Computational Mechanics (2018) 62:835–852

https://doi.org/10.1007/s00466-017-1532-y

ORIGINAL PAPER

Multiple crack detection in 3D using a stable XFEM and global

optimization
Konstantinos Agathos2 · Eleni Chatzi3 · Stéphane P. A. Bordas1,2

Received: 24 February 2017 / Accepted: 18 December 2017 / Published online: 19 February 2018
© The Author(s) 2018. This article is an open access publication

Abstract
A numerical scheme is proposed for the detection of multiple cracks in three dimensional (3D) structures. The scheme is
based on a variant of the extended finite element method (XFEM) and a hybrid optimizer solution. The proposed XFEM
variant is particularly well-suited for the simulation of 3D fracture problems, and as such serves as an efficient solution to the
so-called forward problem. A set of heuristic optimization algorithms are recombined into a multiscale optimization scheme.
The introduced approach proves effective in tackling the complex inverse problem involved, where identification of multiple
flaws is sought on the basis of sparse measurements collected near the structural boundary. The potential of the scheme is
demonstrated through a set of numerical case studies of varying complexity.

Keywords Crack detection · XFEM · Genetic algorithms · CMA-ES

1 Introduction tion (diagnostics) [9]. The damage detection task is one of

particular importance and is often considered as the focus
The advent of low-cost and easily deployable sensor tech- of SHM processes, which may be defined across four lev-
nologies, has in recent years sparked a significant rise in els [10]: (i) detection of damage; (ii) localization of damage;
the deployment of monitoring technologies for large-scale (iii) quantification of the severity and extent of damage; and
structural systems [1]. Due to the flexibility of technologies (iv) estimation of the future performance of the component
involved, Structural Health Monitoring (SHM) methods are (or system) as damage accumulates.
available in various forms, i.e., vibration-based [2] or static While the first tasks of damage detection, and potentially
monitoring [3], periodic and short-term versus continuous localization, may be often achieved on the basis of data
and long-term deployments, visual inspections versus non- processing alone, the more refined diagnostic levels typi-
destructive evaluation [4,5]. etc. cally require the combined use of a simulation model for the
Availability of monitoring data may be exploited in a monitored system. Availability of a system model enables
number of tasks pertaining to the life-cycle assessment and formulation of a so-called inverse problem procedure [11],
management of infrastructure systems including condition where the task lies in updating the system’s representation in
and reliability assessment [6], updating/calibration of sim- a way which reveals its current status, and is thereby infor-
ulation models [7], prediction of performance and residual mative with respect to the nature of the induced damage, e.g.
life (prognostics) [8], damage identification and fault detec- fatigue, cracking or component failure. Availability of mon-
itoring data drives the inverse problem formulation, which
B Stéphane P. A. Bordas aims to minimize the difference between the model predic-
[email protected];
[email protected]
tion and the structural response data acquired via monitoring.
This may often be solved by means of optimization methods
1 Institute of Theoretical, Applied and Computational based on least squares or based on Bayesian analysis [12,13].
Mechanics, Cardiff University, Cardiff CF24 3AA, UK In an optimization setting, monitoring data such as accel-
2 Research Unit in Engineering Science, Luxembourg eration [14], strain [15], acoustic emission, wave propagation
University, 6 rue Richard Coudenhove-Kalergi, [16], or impedance [17] data essentially establish the tar-
1359 Luxembourg, Luxembourg
get function to be optimized, while structural properties and
3 Institute of Structural Engineering, ETH Zürich, the characteristics of potential damage (geometry, location,
Stefano-Franscini-Platz 5, 8093 Zurich, Switzerland

123
836 Computational Mechanics (2018) 62:835–852

extent of flaw) form the optimization variables. The inverse ple shaped inclusions/void and cracks with XFEM [32,33],
problem solution calls for multiple analyses of the so-called as well as in the modeling of arbitrarily-shaped objects as
forward problem, i.e., the simulation of the system given demonstrated by Benowitz and Waisman [34], and Jung and
prescribed structural and flaw properties. In this sense, it is Taciroglu [35].
evident that the problem may become computationally tax- Complimentary to the forward problem, an appropriate
ing when forward analysis of complex systems is involved, optimization procedure need be enforced. Heuristic opti-
including analyses in the three dimensional space. Since it mization [36] is particularly suited to such an end, since
is oftentimes desired to perform the diagnostic tasks in the it allows for flexibility in the formulation of the forward
short time that follows an initial indication of damage, the problem, which need not be linear, convex, or smooth. Due
corresponding analysis tools ought to ensure rapid computa- to this feature, different forms of heuristic procedures have
tion. been adopted in the context of Structural Health Monitoring.
Within this context, a number of techniques have been Hunaidi [37] employs evolution-based Genetic Algorithms
proposed in recent literature for cutting down on computa- (GAs) for non-destructive assessment of pavements on the
tion while maintaining estimation accuracy, the majority of basis of surface waves tests; Farley et al. [38] adopt an
which rely on reduced order representations. A first approach artificial neural network approach for defect detection via
pertains to the use of surrogate models [18,19], which are ultrasonic signals; Lee et al. [39] formulate an inverse scat-
often data-driven albeit not necessarily linked to first prin- tering problem on the basis of Particle Swarm Optimization;
ciples (physical) information. A second alternative however, while Bernieri et al. [40] reconstruct cracks via eddy current
pertains to reduced representations that are founded on the testing and a machine learning approach.
principles of computational mechanics, such as multiscale For the solution of the inverse problem in the particular
schemes [20] for composite systems, component mode syn- domain of flaw/crack detection, Rabinovich et al. [41,42],
thesis methods [21] for structural dynamics problems, or the combine and XFEM approach with GAs for crack identifica-
extended Finite Element Method (XFEM) for linear elastic tion in static and dynamic 2D problems. Waisman et al. [43]
fracture mechanics [22]. and Chatzi et al. [44], extend and experimentally validate
In the case of fracture, a significant challenge faced by the XFEM–GA scheme for identification of generalized flaw
reduced order representations lies in the tackling of the non- types. Sun at al. [45] presented an adaptive algorithm, once
linearities that are associated with the fracture or damage again relying on XFEM, able to detect multiple flaws without
process. These typically require inclusion of a large number prior knowledge on their number by means of an Enhanced
of modes for accurately capturing the high-frequency nature Artificial Bee Colony (EABC) algorithm [46] and a sweeping
of the solution around the damage zone. The latter may in window method for dynamic problems [47]. Yan et al. [48]
general not be entirely precomputed due to the non-linearities introduce a guided bayesian inference approach for detec-
in the damage and fracture processes. tion of multiple flaws. Jung and Taciroglu employ XFEM
To address this, a number of possibilities exist, mainly for identification of an arbitrarily shaped scatterer embed-
relying on updating the reduced space on-the-fly. The inter- ded in elastic heterogeneous media [35]. Nanthakumar et
ested reader is referred to the work of Kerfriden and cowork- al. [49] combine XFEM to the Multilevel Coordinate Search
ers and the publications therein, where Newton–Krylov (MCS) method to detect cracks and voids in piezoelectric
[23], local–global [24] domain-wise model order reduction materials, while in a later work [50] they employ deriva-
[25], and Bayesian approaches [26] are proposed. Those tives of the level sets for the optimization step in order to
algebraic-based model order reduction techniques may be increase the robustness and efficiency of their method. In
complemented by multiscale approaches, as in [27], where a more recent work [51], the same authoring team applies
a scale-selection approach is proposed for determining the the XFEM–MCS scheme to the detection of multiple cracks
optimal model for a given region. Finally, statistical-based in piezoelectric structures under dynamic electric loads. In
approaches have been proposed [28] in order to determine Ma et al. [52] XFEM is incorporated in a three step algo-
the fracture process zone based on the lack of ability of rithm for the detection of multiple flaw clusters. Finally,
reduced order models to represent the failure of the sys- XFEM is employed in damage detection schemes for dams
tem. in the works of Alalade et al. [53] and Pirboudaghi et
In this paper, and motivated by previous works of the al. [54].
authoring team in the two-dimensional domain, we rely A characteristic feature of the aforementioned works is
on XFEM for solution of the forward problem. XFEM their confinement and demonstration in the two-dimensional
alleviates the need for remeshing [29,30] for diverse flaw domain. The extension in the third dimension comes with a
locations and geometries thereby significantly cutting down number of challenges, some of which have recently been
on the computational toll of the forward analysis [31]. tackled in a robust 3D XFEM scheme introduced by the
XFEM has been proven adept in the modeling of multi- authoring team [55,56]. This XFEM scheme was coupled

123
Computational Mechanics (2018) 62:835–852 837

with a Covariance Matrix Adaptation Evolution Strategy

(CMA-ES) [57] in a first attempt to apply XFEM based
crack detection to 3D problems [58]. In the present work, the
proposed XFEM variant is combined to a multiscale opti-
mization strategy consisting of a discontinuous step utilizing
genetic algorithms and a continuous step utilizing the CMA-
ES algorithm [57] in order to detect multiple cracks in 3D
solids.

2 Inverse problem formulation

Inverse problems aim at identifying the latent and unknown

parameters of a system given measured information on its Fig. 1 Cracked body and boundary conditions
response and commonly, albeit not necessarily, a compu-
tational model of the system. The estimation of structural
response for a prescribed set of model parameters using an 3 Solution of the forward problem using
available model structure may be considered as the forward XFEM
problem. In the present case, this forward problem is solved
via a 3D XFEM approach. For the solution of the optimization problem posed in the
For the specific case of detection of multiple flaws in previous section, several evaluations of the fitness function,
the form of cracks, the unknown parameter set comprises for different values of the design variables, are required.
the number, location, shape, size and orientation of existing These evaluations correspond to solutions of the forward
cracks in a structure. While various sensors may be employed problem, for different crack numbers, shapes, sizes and loca-
for monitoring structural response, we here assume avail- tions, and should be obtained in a robust and efficient way,
ability of strain information at specific locations along the ensuring minimization of the associated computational toll.
structure obtained via conventional strain gauges. Due to its In the present work, the extended finite element method
low cost and ease of deployment this monitoring option is (XFEM) [30], and in particular the variant introduced in
often adopted, albeit distributed sensing alternatives, such as Reference [56], is employed for the solution of forward prob-
fiber optics solutions [59], may also be adopted. lems. The method has already been successfully used in 2D
The inverse formulation may then be summarized as the crack detection schemes [41,43] due to its ability to repre-
following optimization problem [41,43]: sent discontinuities without requiring any modification of the
finite element mesh, a feature which is crucial for this cat-
egory of applications where the forward problem has to be
Find θi such that
solved for a very large number of different crack configura-
F (θi ) → min (1) tions. In the following subsections, the forward problem is
mathematically formulated and the solution method is pre-
where θi is a set of parameters used to describe the number, sented.
location, shape, size and orientation of the cracks and F is
the objective function given by: 3.1 Problem statement

 h  The problem consists of a linear elastic solid Ω (Fig. 1)

¯ (θi ) − ¯ m 
F (θi ) = (2) cracked at several locations and bounded by the boundary
¯ m  Γ where:

where ¯ h (θi ) are the numerically computed strains at the Γ = Γ0 ∪ Γu ∪ Γt ∪ Γc (3)

sensor locations and ¯ m are the measured strains at the same
locations. The strain components for all sensor locations con- Γ0 is the part of the boundary where no boundary conditions
sidered are arranged in vectors containing n c × n s elements, are applied. Γu is the part of the boundary where displace-
where n c is the number of components of the strain tensor ments ū are imposed as Dirichlet boundary conditions. Γt is
(n c = 9 for the 3D case), and n s pertains to the number of the part of the boundary where surface tractions t¯ are applied
sensors. as Neumann conditions. Γc is the surface of the cracks.

123
838 Computational Mechanics (2018) 62:835–852

The equilibrium equations are formulated in weak form Moreover, it enables the use of higher order enrichment func-
as: tions in 3D linear elastic fracture mechanics.
Find a kinematically admissible displacement field u ∈ U Perhaps the most critical feature of extended and general-
such that ∀v ∈ V ized finite element methods (X/GFEM) is the enrichment of
   the FE approximation with functions which are able to rep-
(u) : D : (v) dΩ = b · v dΩ + t̄ · v dΓ (4) resent known features of the solution. Enrichment is realized
Ω Ω Γt by employing the partition of unity (PU) method [63]:
where:
u (x) = N I (x) u I + N I∗ (x) Ψ (x) b I (9)
  3   ∀I
   ∀I  
U = u|u ∈ H 1 (Ω) , u = ū on Γu (5) FE approximation enriched part

and where N I (x) are the FE interpolation functions, u I are FE

degrees of freedom (dofs), N I∗ (x) is a basis of functions that
  3 
form a partition of unity, Ψ (x) are the enrichment functions
V = v|v ∈ H (Ω) , v = 0 on Γu
1
(6)
and b I are the enriched degrees of freedom.
While in PU-FEM enrichment is applied globally to all
Functions of H 1 (Ω) are implicitly discontinuous along the FE nodes, in XFEM enrichment is only applied locally
the crack surfaces. to approximate local phenomena such as cracks and discon-
In the above,  is the small strain field, D is the elasticity tinuities. This can be achieved by appropriately defining the
tensor and b is the applied body force per unit volume. set of enriched nodes, as will be done in the following.
In linear elastic fracture mechanics two different enrich-
3.2 Crack representation ment functions are employed, the modified Heaviside or jump
enrichment functions:
As is commonly the case in XFEM [60–62], cracks are rep- 
resented implicitly using the level set method. Level set 1 for φ ≥ 0
H (φ) = (10)
functions, denoted as φ and ψ, are defined for an arbitrary −1 for φ < 0
point x as follows:
which are used to represent the displacement jump along the
– φ is the signed distance from the crack surface defined crack surfaces, and the asymptotic or tip enrichment func-
as: tions:
 
φ (x) = min x − x̄ sign n+ · (x − x̄) (7) √ θ √ θ √ θ
x̄∈Γc
F j (r , θ ) = r sin , r cos , r sin sin θ,
2 2 2

√ θ
where n+ is the outward normal to the crack surface and r cos sin θ (11)
sign () is the sign function. 2
– ψ is a signed distance function such that ∇φ · ∇ψ = 0
which are used to represent the asymptotic fields around the
and φ (x) = 0 and ψ (x) = 0 defines the crack front.
crack front.
Since enrichment is applied locally, the nodal sets where
Additionally, a polar coordinate system is defined along
each enrichment function is used have to be appropriately
the crack front with coordinates [60–62]:
selected:

 
φ
r= φ2 + ψ 2, θ = arctan (8) – Jump enrichment is used for nodes belonging to elements
ψ
that are divided in two parts by the crack surface.
These coordinates refer to a plane normal to the crack front. – Tip enrichment is used for nodes belonging to elements
that contain the crack front (topological enrichment), or
3.3 Discretization for nodes that lie in a certain distance (enrichment radius)
from the crack front (geometrical enrichment). In the first
The weak form is discretized using a stable XFEM vari- case sub optimal convergence rates are obtained [64,65].
ant introduced in the authors’ previous works [55,56]. In the second, while optimal convergence is achieved,
The method was shown to provide increased accuracy and conditioning problems are caused, for the solution of
improved conditioning when compared to standard XFEM. which, special techniques are necessary [65–67].

123
Computational Mechanics (2018) 62:835–852 839

blending area, along which the weight function fades to zero,

can consist of one or several layers of elements [56].
The displacement approximation for the method is:

u (x) = N I (x) u I
I ∈N

+ ϕ̄ (x) N J (x) (H (x) − H J )b J
J ∈N j
⎛

+ ϕ (x) ⎝
g
N K (x) F j (x)
Fig. 2 Front elements, nodes and boundaries K ∈N s j
⎞

F j (xT )⎠ c K j
g
− N T (x) NK (xT )
T ∈N t K ∈N s j
(13)

where N is the set of all nodes in the FE mesh. N j is the

set of jump enriched nodes. This nodal set includes all nodes
whose support is split in two by the crack and in addition
belong to elements where the weight function ϕ̄ (x) assumes
values greater than zero. N t is the set of tip enriched nodes.
This nodal set includes all nodes that belong to an element
with at least one node inside the enrichment radius. N s is the
set of nodes in the superimposed mesh.
Fig. 3 Local coordinate ξ of the front elements

Functions N I∗ (x) used for the partition of unity enrich- 4 Parametrization and constraints
ment are typically selected to coincide with the FE shape
functions N I (x) ≡ N I∗ (x) . In the variant used herein how- Since the present work is only one of the first attempts to
ever, an alternative definition is used which has been shown extend flaw detection schemes [41,43,44] in 3D, some sim-
[55,56] to provide improved conditioning of the resulting plifications are made in order to reduce the complexity of the
stiffness matrices. More specifically, a superimposed mesh general problem. Two main simplifications are made, with
of special elements discretizing the crack front is introduced, regard to the crack geometries and interactions.
as illustrated in Fig. 2 and the shape functions correspond- The first aims at reducing the number of parameters used
ing to those front elements are used as a basis for the PU to represent crack geometries by only employing elliptical
enrichment. cracks for the forward problem, and approximating cracks of
The shape functions of the front elements are defined as different shapes by appropriately varying the ellipse param-
simple 1D FE shape functions: eters. Although this approach may seem somehow limited, it
provides the possibility to model a variety of crack shapes,
 
1−ξ 1+ξ while requiring a relatively small number of parameters to
Ng (ξ ) = (12)
2 2 describe each crack.
A second simplification is assumed with respect to the
where ξ is the local coordinate of the superimposed element interactions between different cracks. Although multiple
(Fig. 3). This parameter is defined in detail in References [55, cracks are considered, we herein only investigate cases where
56]. the minimum distances between the different cracks are
Blending problems between the standard and the enriched larger than some predefined value. The above approach is
part of the approximation [68–70] are dealt with using the necessary in order to avoid crack intersections which would
techniques developed in the works of Fries [68] and Ventura pose problems in the solution of the forward problem with
et al. [69] and applied in a 3D framework in our previous XFEM, since the treatment of intersecting cracks in 3D can
works [55,56]. These introduce a weight function ϕ (x) that be problematic.
assumes a value of unity for the fully enriched elements, The above simplifications can be overviewed as follows.
and linearly fades to zero for the blending elements. The The scheme developed in this work aims at determining the

123
840 Computational Mechanics (2018) 62:835–852

where p is a low degree polynomial: p (x) = {a1 , a2 , . . . , al }·

{ p1 , p2 , . . . , pl }T where ai are coefficients to be determined,
pi are the elements of the polynomial basis and l is the num-
ber of polynomial terms used R is the basic function, common
choices for this function are:

– The thin plate spline: R (r ) = r 2 log (r )

The Gaussian: R (r ) = e−cr√
2
–
– The multiquadric: R (r ) = r 2 + c2
Fig. 4 Parametrization of an elliptical crack – The biharmonic spline: R (r ) = r
– The triharmonic spline: R (r ) = r 3
number and locations of existing cracks and roughly esti-
mating their sizes and shapes. The accurate determination of in the above r = x − xi . The variable r in this case is not
the geometrical shapes of the cracks, including crack inter- to be confused with the polar coordinate used in Sect. 3.
sections, exceeds the above aim and in addition would be
limited by the amount and accuracy of the available struc- λi are coefficients to be determined.
tural response measurements.
By employing the known values of the function si at points
4.1 Parametrization xi a system of equation can be created:
    
The parameters involved in the definition of each elliptical A P λ s
= (15)
crack are the coordinates of its center point x0 ({x0 , y0 , z 0 }), PT 0 a 0
the angles of rotation about the three axes φx , φ y and φz and
lengths a and b. Angles φx , φ y and φz are used to produce where
vectors n, t1 and t2 by rotating unit vectors e1 , e2 and e3 . All  
of the above parameters are illustrated in Fig. 4. Ai j = R xi − x j  , i, j = 1, . . . , N
Pi j = p j (xi ) , i = 1, . . . , N , j = 1, . . . , l
4.2 Constraints for a single crack
The solution of the above system yields the values of the
coefficients ai and λi which in term make possible the eval-
While the range of values assumed by the design variables
uation of the RBF at any given point.
may be restricted by upper and lower bounds, for complex
structure geometries invalid crack locations may still be gen-
4.2.2 Determination of invalid cracks
erated which would result in unnecessary solutions of the for-
ward problem. In order to avoid such occurrences, a method
Once the RBF representation of the structure has been con-
of determining the relative position of the cracks with respect
structed, a set of control points lying on the crack surface is
to the structure is introduced herein. This method represents
generated for each candidate crack. For the elliptical cracks
the boundaries of the structure via implicit functions and eval-
considered in the present work, those points are generated
uates this function for several points on the crack surface.
according to the pattern illustrated in Fig. 5. The relative
position of the crack with respect to the structure can be
4.2.1 Radial basis functions
determined from the signs of the RBF at the control points.
For instance, if the sign of the RBF is negative for all the
The implicit functions used in the present work are radial
control points then crack lies entirely inside the structure.
basis functions [71] (RBF) and they are defined so as to
Moreover, from the signs and values of the RBF at the
assume negative values in the interior of the structure, pos-
control points some other cases can be identified:
itive values in the exterior and a value equal to zero on the
structure boundaries.
– When only a fraction of the control points assume pos-
Radial basis functions are constructed from a set of points
itive values, then the corresponding crack intersects the
xi , i = 1, . . . , N lying on and off the surface to be described.
structure boundary and, since such cases are also of inter-
In general they assume the form:
est, it is considered valid. However, the percentage of
positive values should lie within certain bounds in order

N
s (x) = p (x) + λi R (x − xi ) (14) to avoid situations where a very small or a very large
i=1 part of the crack lies outside the structure. Those cracks,

123
Computational Mechanics (2018) 62:835–852 841

Fig. 6 Bounding box used to prevent crack intersections

Fig. 5 Control points on the crack surface

of the bounding boxes are given the values 2ai + 2c, 2bi + 2c
besides being physically meaningless, could potentially and 2c where a and b are the lengths defining the correspond-
cause numerical problems and should be discarded. ing crack. Parameter c should be attributed a large enough
– If the majority of points assume low values (below a pre- value in order to ensure that enriched elements belonging to
defined tolerance), the crack is considered invalid since different cracks do not overlap. Subsequently, the separat-
an actual crack would not lie on the structural boundary. ing axis theorem [72] is employed to determine whether two
This case could also cause numerical problems. bounding boxes intersect.
– If the RBF repeatedly alternates in sign along a line of For a given set of candidate cracks the detection of inter-
points (Fig. 5), then the crack is discarded since only sections is achieved by investigating all possible crack pairs
simple intersections of the crack with the boundary are and determining weather the corresponding bounding boxes
considered. intersect. If two bounding boxes are found to intersect, then
one of the corresponding cracks is discarded. In our current
Considering the above, some further remarks can be made implementation the selection of the crack to be discarded
regarding the definition and use of the RBF in the present is being done arbitrarily since it is assumed the cracks will
application: be either overlapping or in close proximity therefore either
of the cracks will eventually converge to the actual crack if
– Since the RBF values of several points are taken into retained.
account in order to determine the position of the crack A more refined method for performing the above selec-
with respect to the structure, the zero iso-surface of the tion would consist of evaluating the fitness function for both
function does not need to coincide very accurately with cracks and eliminating the crack leading to the worst value.
the structure boundaries. As a result the number of points Such a criterion might lead in faster convergence of the opti-
needed to define the RBF can be kept relatively small, mization process in the expense of increasing the numerical
making the generation and evaluation of the function cost of the evaluation of individual crack configurations. Nev-
faster. ertheless, a detailed comparison would exceed the purposes
– The RBF function can be modified in order to restrict the of the present work.
search space in a part of the structure where the cracks are
expected to be lying, thus making the whole procedure
faster. 5 Inverse problem solution

4.3 Constraints for multiple cracks For the solution of the inverse problem a multiscale strategy,
similar to Reference [46] is employed, which utilizes two
The procedure described above for a single crack has to different optimization algorithms. In what follows, the two
be applied for each individual crack in the case of multi- algorithms are first briefly described with the proposed hybrid
ple cracks. Additionally, overlapping or intersecting cracks strategy introduced next.
have to be detected and discarded.
5.1 Genetic algorithms
4.3.1 Detection of overlapping cracks
Genetic algorithms (GAs) are a category of optimization
In order to detect overlapping or intersecting cracks a bound- tools inspired by biological evolution [73,74]. Solutions to
ing box is first defined for each crack as in Fig. 6. The sides optimization problems are obtained by iteratively improving

123
842 Computational Mechanics (2018) 62:835–852

a set of candidate solutions in an attempt to mimic natural increased. This is achieved by setting the mean in each
evolution processes. Following the GA terminology, the set iteration equal to the weighted average of a predefined
of candidate solutions is termed the population of individuals, number of candidates with the best fitness values from
while each iteration corresponds to a generation. Each indi- the previous iteration.
vidual in the population is represented by set of genes whose Covariance matrix The covariance matrix is updated so
numerical equivalent is a binary array. Moreover, in order that the probability of successful search directions is
to mimic natural selection through survival of the fittest, a increased and in addition information from previous gen-
fitness value is assigned to each individual by evaluating the erations is utilized.
fitness function which usually coincides with the objective Step size The step size is adjusted in order to avoid prema-
function [75]. Typically the following steps take place in a ture convergence while yet ensuring that the algorithm
genetic algorithm: converges fast enough.

Initialization Once the number of parameters and the pop- 5.3 The proposed multiscale strategy
ulation size have been set the initial population is
generated, usually randomly. The basic idea behind the strategy proposed herein, is sim-
Selection The fitness function is evaluated for all the indi- ilar to the one introduced in Sun et al. [46]. In particular, a
viduals in each generation and only a percentage of the two step procedure is adopted where in the first step a dis-
population, corresponding to the highest fitness values, crete optimization algorithm is used to obtain the number
is selected to form the next generation. and approximate location, size, and orientation of the cracks
Reproduction During this step the fittest individuals from while in the second step a continuous optimization algorithm
each generation reproduce to form the next generation, is employed to refine the values of the parameters obtained in
two processes are involved in this reproduction: the first step. The discrete optimization step is employed in
Crossover The genes of two individuals (parents) are com- order to reduce the complexity of the original problem and
bined, through recombination of the bits corresponding obtain an approximate solution which is used as an initial
to their bit representation, to form an offspring. guess for the continuous step where a more accurate solution
Mutation During the reproduction procedure, some bits are can be obtained. The two steps are described in detail in the
randomly flipped in order to simulate mutations that following.
occur in the biological reproduction process.
5.3.1 Discrete optimization step
During this step another practice, called elitism, is commonly
used which consists of allowing the fittest individual or indi- In the first step of the procedure, in which Genetic Algorithms
viduals to survive, unaltered, in the next generation. are used as an optimization tool, the number of cracks is iden-
The above steps are repeated until some prescribed ter- tified, therefore topological variables [45] are employed to
mination criteria are met. The parameters involved in the activate/deactivate candidate cracks. Moreover, the original
above steps are user-defined and include the population size, identification problem is simplified in order to minimize the
crossover rate, mutation rate and the termination criteria. The number of parameters to be identified thus accelerating the
most widely adopted termination criteria include the def- convergence to the approximate solution. The reduction of
inition of a maximal number of generations, a predefined the number of parameters is achieved in two ways, firstly by
target value for the fitness function, as well as a maximal assuming the shape of the cracks to be detected circular rather
consecutive number of generations without improvement in than elliptical and secondly by reducing the number of binary
the fitness values. digits used to represent each of the parameters involved in
the optimization process.
5.2 Covariance matrix adaptation evolution strategy At this stage, the parameters described in Sect. 4.1 are
(CMA-ES) encoded as follows:
In this method [57,76] candidate solutions are generated from
a multivariate normal distribution whose parameters, namely x0 , y0 , z 0 The coordinates of the center of the ellipses are
the distribution mean, covariance matrix and step size, are encoded as:
updated such that the probability of obtaining improved solu-
tions is increased. pi = pimin + ( pimax − pimin ) θi (16)

Distribution mean The distribution mean is updated so that where pi are the coordinates, pimin and pimax are the min-
the probability of successful candidate solutions is imum and maximum values allowed for these coordinates

123
Computational Mechanics (2018) 62:835–852 843

and θi are the design variables used in the genetic algo- space around the initial values. This is achieved by using the
rithm. In order to represent variables θi , n binary digits following encoding for the parameters of the problem.
are used and the resulting values are divided by 2n so
that the variables assume n possible values between 0 θi
pi = pi0 + dpi (17)
and 1. The number of digits defines the total number of 10
possible crack locations and should be chosen according
to the geometry of the solid. It should be noted that a where pi0 are the initial values of the parameters obtained in
different number of digits for each variable can be used. the previous step, dpi are the half lengths of the search space
The minimum and maximum values define a box which (in the direction of each parameter) and θi are the design
should contain the whole domain of interest. variables used in the algorithm.
a, b In the first step of the procedure cracks are considered to In the above, half lengths of the search spaces are given a
be circular, therefore the two ellipse parameters are equal value equal to the distance of two consecutive possible values
and a single variable is required for their representation. (length of the search space for each variable divided by 2n )
The encoding used for this variable is the same as the one of the previous step of the algorithm. The design variables
used for the previous variables (Eq. 16) and the minimum are initialized to zero and the step size is set to σ = 3 which
and maximum values should be chosen according to the implies that the final solution lies in the interval 0 ± 2σ =
expected size of the cracks to be detected. 0 ± 6.
φx , φ y , φz Since cracks are assumed to be circular only the For the parameters that were omitted in the first step (a
first two angles are used at this stage. The encoding used and θz ) a slightly different approach is used. Regarding the
is again that of Eq. (16) and the minimum and maximum parameters of the ellipse, the value computed in the first step
values are set to 0 and π respectively while the number (were the two parameters were assumed equal) is used as
of digits used is n = 2 which results in 4 possible val- an initial value for parameter b for which the encoding of
ues for each angle. For these parameters variables are Eq. (17) is used. Parameter a which should be larger than or
divided by 2n + 1 so that the value π is not included equal to parameter b is obtained as the sum of parameter b
in the possible values for the angles since it is equiv- and an additional parameter ainc :
alent to the value 0. For the specific choice n = 2
the possible values for each angle are 0, π/4, π/2 and a = b + ainc (18)
3π/4.
where the additional parameter is computed as:
At this step of the algorithm, candidate solutions that vio-
|θa |
late the constraints described in Sect. 4 are penalized by ainc = da (19)
being assigned large fitness values. For those solutions the 10
forward problem does not have to be solved. Also, it is pos- In the above, θa is the design variable corresponding to ainc ,
sible that the solution produced by this step contains two or da is the maximum allowed difference between a and b and
more overlapping cracks. Although those cracks might be the absolute value is used to prevent ainc from assuming neg-
activated through their corresponding topological variables, ative values and therefore b from assuming larger values than
the procedure described in Sect. 4.3 will discard all but one of a.
the overlapping cracks and therefore the value of the fitness Regarding angle φz , the following encoding is employed:
function obtained will correspond to a single crack at that
specific location. At the end of the step cracks that have been θφz π
discarded through the above process are considered inactive φz = (20)
10 2
and as a result are not considered in the next step of the
optimization procedure. in order to restrict the possible values of the angle in the
interval [−π/2, π/2].
5.3.2 Continuous optimization step At this step candidate solutions that violate constraints are
re-sampled.
In the second part of the multiscale strategy, the results
obtained in the previous stage are used as an initial solu- 5.3.3 Discussion
tion for the CMA-ES algorithm. The number of cracks is
assumed to have been correctly determined in the previous In the strategy described above, the problem to be solved
step. Furthermore, the scaling of the parameters and the ini- in each individual step is of reduced complexity in com-
tial step size used in the algorithm are chosen so that the parison to the original problem definition. In the first step,
search space is confined in a small part of the original search the dimension of the search space is significantly reduced

123
844 Computational Mechanics (2018) 62:835–852

by removing some of the problem parameters, and reducing presence of at least one standard element between two
the number of binary digits used to represent the remaining enriched elements.
ones. In the second step, the search space is restricted in a – It is considered that cracks are far enough so that no inter-
small region around the solution obtained in the previous action between cracks takes place and that cracks which
step. Moreover, in the second part of the algorithm the num- are in close proximity can be approximated by a single
ber of cracks is considered to have already been determined, larger crack. As a result, the allowed distance between
and as a result the problem is further simplified. Without cracks can be further increased to prevent evaluations of
these simplifications, the complexity of the problem would the forward problem for cases that are not of interest and
render convergence extremely slow, or even impossible. to avoid the aforementioned numerical problems.

Due to the stochastic nature of the algorithms used, the

problems were solved 10 times and in the following repre-
6 Numerical examples
sentative runs from each problem are presented.
The method used for the forward problem was imple-
The potential of the proposed method is demonstrated in
mented in a C++ code utilizing the Gmm++ library [77]
three numerical examples involving the detection of multiple
for linear algebra operations. The unstructured meshes used
cracks in solids of varying geometrical complexity.
were generated using the gmsh mesher [78] and results were
The forward problem is solved using topological enrich-
visualized using Paraview [79,80].
ment in order to reduce the computational cost associated
For the optimization algorithms the MATLAB ga function
with the numerical integration of the asymptotic enrichment
and the MATLAB implementation of the CMA-ES algorithm
functions. The XFEM variant used for the solution of the
[57,81] developed by the Koumoutsakos group (CSE Lab),
forward problems is still advantageous in this case since as
at ETH Zurich were used.
shown in Reference [56] it provides improved conditioning
and accuracy compared to standard XFEM.
6.1 Detection of two edge cracks in a unit cube
The additional meshes required to discretize the crack
front are automatically generated by dividing the circum-
The first example involves the detection of two edge cracks
ference of the candidate cracks in segments of equal length,
in a unit cube. The cube is fixed at one side and subjected to
the approximate length of those segments is set to 2h, where
a uniform load at the other side as illustrated in Fig. 7a. The
h is the mesh parameter. Since edge cracks are also con-
geometry parameters are defined as L x = L y = L z = 1 unit
sidered it is possible that several of the elements created lie
and the load has a unit value (P = 1 unit). Academic material
entirely outside the solid considered therefore resulting in
properties E = 200,000 units and ν = 0.2 are used. The cube
zero stiffness matrix entries which of course do not affect the
is meshed with a structured mesh consisting of 51 × 51 × 51
solution.
tetrahedral elements. A network of 5 × 5 sensors is assumed
For the evaluation of constraints, as mentioned in Sect. 4.2,
to be located in each free face of the cube (Fig. 7b)). The
the structure boundaries are represented using radial basis
strains measured by those sensors are simulated using a finer
functions based on biharmonic functions and linear polyno-
mesh of 101 × 101 × 101 tetrahedral elements. The location
mials. Moreover, parameter c used in the definition of the
of the cracks is shown in Fig. 10.
bounding boxes described in Sect. 4.3.1 is given the value
The RBF representation of the cube used for the evaluation
5h. This value may seem large compared to the one required
of constraints in created using a set of 10 × 10 points on
for standard XFEM in 2D, however the following factors
each edge of the cube. In Fig. 8 the zero iso surface of this
need to be taken into account:
RBF representation is illustrated. As can be seen the zero iso
surface is not an accurate representation of the boundaries
– In the XFEM variant used, the set of tip enriched nodes of the cube since it is only used to determine the relative
is larger than in standard XFEM since it involves nodes location of candidate cracks with respect to the structure.
belonging to elements which lie along the layer surround- The limits for the parameters used in the first step of the
ing the elements containing the crack front. Although optimization procedure of Sect. 5.3 were set to x0min = 0
those additional nodes and elements do not result in addi- units, x0max = 1 unit, y0min = 0 units, y0max = 1 unit,
tional dofs, they are considered enriched and therefore z 0min = 0 units, z 0max = 1 unit, amin = bmin = 0.15 units
have to be associated to one of the cracks. In addition, for and amax = bmax = 0.30 units. Three binary digits (8 possi-
curved crack fronts and unstructured meshes, additional ble values) where used for the representation of parameters
elements might be characterized as enriched due to some x0 , y0 , z 0 and two (4 possible values) for the rest of the
of their nodes being enriched. Therefore the distance was parameters. The maximum number of cracks allowed in the
extended to avoid such occurrences and to ensure the medium is set to four and through the use of topological vari-

123
Computational Mechanics (2018) 62:835–852 845

Fig. 7 Unit cube. a Geometry

and loading, b sensor locations.
The geometry and load
parameters are defined as
L x = L y = L z = 1 unit and
P = 1 unit

(a) (b)

Fig. 8 Unit cube, zero iso surface of the radial basis function used to
describe the cube boundaries for the imposition of constraints
Fig. 9 Convergence of the proposed multiscale strategy for the problem
of a unit cube with multiple cracks
ables can be adjusted to the actual number of cracks (two).
The population size was set to 40 individuals, the mutation
rate was set to 0.4 in order to prevent the algorithm from ations is illustrated. In Table 1 the actual and detected values
converging to local minima and the optimization was set to of the parameters describing the crack geometry are pro-
run for 2000 evaluations of the fitness function. vided.
For the second part of the procedure the default parameters
of the CMA-ES algorithm are adopted resulting in a popu- 6.2 Detection of three edge cracks in a beam under
lation of 12 individuals. The scaling of parameters defined three point bending
in Subsection 5.3.2 results in each variable assuming values
in the interval [−10, 10] which would require an initial step In this example a beam under three point bending,as illus-
size equal to σ0 = 6. However, since the initial values of trated in Fig. 11, is considered with edge cracks in three
the parameters should already be close to the actual solution different locations. The geometry parameters are defined as
the initial step size is given a smaller value equal to σ0 = 3. L x = 0.6 units, L y = 0.15 units, L z = 0.15 units and the
The maximum allowed difference between the two param- load is given a unit value (P = 1 unit). Academic material
eters of the ellipse is set to da = 0.10 units. The CMA-ES properties E = 200, 000 units and ν = 0.3 are used. A net-
algorithm is set to run for 2000 evaluations of the forward work of 4 × 8 sensors is assumed to be located in each of the
problem. long sides of the beam (Fig. 11).
In Fig. 9 the fitness function value achieved by the best The beam is meshed with an unstructured mesh consist-
individual of the population is given as a function of the ing of 68,439 tetrahedral elements and 14,039 nodes. For
number of evaluations of the fitness function, while in simulating measurements a finer mesh consisting of 491,244
Fig. 10 the best solution after different numbers of evalu- tetrahedral elements and 89,757 nodes is used. Both meshes

123
846 Computational Mechanics (2018) 62:835–852

Fig. 10 Best candidate solution

after different numbers of
evaluations for the problem of a
unit cube with multiple cracks

Table 1 Actual and detected values for the parameters describing crack
geometries for the problem of a unit cube with multiple cracks
Parameter Actual value Identified value
Crack 1 Crack 2 Crack 1 Crack 2

x0 0.95 − 0.05 1.0075 − 0.0751

y0 0.64 0.35 0.6337 0.3538
z0 0.65 0.40 0.6606 0.4062
φx 0.1667 π − 0.125 π 0.1625 π − 0.0060 π
φy 0.125 π 0 0.1320 π 0.8750 π
φz 0 0.5 π 0.02 π 0.01 π
a 0.29 0.33 0.4043 0.3320
b 0.29 0.29 0.2831 0.3340

Fig. 12 Meshes used for the three point bending problem. a Coarse
mesh for the solution of the forward problem and b fine mesh for sim-
ulating measurements

used are illustrated in Fig. 12 while the locations of the cracks

Fig. 11 Geometry, loading and sensor locations for the beam under
three point bending. The geometry and load parameters are defined as are shown in Fig. 14.
L x = 0.6 units, L y = 0.15 units, L z = 0.15 units and P = 1 unit The limits for the parameters used in the first step of the
optimization procedure were set to x0min = −0.3 units,

123
Computational Mechanics (2018) 62:835–852 847

the best solution after different numbers of evaluations is

shown.
As illustrated in Fig. 14, a quite accurate fit can be achieved
for all three cracks, nevertheless the number of evaluations
required (6000) would be prohibitive for larger models. In
addition, an increased number of evaluations was required
in the first step compared to the previous example due to the
increased number of cracks. However, the number of cracks
would not be known in the general case, therefore a large
number of evaluations (probably larger than the one used
herein) might be necessary.

6.3 Detection of two edge cracks in a wind turbine

Fig. 13 Convergence of the proposed multiscale strategy for the prob- blade
lem of a beam under three point bending
In the last example a more complicated geometry is used to
test the proposed scheme. More specifically, the geometry of
x0max = 0.3 unit, y0min = −0.075 units, y0max = 0.075 a wind turbine blade with two edge cracks is considered. It
unit, z 0min = −0.075 units, z 0max = 0.075 unit, amin = should be noted that the example is only of academic inter-
bmin = 0.04 units and amax = bmax = 0.08 units. For the est since several simplifications are made which render the
representation of parameters x0 , y0 and z 0 , 4, 2 and 2 binary problem quite unrealistic. The most important of those sim-
digits were used respectively, two (4 possible values) for the plifications are the following:
angles defining the plane of the ellipse and one for the ellipse
parameter. The population was set to 40 individuals and the – A uniform material is considered for the whole blade. In
mutation rate to 0.4 as in the previous example, however the reality the blade is hollow and made of a composite mate-
the optimization was set to run for 4000 evaluations of the rial, whose modeling complexity lies beyond the scope
fitness function due to the increased complexity. of this initial investigation.
For the second part of the procedure the default parameters – Static loading is considered.
of the CMA-ES algorithm are adopted and the algorithm is – The crack locations considered are not consistent with
set to run for 2000 evaluations of the forward problem. The the ones observed in actual turbine blades.
maximum allowed difference between the parameters of the
ellipse is set to da = 0.04 units. In Fig. 15 the geometry of the blade as well as the sen-
The fitness function value achieved by the best individual sor locations and applied boundary conditions are illustrated.
of the population is given as a function of the number of Sensors are placed following the geometry of the blade,
evaluations of the fitness function in Fig. 13, while in Fig. 14 moreover one end of the blade is considered fixed (Fig. 15)

Fig. 14 Best candidate solution

after different numbers of
evaluations for the problem of
beam under three point bending

123
848 Computational Mechanics (2018) 62:835–852

unit, z 0min = 0.2 units, z 0max = 1.6 unit, amin = bmin =

0.08 units and amax = bmax = 0.20 units. For the repre-
sentation of parameters x0 , y0 , z 0 1, 2 and 4 binary digits
(2, 4 and 16 possible values) were used respectively while
two digits (4 possible values) were used for the rest of the
parameters. In Fig. 17 all possible crack locations resulting
from the above parameters are depicted. As in the previous
example the possible number of cracks was set to four, the
population size was set to 40 individuals, the mutation rate
was set to 0.4 and the optimization was set to run for 2000
evaluations of the fitness function.
The default parameters of the CMA-ES algorithm are
again adopted for the second part of the procedure. The max-
imum allowed difference between the two parameters of the
ellipse is set to da = 0.10 units.
In Fig. 18 the fitness function value achieved by the best
individual of the population is given as a function of the num-
ber of evaluations of the fitness function, while in Fig. 19 the
optimal solution after successive evaluations is illustrated. It
Fig. 15 Wind turbine blade geometry, sensor locations and boundary
should be noted that because of the more complicated geome-
conditions try, the whole ellipses are plotted rather than only the parts of
the ellipses that lie within the structure as in previous exam-
ples. In Fig. 20 the deformed shape of the blade with the actual
while a uniform pressure P = 10 is applied in a small area
and predicted cracks is plotted and in Fig. 20 the deformed
on the edge of the other end (Fig. 15). Academic material
shape of the blade is given with the actual and the detected
properties E = 200,000 units and ν = 0.3 units were used.
cracks. Although the accuracy is decreased compared to the
Two unstructured meshes were used for the problem, a
previous examples, the number of cracks and rough locations
fine mesh for the simulation of the measured response of
can still be obtained. This reduced accuracy can be attributed
the blade consisting of 1,154,327 linear tetrahedral elements
mostly to the fact that the applied loading does not activate
and 212,325 nodes (Fig. 16b), and a coarser mesh for the
both cracks equally making it harder to accurately detect the
solution of forward problems consisting of 174,580 elements
upper crack.
and 36,325 nodes (Fig. 16a).
It should be remarked that due to the increased complex-
The limits for the parameters used in the first step of the
ity of the present problem and the stochastic nature of the
multiscale scheme of Sect. 5.3 were set to x0min = −0.035
optimization procedure it is not always possible to detect
units, x0max = 0.05 unit, y0min = −0.12 units, y0max = 0.12

Fig. 16 Meshes used for the

wind turbine blade problem. a
Coarse mesh for the solution of
forward problems and b fine
mesh for simulating
measurements

(a) (b)

123
Computational Mechanics (2018) 62:835–852 849

Fig. 17 Possible crack locations for the first step of the multiscale crack
detection scheme for the case of a wind turbine blade
Fig. 19 Best candidate solution after different numbers of evaluations
for the problem of a wind turbine blade with two edge cracks

Fig. 18 Convergence of the proposed multiscale strategy for the prob-

lem of a wind turbine blade with multiple cracks
Fig. 20 Deformed shape of the blade with the actual (left) and detected
cracks (right)
both of the cracks, this is illustrated in Fig. 21 where the
best candidate obtained at the first step of the procedure is cases, however in future works those cases can be dealt with
given for alternative runs. More specifically, in the first case either by locating and penalizing those cracks or by including
(Run 1) both cracks are detected while in the following two contact which would prevent negative crack openings.
cases (Run 2 and Run 3) only one of the cracks is accurately
detected. The second run in particular is of special interest
since one of the detected cracks (the upper crack) would 7 Conclusions
result in zero or negative crack opening displacements and
therefore would be physically meaningless. In the present A methodology for the detection of multiple cracks in 3D
version of the method no particular care was taken for those solids of arbitrary geometries was presented, resulting via

123
850 Computational Mechanics (2018) 62:835–852

The proposed method offers a highly promising tool

towards the accurate detection of multiple cracks in com-
plex engineered systems, simulated in the three dimensional
domain.

Acknowledgements Stéphane Bordas thanks funding provided by

the European Research Council Starting Independent Research Grant
(ERC Stg Grant agreement No. 279578)“RealTCut towards real
time multiscale simulation of cutting in non-linear materials with
applications to surgical simulation and computer guided surgery”.
Stéphane Bordas also thanks the financial support of the Fonds
National de la Recherche Luxembourg FNRS-FNR grant INTER/
FNRS/15/11019432/EnLightenIt/Bordas and of the Engineering and
Physical Sciences Research Council EP/G042705/1. Eleni Chatzi would
like to acknowledge the support of the Swiss National Science Founda-
tion under research Grant # 200021_153379 “A Multiscale Hysteretic
XFEM Scheme for the Analysis of Composite Structures”.
Fig. 21 Best candidate from the first step of the solution for three alter- Open Access This article is distributed under the terms of the Creative
native runs Commons Attribution 4.0 International License (http://creativecomm
ons.org/licenses/by/4.0/), which permits unrestricted use, distribution,
and reproduction in any medium, provided you give appropriate credit
fusion of a recently introduced XFEM variant [55,56] with to the original author(s) and the source, provide a link to the Creative
a multiscale optimization strategy. The latter comprises a Commons license, and indicate if changes were made.
discrete step, where genetic algorithms are employed, and a
continuous step employing the CMA-ES algorithm [57].
The method was tested in numerical examples involv-
ing the detection of multiple cracks in solids of non-regular References
geometries and promising results were obtained. Neverthe-
less, before the method may be implemented onto practical 1. Farrar CR, Worden K (2012) Structural health monitoring: a
machine learning perspective. John Wiley & Sons, Ltd, Hoboken
problems several improvements have been identified as
2. Carden EP, Fanning P (2004) Vibration based condition monitor-
future work, for alleviating certain methodological limita- ing: a review. Struct Health Monit 3(4):355–377
tions. More specifically: 3. Inaudi D (2010) Long-term static structural health monitoring. In:
ASCE structures congress. Orlando
4. Chang PC, Liu S C (2003) Recent research in nondestructive eval-
uation of civil infrastructures. J Mater Civ Eng 15(3):298–304
– The computational cost associated with the solution of 5. Grosse CU, Ohtsu M (2008) Acoustic emission testing. Springer,
the forward problems is high, which in turn increases the Berlin
total computational cost since those problems need to be 6. Catbas FN, Susoy M, Frangopol DM (2008) Structural health mon-
itoring and reliability estimation: long span truss bridge application
solved thousands of times. This could be a prohibiting
with environmental monitoring data. Eng Struct 30(9):2347–2359
factor for several applications, thus special techniques, 7. Moaveni B, Stavridis A, Lombaert G, Conte JP, Shing PB (2012)
such as model order reduction [24,25,82], would be Finite-element model updating for assessment of progressive dam-
required to extend the method’s applicability. age in a 3-story infilled RC frame. J Struct Eng 139(10):1665–1674
8. Neerukatti RK, Liu KC, Kovvali N, Chattopadhyay A (2014)
– In some cases the cracks detected by the method are phys- Fatigue life prediction using hybrid prognosis for structural health
ically meaningless since they involve zero or negative monitoring. J Aerosp Inf Syst 11(4):211–232
crack opening displacements. This can be dealt with by 9. Saadat S, Noori MN, Buckner GD, Furukawa T, Suzuki Y (2004)
detecting and penalizing those cracks or by including Structural health monitoring and damage detection using an intel-
ligent parameter varying (IPV) technique. Int J Non-Linear Mech
contact in the model of the forward problem. 39(10):1687–1697
– Despite adoption of a rather high number of sensors, the 10. Adams D (2007) Health monitoring of structural materials and
inverse problem may result as ill-conditioned, especially components: methods with applications. Wiley, Hoboken
for the more complex geometry and loading conditions 11. Friswell MI (2007) Damage identification using inverse methods.
Philos Trans R Soc Lond A Math Phys Eng Sci 365(1851):393–410
of the third problem tested. A possible remedy to this 12. He J, Yang J, Wang Y, Waisman H, Zhang W (2016) Probabilistic
problem could result via use of multiple loading cases, model updating for sizing of hole-edge crack using fiber bragg
as in the work of Rabinovich et al. [41]. Such a rem- grating sensors and the high-order extended finite element method.
edy would only be possible once the size of the forward Sensors 16(11):1956
13. Rappel H, Beex LAA, Hale JS, Bordas SPA (2016) Bayesian
problems has been reduced, as mentioned above, since it inference for the stochastic identification of elastoplastic mate-
would further increase the total number of evaluations of rial parameters: introduction, misconceptions and insights. arXiv
the forward problem. preprint arXiv:1606.02422

123
Computational Mechanics (2018) 62:835–852 851

14. Xu ZD, Wu Z (2007) Energy damage detection strategy based on 33. Sukumar N, Chopp DL, Moës N, Belytschko T (2000) Modeling
acceleration responses for long-span bridge structures. Eng Struct holes and inclusions by level sets in the extended finite element
29(4):609–617 method. Comput Methods Appl Mech Eng 190(46–47):6183–6200
15. Abdalla MO, Zimmerman DC, Grigoriadis KM (1999) Structural 34. Benowitz BA, Waisman H (2013) A spline-based enrichment func-
damage detection using strain data via linear matrix inequality tion for arbitrary inclusions in extended finite element method
based methods. In: American control conference, 1999. Proceed- with applications to finite deformations. Int J Numer Methods Eng
ings of the 1999, vol 2. IEEE, pp 1114–1118 95(5):361–386
16. Ostachowicz WM (2008) Damage detection of structures using 35. Jung J, Taciroglu E (2014) Modeling and identification of an arbi-
spectral finite element method. Comput Struct 86(3–5):454–462 trarily shaped scatterer using dynamic XFEM with cubic splines.
17. Park S, Yun CB, Roh Y, Lee JJ (2005) Health monitoring of steel Comput Methods Appl Mech Eng 278:101–118
structures using impedance of thickness modes at PZT patches. 36. Maringer DG (2006) Portfolio management with heuristic opti-
Smart Struct Syst 1(4):339–353 mization, vol 8. Springer, Berlin
18. Spiridonakos MD, Chatzi EN (2015) Metamodeling of dynamic 37. Hunaidi O (1998) Evolution-based genetic algorithms for analysis
nonlinear structural systems through polynomial chaos NARX of non-destructive surface wave tests on pavements. NDT & E Int
models. Comput Struct 157:99–113 31(4):273–280 (Non-destructive Testing in Civil Engineering)
19. Papadimitriou C, Angelikopoulos P, Koumoutsakos P, Papadioti 38. Farley SJ, Durodola JF, Fellows NA, Hernndez-Gmez LH (2012)
DC (2013) Efficient techniques for bayesian inverse modeling of High resolution non-destructive evaluation of defects using artifi-
large-order computational models. In: Safety, reliability, risk and cial neural networks and wavelets. NDT & E Int 52:69–75
life-cycle performance of structures and infrastructures: proceed- 39. Lee W, Sun C, Chiu C (2014) Nondestructive evaluation of buried
ings of the 11th International Conference on Structural Safety and dielectric cylinders by asynchronous particle swarm optimization.
Reliability, New York, USA, 16–20 June 2013. CRC Press J Test Eval 43(1):212–220
20. Triantafyllou SP, Chatzi EN (2014) A hysteretic multiscale for- 40. Bernieri A, Ferrigno L, Laracca M, Molinara M (2008) Crack shape
mulation for nonlinear dynamic analysis of composite materials. reconstruction in eddy current testing using machine learning sys-
Comput Mech 54(3):763–787 tems for regression. IEEE Trans Instrum Meas 57(9):1958–1968
21. Papadimitriou C, Papadioti DC (2013) Component mode synthe- 41. Rabinovich D, Givoli D, Vigdergauz S (2007) XFEM-based crack
sis techniques for finite element model updating. Comput Struct detection scheme using a genetic algorithm. Int J Numer Meth Eng
126:15–28 71(9):1051–1080
22. Wang Y, Waisman H (2015) Progressive delamination analysis 42. Rabinovich D, Givoli D, Vigdergauz S (2008) Crack identification
of composite materials using XFEM and a discrete damage zone by arrival time using XFEM and a genetic algorithm. Int J Numer
model. Comput Mech 55(1):1–26 Methods Eng 77(3):337–359
23. Kerfriden P, Gosselet P, Adhikari S, Bordas SPA (2011) Bridg- 43. Waisman H, Chatzi E, Smyth AW (2010) Detection and quantifi-
ing proper orthogonal decomposition methods and augmented cation of flaws in structures by the extended finite element method
Newton–Krylov algorithms: an adaptive model order reduction and genetic algorithms. Int J Numer Meth Eng 82(3):303–328
for highly nonlinear mechanical problems. Comput Methods Appl 44. Chatzi EN, Hiriyur B, Waisman H, Smyth AW (2011) Experimental
Mech Eng 200(5):850–866 application and enhancement of the XFEM–GA algorithm for the
24. Kerfriden P, Passieux JC, Bordas SPA (2012) Local/global model detection of flaws in structures. Comput Struct 89(7):556–570
order reduction strategy for the simulation of quasi-brittle fracture. 45. Sun H, Waisman H, Betti R (2013) Nondestructive identification of
Int J Numer Meth Eng 89(2):154–179 multiple flaws using XFEM and a topologically adapting artificial
25. Kerfriden P, Goury O, Rabczuk T, Bordas SPA (2013) A parti- bee colony algorithm. Int J Numer Methods Eng 95(10):871–900
tioned model order reduction approach to rationalise computational 46. Sun H, Waisman H, Betti R (2014) A multiscale flaw detection
expenses in nonlinear fracture mechanics. Comput Methods Appl algorithm based on XFEM. Int J Numer Methods Eng 100(7):477–
Mech Eng 256:169–188 503
26. Goury O, Amsallem D, Bordas SPA, Liu WK, Kerfriden P 47. Sun H, Waisman H, Betti R (2015) A sweeping window method for
(2016) Automatised selection of load paths to construct reduced- detection of flaws using an explicit dynamic XFEM and absorbing
order models in computational damage micromechanics: from boundary layers. Int J Numer Methods Eng 105(13):1014–1040
dissipation-driven random selection to bayesian optimization. 48. Yan G, Sun H, Waisman H (2015) A guided bayesian inference
Comput Mech 58(2):213–234 approach for detection of multiple flaws in structures using the
27. Rahimabadi A Akbari, Kerfriden P, Bordas S, Bordas S (2015) extended finite element method. Comput Struct 152:27–44
Scale selection in nonlinear fracture mechanics of heterogeneous 49. Nanthakumar SS, Lahmer T, Rabczuk T (2013) Detection of flaws
materials. Philos Mag 95(28–30):3328–3347 in piezoelectric structures using extended FEM. Int J Numer Meth-
28. Kerfriden P, Schmidt KM, Rabczuk T, Bordas SPA (2013) Sta- ods Eng 96(6):373–389
tistical extraction of process zones and representative subspaces 50. Nanthakumar SS, Lahmer T, Rabczuk T (2014) Detection of mul-
in fracture of random composites. Int J Multiscale Comput Eng tiple flaws in piezoelectric structures using XFEM and level sets.
11(3):253–287 Comput Methods Appl Mech Eng 275:98–112
29. Belytschko T, Black T (1999) Elastic crack growth in finite ele- 51. Zhang C, Nanthakumar SS, Lahmer T, Rabczuk T (2017) Mul-
ments with minimal remeshing. Int J Numer Methods Eng 620(July tiple cracks identification for piezoelectric structures. Int J Fract
1998):601–620 206(2):151–169
30. Moës N, Dolbow J, Belytschko T (1999) A finite element method 52. Ma C, Yu T, Van Lich L, Bui TQ (2017) An effective computa-
for crack growth without remeshing. Int J Numer Meth Eng tional approach based on XFEM and a novel three-step detection
46(1):131–150 algorithm for multiple complex flaw clusters. Comput Struct
31. Sukumar N, Moës N, Moran B, Belytschko T (2000) Extended 193:207–225
finite element method for three-dimensional crack modelling. Int J 53. Alalade M, Nguyen-Tuan L, Wuttke F, Lahmer T Damage identi-
Numer Methods Eng 48(November 1999):1549–1570 fication in gravity dams using dynamic coupled hydro-mechanical
32. Daux C, Moës N, Dolbow J, Sukumar N, Belytschko T (2000) XFEM. Int J Mech Mater Des 1–19. https://doi.org/10.1007/
Arbitrary branched and intersecting cracks with the extended finite s10999-017-9367-4
element method. Int J Numer Meth Eng 48(12):1741–1760

123
852 Computational Mechanics (2018) 62:835–852

54. Pirboudaghi S, Tarinejad R, Alami MT (2017) Damage detection 69. Ventura G, Gracie R, Belytschko T (2009) Fast integration and
based on system identification of concrete dams using an extended weight function blending in the extended finite element method.
finite element–wavelet transform coupled procedure. J Vib Control. Int J Numer Methods Eng 77(July 2008):1–29
https://doi.org/10.1177/1077546317722428 70. Loehnert S, Mueller-Hoeppe DS, Wriggers P (2011) 3D corrected
55. Agathos K, Chatzi E, Bordas SPA, Talaslidis D (2015) A well- XFEM approach and extension to finite deformation theory. Int J
conditioned and optimally convergent XFEM for 3D linear elastic Numer Methods Eng 86(October 2010):431–452
fracture. Int J Numer Methods Eng 105(9):643–677 71. Carr JC, Beatson RK, Cherrie JB, Mitchell TJ, Fright WR, McCal-
56. Agathos K, Chatzi E, Bordas S (2016) Stable 3D extended finite lum BC, Evans TR (2001) Reconstruction and representation of
elements with higher order enrichment for accurate non planar frac- 3D objects with radial basis functions. In: Proceedings of the
ture. Comput Methods Appl Mech Eng 306:19–46 28th annual conference on Computer graphics and interactive tech-
57. Hansen N, Muller SD, Koumoutsakos P (2003) Reducing the time niques. ACM, pp 67–76
complexity of the derandomized evolution strategy with covariance 72. Gottschalk S, Lin MC, Manocha D (1996) OBBtree: A hierarchi-
matrix adaptation (CMA-ES). Evol Comput 11(1):1–18 cal structure for rapid interference detection. In Proceedings of
58. Agathos K, Chatzi E, Bordas S (2016) 3D crack detection using the 23rd annual conference on Computer graphics and interactive
an XFEM variant and global optimization algorithms. In: Proceed- techniques. ACM, pp 171–180
ings of the 9th International Conference on Fracture Mechanics of 73. Holland JH (1992) Adaptation in natural and artificial systems:
Concrete and Concrete Structures FraMCoS-9 an introductory analysis with applications to biology, control and
59. Glisic B, Inaudi D (2008) Fibre optic methods for structural health artificial intelligence. MIT Press, Cambridge
monitoring. Wiley, Hoboken 74. Goldberg DE (1989) Genetic algorithms in search, optimization
60. Moës N, Gravouil A, Belytschko T (2002) Non-planar 3D crack and machine learning, 1st edn. Addison-Wesley Longman Pub-
growth by the extended finite element and level sets-part I: mechan- lishing Co. Inc, Boston
ical model. Int J Numer Methods Eng 53(11):2549–2568 75. Banzhaf Wolfgang, Francone Frank D, Keller Robert E, Nordin
61. Gravouil A, Moës N, Belytschko T (2002) Non-planar 3D crack Peter (1998) Genetic programming: an introduction: on the auto-
growth by the extended finite element and level sets-part II: level matic evolution of computer programs and its applications. Morgan
set update. Int J Numer Methods Eng 53(11):2569–2586 Kaufmann Publishers Inc., San Francisco
62. Duflot M (2007) A study of the representation of cracks with level 76. Auger A, Hansen N (September 2005) A restart CMA evolution
sets. Int J Numer Methods Eng 70(November 2006):1261–1302 strategy with increasing population size. In: 2005 IEEE congress
63. Melenk JM, Babuska I (1996) The partition of unity finite ele- on evolutionary computation, vol 2. pp 1769–1776
ment method: basic theory and applications. Comput Methods Appl 77. Renard Y, Pommier J (2004–2014) Gmm++ library. http://
Mech Eng 139(1–4):289–314 download.gna.org/getfem/html/homepage/gmm/index.html
64. Stazi FL, Budyn E, Chessa J, Belytschko T (2003) An extended 78. Geuzaine C, Remacle JF (2009) Gmsh: A 3-D finite element mesh
finite element method with higher-order elements for curved generator with built-in pre-and post-processing facilities. Int J
cracks. Comput Mech 31(1–2):38–48 Numer Methods Eng 79(11):1309–1331
65. Laborde P, Pommier J, Renard Y, Salaün M (2005) High-order 79. Ahrens J, Geveci B, Law C, Hansen CD, CR Johnson (2005) An
extended finite element method for cracked domains. Int J Numer end-user tool for large-data visualization. Paraview
Methods Eng 64(3):354–381 80. Ayachit U (2015) The paraview guide: a parallel visualization appli-
66. Béchet E, Minnebo H, Moës N, Burgardt B (2005) Improved imple- cation. https://dl.acm.org/citation.cfm?id=2789330
mentation and robustness study of the X-FEM for stress analysis 81. Hansen N, Kern S (2004) Evaluating the CMA evolution strategy
around cracks. Int J Numer Methods Eng 64(8):1033–1056 on multimodal test functions. In: Parallel problem solving from
67. Menk A, Bordas SPA (2011) A robust preconditioning technique nature-PPSN VIII. Springer, pp 282–291
for the extended finite element method. Int J Numer Methods Eng 82. Niroomandi S, Alfaro I, Gonzalez D, Cueto E, Chinesta Francisco
85(October 2010):1609–1632 (2012) Real-time simulation of surgery by reduced-order model-
68. Fries TP (2008) A corrected XFEM approximation without prob- ing and X-FEM techniques. Int J Numer Methods Biomed Eng
lems in blending elements. Int J Numer Methods Eng 75(November 28(5):574–588
2007):503–532

123