Mechanical Systems and Signal Processing (1998) 12(1), 163–186

DAMAGE DETECTION BASED ON MODEL

UPDATING METHODS
C.-P. F, D. J
Institute of Mechanics and Control Engineering, University of Siegen, Germany

T. K
Institute of Applied Mechanics, University of Kaiserslautern, Germany

(Received January 1997, accepted after revisions May 1997)

The paper examines the problem of detecting the location and extent of structural
damage from measured vibration test data. The method is based upon a mathematical
model representing the undamaged vibrating structure and a local description of the
damage, e.g. a finite element for a cracked beam. The problem of modeling errors and their
influence to damage localisation accuracy is discussed and an approach to obtain reliable
results in this case is presented. The concept of inverse sensitivity equations is used which
can be based on any type of data, e.g. modal data, FRFs, time series, or a combination
of these. The resulting inverse problem usually is ill-posed, and therefore special attention
is required for an accurate solution. The application to damage detection problems requires
the reduction of a large set of damage parameter candidates to a small subset of one or
two parameters that actually describes the local change of the system. An orthogonalisation
strategy is given to reduce the parameter set. The method is demonstrated through
application to laboratory structures in the frequency domain using frequency response
functions and in the time domain.
7 1998 Academic Press Limited

1. INTRODUCTION
Recently, the possibility of using measured vibration data to detect changes in structural
systems due to damage has gained increasing attention [1–6]. The aim is to proceed from
routine inspections with fixed intervals to continuous monitoring of the system, giving the
operator constant information on safety, reliability or remaining lifetime of the structure,
machine or plant. Methods, which are based on pure signal processing (such as the
evaluation of means, rms, trends or spectral densities, etc.) have only a limited capability
for the early detection of damage and often do not allow unique conclusions to be drawn
on the source of the damage. The model-based methods use a mathematical model of the
system in which some particular failure types can be integrated. In combination with the
measurement of the real structure, the model is able to provide much more detailed
information about the state of the system than the measurement alone. In many practical
cases, it is sufficient that the location of a structural damage is indicated and this damage
is investigated subsequently by visual inspection or non-destructive testing methods like
ultrasonics, X-rays, etc., for further quantification.
Damage detection methods which are based on time domain data were developed mainly
from control theory and process automation, e.g. [7]. The application of these concepts

0888–3270/98/010163 + 24 $25.00/0/pg970139 7 1998 Academic Press Limited

164 .-.   .
to problems of structural dynamics is difficult due to the high number of degrees of
freedom and therefore suitable reduction methods are required.
Detection schemes can be based on observer or filter techniques (failure sensitive filters,
innovation tests, filter banks) or on parameter identification or on a combination of both.
A characteristic of all the approaches is that certain residuals between the measurement
and the analytically redundant system are generated, which have to be evaluated in an
appropriate way to obtain information about the damage. Methods which are based on
frequency domain data have been developed mainly by vibration engineers and most
model-based damage detection procedures in structural mechanics resulted from model
updating and error localisation methods [8–16]. An outline of damage detection in the
context of model updating is shown in Fig. 1.
Another approach for structural damage detection is the use of neural networks as more
abstract mathematical models [17, 18]. Besides these quantitative models, knowledge-
based, heuristic models can be used in combination with methods of artificial intelligence
to evaluate the measured data [7].
In addition to fundamental work to develop the methods, different applications in
mechanical and civil engineering have been examined such as damage detection in offshore
structures [19, 20], composite materials [21] or rotating machinery [22–24].

2. MATHEMATICAL MODEL OF THE SYSTEM

2.1.   
The motion of a linear system with n degrees of freedom is described by
[M0 ]{q̈} + [C0 ]{q̇} + [K0 ]{q} = {f(t)} (1)
where the index ‘0’ is related to the original undamaged system. Due to a damage the
system changes its characteristics which may be expressed in the dynamic model by the
matrix changes [DM], [DC] and [DK]:
[M0 + DM]{q̈} + [C0 + DC]{q̇} + [K0 + DK]{q} = {f(t)} (2)
In general, the system may become non-linear or time variant. In this case, the D-matrices
depend on the system state {q}, {q̇}, explicitly on the time t or both.
For example, in the case of an opening and closing crack (as in rotating machinery
[22–24]), the stiffness varies due to the actual displacements [DK] = [DK{q(t)}] which can

Figure 1. Basic concept of damage localisation by model updating.

       165
be reduced to a simpler periodic time dependence [DK] = [DK(t)] under stationary
operating conditions of the rotor for one revolution of the shaft.
To describe the damage by physical parameters like crack depth, crack position,
delamination area of composite materials, loss of mass, etc., the D-matrices must be related
to changes of the physical parameters {Dp} which are made dimensionless for numerical
reasons. In general, a certain parameter may influence each matrix. A linear Taylor
expansion of the D-matrices yields:

[DM] = s [Mj ]Daj [DC] = s [Cj ]Daj [DK] = s [Kj ]Daj (3)
j j j

where the submatrices [Mj ], [Cj ] and [Kj ] result from the partial derivatives of the system
matrices with respect to the np dimensionless parameters

Dpj pj − poj pj
Daj = = aj = j = 1, 2, . . . , np (4)
poj poj poj

For this paper, it is assumed that the D-matrices are constant during the short time interval
of one data acquisition.
The state space formulation as a general description of a linear dynamical system is

{ẋ} = [As ]{x} + [Bs ]{u(t)} (5a)

{y} = [Hs ]{x} + [Gs ]{u(t)} (5b)

where for linear mechanical systems [As ], [Bs ], {x}, {u}, take the well-known form

$ % $ %
[0] [I] [0]
[As ] = [Bs ] = (6)
−[M]−1[K] −[M]−1[C] [M]−1

{x}T = {{q}T, {q̇}T} {u(t)} = {f(t)}

The matrices of the measurement equation [Hs ], [Gs ] depend on the nature of the measured
quantities (displacements, velocities, accelerations, strains, forces etc.). In previous
investigations, the matrix [Gs ] is often set to zero in advance, which excludes the possibility
to formulate the state space equations for accelerations as measurement data for example
(see Section 5.2).
If the number of state variables is high, a reduction of the system equations, e.g.
by modal transformation [25, 26], is necessary. Transformation of the displacements,
e.g. by means of the real mode shapes [F] of the related conservative system (normalised
to unit modal mass)

{q} = [F]{j} (7)

yields the transformed state space and measurement equations

{z } = [Amod ]{z} + [Bmod ]{u(t)} (8a)

{y} = [Hmod ]{z} + [Gmod ]{u(t)} (8b)

166 .-.   .
where
{z}T = {{j}T, {j }T} (8c)

$ % $ %
[0] [I] [0]
[Amod ] = [Bmod ] = (8d)
−[L] −[CF ] [F]T

with
[L] = diag (v12, v22, . . . , vn2 ) [CF ] = [F]T[C][F] (8e)
where the vi are the eigenfrequencies of the undamped system. The reduction of order is
achieved by considering only the r lower eigenfrequencies and lower mode shapes of the
modal matrix [F]n × n . The matrices [Hmod ] and [Gmod ] of the measurement equation again
result from the type of the measured data.
The transformed damping matrix becomes diagonal only in the case of proportional
damping. The transformation by means of the normal mode shapes yields good results if
the damping behaviour is near proportional. The transformation by means of the complex
modes is treated in [23, 25].

2.2.     

To describe the influence of damage to the dynamical behaviour of the system, a local
model must be formulated. The simplest way to consider damage in a structure is to use
stiffness, damping or mass parameters directly. The loss of a mass can be expressed by
reduction of the nodal mass and/or mass moment of inertia. For example, a crack causes
a local loss of stiffness so that the decrease of stiffness in a finite element can be used as
simple damage model. A more detailed crack model describes the relation between the
stiffness reduction and crack depth and crack position in the element used in Section 5.2
[27–29]. Moreover, other damage models have been described also: for delaminations in
sandwich plates [21], or multiple cracks in concrete beams [30].

2.3.    

Assume that the analytical redundant reference model [Mo ], [Co ], [Ko ] perfectly
represents the undamaged system. Then the model (1) allows the calculation of all the
quantities of interest, e.g. modal data (either real or complex), frequency response
functions (FRFs), impulse response functions or transient or stationary responses {q(t)}
to a known excitation {f(t)}.
In order to detect a system change due to the damage, measured quantities of the real
structure must be available. The residual rj characterising the differences between damaged
and undamaged state can be formulated for real eigenvalues li = vi2 , responses y(t) or
FRFs Hkl (v), etc., as
rj = wj (lidam − li0 ),
rj = wj (yidam (tk ) − yi0 (tk )), (9)
rj = wj (Hkldam (vv ) − Hkl0 (vv )), etc.
where wj is an individual weighting factor for the jth residual.
A linear or sequentially linearised relation is required of the form
[S]{Da} = {r} (10)
   167
which expresses the effect of parameter changes due to the changes of the measurement
data included in the residual vector. [S] is the sensitivity matrix which is used here
in a generalised sense. Usually, [S] is calculated from the partial derivatives of the
residuals
1r
Sjk = − j (11)
1ak
with respect to the dimensionless parameters. In some cases more accurate relations can
be obtained by fulfilling identically the equation of motion without linearisation [14, 16].
This is an essential feature—especially in those cases where small parameter variations
cause strongly non-linear changes of the residuals. An example is the change of the
magnitude of FRFs for frequencies near the resonances due to a parameter variation for
slightly damped systems (see Section 5.1).
The sensitivities of the real eigenfrequencies and mode shapes are given in [8, 10]. The
use of FRFs is treated in [8, 10, 13, 15, 16].
The general expression for sensitivities of the time domain output can be obtained by
differentiation of the state space equation with respect to the parameters [3, 25, 26].
In general, the resulting sensitivity equation system can be built up from any type of
measurement data such as modal data, frequency response functions or time domain data
or combinations of them. In the case of combining different types of measurement data,
a blockwise global scaling of the measurement sets is necessary according to their
numerical order.
From the viewpoint of damage detection, it is desirable that small parameter changes
(small damages) cause large deviations in the system characteristics or response data, that
means that the residuals are highly sensitive to a parameter change. Otherwise if the system
is insensitive the system change cannot be observed by the chosen set of measurement data.
Furthermore, in the case of insensitivity small measurement errors cause large parameter
errors when solving the inverse problem [31].
However, in some situations it is necessary to generate high sensitivity for some special
parameters and low or zero sensitivity for the other parameters by appropriate choice of
experimental design [32] or by weighting of existing measurement data. This concept is
called selective sensitivity [8, 33].
However, sensitivity alone is not a sufficient condition because the parameter describing
the actual damage is in general not the parameter with the highest sensitivity. Besides the
necessary condition of sensitivity recently in [34], a sufficient condition was formulated
from which follows that the changes of the system response between the damaged and
undamaged system must show a non-vanishing projection into the local modes of the
substructure which is damaged.

3. SOLUTION OF THE INVERSE PROBLEM

The direct solution of equation (10) yields very poor results if the data are polluted by
measurement noise or other errors, since small disturbances are amplified due to to the
nature of the ill-posed problem. This is because changes in two neighbour elements of a
finite element model may have nearly the same effects on the dynamics of the structure,
so almost linearly dependent columns of the sensitivity matrix are obtained. The finer the
discretisation, the stronger this effect. Therefore, it is important to use regularisation
methods to solve equation (10) [31]. In the context of damage detection, it is important
to keep the number of unknowns to a minimum. Hence, the high number of possible
damage candidates must be reduced to a minimum to ensure good results.
168 .-.   .
3.1.     -
The regularisation effect using orthogonal decomposition techniques like the
QR-algorithm is obtained by reduction of the original parameter space to a smaller
subspace so that the problem (10) is numerically much more stable but also yields a
sufficiently accurate solution. Therefore, this subspace needs to be defined.
The sensitivity matrix [S] is decomposed into a product of an orthogonal matrix [Q] and
an upper triangular matrix [R], while performing a column interchange for [S] which is
stored in the permutation matrix [P].

[S][P] = [Q][R] (12)

Introducing new instrumental co-ordinates bi describing the contribution of the ith

normalised orthogonal vector {qi } which are calculated by means of Gram–Schmidt
orthogonalisation [38].

np
[Q][R]{P TDa} = {r} or s {qi }bi = {r} with >{qi }> = 1 (13)
i=1
zXcXv
{b}

Due to the orthogonality of {qi }, the factors bi can be determined directly from the scalar
product

bi = {qiT }{r} (14)

So far, one is free to choose the order of the column permutation to reach the vectors {qi }.
The standard approach is sensitivity based and the permutation is performed in such a
way that the magnitudes of the diagonal elements of the matrix [R] have descending order
[38]. The algorithm (called QRD-S) was applied in [14, 16].
To understand how much each parameter can contribute to the right-hand side (rhs)
of equation (10), a correlation coefficient ki , i = 1, . . . , np is calculated. The scalar product
of {si } and {r} is the projection of an infinitesimal change of the residuals due to a change
of parameter i onto the residuals, expressing the similarity of the ith column of the
sensitivity matrix and the residual vector:

{si }T{r}
ki = −1 E ki E 1 (15)
>{si }>>{r}>

The norm >{si }> is a measure of sensitivity. =ki = = 1 means that the ith parameter alone
is sufficient to represent the rhs, while ki = 0 indicates that there is no correlation between
this parameter and the residuals.
The column permutation in [S] is performed so that a column {sj }, can—after an
orthogonalisation step—contribute most to the reduction of the norm of the rhs >{r}>.
The procedure is started with the parameter ai whose corresponding column {si } yields
kmax = max {=ki =, i = 1, 2, . . . , np } so that {q1 } = {si }/>{si }>.
Finally, evaluation of the error using equations (13) and (14) leads to

B B 0 1
np 2 np np
>{o}>2 = {r} − s {qi }bi = >{r}>2 − s bi2 = >{r}>2 1 − s 
b2i (16a)
i=1 i=1 i=1
   169
with
{qiT }{r}
i =
b (16b)
>{r}>
From equation (14) it can be deduced that {qi } yields no contribution to the error
reduction, if it is orthogonal to the rhs, because then bi = 0. Small bi yield small
contributions for the error reduction. Hence, it makes sense only to choose columns having
large bi . The column interchange during orthogonalisation is performed in such a way that
=b1 = e =b2 = e =b3 = e · · · e =bnp = where only the first k E np columns with the largest bi are
used so that =bi =/>r> = =b i = e t where t is a user-defined threshold. With this reduced set
the equation system is solved

6 7
T

[R]{P TDared,k } = {bred,k } with {bred,k } = b1 , b2 , . . . , bk , 0, 0, . . . , 0 (17)

zXcXv
np − k

finally yielding the physical parameter changes {Dared,k } after backward substitution, where
the remaining np − k parameter changes, which do not contribute to the solution of the
problem, are zero.
The remaining dimensionless error ek for truncation after k follows from

k
>{r} − [S][P]{Dared,k }>2
ek2 = 1 − s 
b2i = (18)
i=1
>{r}>2

This method is abbreviated to QRD-E. Details of the algorithm are presented in Appendix
A. The idea of the error reduction ratio was presented in [39] but it is given here in the
context of the truncated QR decomposition.

3.2.   

Another possible way to solve equation (10) is the use of either the singular value
decomposition (SVD) or the iterative method of conjugate gradients (CG), both of which
have been applied to detect damage in a uniform beam [16]. However, both methods yield
a full parameter set and the localisation result is often not as clear as QRD.
The simulated annealing method is based on random search which finds the global
minimum of the problem by admitting not only steps towards the minimum but also to
larger function values within a certain probability so that local extrema can be passed to
run finally into the global minimum. In [40], the method is used to solve the damage
detection problem by a random choice of small subsets of updating parameters. The
simulated annealing method is, however, very time consuming.

4. DAMAGE DETECTION WITH INACCURATE STRUCTURAL MODELS

Up to now, the model has been assumed to be correct. The true model representing the
undamaged system is denoted by [M0 ], [C0 ], [K0 ]. However, in practice this model is not
known but only an approximation [M 0 ], [C 0 ], [K 0 ] where
[M 0 ] = [M0 ] + [dM]; [C 0 ] = [C0 ] + [dC]; [K 0 ] = [K0 ] + [dK] (19)
The d-matrices express the mismodeling caused by inaccurate discretisation, the use of
inappropriate theories, neglection of important physical effects, wrong parameter values,
170 .-.   .
etc. Some systematic investigations on model uncertainty have been carried out [37]. The
use of the incorrect model yields errors dli , {d8i }, dHkl (v) of eigenvalues, eigenvectors,
FRFs, etc. On the other hand, the true system changes due to a damage are given by the
D-matrices, see equation (3). Now, working with the differences between the damaged
system data and those calculated by means of the approximate model, the problem arises
that modeling errors and changes due to the damage are superimposed:
[Mdam ] = [M 0 ] + ([DM] − [dM])
[Cdam ] = [C 0 ] + ([DC] − [dC]) (20)
[Kdam ] = [K 0 ] + ([DK] − [dK])
The algorithm is not able to separate the two error sources and the localisation result
may become very poor. Proceeding as before, one obtains the equation system
([S] − [dS])({Da} − {da}) = ({r} − {dr}) (21)
Due to the mismatch of the model leading to errors of the sensitivity matrix [dS] and
the residual vector {dr}, an additional parameter error {da} is obtained. If the sensitivity
matrix is ill-conditioned due to the ill-posed nature of the problem, the errors are amplified
in such a way that unreasonable results are obtained. The estimation for the maximum
of >da> can be found, from [42]

0 1
>{da}> >{o}> >[dS]>
E 1 + cond ([S]) cond ([S])
>{Da}> >[S]>>{Da}> >[S]>

>{r}> >{dr}>
+ cond ([S]) (22)
>[S]>>{Da}> >{r}>
From equation (22) it can be seen that the accuracy is strongly influenced by the
condition number of the sensitivity matrix cond ([S]) amplifying the relative errors of the
modeling errors of [dA] and [dr]. Localisation problems, as given in Section 5, are usually
characterised by high condition numbers. The condition number can be improved by
reduction of the parameter set using the truncated QR scheme. Equation (22) can also be
used to check the effects of measurement noise on the parameter vector.
One possibility to improve this situation is to update the reference model by means of
the measurement data of the undamaged system. However, this may fail if the mismodeling
cannot be compensated by simply correcting some parameter values.

4.1.    

One method to suppress the effects of the incorrect model is to replace the faulty
analytical data of the original model by the corresponding measured counterpart of the
undamaged system. For example, for eigenfrequencies, the new residual is
meas
ri = wi (lm,dam − lm,0
meas
) (23)
and the sensitivities are
sij = wi {8m,0 }T([Kj ] − lm,0
meas
[Mj ]){8m,0 } (24)
The model is used to provide the derivatives of the system matrices [Kj ], [Mj ]. In many
cases the mode shapes do not change significantly as the parameters change. Thus, the
mode shapes can be calculated completely from the model. Alternatively, the measured
   171
modes are used, expanding them to the full length n by any suitable expansion method
[43].
Working in the time domain with an arbitrary time series yk (t), the residual is

ri = wi (yk,dam (tv ) − yk,0 (tv )) (25)

meas
Either the measured time series yk,0 (t) = yk,0 (t) can be used directly—this requires the
identical excitation signal u(t) for both, the damaged and the undamaged system so that
they can be compared—or yk,0 (t) can be obtained by integration of the modal state space
model (8) replacing the analytical eigenfrequencies by the measured ones. Additionally, the
damping behaviour can be fitted to the measured modal damping calculating the diagonal
elements of [CF ] by means of the measured modal dampings.

$ %
[0] [I]
[Amod ] = (26)
−[Lmeas] −[CFmeas ]

The sensitivities can be approximated by

1yi (t) Dyi (t) yi (t, {a+

j }) − yi (t, {a})
1 = (27)
1aj o (aj + o) − aj

for numerical computation, where {a+ j } results from the parameter vector {a} whose
component aj is disturbed by a small value o. The change of the matrices [DAmod ], [DBmod ]
according to the parameter disturbance o is calculated by means of the model resp. by the
measurement. For example, the changes of the eigenvalues in equation (26) can be
estimated from

li 1 limeas + {8i,0 }T([Kj ] − li,0

meas
[Mj ]){8i,0 }o (28)

while the changes of the mode shapes are calculated by the model alone.
The application of this concept to FRFs is demonstrated in Section 5.1.

4.2.    

This is another possibility to make the damage detection procedure less sensitive to
model inaccuracies. Furthermore, a state observer improves the situation if additional
unknown excitations are present and the initial conditions are uncertain [24, 26, 35]. The
state observer has the form

{x̂ } = [As ]{x̂} + [Bs ]{u} + [L]({ydam

meas
(t)} − {ŷ(t)}) (29a)

{ŷ} = [Hs ]{x̂} + [Gs ]{u} (29b)

Inserting equation (29b) in (29a) leads to

{x̂ } = [As − LHs ]{x̂} + [Bs − LGs ]{u(t)} + [L]{ydam

meas
(t)} (29c)

The matrix [L] is the time invariant observer matrix, weighting the residuals between
the measured data and the corresponding predicted values {ŷ}, {x̂} is the estimated
response. When the gain matrix is zero, then {x̂} = {x} and no feedback of the residuals
is introduced. On the other hand, if [L] becomes very large, the model dynamics
represented by [As ], [Bs ] are of less significance. The influence of [L] can be seen in
Fig. 2.
172 .-.   .

Figure 2. Effect of state observers in connection with an inaccurate model: (a) no observer; (b) ‘slow’ observer;
(c) ‘fast’ observer; ——, measurement; –––, inaccurate model; –·–·, residuals.

For the calculation of the observer matrix, one can refer to the control literature [35].
Here, a Riccati observer design is used [41], solving the stationary Riccati equation
[0] = [A][P] + [P][A]T − [P][C]T[R]−1[C][P] + [Q] (30a)
[L] = [P][C]T[R]−1 (30b)
After modal transformation equation (29) becomes
{z
 } = [Amod − Lmod Hmod ]{z
} + [Bmod − Lmod Gmod ]{u(t)} + [Lmod ]{ydam
meas
(t)}
{ŷ} = [Hmod ]{z
} + [Gmod ]{u(t)} (31)
[Lmod ] is calculated directly from the reduced system.
According to equation (27), the partial derivatives are obtained numerically by Dŷi (t)/o.
It is important to remember that the observer matrix depends also on the system
parameters, as it is derived from equation (30b). [P] depends on the parameter vector {a}
as it is the solution of the Riccati equation (30) and the system matrix [A] is a function
of {a}. Therefore, [L] depends also on {a}. That means that the observer matrix
[L] = [L({a+ j })] (resp. [Lmod ]) has to be recalculated for each parameter sensitivity. Due to
the reduced size of the transformed system the additional computing time is negligible,
however, this improves the accuracy of the procedure.
The matrices [R] and [Q] in equation (30) have to be chosen to be diagonal and so that
damage
the model (including the observer gain) calculates an output signal ŷ being close to ymeas
   173

Figure 3. Finite element model, measurement points, damage location and element numbers in parentheses.

(but not too close by choosing a too large [L]; there is no theorem for an accurate
choice). Increasing [R] means decreasing the influence of measurement. Increasing [Q]
decreases the influence of model. The calculations have shown that the remaining
damage
differences between ŷ and ymeas serve to find the damage parameter if the model
inaccuracies are not too large.

5. EXAMPLES
5.2.      
5.1.1. System description, model and measurement
As an application for the described damage detection procedure, a rectangular
aluminum plate (300 × 600 × 6 mm) is investigated. Based on the measured weight, the
density is calculated to r = 2800 kg/m3. An initial guess of 70 000 MPa for Young’s
modulus (E) is used. The damage is represented by two orthogonal slots, each having a
length of 170 mm and a width of 14 mm. The location of the damage and the 18 nodes
where the system is excited by an impact hammer are shown in Fig. 3. The acceleration
pick-up is located at node 1. The 18 corresponding FRFs contain the first 10 modes in
the measured frequency range. Free boundary conditions are obtained by fixing the
structure with very soft springs. A comparison of the measured undamaged and damaged
system is given in Fig. 4. The shift of eigenfrequencies due to the damage are in a range
of about 1–2.5%.

Figure 4. Comparison of measured frequency response functions. ——, Undamaged; –––, damaged.
174 .-.   .

Figure 5. Eigenfrequencies and mode shapes of the FE model.

The FE model consists of flat shell elements (plate elements) with nine nodes and three
dof per node. Using an FE model discretised by 3 × 6 = 18 elements, the difference
between measured and analytical eigenvalues for the undamaged case is minimised by
global updating of the Young’s modulus. This leads to a value of E 1 67 000 MPa and
a maximum difference in frequency of 14.5%. For an FE model discretised by 9 × 18
elements (or even more) the remaining maximum difference of the eigenfrequencies after
updating the Young’s modulus is still 13%. Other than measurement errors, another
explanation for these deviations is that the real plate is not perfectly flat.
Despite of this inaccuracy, and with respect to computing time, the flat plate FE model
consisting of 18 elements (Fig. 3) is used further as an original model. The corresponding
eigenvalues and mode shapes are shown in Fig. 5. The difference between the analytical
(3 × 6) and measured FRFs can be seen from Fig. 6. Note that the deviations due to the
incorrect model are of the same order as the deviations between damaged and undamaged
system.
The sensitivity approach using output residuals of measured and analytical FRFs based
on identically fulfilling the equation of motion yields an overdetermined equation system
[S]{Da} = {r} which is built up by applying

−[Cm ] s [[H0 ](s 2[Mj ] + s[Cj ] + [Kj ]){Hl,dam

meas
}]Daj = [Cm ]({Hl,dam
meas
} − {Hl,0 }) (32a)
j

zXXXXXXcXXXXXXv zXXcXXv
column {S*
j } {r*}

with

[H0 (s)] = (s 2[M0 ] + s[C0 ] + [K0 ])−1 (32b)

   175

Figure 6. Comparison of frequency response functions. ——, Measurement, undamaged; –––, FE model, 3×6
elements; –·–·, FE model using measured eigenvalues.

for multiple frequencies si = s + ivi , and splitting up into real and imaginary parts. The
vector {Hl,dam
meas
} represents the measured lth column of the FRF matrix [H] of the damaged
system. {Hl,0 } is the corresponding vector of the original FE model. The matrix [Cm ],
consisting of zeros and ones, reduces the equation system to those rows belonging to a
measured dof. Nevertheless, the elements of {Hl,dam meas
} belonging to non-measured dof are
still needed to build up {S* j }. They are calculated using a dynamic expansion method based
on the original FE model. The derivation of equation (32) and the expansion method are
shown in [14, 16]. In many cases, the applied linearisation [equation (3)] of the system
matrices with respect to the parameters is exact because the element matrices depend
linearly on parameters like Young’s modulus or density. In those cases, equation (32) is
exact in absence of measurement errors and errors due to the expansion.
Calculating the sensitivity matrix [S] according to equation (11) from the partial
derivatives of the residuals from

b
1[H0 ]
= −[H0 ](s 2[Mj ] + s[Cj ] + [Kj ])[H0 ]
1aj 0

yields a formulation which is slightly different from equation (32). The ‘small’ but decisive
difference is that on the left-hand side (lhs) of equation (32), the analytical FRFs {Hl,0 }

Figure 7. Comparison of frequency response functions near a resonance. ——, Measurement, damaged; –––,
FE model, undamaged.
176 .-.   .

Figure 8. Dimensionless error norm, no eigenvalue replacement. –––, QRD-S; ——, QRD-E.

appear instead of {Hl,dam

meas
} [14]. In this case, the linearisation of the residuals (resp. {Hl,0 })
may introduce a large error in the sensitivity {S* j }, because the relation between the FRFs
and the parameter is strongly non-linear near the resonances.
It has been shown that the use of the experimental expressions {Hl,dam meas
} on the lhs of
equation (32) is essential. The obtained columns of {S* j }, and, hence the results applying
the two methods, are very different. This is due to the fact that the peaks of the FRFs
are shifted and the values of {Hl,0 (v*)} and {Hl,dammeas
(v*)} for the same frequency v* (near
a resonance) may vary strongly in magnitude and phase (Fig. 7).
The existing deviations of analytical and measured eigenvectors (undamaged) are small.
This can be proved by using modal quantities for the calculation of the analytical FRFs:

nmod
{8j,0 }{8j,0 }T
[Ho (iv)] = s nmod E n. (33)
j=1
(iv) + vj,0
2 2
+ iv(2vj,0 Dj )

To compensate the effect of the exponential window e−st, iv is replaced by s = s + iv.

If the analytical eigenvalues vj,0
2
are replaced with the measured ones of the undamaged
system, as proposed in Section 4.1, the resulting FRF is nearly identical to the measured
FRF (Fig. 6). This effect provides the possibility to improve the localisation procedure.
To reduce the influence of the inaccurate FE model [H0 ] resp. {Hl,0 } is calculated using
modal quantities with replacement of the analytical eigenvalues by the corresponding
measured eigenvalues of the undamaged system:

nmod
{8 }{8j,0 }T
[Ho (s)] = s 2
j,0
meas 2 meas meas (34)
j=1
s + (v ) + s(2vj,0
j,0 Dj )

The solution of the resulting overdetermined equation system [equation (10)] is done by
means of two different types of the QR decomposition to reduce the number of parameters.

Figure 9. Localisation results for QRD-S, no eigenvalue replacement: (a) k = 3, (b) k = 6.

   177

Figure 10. Localisation results for QRD-E, no eigenvalue replacement. Dimensionless parameter changes for
k = 10.

In the previously described method (QRD-E), the column interchange is based on

(so-called) error reduction ratios. The second method (QRD-S) uses the euclidian norm
of the columns as criterion for column interchanges [14, 16, 38]. The column with the
greatest norm is selected first. A column with a large norm belongs to a parameter which
has a high sensitivity. In both methods, a dimensionless error norm ek can be defined
according to equation (18), where k is the number of parameters taken into account. The
optimal k is reached when no more further significant reduction of the error norm can be
achieved by taking more parameters into account.

5.1.2. Localisation results

The localisation is performed using the described method with and without replacement
of the analytical eigenvalues. Some settings are the same in both cases:
, all 18 measured FRFs are used;
, the Young’s modulus E of each plate element j is selected as a parameter, influencing
the flexural stiffness of this element (the damage is located in element 5);
, at each of the 10 peaks within the measured frequency range, three excitation frequencies
are selected. Frequencies near the resonances yield high sensitivities.
Starting with the localisation without replacement of the inaccurate analytical
eigenvalues, the applied procedure is unable to localise the damage. The dimensionless
error norm belonging to the QRD-S method leads to k = 6 as an optimal value (Fig. 8).
The corresponding parameter changes do not indicate element 5 as the damaged element

Figure 11. Dimensionless error norm, with eigenvalue replacement. –––, QRD-S; ——, QRD-E.
178 .-.   .

Figure 12. Localisation results for QRD-S, with eigenvalue replacement. Dimensionless parameter changes for
k = 4. Q, q0; q, Q0.

(Fig. 9). The parameter a5 is added to the parameter vector for case k = 3 which is also
shown, but contains bad results also. Using the QRD-E method, a5 is taken into account
but not before k = 10. In Fig. 10, the poor results for the localisation are shown.
However, k = 10 would never be chosen by regarding the dimensionless error norm in
Fig. 8.
On the other hand, localisation with replacement of analytical eigenvalues by their
measured counterparts the dimensionless error norm gives clear hints to choose the
optimal number of parameters which is k = 4 for QRD-S and k = 1 for QRD-E (Fig. 11).
The corresponding parameter changes for the QRD-S method (Fig. 12) are obviously
dominated by parameter 5 (Da5 = −26.5%), belonging to the correct location. The
QRD-E method delivers directly element 5 as damaged (Da5 = −30.5%) due to the fact
that this parameter yields the maximum error reduction (Fig. 13).
For both cases, the same calculations are repeated changing the selected frequencies to
five resp. seven per peak and selecting the whole frequency range (200 lines). The
differences between the obtained results are negligible. Furthermore, calculations with a
higher discretised FE model (6 × 12 elements) lead to the same results.

Figure 13. Localisation results for QRD-E, with eigenvalue replacement. Dimensionless parameter changes
for k = 1. Q, q0; q, Q0.
   179

Figure 14. Test structure and nodes of the FE model.

5.2.       

5.2.1. Description of the system, modeling and measurement
This is an example of localisation in the time domain. A T-like frame is used consisting
of two welded straight aluminum bars under free–free boundary conditions realised by a
soft spring (Fig. 14). The system was excited by means of an impulse hammer at node 9
and the system responses are measured by means of three acceleration sensors at nodes
1 and 11 (both y-direction) and node 20 (x-direction).
The system is modeled by means of 19 finite Timoshenko beam elements and three
lumped masses representing the sensors. The welded connection of the two beams must
be considered by higher stiffnesses than the normal beam elements between the nodal
points 5, 6, 7 and 12 (see Fig. 14). To reduce calculation time, modal reduction is used.
Twenty modes (including the rigid body modes) are taken into account. Two sharp
crack-like notches were introduced by means of wire erosion. The equations of motion
remain linear because the cracks do not close (closing cracks are treated in [23, 25]).
The use of acceleration data {y} = [Cm ]{q̈} requires both matrices [Hmod ] and [Gmod ] of
the measurement equation (8b). The matrix [Cm ] is composed by zeros and ones and
expresses that y1 = q̈2 , y2 = q̈30 and y3 = q̈58 have been measured. This leads to the relation:

67
j
{y} = [−[Cm ][F][L] − [Cm ][F][CF ]] + [[Cm ][F][F]T]{u} (35)
j
zXXXXXcXXXXXv zXcXv
Hmod Gmod

Figure 15. 3-D beam element with a crack characterised by crack depth a and position b.
180 .-.   .

Figure 16. Acceleration at node 20 (x-direction) for undamaged (——) and damaged system (–––).

T 1
Results of localisation of structure with one crack
Crack 1
Location a (mm) b (mm)
True 12 6.0 5.0
Identified 12 6.4 6.2

The crack model describes the relation between the stiffness reduction and crack depth
a and crack position b in the element and is formulated as special finite beam element (Fig.
15). The main idea [27–29] is based on principles of the linear–elastic fracture mechanics
connecting the additional compliance of the element due to the damage via Castigliano’s
theorem to the decrease of elastic strain energy which is expressed in terms of the stress
intensity factors [36].

5.2.2. Localisation with an updated model

The model representing the undamaged system is determined by model updating
correcting Young’s modulus, the stiffnesses in the area of the welded joint and the damping
coefficients for proportional damping. Two cases are investigated: the beam with one and
two notches.

Figure 17. Acceleration at node 1 (y-direction): measurement (——), model without filter (–––).
   181
100 100
(a) (b)

Error reduction rate (%) 80 80

60 3 60 12

40 40

20 20
19 8 4 1 7 2 16
0 0
10 5 15 5 10 15
Column of Q Column of Q
Figure 18. Parameter selection based on orthogonalisation process, stages (a) 1, and (b) 2. Error reduction
rate corresponds to b
2i in equations (16) and (18).

Crack 1 is located in element no. 12 (between nodes 12 and 13) and has a depth of
a = 6 mm and b = 5 mm. Crack 2 is located in element 3 (between nodes 3 and 4) with
a = 6 mm and b = 0 mm (Figs 14 and 15).
The procedure for locating the structural damages is as follows.
, The sensitivity equation (10) is built up from acceleration time data. The first step is
localisation of the section(s), characterised by reduced stiffness(es). A section can be a
single finite element (applied here) or a substructure or macro element consisting of
several elements. The resulting parameter changes are obtained by means of orthogonal
decomposition (QRD-E).
, To obtain a geometrical description of the crack/notch the elements where the damages
have been localised are exchanged by the special crack element (Fig. 15) and a, b are
fitted to the measurement data by non-linear unconstrained parameter optimisation.
The BFGS quasi-Newton algorithm with mixed quadratic and cubic line search
procedure from the MATLAB Optimization Toolbox is used [44].
In the first case, only damage in element 12 is present. Figure 16 shows the measured
accelerations y3 for the undamaged and the damaged case. The crack location is found by
searching (QRD-E) for the finite element where a stiffness reduction (DEI/EI) yields the
highest contribution to the error reduction. Element 12 was clearly indicated with an error
reduction rate of about 70%. The result of the subsequent adaptation of a and b is given
in Table 1. The correspondence of the model with crack element (a, b) and measurement
is shown in Fig. 17.
In the next case two instances of damage have to be found. The algorithm determined
element 3 in a first step. Only this element shows a significant contribution to the error
reduction while the other columns (1, 4, 8, 19) lead to non-significant contributions [Fig.

T 2
Results of localisation of structure with two cracks
Crack 1 Crack 2
Location a (mm) b (mm) Location a (mm) b (mm)
True 12 6.0 5.0 3 6.0 0.0
Identified 12 5.8 6.7 3 7.4 −5.3
182 .-.   .
T 3
Results of damage localisation with inaccurate models and methods
Model no.
Method 1 2 3 4 5 6 7 8
Ed Damage localisation? + × + × × × × +/w
Da12 0.33 — 0.31 — — — — 0.38
ERR (%) 74 — 68 — — — — 37
Ed, Obs Damage localisation? + + + w +/w + + +
Da12 0.39 0.49 0.35 0.22 0.37 0.66 0.51 0.43
ERR (%) 74 80 68 29 37 71 71 60
Ed, EO Damage localisation? + + + + + + + +
Da12 0.32 0.32 0.32 0.33 0.35 0.25 0.24 0.25
ERR (%) 62 62 62 61 61 65 62 61
Methods are damage localisation using the model with: ED, experimental data of the damaged system; EO,
experimental data of the original system; Obs, observer. ERR, Error reduction rate. Damage localisation
classified as: +, successful (ERR q 50%); ×, unsuccessful; w, no definite conclusion (ERR E 30%); +/w,
damage found correctly on a medium level (30% Q ERR E 50%).

18(a)]. The threshold t is 0.01 so that five columns are considered. As can be seen, the
other damaged element, 12, does not appear in this step. The calculation of the stiffness
reduction with column 3 leads to Da3 = −(DEI/EI)3 = 27.7%.
After updating the bending stiffness in element 3, the sensitivities and residuals are
recalculated and the procedure repeated. Figure 18(b) shows the result of the
QR-decomposition with respect to the error contributions of the parameters. In this
second step, element 12 is indicated clearly, while the other parameters show no
significance. The result for the stiffness reduction is Da12 = −(DEI/EI)12 = 27.5% and the
model was updated again. Further steps do not indicate the need for any other parameter
correction.
To quantify the damage in terms of crack depths, the local crack model is now built
in at element 3 and 12. First, the usual beam element 3 is replaced by the special crack
element while element 12 is treated as normal element but with reduced stiffness. As free
parameters, a3 and b3 are fitted to the measurement data. The procedure is repeated for
element 12. The final results for the crack data are shown in Table 2. Parallel iteration
of the a and b values for both crack sites did not yield better results.

5.2.3. Damage localisation from inaccurate models

To show the influence of the modeling errors, eight models have been investigated.
(1) This is the result of updating as used in Section 5.2.2 with Young’s modulus of
E = 72 500 MPa and stiffnesses approximately twice as large in the area of the welded
joint than to the normal beam stiffness. Timoshenko beam elements are used and the
acceleration sensors are considered by their mass.
(2) Same as model 1, but E = 73 900 MPa.
(3) Same as model 1, but E = 72 000 MPa.
(4) Same as model 1, but E = 70 000 MPa.
(5) Same as model 1, but no consideration of stiffness increase due to the welded joint.
(6) Same as model 1, but E = 73 000 MPa and rigid elements in the area of the welded
joint.
(7) Same as model 1, but Euler–Bernoulli theory is used for the beam elements.
(8) Same as model 7, but the higher stiffness of the welded joint is not considered.
   183
It can be seen from Table 3 that the model-based damage detection using only
experimental data of the damaged system (which is the straightforward updating
procedure) is very sensitive to the modeling errors. The model must represent a good
starting-point for the localisation process (as could be seen in Section 5.1 as well). Even
the use of Euler–Bernoulli beams instead of Timoshenko theory shifts the higher
eigenfrequencies so that no reasonable results can be obtained. Variations of the Young’s
modulus of 24% as well as mismodeling of the welded joint by neglection of the stiffness
increase yield the same behaviour.
The addition of the filter to the model improves the robustness of the detection
procedure significantly. In most cases, the damage was localised correctly with high error
reduction ratios. However, the parameter (stiffness reduction) is overestimated/underesti-
mated when the global stiffness of the model is too high/too low. As mentioned before,
the filter attracts the model response to the measured data and hence makes it less sensitive
to modeling errors. The observer matrix must be chosen in a way that the sensitivity with
respect to modeling errors is decreased but the sensitivity of the parameters describing the
damage is maintained on a sufficient level.
Robust results are obtained by the last method where the measured data of the
undamaged system are used in addition. The results vary only slightly for the different
modeling errors (Table 3). This arises from the fact that the residuals are calculated from
measured data only and are independent of any model. The integration of the model
equations to obtain the sensitivities is performed in modal form where the eigenfrequencies
of the erroneous model are replaced by their measured counterparts (in the lower frequency
range, as far as they are known). Hence, this source which causes the frequency shifts due
to the faulty model is excluded as well. The only influence of the model errors can be found
in the mode shapes which are taken completely from the analytical model. A further
improvement could be expected if expanded measured mode shapes could be used. For
this procedure, the input signal has to be reproducible. As a hand-held hammer is used
here, this could cause some problems. However, the time series of the force f(t) are always
similar and the equality can be approximately achieved by simply scaling the signals to
a reference maximum magnitude.
Finally, it should be noted that only three output measurements (compared to 60 dof)
had been used to localise the damage.

6. CONCLUSIONS
In principle, model updating methods, especially the sensitivity approach, are suitable
tools to treat the problem of damage localisation. The presented methods allow the
processing of time and frequency domain data. Special attention was paid to two points:
the solution of the equation system, and the problem of inaccurate models. Accurate
solutions can be obtained by QR orthogonal decomposition with error-based truncation
so that a reduction to those parameters which are actually responsible to the system
changes is realised. As demonstrated by the examples, the accuracy of the original model
is of great importance. In many cases, the errors between model and measurement of the
undamaged system lie within the order of the changes which are caused by the damage
especially in the early phase of the damage evolution.
For a damage localisation to be successful, the reference model must be as accurate as
possible, otherwise the procedure will lead to incorrect results. To improve the model
quality, a prior updating of the model to the measurement data of the undamaged system
is recommended. If this does not lead to the desired improvement, it was shown that the
substitution of faulty analytical data by corresponding measurement data increases the
184 .-.   .
robustness of the procedure with respect to model inaccuracies. In the time domain, the
application of a state observer turned out to reduce the dependence on modeling errors.
The observer matrix must be adapted to each individual problem.
Further investigations need to consider the problem of inaccurate models, so that
conditions under which a damage diagnosis is successful can be understood better.
Realistic applications are necessary to establish the model based damage detection in
structural mechanics as a reliable tool.

REFERENCES
1. H. G. N and J. T. P. Y (eds) 1988 Structural Safety Evaluation Based on System
Identification. Braunschweig/Wiesbaden: Friedr. Vieweg.
2. P. H and F. J. S 1990 Structural Optimization 2, 1–10. Recent developments in
damage detection based on system identification methods.
3. H. G. N and C. C 1991 Mechanical Systems and Signal Processing 5, 345–356. Fault
detection and localisation in structures: a discussion.
4. C. C and H. G. N 1991 Structural Dynamics 1, 393–398. Fault diagnosis in mechanical
structures by means of vibration due to impact.
5. P. S (ed.) 1992 Structural Integrity Assessment. New York: Elsevier Applied Science.
6. H. G. N, R. T and J. T. P. Y (eds) 1993 Structural Safety Evaluation Based
on System Identification Approaches. Braunschweig, Wiesbaden: Fr. Vieweg.
7. P. M. F 1990 Automatica 26, 459–474. Fault diagnosis in dynamic systems using analytical
and knowledge-based redundancy—a survey and some new results.
8. M. I. F and J. E. M 1995 Finite Element Model Updating in Structural
Dynamics. Kluwer: Academic Press.
9. J. E. M and M. I. F 1993 Journal of Sound and Vibration 167, 347–345.
Model updating in structural dynamics: a survey.
10. H. G. N 1992 Einfuehrung in Theorie und Praxis der Zeitreihen und Modalanalyse, 3rd edn.
Braunschweig: Vieweg.
11. M. L 1991 Structural Dynamics 1, 305–314. Localization of errors in computational models
using test data.
12. G. L and H. A 1991 Structural Dynamics 1, 291–297. Localization
of errors in computational models using test data.
13. C.-P. F and S. Z 1991 Computers and Structures 40, 475–486. Updating of finite
element models by means of measured information.
14. C.-P. F and T. K 1992 Proceedings of the 10th IMAC, 1064–1071. Localization and
correction of errors in analytical models.
15. D. C. Z, T. S and M. K 1995 Proceedings of the 13th IMAC,
179–184. Structural damage detection using frequency response functions.
16. C.-P. F and T. K 1992 Proceedings of the 17th International Seminar on Modal
Analysis and Structural Dynamics, 1581–1596. Localization and correction of errors in
finite-element-models based on experimental data.
17. P. T and M. H. H. S 1994 AIAA Journal 32, 176–183. Structural damage detection and
system identification using neural networks.
18. X. W, J. G and J. H. G 1992 Computers and Structures 42, 649–659. Use of
neural networks in detection of structural damage.
19. N. S and R. O 1987 Proceedings of the 6th International Offshore Mechanics and
Arctic Engineering Symposium 2, 517–524. Global nondestructive damage evaluation of offshore
platforms using modal analysis.
20. R. B, P. H. K and P. A 1995 Proceedings of the 13th IMAC, 661–667.
Damage detection in an offshore structure.
21. Z. Z and M. L 1996 Proceedings of the Identification in Engineering Systems, 58–66.
Identification of delamination in sandwich plates using vibration test data.
22. C.-P. F and S. S 1990 Proceedings of the 3rd International IFToMM Conference
on Rotordynamics, 423–429. Identification of mechanical systems by means of the extended
kalman filter.
   185
23. S. S and C.-P. F 1995 International Journal of Rotating Machinery 1, 267–275.
Identification procedures as tools for fault diagnosis of rotating machinery.
24. D. S̈ and J. B 1991 Proceedings of the 8th IFToMM World Congress on the
Theory of Machines and Mechanisms, 771–774. Crack detection of a rotor by state observers.
25. C.-P. F, S. S and D. B 1995 Proceedings of the 13th IMAC, 1874–1881.
Application of filter techniques for damage detection in linear and nonlinear mechanical systems.
26. H. W and R. S 1990 Mechanical Systems and Signal Processing 4, 195–213. The
application of state observers in structural dynamics.
27. G. G and A. D. D 1988 Computers and Structures 28, 309–313. A finite
element of a cracked prismatic beam for structural analysis.
28. W. T 1990 VDI Fortschritt-Berichte Reihe 11, Längs- und Torsionsschwingungen bei quer
angerissenen Rotoren—Untersuchungen auf der Grundlage eines Rißmodells mit 6
Balkenfreiheitsgraden.
29. W. M. O and M. K 1990 Computers and Structures 36, 245–250. Vibration
analysis of a cracked beam.
30. G. E and M. L 1995 Proceedings of the International Symposium on Non-Destructive
Testing in Civil Engineering (NDT-CE) , 327–334. Identification of crack parameters in concrete
beams using modal test data.
31. J. B 1987 Stable Solutions of Inverse Problems. Braunschweig/Wiesbaden: Friedr.
Vieweg.
32. D. J and C.-P. F 1996 Proceedings of the 17th International Seminar on Modal
Analysis and Structural Dynamics, 1287–1298. On the design of optimal input signals for
parameter identification.
33. S. C and G. L and Y. B-H 1994 Proceedings of the 12th IMAC, 515–520.
Updating linear elastic models with modal-based selective sensitivity.
34. H. A, J. E. M and M. I. F 1996 Proceedings of ISMA21 —Noise
and Vibration Engineering 2, 983–991. Damage detection from substructure modes.
35. F. L. L 1986 Optimal Estimation. New York: John Wiley.
36. H. T and P. C. P and G. R. I 1973 The Stress Analysis of Cracks Handbook.
Hellertown, Pennsylvania: Del Research Corp.
37. J.-T. K and N. S 1995 Journal of Structural Engineering, 1409–1417. Model-uncertainty
impact and damage detection accuracy in plate girder.
38. G. H. G and C. F. V L 1983 Matrix Computations. Baltimore, Maryland: John
Hopkins University Press.
39. S. A. B and K. M. T 1989 Mechanical Systems and Signal Processing 3, 319–339.
Spectral analysis for nonlinear systems, part 1: parametric non-linear spectral analysis.
40. C.-P. F, T. K and K. v. H 1993 Proceedings 4 Tagung ueber Dynamische
Probleme—Modellierung und Wirklichkeit, Schriftenreihe des Curt-Risch-Instituts, 287–307.
Schadenserkennung mit Hilfe von Fehlerlokalisationsverfahren.
41. MW I. 1992 MATLAB Control Toolbox Reference Guide. Natick: MathWorks.
42. J. S 1976 Einführung in die Numerische Mathematik. Berlin: Springer Verlag.
43. H. G 1990 Proceedings of the 8th IMAC, 195–204. Comparison of expansion methods for
FE modeling error localization.
44. MW I. 1995 MATLAB Optimization Toolbox Reference Guide. Natick: MathWorks.

APPENDIX A: GRAM–SCHMIDT QR-ALGORITHM TO SOLVE [A]{x} = {y}


Parameters: xk = 0 for k = 1, 2, . . . , np
Correlation coefficient for kth column: kk = 0 for k = 1, 2, . . . , np
Threshold value: t (e.g. 0.01)
Set of parameter indices: Iopt = { }

-
for k = 1, 2, . . . , np
for i = 1, 2, . . . , np
if index i (Iopt , (means ith column has not yet been chosen)
186 .-.   .
k−1
{w̃} = {ai } − s r̃ji {qj } with r̃ji = {qjT }{ai }
j=1

{w̃} b
{q̃} = ; b = {q̃T}{y}; b =

>{w̃}> >{y}>
 ) q kk :
if abs (b
kk = b
 ; bk = b
{qk } = {q̃} (column of [Q]-matrix)
rjk = r̃ji , for j = 1, 2, . . . , k − 1 (elements of [R])
rkk = {q }{ak } T
k

index vector: mk = i
end
end
end (i-loop)
if kk E t (convergence criterion)
kmax = k − 1
terminate k-loop
else
add parameter to index set: Iopt MIopt + {mk }
end
end (k-loop)

 
bk max
xkredmax =
rk max ,k max
for k = (kmax − 1), . . . , 1

0 1
kmax
1
xkred = bk − s rkj xjred
rkk j=k+1

end (k-loop)

      

for k = 1, 2, . . . , kmax
xmk = xkred
end (k-loop)