You are on page 1of 9

Introduction

In this chapter, Bayesian dynamic model updating technique with noisy incomplete
modal data is presented with applications to structural health monitoring and damage
detection. The modal data consists of modal frequencies (eigen values) and partial mode
shapes (eigen vectors). This Linear Optimisation Iterative technique is used to determine
most probable system modal frequencies, full system mode shapes and most probable
dynamical model within a specified class of models. The advantage of this method is that
it does not require matching between measured modes and corresponding model modes.
The eigen value problem is incorporated in the prior probability distribution to provide
soft constraints and hence need not be solved explicitly. The formulating of this
technique is discussed in the forthcoming sections.

Formulation

The existing global structural health monitoring methods adopt minimisation of the
measure of difference between modal data from finite element model and dynamic test
data to determine the local loss of stiffness in the structure. This methodology requires
matching of modes. The generic form of the goodness-of-fit function to be minimized is:
Jg
(θ) =
Nm m=1
w
m λ(m)(θ) − λˆ (m) 2 + Nm
m=1
w
m
φ(m)(θ) − φˆ (m) 2 (5.5)

where λˆ (m) and φˆ (m) are the measured eigenvalue and eigenvector of the mth mode,
λ(m)(θ) and φ(m)(θ) are the eigenvalue and eigenvector, respectively, of the mth mode
from the dynamical model with parameter vector θ that determines the stiffness and
mass matrix, and wm and wm, m = 1, 2, . . . , Nm, are chosen weightings that depend on
the specific method.

In the method adopted in this work, the realistic assumption is made that only the modal
frequencies and partial mode shapes of some modes are measured. The system mode
shapes are introduced to avoid mode matching between the measured modes and those
of the dynamical model. The system frequencies are also introduced as parameters to be
identified in order to represent actual modal frequencies of the dynamical system. The
eigen equations of the dynamical model are used only in the prior probability distribution
to provide soft constraints.

A class of dynamical models C is considered with Nd DOFs that has a known mass matrix
M ∈ RNd×Nd and the stiffness matrix K( θ) ∈ RNd×Nd is parameterized by θ =[θ1, θ2, . . . ,
θNθ]T ∈ RNθ as follows:
K( θ) = K0 +Nθl=1θlKl (5.11)
where the subsystem stiffness matrices Kl, l = 0, 1, . . . , Nθ, are specified, e.g., by a
finite element model of the structure. The scaling parameters in θ allow the nominal
model matrix
given by θ = θη in Equation (5.11) to be updated based on dynamic test data from the
system.
Assume that Nm(≤ Nd) modes of the system are measured (not necessarily the first Nm
lowest frequency modes), which have eigenvalues λ(m), m = 1, 2, . . . , Nm, and real
eigenvector components φ(m) ∈ RNd, m = 1, 2, . . . , Nm. It is not assumed that these
modal parameters satisfy exactly the eigenvalue problem of any given dynamical model
(M, K( θ)) because there are always modeling approximations and errors. The quantities
λ(m) and φ(m) are referred to as the system eigenvalue and eigenvector, respectively of
the mth mode to distinguish them from the corresponding modal parameters given by
any dynamical model specified by θ.

Formuation of Prior PDF

The prior probability density function


(PDF) for λ = λ(1), λ(2), . . . , λ(Nm ) T and φ = φ(1) T, φ(2) T, . . . , φ(Nm ) T T is chosen as:
p(λ, φ| θ, C) = κ0 exp − 21 Jg(λ, φ; θ) (5.12)
where κ0 is a normalizing constant and the goodness-of-fit function is given by:
Jg
(λ, φ; θ) =
⎡⎢⎣
(K(θ) − λ(1)M)φ(1)
(K(θ) − λ(2)M)φ(2)
.
(K(θ) − λ(Nm )M)φ(Nm )
⎤⎥⎦
T
−1
eq
⎡⎢⎣
(K(θ) − λ(1)M)φ(1)
(K(θ) − λ(2)M)φ(2)
.
(K(θ) − λ(Nm )M)φ(Nm )
⎤⎥⎦

This prior PDF is based on choosing a Gaussian PDF as a probability model for the eigen
equation errors, where the prior covariance matrix eq controls the size of these equation
errors. The uncertainty in the equation errors for each mode are modeled as independent
and
identically distributed, so:

eq = σeq2 INdNm (5.14)


where INdNm denotes the NdNm × NdNm identity matrix and σeq2 is a prescribed equation-
error
variance. The usage of this variance parameter allows for explicit treatment of modeling
error
as the parametric models for the stiffness matrix, and hence the eigen equation, is never
exact
in practice. If this error level can be estimated, the mathematical constraint given by the
eigen
equation will become a soft constraint instead of a rigid constraint. In other words, errors
of the eigen equation in the level corresponding to σeq is allowed. In the numerical
examples
presented later, the value of σ2
eq is chosen to be very small so that the eigen equations are nearly
satisfied. This means that the system modal frequencies and mode shapes will
correspond
closely to modal parameters of the identified dynamical model. For modal data from a
real
structure, this would be a reasonable strategy to start with. If the measured modal
parameters
did not agree well with those corresponding to the identified (most probable) dynamical
model,
implying considerable modeling errors, then σ2
eq could be increased. This procedure allows
explicit control of the inherent trade-off between how well the measured modal
parameters
are matched and how well the eigen equations of the identified dynamical model are
satisfied.
This additional modeling flexibility is an appealing feature of this method.
The prior PDF p( λ, φ| θ, C) implies that, given a class of dynamical models and before
using
the dynamic test data, the most probable values of λ and φ are those that minimize the
Euclidean
norm (2-norm) of the error in the eigen equation for the dynamical model. This implies
that
the prior most probable values of λ and φ are the squared modal frequencies and mode
shapes
of a dynamical model, but these values are never explicitly required. This prior PDF will
have
multiple peaks because there is no implied ordering of the modes here.
The prior PDF for all the unknown parameters is given by:
p( λ, φ, θ| C) = p( λ, φ| θ, C)p( θ| C) (5.15)
where the prior PDF p( θ| C) can be taken as a Gaussian distribution with mean θη
representing
the nominal values of the model parameters and with covariance matrix θ. In the
examples
later, the prior covariance matrix θ is taken to be diagonal with large variances, giving
virtually
a non-informative prior.

To construct the likelihood function, the measurement error is introduced:


ψλˆ = Lλoφ + (5.16)
and a Gaussian probability model is chosen for ∈ RNm (No+1) with zero mean and
covariance matrix
, which can be obtained by Bayesian modal identification methods
[136, 285, 290, 291], ψˆ = ψˆ (1) T, ψˆ (2) T, . . . , ψˆ (Nm ) T T and λˆ = [λˆ (1), λˆ (2), . . . ,
λˆ (Nm )]T,
where ψˆ (m) ∈ RNo gives the observed components of the system eigenvector of the mth
mode
and λˆ (m) gives the corresponding observed system eigenvalue from dynamic test data.
Finally,

L
o is an NmNo × NmNd observation matrix of ‘1s’ or ‘0s’ that picks the components of φ
corresponding to the No measured DOFs. The likelihood function is therefore:
p λˆ, ψˆ |λ, φ, θ, C = p λˆ, ψˆ |λ, φ = G λˆ T, ψˆ T T ; λT, (φ)T T , (5.17)
that is, a Gaussian distribution with mean λT, ( Loφ)T T and covariance matrix .

The posterior PDF for the unknown parameters is given by the Bayes’ theorem:
p λ, φ, θ|λˆ, ψˆ , C = κ1p λˆ, ψˆ |λ, φ, θ, C p(λ, φ|θ, C)p(θ| C)
= κ1p λˆ, ψˆ |λ, φ p(λ, φ|θ, C)p(θ| C)
(5.18)
The most probable values of the unknown parameters can be found by maximizing this
PDF.
To proceed, the objective function is defined as [286]:
J(λ, φ, θ) =
12
(θ − θη)T−θ 1 (θ − θη) + 1
2σ2
eq
Nm m=1
K(θ) − λ(m) M φ(m) 2
+
12
ψˆ λˆ−−Lλoφ T − 1 ψˆ λˆ−−Lλoφ
(5.19)
which is the negative logarithm of the posterior PDF without including the constant that
does not
depend on the uncertain parameters. Here, ||.|| denotes the Euclidean norm. Then, the
function
J(λ, φ, θ) is minimized instead of maximizing the posterior PDF. The objective function J is
not a quadratic function for the uncertain parameters. However, this function is quadratic
for
any of the uncertain parameter vectors of λ, φ or θ if the other two are fixed. Therefore,
the
original nonlinear optimization problems can be done iteratively through a sequence of
linear
optimization problems, as shown in the next section.

Linear Optimisation Problem

The mode shapes are usually measured with incomplete components, i.e., with missing
DOFs
but the modal frequencies are measured with relatively high accuracy. Therefore, the
sequence
of optimization starts from computing the missing components of the mode shapes. First
set
the updated model parameters at their nominal values:
θ = θη (5.20)
and the eigenvalues at their measured values:
λ = λˆ

Then, perform a sequence of iterations comprised of the following linear optimization


problems:
φ = arg min
φ
J( λ, φ, θ)
λ = arg min
λ
J( λ, φ, θ)
θ = arg min
θ
J( λ, φ, θ)
(5.22)
until the prescribed convergence criterion is satisfied. Each of these three optimization
problems
is explained in more detail in the next three subsections.
5.3.1 Optimization for Mode Shapes
By minimizing the objective function J( λ, φ, θ) in Equation (5.19) with respect to φ, the
optimal vector φ can be obtained:
φ = [σ−2
eq Gφ + LTo ( − 1 )22Lo]−1LTo [( − 1 )21 ( λˆ − λ) + ( − 1 )22 ψˆ ] (5.23)
where ( − 1 )21 and ( − 1 )22 are the NmNo × Nm left bottom and NmNo × NmNo right
bottom
sub-matrices of −1
, and the symmetric matrix Gφ is given by:
G
φ=
⎡⎢⎣
(λ(1)M − K)2 0
(λ(2)M − K)2
.
.
.
0 (λ(Nm )M − K)2
⎤⎥⎦
NdNm×NdNm
(5.24)
where the updated stiffness matrix K = K( θ).
5.3.2 Optimization for Modal Frequencies
By minimizing the objective function J( λ, φ, θ) in Equation (5.19) with respect to λ, the
updated parameter vector λ is given by:
λ = [Gλ + ( − 1 )11 ]−1σeq−2
⎡⎢⎣
φ(1)TMKφ(1)
φ(2)TMKφ(2)
.
φ(Nm )TMKφ(Nm )
⎤⎥⎦
+ (− 1 )11 λˆ + (− 1 )12(ψˆ − Loφ)

where ( − 1 )11 and ( − 1 )12 are the Nm × Nm left top and Nm × NmNo right top sub-
matrices
of −1
, and the matrix Gλ is given by:
Gλ = σ−2
eq
⎡⎢⎣
φ(1) TM2 φ(1) 0
φ(2) TM2 φ(2)
.
.
.
0 φ(Nm ) TM2 φ(Nm )
⎤⎥⎦
Nm
×N
m

(5.26)
5.3.3 Optimization for Model Parameters
By minimizing Equation (5.19) with respect to θ, the updated model parameter vector θ
is
given by:
θ = (σ−2
eq GθTGθ + − θ 1 )−1 (σeq−2GθTb + − θ 1 θη) (5.27)
where the matrix Gθ is given by:
Gθ =
⎡⎢⎣
K1 φ(1) K2 φ(1) · · · KNθ φ(1)
K1 φ(2) K2φ(2) · · · KNθ φ(2)
.
.
.
.
.
.
K1 φ(Nm ) K2φ(Nm ) · · · KNθ φ(Nm )
⎤⎥⎦
NdNm× Nθ
(5.28)
and the vector b ∈ RNdNm is given by:
b=
⎡⎢⎣
(λ(1) M − K0)φ(1)
(λ(2) M − K0)φ(2)
.
(λ(Nm ) M − K0)φ(Nm )
⎤⎥⎦
NdNm× 1
(5.29)
5.4 Iterative Algorithm

The three optimization problems are coupled. In other words, the expressions of the
optimal
vectors depend on the optimal values of other parameters. In order to search for the
overall
optimal parameters efficiently, the proposed methodology updates the system modal
frequencies, system mode shapes and stiffness scaling parameters in an iterative
manner by using
successively the optimization results of Section 5.3. The iterative procedure consists of
the
following steps:
1. Take the initial values of the model parameters as the nominal values, θ = θη, and the
eigenvalues as the measured values, λ = λˆ . Then, K = K(θ).

2. Update the estimates of the system eigenvectors φ(m), m = 1, 2, . . . , Nm, using


Equation (5.23).
3. Update the estimates of the system eigenvalues (squared modal frequencies) λ(m), m
=
1, 2, . . . , Nm, using Equation (5.25).
4. Update the estimates of the model parameters θ by using Equation (5.27).
5. Iterate the previous Steps 2, 3 and 4 until the model parameters in θ satisfy some
convergence criterion, thereby giving the most probable values of the model parameters
based on
the modal data.
5.5 Uncertainty Estimation
The posterior PDF in Equation (5.18) can be well approximated by a Gaussian distribution
centered at the optimal (most probable) parameters ( λ, φ, θ) and with covariance
matrix ( λ, φ, θ) equal to the inverse of the Hessian of the objective function J( λ, φ, θ) =
− ln p( λ, φ, θ| λˆ, ψˆ , C) calculated at the optimal parameters [19]. This covariance
matrix is
given by

0.42235 1.2646 2.0994 2.9218 3.7272

Three-dimensional Six-story Braced Frame


In this example, the Bayesian model updating method is applied to update the finite-
element
model of a three-dimensional six-story braced frame, which is based on a model of an
actual
laboratory test structure. It is square in plan with width a = 5 m. There are four columns
for
each floor, one at each corner. Each of them have interstory stiffnesses of 10 MN/m and
15
MN/m in the x and y directions, respectively. Furthermore, each face in each floor is
stiffened
by a brace and its stiffness is taken to be 20 MN/m. As a result, the interstory stiffness is
80
MN/m and 100 MN/m in the x and y directions, respectively. The floor mass is taken to be
10 metric tons for each floor. As a result, the first five modal frequencies of the structure
are
3.432, 3.837, 6.305, 10.10 and 11.29 Hz.
In order to locate the face(s) that sustain damage, four stiffness parameters are used for
each story to give twenty-four stiffness parameters, θ4(l−1)+1 = Kl,+x, θ4(l−1)+2 = Kl,+y,
θ
4(l−1)+3 = Kl,−x and θ4(l−1)+4 = Kl,−y, l = 1, 2, . . . , 6, where the index l represents the
story number and ‘+x’, ‘+y’, ‘−x’ and ‘−y’ represent the direction of the outward normal
of the face. The actual values of these stiffnesses are Kl,+x = Kl,−x = 50 MN/m and
K
l,+y = Kl,−y = 40 MN/m, l = 1, 2, . . . , 6. In other words, θ1 = θ3 = · · · = θ23 = 50 MN/m
and θ2 = θ4 = · · · = θ24 = 40 MN/m. The floor plan is shown in Figure 5.3. The point
O
l(¯xl, y¯l) is the stiffness center of floor l, where ¯xl and ¯yl, l = 1, 2, . . . , 6, are given by:
x¯l =
a(Kl,+x − Kl,−x)
2(Kl,+x + Kl,−x)
; ¯yl =
a(Kl,+y − Kl,−y)
2(Kl,+y + Kl,−y

The structure is a 4-story, 2-bay by 2-bay steel frame scale-model structure in


the Earthquake Engineering Research Laboratory at the University of British
Columbia (UBC). It has a 2.5 m 2.5 m in plan and is 3.6 m tall. The members are
hot rolled grade 300W steel with a nominal yield stress 300 MPa (42.6 kpsi). The
sections are unusual, designed for a scale model, with properties as given in
Table 1. The columns are all oriented to be stronger bending toward the x-
direction (i.e., about the y-axis). The floor beams are oriented to be stronger
bending vertically, i.e., about the y-axis (x-axis) for those oriented with
longitudinal axis parallel to the x-axis (y-axis). The braces have no bending
stiffness, so their orientation is irrelevant. There is one floor slab per bay per
floor: four 800 kg slabs at the first level, four 600 kg slabs at each of the second
and third levels, and, on the fourth floor, either four 400 kg slabs. A finite
element model based on this structure was developed to generate the simulated
response data. It is a 12DOF shear-building model that constrains all motion
except two horizontal translations and one rotation per floor. The columns and
floor beams are modeled as Euler-Bernoulli beams in the finite element model.
The braces are bars with no bending stiffness. The x-direction (i.e., bending
about the y-axis) is the strong direction due to the orientation of the columns.
Further, to be consistent with the axes used in later experimental, the compass
directions associated with the axes are South for the positive y (weak) direction,
and West for the positive x (strong) direction.

The discussion above uses the most probable values of the stiffness parameters based
on
modal data from the undamaged and damaged structure. In order to further portray the
damage,
these are the most probable values and the standard deviations for the stiffness
parameters are
used to compute the probability that a given stiffness parameter θl has been reduced by
a certain
fraction d compared to the undamaged state of the structure [18, 268]. An asymptotic
Gaussian
approximation [197] is used for the integrals involved to give:
Pdam
l (d) = P(θlpd < (1 − d)θlud| C)
= −∞∞ P(θlpd < (1 − d)θlud|θlud, C)p(θlud| C)dθlud
≈ ⎡⎣(1(1−−dd)2)(θσllud)−2 +θl(pdσlpd)2 ⎤⎦

where (·) is the cumulative distribution function of the standard Gaussian random
variable,
θ
l
ud
and θ
l
pd denote the most probable values of the stiffness parameters for the undamaged
and (possibly) damaged structures respectively, and σlud and σlpd are the corresponding
standard
deviations of the stiffness parameters.
The probabilities of damage for the twenty-four θl are shown in Figure 5.6. It can be
clearly
seen that the +x face of the first story and the +y face of the fourth story have damage
with a
probability of almost unity. The means of the damage are 15.5% and 10.5% while their
actual
values are 15% and 10%, respectively. Furthermore, the COVs of these estimates are
0.6% and
1.0%, respectively. These types of plots can be interpreted as follows. Consider the curve
for
the probability of damage of the +x face of the first story. The damage is 13.5% or more
with a
probability of almost 1.0 and has a probability of only 0.05 to exceed 16.3%. For the
twenty-two
faces that are undamaged, the difference between the mean stiffness parameter values
of the
undamaged and damaged structures is less than 2.5%, which is within the expected
uncertainty
level.

You might also like