 7, 63–79.
Efstathiades, Ch., Baniotopoulos, C.C., Nazarko, P., Ziemianski, L., Stavroulakis, G.E., 2007. Application of neural networks for the structural health monitoring in curtain-wall
systems. Eng. Struct. 29, 3475–3484.
Friswell, M.I., Motterhead, J.E., 1996. Finite Element Model Updating in Structural Dynamics. Dordrecht: Kluwer Academic Publisher.
Furukawa, T., Yagawa, G., 1997. Inelastic constitutive parameter identification using an evolutionary algorithm with continuous individuals. Int. J. Num. Methods Eng. 40,
1071–1090.
Furukawa, T., Yagawa, G., 1998. Implicit constitutive modelling for viscoplasticity using neural networks. Int. J. Num. Methods Eng. 43, 195–219.
Furukawa, T., Sugata, T., Yoshimura, S., Hoffman, M., 2002. An automated system for simulation and parameter identification of inelastic constitutive models. Comp. Methods
Appl. Mech. Eng. 191, 2235–2260.
Fuzzy Logic Toolbox for Use with MATLAB, 2001. User's Guide Version 2. Natick, MA: The Math Works.
Gallego, R., Suarez, J., 2000. Solution of inverse problems by boundary integral equations without residual minimization. Int. J. Solids Struct. 37 (40), 5029--5052.
Ghaboussi, J., Lin, C.-C.J., 1998. New methods of generating spectrum compatible accelerograms using neural networks. Earth. Eng. Strut. Dyn. 27, 377–396.
Gioda, G., Maier, G., 1980. Direct solution of an inverse problem in elasto-plasticity: Identification of cohesion, friction angle and in situ stress by pressure tunnel tests. Int. J.
Num. Methods Eng. 15, 1823–1848.
Goldberg, D.E., 1989. Genetic algorithms in search, optimization and machine learning. Reading, MA: Addison-Wesley.
Author's personal copy
38 Inverse Analysis
Gosh, R., Joshi, Y., 2013. Error estimation in POD-based dynamic reduced order thermal modeling of data centers. Int. J. Heat Mass Transfer 57, 698–707.
Guo, Z.K., Kobayashi, A.S., Hawkins, N.M., 1994. Mixed modes I and II concrete fracture: An experimental analysis. ASME J. Appl. Mech. 61, 815–821.
Gurson, A.L., 1977. Continuum theory of ductile rupture by void nucleation: Part I. Yield criteria and flow rules for porous ductile media. ASME J. Eng. Mater. Technol. 99,
2–15.
Haykin, S., 1999. Neural Networks—A Comprehensive Foundation. Upper Saddle River, NJ: Prentice-Hall.
Hillerborg, A., 1985. Results of three comparative test series for determining the fracture energy Gf of concrete. Mater. Struct. 18, 407–413.
Hillerborg, A., 1991. Application of the fictitious crack model to different types of materials. Int. J. Fract. 51, 95–102.
Horii, H., Ichinomiya, T., 1991. Observation of fracture process zone by laser speckle technique and governing mechanism in fracture of concrete. Int. J. Fract. 51, 19–29.
Hurtado, J.E., 2001. Neural networks in stochastic mechanics. Arch. Comp. Methods Eng. 8, 303–342.
Ingham, D.B., Wrobel, L.C., 1997. Boundary Integral Formulations for Inverse Analysis. Southampton: Computational Mechanics Publications.
Ishak, S.I., Liu, G.R., Shang, H.M., Lim, S.P., 2002. Non-destructive evaluation of horizontal crack detection in beams using transverse impact. J. Sound Vib. 252 (2),
343–360.
Jang, J.-S.R., Sun, Ch.T., Mizutani, E., 1997. Neuro-Fuzzy and Soft Computing. Upper Saddle River, NJ: Prentice-Hall.
Juang, C.H., Ni, S.H., Lu, Ping C., 1999. Training artificial neural networks with the aid of fuzzy sets. Comput. Aided Civil Infrastrut. Eng. 14, 407–415.
Kailath, T., Sayed, A.H., Hassibi, B., 2000. Linear Estimation. London: Prentice-Hall.
Kaliszuk, J., Urbanska, A., Waszczyszyn, Z., Furtak, K., 2001. Neural analysis of concrete fatigue durability on the basis of experimental evidence. Arch. Civil Eng. 47, 327–339.
Kaliszuk, J., Waszczyszyn, Z., 2003. Reliability analysis of structures by neural network supported Monte Carlo methods. In: Rutkowski, L., Kacprzyk, J. (Eds), Neural Networks
and Soft Computing, Heidelberg, Germany: Physica-Verlag, A Springer-Verlag Co, pp. 754–759.
Kalman, R.E., 1960. A new approach to linear filtering and prediction problems. ASME J. Basic Eng. 82D, 34–45.
Kleiber, M., 1993. Shape and non-shape sensitivity analysis for problems with any material and kinematic nonlinearity. Comp. Methods Appl. Mech. Eng. 108, 73–97.
Kleiber, M., Antunez, H., Hien, T.D., Kowalczyk, P., 1997. Parameter Sensitivity in Nonlinear Mechanics. Theory and Finite Element Computations. Chichester: Wiley.
Kojima, F., Kubota, N., Hashimoto, S., 2001. Identification of crack profiles genetic programming and fuzzy inference. J. Mater. Process. Technol. 108 (2), 263–267.
Kortesis, S., Panagiotopoulos, P.D., 1993. Neural networks for computing in structural analysis: methods and prospects of applications. J. Num. Methods Eng. 36, 2305–2318.
Kuzniar, K., Maciag, E., Waszczyszyn, Z., 2000. Computation of natural periods of vibrations of medium-height prefabricated buildings by neural networks. Arch. Civil Eng. 46,
512–523.
Kuzniar, K., Waszczyszyn, Z., 2001. Neural simulation of dynamic response of prefabricated buildings subjected to paraseismic excitations. In: Waszczyszyn, Z., Pamin, J.
(Eds.), CD-ROM, Proc. 2nd Euro. Conf. Comp. Mech., Cracow, Poland. Cracow University of Technology, Cracow, Poland 13.
Kuzniar, K., Waszczyszyn, Z., 2002. Neural analysis of vibration problems of real flat buildings and data pre-processing. Eng. Struct. 24, 1327–1335.
Larson, M.C., Verges, M.A., Keat, W.D., 1999. Nondestructive identification of three-dimensional embedded cracks in finite bodies by inversion of surface displacements. Eng.
Fract. Mech. 63, 611–629.
Lee, I.M., Kim, D.-H., 1999. Parameter estimation using extended Bayesian method in tunneling. Comp. Geotech. 24, 109–124.
Liang, Y.C., Hwu, Ch., 2001. On-line identification of holes/cracks in composite structures. Smart Mater. Struct. 10, 599–609.
Liang, Y.C., Lee, H.P., Lim, S.P., et al., 2002. Proper orthogonal decomposition and its applications−Part I: Theory. J. Sound Vib. 252, 527–544.
Liu, G.R., Chen, S.C., 2001. Flaw detection in sandwich plates based on time-harmonic response using genetic algorithm. Comp. Methods Appl. Mech. Eng. 190, 5505–5514.
Luo, Z.Q., Pang, J.S., Ralph, D., 1996. Mathematical Programs with Equilibrium Constraints. Cambridge: Cambridge University Press.
Ly, H.V., Tran, H.T., 2001. Modeling and control of physical processes using proper orthogonal decomposition. Math. Comput. Model. 33, 223–236.
Maciag, E., 1986. Experimental evaluation of changes of dynamic properties on different grounds. Earth. Eng. Struct. Dyn. 14, 925–932.
Madsen, H.O., Krenk, S., Lind, N.C., 1986. Methods of Structural Safety. Englewood Cliffs, NJ: Prentice-Hall.
Mahnken, R., 1999. Aspects of the finite element implementation of the Gurson model including parameter identification. Int. J. Plast. 15, 1111–1137.
Mahnken, R., 2000. An inverse finite-element algorithm for parameter identification of thermoelastic damage models. Int. J. Num. Methods Eng. 48, 1015.
Mahnken, R., 2002. Theoretical, numerical and identification aspects of a new model class for ductile damage. Int. J. Plast. 18, 801–831.
Mahnken, R., Khul, E., 1999. Parameter identification of gradient enhanced damage models with the finite element method. Euro. J. Mech. A/Solids 18, 819.
Mahnken, R., Stein, E., 1996. A unified approach for parameter identification of inelastic material models in the frame of the finite element method. Comp. Methods Appl.
Mech. Eng. 136 (3), 225–258.
Mahnken, R., Steinmann, P., 2001. A finite element algorithm for parameter identification of material models for fluid saturated porous media. Int. J. Num. Anal. Methods
Geomech. 25, 415–434.
Marek, P., Brozetti, J., Gustar, M., 2001. Probabilistic Assessment of Structures Using Monte Carlo Simulation. Czech Republic: TERECO, ITAM CAS Prague.
Matlab, 2000. Version 6, User's Guide & Optimization Toolbox User's Guide. USA: The Math Works.
Miller, B., Ziemianski, L., 2000. Neural networks in updating of a twelve storey frame mode1. In: Topping, B.H.V. (Ed.), Computational Engineering Using Metaphors from
Nature. Edinburgh, Scotland: Civil-Comp, pp. 31–36.
Miller, B., Ziemianski, L., 2001. Application of neural networks to the structural model updating. In: Waszczyszyn, I., Pamin, J. (Eds.), CD-ROM, Proc. 2nd European
Conference on Computational Mechanics, Cracow, Poland. Cracow University of Technology, Cracow, Poland 10.
Mroz, Z., Stavroulakis, G.E. (Eds.), 2005. Parameter Identification of Materials and Structures. CISM Lecture Notes, vol. 469. Wien: Springer.
Nag, A., Mahapatra, D.R., Gopalakrishnan, S., 2002. Identification of delamination in composite beams using spectral estimation and a genetic algorithm. Smart Mater. Struct.
11, 899–10908.
Narendran, N., Shukla, A., Letcher, S.V., 1991. Determination of fracture parameters using embedded fiber-optic sensors. Exp. Mech. 4, 360–365.
Neural Network Toolbox for Use with MATLAB, 2000. User's Guide, Version 4. Natick, MA: The Math Works.
Nishimura, N., 1993. A boundary integral equation method for solving elasto-dynamic crack determination problems. In: Kleinman, R., Angell, T., Colton, D., Santosa, I.,
Stakgold, F. (Eds.), Mathematical and Numerical Aspects of Wave Propagation. SIAM, pp. 390–397. (Chapter 41).
Oishi, A., Yamada, K., Yoshimura, S., Yagawa, G., 1995. Quantitative nondestructive evaluation with ultrasonic method using neural networks and computational mechanics.
Comp. Mech. 15 (6), 521–533.
Oishi, A., Yamada, K., Yoshimura, S., et al., 2001. Neural network-based inverse analysis for defect identification with laser ultrasonics. Res. Nondestruct. Eval. 2, 79–95.
Ostrowski, Z., Bialecki, R.A., Kassab, A.J., 2005. Estimation of constant thermal conductivity by use of proper orthogonal decomposition. Comput. Mech. 37, 52–59.
Ostrowski, Z., Bialecki, R.A., Kassab, A.J., 2008. Solving inverse heat conduction problems using trained POD-RBF network. Inv. Probl. Sci. Eng. 16, 705–714.
Pabisek, E., Waszczyszyn, Z., 2000a. Neural networks in the analysis of elastoplastic plane stress problem with unilateral constraints. In: Topping, B.H.V. (Ed.), Computational
Engineering Using Metaphors from Nature. Edinburgh: Civil-Comp Press, pp. 1–6.
Pabisek, E., Waszczyszyn, Z., 2000b. Neural analysis of elastoplastic plane stress problem with unilateral constraints. In: Waszczyszyn, Z., Pamin, J. (Eds.), CD-ROM,
Proceeding of 2nd European Conference on Computational Mechanics, Cracow, Poland. Cracow University of Technology, Cracow, Poland 15.
Paez, Th.L., 1993. Neural Networks in Mechanical System Simulation, Identification and Assessment. Shock Vibrat. 1, 177–199.
Pang, J.S., Tin-Loi, F., 1998. A penalty interior point algorithm for a parameter identification problem in elastoplasticity. Mech. Struct. Machines 29, 85–99.
Papadrakakis, M., Papadopoulos, V., Lagaros, N.D., 1996. Structural reliability analysis of elastic−plastic structures using neural networks and Monte Carlo simulation. Comp.
Methods Appl. Mech. Eng. 136, 145–163.
Author's personal copy
Inverse Analysis 39
Piatkowski, G., Ziemianski, L., 2001. Damage detection in rods using wave propagation and neural networks. In: Waszczyszyn, Z., Pamin, J. (Eds.), CD-ROM, Proceeding of
2nd European Conference on Computational Mechanics, Cracow, Poland. Cracow University of Technology, Cracow, Poland 8.
Pulido, J.E., Jacobs, T.L., Prates de Lima, E.C., 1992. Structural reliability using Monte Carlo with variance reduction techniques on elastoplastic structures. Comp. Struct. 43,
419–430.
RILEM Draft Recommendation, 1985. Determination of the fracture energy of mortar and concrete by means of three-point bend tests on notched beams. Mater. Struct. 18,
285–290.
RILEM Report 5, 1991. Fracture Mechanics Test Methods for Concrete. Shah, S.P., Carpinteri, A. (Eds.) Chapman and Hall.
Rojas, R., 1996. Neural Networks—A Systematic Introduction. Berlin, Heidelberg: Springer.
Rubinstein, R.Y., 1981. Simulation and Monte Carlo Method. New York: Wiley.
Rus, G., Gallego, R., 2002. Optimization algorithms for identification inverse problems with the boundary element method. Eng. Anal. Boundary Elem. 26 (4), 315–327.
Sansalone, M.J., Streett, W.B., 1997. Impact—Echo. Nondestructive evaluation of concrete and masonry. Ithaca, NY: Bullbrier Press.
Simoni, L., Schrefler, B.A., 2001. Parameter identification for a suction-dependent plasticity model. Int. J. Num. Anal. Methods Geomech. 25, 273–288.
Slowik, V., Saouma, V.E., 2000. Water pressure in propagating concrete cracks. ASCE J. Struct. Eng. 126, 235–242.
SNNS—Stuttgart Neural Network Simulator, 1998. In: Zell, A. (Ed.), User's Manual, Version 4.2. Germany: University Stuttgart.
Stavroulakis, G.E., 1999. Impact-echo from a unilateral interlayer crack. BEM modelling and neural identification. Eng. Fract. Mech. 62 (2−3), 165–184.
Stavroulakis, G.E., 2001. Inverse and crack identification problems in engineering mechanics. Dordrecht, Boston, London: Kluwer Academic (currently Springer).
Stavroulakis, G.E., Antes, H., 1997. Nondestructive static unilateral crack identification A BEM—Neural network approach. Comp. Mech. 20 (5), 439–451.
Stavroulakis, G.E., Antes, H., 1998a. Flaw identification in elastomechanics: BEM simulation with local and genetic optimization. Struct. Optim. 16 (2/3), 162–175.
Stavroulakis, G.E., Antes, H., 1998b. Neural crack identification in steady state elastodynamics. Comp. Methods Appl. Mech. Eng. 165 (1/4), 129–146.
Stavroulakis, G.E., Antes, H., 2000. Unilateral crack identification: A filter-driven, iterative, boundary element approach. J. Global Optim. 17 (1−4), 339–352.
Stavroulakis, G.E., Antes, H., Panagiotopoulos, P.D., 1999. Transient elastodynamics around cracks including contact and friction. In: Martins, J.A.C., Klarbring, A. (Eds.),
Computer Methods in Applied Mechanics and Engineering. Special Issue: Computational Modeling of Contact and Friction, vol. 177(3/4). , pp. 427–440.
Stavroulakis, G.E., Engelhardt, M., Gallego, R., 2002. Crack identification in elasticity. In: Mang, H.A., Rammerstorfer, F.G., Eberhardsteiner, J. (Eds.), Proceeding of 5th World
Congress on Computational Mechanics WCCM V, Vienna, July 7−12, 2002. Austria: Vienna University of Technology. ISBN 3-9501554-0-6. Available at: http://ccm.tuwien.
ac.at
ST Neural Network, 1998. STATISTICA Polish Version. Cracow, Poland: StatSoft.
Suh, M.W., Shim, M.B., Kim, M.Y., 2002. Crack identification using hybrid neuro-genetic technique. J. Sound Vib. 238 (4), 617–635.
Sun, N.-Z., 1994. Inverse Problems in Groundwater Modeling. Dordrecht: Kluwer.
Tanaka, M., Nakamura, M., Nakano, T., 1990. Detection of cracks in structural components by the elastodynamic boundary element method. In: Tanaka, M., Brebbia, C.A.,
Homna, T. (Eds.), Boundary Elements XII, vol. 2. Southampton, UK: Computational Mechanics Press, pp. 413–424.
Tarantola, A., 1994. Inverse Problems Theory: Methods for Data Fitting and Model Parameter Estimation. Amsterdam: Elsevier Applied.
Theocaris, P.S., Panagiotopoulos, P.D., 1993. Neural networks for computing in fracture mechanics: Methods and prospects of application. Comp. Methods Appl. Eng. Sci.
106, 213–228.
Tiago, C.M., Leitão, V.M.A., 2006. Application of radial basis functions to linear and nonlinear structural analysis problems. Comput. Math. Appl. 51, 1311–1334.
Tin-Loi, F., Que, N.S., 2001. Parameter identification of quasibrittle materials as a mathematical program with equilibrium constraints. Comp. Methods Appl. Mech. Eng. 190,
5819–5836.
Topping, B.H.V., Bahreininejad, A., 1997. Neural Computing for Structural Mechanics. Edinburgh, UK: Saxe-Coburg.
Tosaka, N., Utani, A., Takahashi, H., 1995. Unknown defect identification in elastic field by boundary element method with filtering procedure. Eng. Anal. Boundary Elem. 15,
207–215.
Tsompanakis, Y., Lagaros, N.D., Stavroulakis, G.E., 2008. Soft computing techniques in parameter identification and probabilistic seismic analysis of structures. Adv. Eng.
Software 39, 612–624.
Tvergaard, V., 1990. Material failure to void growth to cohalescence. Adv. Appl. Mech. 27, 83–151.
Tvergaard, V., Needleman, A., 1995. Effects of nonlocal damage in porous plastic solids. Int. J. Solids Struct. 32, 1063–1077.
van Mier, J.G.M., van Vliet, M.R.A., 2002. Uniaxial tension test for the determination of fracture parameters of concrete: State of the art. Eng. Fract. Mech. 69, 235–247.
Waszczyszyn, Z., 1996. Some recent and current problems of neurocomputing in civil and structural engineering. In: Topping, B.H.V. (Ed.), Advances in Computational
Structures Technology. Edinburgh, UK: Civil-Comp Press, pp. 43–58.
Waszczyszyn, Z., 1998. Some new results in applications of backpropagation neural networks in structural and civil engineering. In: Topping, B.H.V. (Ed.), Advances in
Computational Structures Technology. Edinburgh, UK: Civil-Comp Press, pp. 173–187.
Waszczyszyn, Z. (Ed.), 1999. Neural Networks in the Analysis and Design of Structures. In: CISM Courses and Lectures No. 404. Wien−New York: Springer. 161−196 (Chapter
1, Fundamentals of artificial neural network, Chapter 4, The neural network approach in plasticity and fracture mechanics, together with P.D. Panagiotopoulos).
Waszczyszyn, Z., 2000. Neural networks in structural engineering: Some recent results and prospects of application. In: Topping, B.H.V. (Ed.), Computational Mechanics for the
Twenty-First Century. Edinburgh, UK: Saxe-Coburg Publ, pp. 479–515. (Chapter 23).
Waszczyszyn, Z., 2001. Neural networks in structural engineering: Some new results. In: Valliapan, S., Khalili, N. (Eds.), Computational Mechanics—New Frontiers for New
Millenium, vol. 2. Amsterdam: Elsevier, pp. 1311–1320.
Waszczyszyn, Z., Bartczak, M., 2002. Neural prediction of buckling loads of cylindrical shells with geometrical imperfections. Int. J. Non-Lin. Mech. 37, 763–775.
Waszczyszyn, Z., Kuzniar, K., Obiala, R., Slonski, M., 1999. Some new results and prospects of neural analysis of building vibration problems. In: Duch, W. (Ed.), Engineering
Applications of Neural Networks. Torun, Poland: Wydawnictwo Adam Marszalek, pp. 265–272.
Waszczyszyn, Z., Pabisek, E., 1999. Hybrid NN/FEM analysis of the elasto-plastic plane stress problem. Comp. Ass. Eng. Sci. 6, 177–188.
Waszczyszyn, Z., Pabisek, E., 2000. Neural network supported FEM analysis of elastoplastic plate bending. Hungary: TU Budapest. Research News, Special Issue 2000/4, 12−19.
Waszczyszyn, Z., Pabisek, E., 2002. Elastoplastic analysis of plane steel frame by a new superelement. Arch. Civil Eng. 48, 159–181.
Waszczyszyn, Z., Slonski, M., 2000. Analysis of some problems of experimental mechanics and biomechanics by means of the ANFIS neuro-fuzzy system. J. Theor. Appl.
Mech. 38, 429–445.
Waszczyszyn, Z., Ziemianski, L., 2001. Neural networks in mechanics of structures and materials—New results and prospects of application. Comp. Struct. 79, 2261–2276.
Wu, Z.P., Liu, G.R., Han, X., 2002. An inverse procedure for crack detection in anisotropic laminated plates using elastic waves. Eng. Anal. Comp. 18, 116–223.
Xiang, Z., Swoboda, G., Cen, Z., 2002. Identification of damage parameters for jointed rock. Comp. Struct. 80, 1429–1440.
Xu, Y.G., Liu, G.R., Wu, Z.P., Huang, X.M., 2001. Adaptive multilayer perceptron networks for detection of cracks in anisotropic laminated plates. Int. J. Solids Struct. 38,
5625–5645.
Yagawa, G., Okuda, H., 1996. Neural networks in computational mechanics. Arch. Comp. Meth. Eng. 4, 435–512.
Yoshida, F., Urabe, M., Toropov, V.V., 1998. Identification of material parameters in constitutive model for sheet metals from cyclic bending tests. Int. J. Mech. Sci. 40,
237–249.
Ziemiański, L., 1999. Neural Networks in Structural Dynamics. Rzeszow: Oficyna Wyd. Politechniki Rzeszowskiej (in Polish).
Ziemianski, L., Piatkowski, G., 2000. Use of neural networks for damage detection in structural elements using wave propagation. In: Topping, B.H.V. (Ed.), Computational
Engineering Using Metaphors from Nature. Edinburgh, Scotland: Civil-Comp, pp. 25–30.