Sie sind auf Seite 1von 23

1523_Frame_C12.

fm Page 427 Thursday, August 28, 2003 5:54 PM

12
Inverse Detection of Flaws in Structures

Computational inverse techniques for flaw detection in beams or plates with


applications to sandwich structures are introduced in this chapter. Sandwich
structures are a special class of laminates that require special considerations
and treatments. Two approaches are introduced for the characterization of
flaws in sandwich structures; in both the finite element model is used for
the forward analysis. The first approach is the parameter identification to
determine the presence, location, size, and degree of the damage of flaws in
the core layer; the GA is used for the inverse analysis. The second approach
is to relate the flaw with the element stiffness factors; gradient-based meth-
ods, such as Newton’s root finding method as well as the Levenberg–Mar-
quardt method, are used for the inverse analysis. A number of numerical
examples are provided to demonstrate the application of these computa-
tional inverse techniques. It is also revealed that gradient-based methods are
much more efficient in dealing with inverse problems with large numbers
of parameters.

12.1 Introduction
Damages of flaws in any structures are always a big concern for structure
systems. Sandwich structures have been widely used in various industrial
applications, especially in the aerospace and ship building industries. Sand -
wich structures are a special class of laminates that consist of three layers:
two thin, high-strength, stiff face (outer) layers and one thick, low-density,
flexible inner core layer. Structural efficiency in terms of economy, high
stiffness, and low weight is achieved by combining the stronger facings with
a thicker, lightweight core material. Structural reliability is highly dependent
on the support of the core material, and the stiffness and strength of sand-
wich structures is affected greatly by failure of the core material. Usually,
the core material is more easily damaged than the facings during the man-
ufacturing process and in practical use by local concentrated loads and/or

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 428 Thursday, August 28, 2003 5:54 PM

impacted loads. It is thus very important to locate the flaw and detect its
degree of damage.
Nondestructive testing using ultrasonic techniques plays a very important
role in detecting flaws in structures, determining the flaw according to the
echogram of the ultrasonic bulk waves. It is often not efficient for the case
of sandwich plates due to the material inhomogeneous layered structures.
Also, it is necessary to scan the whole surface of the structure to locate the
flaws. Liu and his co-workers (Liu and Lam, 1994; Liu and Achenbach, 1995;
Liu et al., 1995b, 1996; Lam et al., 1997; Wang et al., 1998) applied the SEM
method (see Section 10.2.1) to investigate the scattering of lamb waves by
rectangular flaws in sandwich plates and anisotropic laminated plates, where
the plates were treated as a plane problem and the scattered wave field in
the frequency domain computed for characterization of the flaws. Another
very important direct methodology is to determine the damage quantities
from changes in dynamic properties of a structure using the measurements
of natural fundamental frequencies shift and modal shapes change in such
structures (Salawu, 1997; Doebling et al., 1996). However, for local small
flaws in a sandwich structure, the changes in lower natural fundamental
frequencies and modal shapes are too small to be detected, especially with
the presence of the measurement error.
The harmonic response of the plate structures excited with a load of a certain
frequency can be significantly different for damaged and undamaged plates.
Practically, it is not too difficult to excite the structure with a harmonic load
and measure the corresponding response at certain points using the available
testing equipment. The recorded responses at certain positions of the plates
to the excitation are different for the plates with different flaws. However, it
is very difficult and complicated to find the direct relationship between the
flaw parameters and the responses to the excitation. Therefore, computational
inverse techniques are required to determine the quantities (parameters) of
the flaw based on the response of structures to the dynamic excitation.
In recent years, Liu and Chen (2000, 2001, 2002) have proposed several
computational inverse techniques to detect flaws in sandwich structures.
The genetic algorithm and Newton’s root finding methods are employed in
the inverse procedure. Based on their study, several computational inverse
techniques are presented in this chapter to detect flaws in sandwich beams
and plates quantitatively using the response of structures to the given har-
monic excitation.

12.2 Inverse Identification Formulation


Consider a general finite element model of a linear-elastic structure. The
dynamic governing equation found in many textbooks on solid mechanics
or FEM (see, e.g., Liu and Quek, 2003) is given for a harmonic excitation:

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 429 Thursday, August 28, 2003 5:54 PM

[K − ω 2 M]w = F (12.1)

where K and M are the stiffness and mass matrices, respectively, ω is the
frequency of the harmonic excitation, w is the vector of nodal displacement
amplitude, and F the vector of the amplitudes of the forces.
In the next sections, two approaches for the characteristics of flaw in
structures, i.e., damaged element identification and stiffness factor identifi-
cation, are introduced.

12.2.1 Damaged Element Identification


Based on the finite element model of a sandwich structure, the unknown
parameters are defined as

• A reference element number, nr , indicating the reference location of


the damaged area
• Damaged element numbers and/or damage profile cases, nd , deter-
mining the type of the damaged area
• The damage factor β f , indicating that the Young’s modulus of the
damaged layer in these elements is reduced to

Ekf = (1 − β f )Ek (12.2)

The input of the inverse analysis is taken as the responses (deflection) sam-
pled at certain points of the structure (Liu and Chen, 2001) excited using a
time-harmonic point load with frequency ω.
To predict the response of the damaged plate, the stiffness matrix of the
entire structure is updated according to the changes in the element stiffness
matrix arising from deduction of the modulus of the damaged layer. How-
ever, the internal change of a structure generally does not result in a loss of
material; therefore, it is assumed that the mass matrix does not change. In
the forward analysis, the stiffness matrix is updated for different trials of
different parameters of flaw. The responses are then solved for constructing
the error function for the inverse analysis.
The objective function is constructed using the sum of squares of the
difference between the responses computed for the plate with actual damage
parameters and those with the assumed parameters by FEM model in the
form of L2 norm as defined in Equation 2.123:

ns

ferr (p) = ∑ {w (p) − w


i =1
c
i
m
i (pt )} 2 (12.3)

where p is the vector of unknown parameters (nr , nd , β f ) describing the flaw


characteristics defined previously, ns is the number of points where deflec-

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 430 Thursday, August 28, 2003 5:54 PM

tions are sampled for inverse analysis, wic is the computed deflection of the
structure with trial damage parameters, and wim the deflection of the plate
with true damage parameters at the sampling point. Constraints to the
parameters are given for different finite element models accordingly, which
will form the search space.
To determine the unknown parameters of the flaw in inverse analysis, it
is necessary to minimize the object function computationally with the FEM
for forward analysis.

12.2.2 Stiffness Factor Identification


The global stiffness matrix of a structure is the assembly of all the elements’
stiffness matrices. For isotropic elastic material, the element stiffness matrix
is always proportional to the elastic modulus of the material and the
geometric coefficient, which are unknown parameters in inverse analysis.
Thus the global stiffness matrix can often be expressed as (see, e.g., Liu
and Chen, 2002)

ne

K= ∑e K
i =1
i i
e
(12.4)

where ne is the total number of elements, ei (i = 1, 2,  , ne ) are the unknown


parameters of elastic modulus or element stiffness factor, and the element
stiffness K i e is obtained by assuming that the element is perfect with a unit
factor. Therefore, the element stiffness factor, ei (i = 1, 2,  , ne ) , reflects the
degree of damage in the element in the damaged structure. Substituting
Equation 12.4 into Equation 12.1 yields

 ne




i =1
ei K ie − ω 2 M w = F

(12.5)

Note that ω and F are known for given excitation force. Mass matrix M is
known for given material for the structure and the FE mesh. K i e is also
known once the finite element mesh is given. For an assumed set of ei , w
can be therefore computed without any difficulty. It is also clear that Equa-
tion 12.5 is a linear function of the parameters ei in an explicit form. This
form facilitates efficient computation of the gradients required in the gradi-
ent-based search techniques.

• In the forward analysis, the displacement response of a finite element


system can be predicted using Equation 12.5 with a given set of
parameters ei (i = 1, 2,  , ne ) .

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 431 Thursday, August 28, 2003 5:54 PM

• In the inverse analysis, the parameters need to be identified using


the measured value of the displacement response or modal param-
eters. That is, parameters are chosen to best fit the experiment data.
Two methods are used to fit these data:
• The least squares method to minimize the error
• The sensitivity-based analysis method, which has different for-
mulations for different problems

12.2.2.1 Objective Function with Weight


The objective function is defined using the weighted sum of squared differ-
ences between the measured data and the corresponding simulated value of
the dynamic properties of structures.

ns

ferr (e) = ∑W ( f (e) − f


i =1
i i
c
i
m
(e t ))2 (12.6)

where e is the vector of unknown parameters (e1 , e 2 ,  , ene )T , fim is the


measured value, and fic is the corresponding simulated value using a trial
e, and Wi is the weight factor used to provide some measurements with more
or less weight. The measured values of a structure can be responses, natural
frequencies, and values of modal assurance criterion (MAC) (Ewins, 1985)
related to the mode shapes.

12.2.2.2 Direct Formulation


A direct formulation is employed by Liu and Chen (2002) in determining
stiffness factors using harmonic response. For a finite element model with
ne elements, ns displacements at different nodes on the structure can be
measured and expressed in a vector form of w . The identification problem
is to determine the element stiffness factor vector e in Equation 12.5 using
the measured response w , i.e., to find e that satisfies the following equation
based on the simple matrix operation:

Qw = w (12.7)

where Q is a constant row vector with elements of zeros or ones that can
always be formed to select the degrees of freedom corresponding to the
measured displacement components. For example, if w = {w3 , w5 , w8 }T , the
row vector Q should be

Q = {0 0 1 0 1 0 0 1 0  0} (12.8)

Vector w is solved from Equation 12.5 for a given e.

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 432 Thursday, August 28, 2003 5:54 PM

Define an error function of

f(e) = Qw − w = 0 (12.9)

where

 f1 (e) 
   e1 
 f2 (e)  e 
   2
f(e) =  , e=  (12.10)
 
   
  ene 
 fns (e)  

Here, f(e) is a set of nonlinear implicit equations with respect to the param-
eters. The value of f(e) and its derivation can be evaluated making use of
the linear property of Equation 12.5. Thus, the solution of the Equation 12.9
can be found directly using a root finding method numerically.

12.3 Use of Uniform µ GA


The forward analysis for obtaining the response of the sandwich plate with
different flaws is conducted using an FEM code developed based on the
Mindlin theory for laminated plates. The flaw in the core layer of the sand-
wich plates is characterized based on the finite element level with damage
represented as a deduction in modulus of the material in the damaged layer.
In the following inverse analysis, a uniform µGA is employed to solve the
optimization problem as defined by Equation 12.3 to find the actual param-
eters of the flaws in sandwich structures. The µGA uses a population size
of 5, tournament selection, no mutation, niching, elitism, and uniform cross-
over of pcross = 0.5.

12.3.1 Example I: Sandwich Beam


Consider a simply supported sandwich beam shown in Figure 12.1. The
material constants and size of the beam are given in Table 12.1. The beam is
discretized into 20 elements, and it is assumed that the elements 14, 15, 16,
and 17 are the damaged elements containing a flaw in the core layer with a
damage factor βf = 0.5. The Young’s modulus of the core material for the
elements is computed using Equation 12.2. A time-harmonic load with fre-
quency ω = 3000 rad/s is applied to excite the beam. The responses are
sampled at five points (marked with dots in the figure). The unknown

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 433 Thursday, August 28, 2003 5:54 PM

Element i
z
14 15 16 17 x

y
x

FIGURE 12.1
Simply supported sandwich beam with a flaw. The beam is divided into 20 elements. The
r
elements numbered 14, 15, 16, and 17 are damaged elements containing a flaw in the core layer.
A time-harmonic load with frequency ω = 3000 ad/s is applied to excite the beam. The
responses are sampled at five points marked with dots in the figure. (From Liu, G.R. and Chen,
S.C, Comput. Methods Appl. Mech. Eng., 190, 5505–5514, 2001. With permission.)

TABLE 12.1
Material Constants and Geometric Size of Sandwich Beam
E ν Densit y Length Width Thickness
Layer 1 16.7 GPa 0.3 1760kg/m3 — 0.005 m
Layer 2 (core) 13.0 GPa 0.3 1000 kg/m3 1m 0.05 m 0.01 m
Layer 3 16.7 GPa 0.3 1760 kg/m3 — 0.005 m
Source: Liu, G.R. and Chen, S.C., Comput. Methods Appl. Mech. Eng., 190, 5505–5514,
2001. With permission.

parameters set to be identified are the reference element number to indicate


the location of the flaw, the numbers of the damaged element to represent
the size of the flaw, and degree of the damage to describe the severity of the
damage, respectively. The range of the reference element number is set to
be 1 to 16, having 16 possibilities. The number of damaged elements ranges
from 1 to 4 with four possibilities. The degree of the damage is in the range
of 0.1 to 0.8, with eight possibilities. Thus, the searching space has a total of
512 possibilities. The characteristics of the actual flaw were successfully
detected after 30 generations with a total of 150 function evaluations that
invoke the FEM code. Figure 12.2 shows the convergence process of the
fitness of the best individual of the population in each generation.

12.3.2 Example II: Sandwich Plate


The second example considered is a simply supported square sandwich
plate. The material constants and size of the plate are given in Table 12.2.
Figure 12.3 shows the finite element model (discretized into 100 elements)
of the plate. The plate is assumed to contain a flaw in shaded elements with
a damage factor βf = 0.5 in the core layer. A time-harmonic load with fre-
quency ω = 3000 rad/s is applied at the center of the plate (marked with a
circle) to excite the plate. The responses are sampled at six points (marked
with dots in the figure). The reference element number that indicates the

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 434 Thursday, August 28, 2003 5:54 PM

Generation

0.00E+00
0 1
10 20 30 40 50

Fitness
-1.00E-01

-2.00E-01

-3.00E-01

FIGURE 12.2
Fitness value of the best individual for the simply supported beam shown in Figure 12.1.
(From Liu, G.R. and Chen, S.C., Comput. Methods Appl. Mech. Eng., 190, 5505–5514, 2001. With
permission.)

TABLE 12.2
Material Constants and Geometric Size of Sandwich Square Plate
E ν Density Length of Sides Thickness
Layer 1 16.7 GPa 0.3 1760 kg/m3 — 0.005 m
Layer 2 (core) 13.0 GPa 0.3 1000 kg/m3 0.6 m 0.01 m
Layer 3 16.7 GPa 0.3 1760 kg/m3 — 0.005 m
Source: Liu, G.R. and Chen, S.C, Comput. Methods Appl. Mech. Eng., 190, 5505–5514,
2001. With permission.

FIGURE 12.3
Simply supported square sandwich plate with a flaw (shaded) in the core layer. The plate is
discretized into 100 plate elements. A time-harmonic load with frequency ω = 3000 rad/s is
applied at the center of the plate. The frequency responses are sampled at six points marked
with dots in the figure. (From Liu, G.R. and Chen, S.C., Comput. Methods Appl. Mech. Eng., 190,
5505–5514, 2001. With permission.)

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 435 Thursday, August 28, 2003 5:54 PM

location of the damage can vary from 1 to 128. Figure 12.4 gives four possible
damage patterns. The degree of the damage is in the range of 0.1 to 0.8
discretized into eight grades. To those individuals with parameters exceed-
ing the domain of the problem, such as the element number of more than
100 in this example, no evaluation is performed but a small value of fitness
is assigned to throw out these individuals. It takes 244 generations or
244 × 5 = 1220 function evaluations that invoke the FEM code to converge
to the true solution. The fitness of the best individual in each generation is
shown in Figure 12.5. Due to the random nature of the GA, it converges to
the true solution slowly after the best individual in the population falls into
the region near the true optima.
To improve the convergence rate, a two-stage searching method (Liu and
Chen, 2001) is suggested. This method consists of global searching at the
first stage and local searching at the second. The local searching is performed
by reducing the search space after global searching locates the likelihood of
the optimal region. For this example, the global searching means that the
reference element number ranges over all the elements, i.e., the flaw may be
possibly everywhere in the plate. The reference element number can be
approximately located after a certain number of generations, for example,
50 generations at the first stage; global searching is then stopped and the

Case 1 Case 2 Case 3 Case 4

FIGURE 12.4
Four possible damage patterns. (From Liu, G.R. and Chen, S.C., Comput. Methods Appl. Mech.
Eng., 190, 5505–5514, 2001. With permission.)

Generation number
2.00E-04

0.0
0 100 200 300
-2.00E-04
Error

-6.00E-04

-1.00E-03

FIGURE 12.5
µGA convergence in searching for flaws in the simply supported plate. Error function is defined
by Equation 12.3. (From Liu, G.R. and Chen, S.C., Comput. Methods Appl. Mech. Eng., 190,
5505–5514, 2001. With permission.)

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 436 Thursday, August 28, 2003 5:54 PM

FIGURE 12.6
Domain for local search at the second stage. (From Liu, G.R. and Chen, S.C., Comput. Methods
Appl. Mech. Eng., 190, 5505–5514, 2001. With permission.)

FIGURE 12.7
Possible damage patterns in a 3 × 3 mesh. (From Liu, G.R. and Chen, S.C., Comput. Methods
Appl. Mech. Eng., 190, 5505–5514, 2001. With permission.)

damage pattern redefined for the reference elements, as shown in Figure


12.6. Starting the GA search in this specified local domain, it takes only 10
generations to obtain the true solution. The two-stage searching method
improves search efficiency greatly because local searching is performed in a
much smaller parameter space that includes the true solution.
For the same example with a larger flaw area, the preceding procedure
can also be applied by changing the damage profiles to include more cases,
as shown in Figure 12.7. The same plate with a larger flaw area as shown in
Figure 12.8 is analyzed. Global search is stopped at the 100th generation to
determine the subregion that contains the flaw. After restarting the GA
search in this subregion, it takes only another 18 generations to obtain the
true solution.

12.4 Use of Newton’s Root Finding Method


In this section, the element stiffness factors defined in Equation 12.4, for all
the elements of the finite element model of a structure, are taken to be
parameters and explicitly expressed in a linear form in the system equation
for forward analysis of the harmonic response of the structure, as shown in
Equation 12.5. This offers great convenience in applying Newton’s root find-

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 437 Thursday, August 28, 2003 5:54 PM

FIGURE 12.8
Simply supported square sandwich plate with a rectangular larger flaw in the core layer.
(From Liu, G.R. and Chen, S.C., Comput. Methods Appl. Mech. Eng., 190, 5505–5514, 2001. With
permission.)

ing method to search for parameters of the stiffness factor inversely because
the Jacobian matrix can be obtained by solving sets of linear algebraic equa-
tion derived from the system equation. Newton’s root searching method is
briefly discussed in Chapter 4.

12.4.1 Calculation of Jacobian Matrix


Newton’s root finding method and modified Newton’s method require cal-
culation of the Jacobian matrix that contains the derivatives of displacements
with respect to the unknown parameters, element stiffness factors e. The
Jacobian matrix can be obtained efficiently by taking advantage of the linear
expression of ei in Equation 12.5. Performing differentiations on the both
sides of Equation 12.5 with respect to each parameter ei leads to

∂w
K ie w + (K − ω 2 M) = 0 (i = 1, 2, … , ne ) (12.11)
∂ei

In Equation 12.11, vector w is solved from Equation 12.5 in forward anal-


ysis. So Equation 12.11 can be written as

∂w
(K − ω 2 M ) = −K ie w (i = 1, 2, … , ne ) (12.12)
∂ei

Thus, the derivative ∂w / ∂ei can be solved from the preceding linear
algebraic equation system, which is in the same form as Equation 12.5. The
only difference is on the right-hand side of the equation. For (i = 1, 2,  , ne )

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 438 Thursday, August 28, 2003 5:54 PM

the Jacobian matrix is obtained by multiplying matrix Q as defined by


Equation 12.7.

12.4.2 Iteration Procedure


Following the procedure as schematically given in Figure 4.8, the stiffness
factors e can be identified. Starting with an initial guess, e( 0 ) , the iteration
procedure is given as:

• Step 1 — solve Equation 12.5 at e( k ) for w, and then compute the


value of error function

f(e( k ) ) = Qw − w (12.13)

• Step 2 — solve Equation 12.12 at e( k ) for {∂w / ∂ei } , and obtain the
Jacobian matrix. In solving Equation 12.12, the right-hand side vector
is formed as “a pseudo load vector” first, using the response
obtained previously. The coefficient matrix has been factorized in
Step 1 so that forward analysis can be utilized and the derivation
vector can be obtained only by back-substitution with the pseudo
load vector.
• Step 3 — find e( k+1) by Newton’s method using Equation 4.74. In
practice, e( k+1) is obtained by solving the linear equation system of

∇fk (e( k +1) − e( k ) ) = −f(e( k ) ) (12.14)

• Step 4 — repeat Step 1 through Step 3 until the required tolerance


is satisfied.

12.4.3 Example I: Cantilever Beam


12.4.3.1 Stiffness of Cantilever Beam
In order to verify the technique, the cantilever beam shown in Figure 12.9
is considered. It is discretized into 20 beam elements; therefore, 20 unknown

F
z
1 2 20 x

FIGURE 12.9
Nonuniform stiffness beam and its finite element model. The beam is discretized into 20 beam
elements and the excitation is a time-harmonic load at the free tip of the beam with a frequency
of ω = 100 rad/s. (From Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5), 823–835, 2002. With
permission.)

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 439 Thursday, August 28, 2003 5:54 PM

TABLE 12.3
Element Stiffness Factors for the Cantilever
Beam as Shown in Figure 12.9 (Case 1, e0 = 2.1)
Element number 1–5 6–10 11–15 16–20
Stiffness factor e 2.1 1.5 2.1 1.5
Source: Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5),
823–835, 2002. With permission.

TABLE 12.4
Element Stiffness Factors for the Cantilever Beam
as Shown in Figure 12.9 (Case 2, e0 = 2.1)
Element number 1–2 3–4 5–6 7–9 10–20
Stiffness factor e 2.1 1.8 2.1 1.05 2.1
Source: Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5),
823–835, 2002. With permission.

TABLE 12.5
Element Stiffness Factors for the Cantilever Beam as Shown in
Figure 12.9 (Case 3, e0 = 2.1)
Element number 1–2 3 4 5 6–20 21–25 26–50
Stiffness factor e 2.1 1.8 1.5 1.2 2.1 1.5 2.1
Source: Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5), 823–835, 2002.
With permission.

parameters represent the stiffness factors of elements. It is related to the


material constant and/or second moment of the section area. The element
stiffness factor to be identified is given in Table 12.3 through Table 12.5. The
mass density is ρ = 7.8 × 103 kg/m3, and the second moment of section area
is I z = 0.8 × 10–8 m4. The excitation is a time-harmonic load at the free tip
of the beam with a frequency of ω = 100 rad/s. The measured deflection
amplitude at 20 nodes is simulated using computational analysis results for
given true parameters. The damage factor β if of the ith element is defined
as the deduction of the element stiffness and can be obtained from the
stiffness factor ei:

 e 
β if =  1 − i  (12.15)
 e0 

where e0 is the undamaged stiffness factor.

• In case 1, a piecewise uniform stiffness distributed beam is consid-


ered; the true stiffness factors are given in Table 12.3. The iteration
is started from an initial guess x ( 0 ) that takes uniform value of
2.1(undamaged stiffness factor) for all the parameters. It converges

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 440 Thursday, August 28, 2003 5:54 PM

2.5

Stiffness Factor 1.5

0.5

0
0 2 4 6 8 10 12 14 16 18 20

Element Number
FIGURE 12.10
Inversely determined stiffness distribution in elements (Case 1) using Newton’s root finding
method. (From Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5), 823–835, 2002. With permission.)

to the solution very fast; the results are shown in Figure 12.10 and
are in very good agreement with the true values given in Table 12.3.
The same results can be obtained for different values of the angular
frequency ω .
• In case 2, a damaged beam with two damaged locations is consid-
ered. The true stiffness factors are given in Table 12.4. The damage
in element 3 and element 4 and element 7 through element 9 is
successfully detected as shown in Figure 12.11.
• In case 3, the same beam is now discretized into 50 elements and
the stiffness distribution is represented by 50 parameters. Table 12.5
shows the true element stiffness factors. The results are obtained
very quickly and accurately as shown in Figure 12.12. The results
also indicate clearly the damage status or the stiffness distribution
of the beam in terms of the stiffness factor.

The examples have shown that the inverse technique is suitable for prob-
lems with large numbers of parameters to be identified. It takes only seconds
of CPU time to obtain a very accurate result for the beam structure consid-
ered. It can be applied accurately to damage detection problems involving
multiple distributed defects with arbitrary degrees of damage. However, like
any gradient-based optimization algorithm, the initial guess will affect iter-
ation progress. For suitable initial guesses, the identical result can be
obtained; otherwise, the solution converges to a local minimum. It should
also be pointed out that the frequency of exciting force should not be too
close to the natural frequency of the structure because, in such a case, Equa -

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 441 Thursday, August 28, 2003 5:54 PM

2.5

Stiffness Factor 1.5

0.5

0
0 2 4 6 8 10 12 14 16 18 20

Element Number
FIGURE 12.11
Inversely determined stiffness distribution in elements (Case 2) using Newton’s root finding
method. (From Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5), 823–835, 2002. With permission.)

2.5

2
Stiffness Factor

1.5

0.5

0
0 5 10 15 20 25 30 35 40 45 50

Element Number
FIGURE 12.12
Inversely determined stiffness distribution in elements (Case 3) using Newton’s root finding
method. (From Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5), 823–835, 2002. With permission.)

tion 12.12 becomes singular and the inverse procedure fails because no
damping terms are considered.

12.4.3.2 Performance Comparison with µ GA


In order to compare the performance of the direct root searching method
with the genetic algorithm, the same beam as shown in Figure 12.9 is re-

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 442 Thursday, August 28, 2003 5:54 PM

F
z
1 2 20 x

FIGURE 12.13
Beam with one damaged location. The beam is discretized into 20 elements; the damage is
located in element 3 and element 4. (From Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5), 823–835,
2002. With permission.)

examined. A µGA is first applied to the stiffness factor identification (damage


detection). The population is taken to be 5 in each generation, while the
probability of uniform crossover is set to be 0.5. The objective function (Equa -
tion 12.6) is employed for fitness evaluation with harmonic response of
deflection taken as input. The stiffness factors of 20 elements are taken as the
parameters to be identified. In the GA, these parameters are required to be
discretized according to the accuracy needed. When all the parameters are
discretized into eight grades in the range of 0.63 to 2.1, the discrete search
space contains 260 ( ≈ 1.15 × 1018 ) candidates. Because of such a great number
of possibilities, the CPU time is excessively long due to the random nature
of GAs and the time-consuming forward analysis of FEM code. In order to
decrease the number of parameters, it is assumed that the beam has only one
damage, which is one of the first four elements. The degrees of damage of
these elements are discretized into eight grades, thus decreasing the discrete
search space to 212 (4096) candidates and making the GA search possible.
As an example, the beam including one damage location shown in Figure
12.13 is considered. The damage is located in element 3 and element 4, with
stiffness deduction factor β f = 0.5 ; it takes 30 generations to obtain the
solution. The CPU time consumed is 40 seconds. Using Newton’s root find-
ing technique to solve the same problem takes only about 1.5 seconds. This
simple test verifies the efficiency of Newton’s root finding technique over
GAs for problems with multiple continuing variables as parameters. Another
advantage of the technique over GAs is that the GAs’ accuracy is limited to
the possibilities to discretize parameters. In order to increase the accuracy
of GAs, more possibilities are required to discretize the parameters and more
generations are required to search the solution.

12.4.3.3 Noise-Contaminated Case


The efficiency and accuracy of the technique has been demonstrated through
the preceding examples without considering random measurement errors.
To study the effect of the measurement noises on parameter identification,
Gaussian noise with zero mean and constant standard deviation is added
to the measured value. Consider again the three beams investigated in
Section 12.4.3.1. Adding these noises to the measured responses obtains a
good result for case 2, as shown in Figure 12.14. For other cases, it fails to
get good results, implying that Newton’s root finding method is very sen-
sitive to noise.

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 443 Thursday, August 28, 2003 5:54 PM

2.5

Stiffness Factor 1.5

0.5

0
0 2 4 6 8 10 12 14 16 18 20

Element Number

FIGURE 12.14
Inversely determined stiffness distribution in elements (Case 2 with measurement noise) using
Newton’s root finding method. (From Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5), 823–835,
2002. With permission.)

12.4.4 Example II: Plate


Newton’s root finding method is extended to plate structures modeled with
finite elements. The plate is simply supported with several elements
deducted in stiffness as given in Figure 12.15. The plate is divided into 100

62 72

1 91
2 92

27

10 100

FIGURE 12.15
Simply supported square plate modeled using 100 eight-node isoparametric quadratic ele-
ments. The six elements deducted in stiffness are shaded; actual element stiffness factors are
listed in Table 12.6. A time-harmonic load with a frequency of ω = 3000 rad/s is applied at
the center of the plate. (From Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5), 823–835, 2002. With
permission.)

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 444 Thursday, August 28, 2003 5:54 PM

TABLE 12.6
Element Stiffness Factors for the Plate Simply Supported as Shown
in Figure 12.15
Element number 27–28 52 53 62 63 Remainder
Stiffness factor e 1.5 1.6 1.8 1.2 1.5 2.1
Damage factor β 0.286 0.238 0.143 0.429 0.286 0
Source: Liu, G.R. and Chen, S.C., J. Sound Vib., 254(5), 823–835, 2002. With
permission.

eight-node isoparametric quadratic plate elements. Numerically simulated


nodal deflection response subjected to a harmonic excitation is used as the
measurement. The true element stiffness factor is given in Table 12.6. Using
the inverse procedure described previously, the stiffness factors of the ele-
ments for plates are identified accurately, as shown in Figure 12.16.
To apply the method to solve practical problems, some consideration and
modification are required. First of all, the forward analysis model of the
structural system should be carefully considered to simulate the practical
system as accurately as possible or corrections to the difference between
simulated and practical responses should be made. For example, damping
terms and support stiffness of the boundary should be considered. Another
important problem considered is the measuring error. To consider measure-
ment error, the Gauss–Newton method should be used, where the number
of measured data is more than the number of parameters (over-posed prob -
lem; see Chapter 2). The Gauss–Newton method gives the estimation of the
parameters based on minimization of the least squares of the error norm. It
is expected to be more robust to random errors of measurement.

FIGURE 12.16
Damage factor identified for the plate using Newton’s root finding method. (From Liu, G.R.
and Chen, S.C., J. Sound Vib., 254(5), 823–835, 2002. With permission.)

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 445 Thursday, August 28, 2003 5:54 PM

F
z
1 2 3 4 5 9 10 11 20 x

FIGURE 12.17
Finite element model of a cantilever beam with damage, divided into 20 elements. The damages
are assumed at element 4 and element 5 and element 9 through element 11 as marked by the
in shaded areas.

12.5 Use of Levenberg–Marquardt Method


Levenberg–Marquart’s root finding method (described in Chapter 4) is
employed for the flaw detection in a cantilever beam as shown in Figure
12.17. The beam is discretized into 20 beam elements. The damages are
located at element 4 and element 5 with damage factor β = 14.3% and
element 9 through element 11 with damage factor β = 28.6% . There are 20
unknown parameters that represent the stiffness factors of elements. The
mass density, Young’s modulus, and Poisson’s ratio are ρ = 7.8 × 10 3 kg/m3,
E = 2.1 × 1011 N/m2, and ν = 0.3 , respectively, and the second moment of
section area is I z = 0.8 × 10 −8 m4. The element stiffness factor before damage
is taken to be 1 as reference or the element damage factor before damage is
0. The excitation is a time-harmonic load at the free tip of the beam with
frequencies of ω 1 = 100 rad/s and ω 2 = 220 rad/s.
The measured deflection amplitude at 16 selected nodes (at points marked
by dots) is used for inverse analysis, which is simulated in forward analysis
using true parameters corresponding to damage status given earlier. The
initial element stiffness factors are taken to be 0.48 for all the elements. Using
the Levenberg–Marquart method, the results are obtained and shown in
Figure 12.18, where the damage locations are clearly indicated and the dam-
age factors are in good agreement with the true values.
To simulate the practical measurements, a Gaussian noise with zero mean
and constant standard deviation is added to the computed responses. Using
the cantilever beam as an example, different levels of Gaussian noise are
considered. The results are shown in Figure 12.19 to Figure 12.21 for noise
levels from 2 to 10%. It is shown from these figures that the damage locations
are clearly indicated at element 4 and element 5 and element 9 through
element 11, while the degrees of damage agree well with the true values. It
can be noted that the results are stable and robust, despite the random
measurement noise, with the noise level up to 10%.
The same beam with two ends fixed is considered, as shown in Figure
12.22. The size and damage status of the beam are the same with the canti-
lever beam. In this case, the harmonic load is applied at the center of the
beam. The frequencies of the load adopted are ω 1 = 100 rad/s and

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 446 Thursday, August 28, 2003 5:54 PM

0.35

0.3

0.25

Damage Factor 0.2 Without noise

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16 18 20
Element Number

FIGURE 12.18
Detected damage factors for the cantilever beam using simulated noise-free measurements using
the Levenberg–Marquardt method.

0.4

0.35

0.3
Damage Factor

0.25 2% noise
0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16 18 20

Element Number

FIGURE 12.19
Detected damage factors for the cantilever beam using simulated measurements with 2% noise
using the Levenberg–Marquardt method.

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 447 Thursday, August 28, 2003 5:54 PM

0.35

0.3

0.25

Damage Factor
5% noise
0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16 18 20

Element Number

FIGURE 12.20
Detected damage factors for the cantilever beam using simulated measurements with 5% noise
using the Levenberg–Marquardt method.

0.4

0.35

0.3
Damage Factor

0.25
10% noise
0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16 18 20

Element Number

FIGURE 12.21
Detected damage factors for the cantilever beam using simulated measurements with 10% noise
using the Levenberg–Marquardt method.

F
z
1 2 20 x

FIGURE 12.22
Beam with two ends fixed.

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 448 Thursday, August 28, 2003 5:54 PM

0.35

0.3

Damage Factor
0.25

0.2

0.15

0.1

0.05

0
0 2 4 6 8 10 12 14 16 18 20

Element Number

FIGURE 12.23
Damage status detected in the fix–fix beam using simulated noise-free measurements using the
Levenberg–Marquardt method.

ω 2 = 220 rad/s. Only deflections at the selected nodes indicated are mea-
sured and used for the inverse analysis. Starting from initial trial parameters
taken to be 1.0 for all element stiffness factor (without any damage initially),
the result is shown in Figure 12.23. It is found that damage location and
damage factor are obtained in good agreement with the true values.

12.6 Remarks

• Remark 12.1 — two approaches are introduced for the characteriza -


tion of flaws in structures; in both approaches the finite element
model is used for forward analysis. The first approach is parameter
identification to determine the presence, location, size, and degree
of flaws in the core layer; a GA is used for inverse analysis. A
uniform µGA has been employed to detect the location, area, and
degree of the flaw in the core layer of sandwich structures from time-
harmonic response of the structure. The characteristics of the flaw
are represented by discrete and continuing variables at the element
level as a set of parameters. The search efficiency of the GA is
improved greatly by using a two-stage searching method.
• Remark 12.2 — the second approach is to relate the flaw with the
element stiffness factors that can be formulated in a linear form in
the system equation, leading to a very efficient way to compute the
gradients. Newton’s root finding method is applied to find the

© 2003 by CRC Press LLC


1523_Frame_C12.fm Page 449 Thursday, August 28, 2003 5:54 PM

inverse solution. It converges to the true solution much faster in


comparison with GAs. The method is very accurate and efficient for
problems with a large number of parameters; however, this tech-
nique is sensitive to measurement noise.
• Remark 12.3 — the Levenberg–Marquardt method is successfully
applied to damage assessment of a finite element model using time-
harmonic responses. Examples demonstrated that location and
degree of damages can be identified simultaneously with a large
number of parameters. However, it should also be noted that the
success of the identification is strongly dependent on the value of the
damping factors (see Equation 4.78), which needs to be selected prop-
erly. The choice of the initial values of parameters is another issue.
• Remark 12.4 — although not yet tested for example in this chapter,
a combined procedure using the GA and the Levenberg–Marquardt
method is expected to work well because the GA can be used to find
a good initial guess, as demonstrated in Section 8.5, Section 9.6, and
Section 11.6.

© 2003 by CRC Press LLC

Das könnte Ihnen auch gefallen