Vous êtes sur la page 1sur 22

EARTHQUAKE ENGINEERING AND STRUCTURAL DYNAMICS

Earthquake Engng Struct. Dyn. 2005; 34:909–930


Published online 31 January 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/eqe.460

LMS-based structural health monitoring of a non-linear


rocking structure

J. Georey Chase1; ∗; † , Hannes A. Spieth2 , C. Francois Blome1 and J. B. Mander2


1 Department of Mechanical Engineering; University of Canterbury; Private Bag 4800;
Christchurch; New Zealand
2 Department of Civil Engineering; University of Canterbury; Private Bag 4800; Christchurch; New Zealand

SUMMARY
A structure’s health or level of damage can be monitored by identifying changes in structural or modal
parameters. This research directly identies changes in structural stiness due to modelling error or
damage for a post-tensioned pre-cast reinforced concrete frame building with rocking beam column
connections and added damping and stiness (ADAS) elements. A structural health monitoring (SHM)
method based on adaptive least mean squares (LMS) ltering theory is presented that identies changes
from a simple baseline model of the structure. This method is able to track changes in the stiness
matrix, identifying when the building is (1) rocking, (2) moving in a hybrid rocking–elastic regime,
or (3) responding linearly. Results are compared for two dierent LMS-based SHM methods using an
L2 error norm metric. In addition, two baseline models of the structure, one using tangential stiness
and the second a more accurate bi-linear stiness model, are employed. The impact of baseline model
complexity is then delineated. The LMS-based methods are able to track the non-linearity of the system
to within 15% using this metric, with the error due primarily to lter convergence rates as the structural
response changes regimes while undergoing the El Centro ground motion record. The use of a bi-linear
baseline model for the SHM problem is shown to result in error metrics that are at least 50% lower
than those for the tangential baseline model. Errors of 5 –15% with this L2 error norm are fairly
stringent compared to the greater than 2 × changes in stiness undergone by the structure, however, in
practice the usefulness of the results is dependent on the resolution required by the user. The impact
of sampling rate is shown to be negligible over the range of 200–1000 Hz, along with the choice of
LMS-based SHM method. The choice of baseline model and its level of knowledge about the actual
structure is seen to be the dominant factor in achieving good results. The methods presented require
2.8 –14:0 Mcycles of computation and therefore could easily be implemented in real time. Copyright
? 2005 John Wiley & Sons, Ltd.

KEY WORDS: structural health monitoring; SHM; adaptive ltering; LMS; damage detection; non-linear
structure; rocking structure

∗ Correspondence to: J. Georey Chase, Department of Mechanical Engineering, University of Canterbury,


Private Bag 4800, Christchurch, New Zealand.
† E-mail: geo.chase@canterbury.ac.nz

Contract=grant sponsor: German Scientic Society (Deutsche Forschungsgemeinschaft DFG)


Contract=grant sponsor: EQC Research Foundation of New Zealand
Received 1 June 2004
Revised 22 November 2004
Copyright ? 2005 John Wiley & Sons, Ltd. Accepted 22 November 2004
910 J. G. CHASE ET AL.

INTRODUCTION

Structural health monitoring (SHM) measures the current state of a structure’s condition rela-
tive to a baseline state, thus determining the existence, location, and degree of damage that may
exist, particularly after a damaging input, such as an earthquake or other large environmental
load. SHM in civil structures is useful for determining the damage state of a structure. In
particular, the ability to assess damage in real-time or immediately after a catastrophic event,
such as an earthquake or terrorist bomb blast, would allow Civil Defence authorities to deter-
mine which structures were safe. In 1999, the International Association for Structural Control
(IASC) and the Dynamics committee of the American Society of Civil Engineers (ASCE)
Engineering Mechanics Division formed the SHM Task Group to study the ecacy of var-
ious SHM methods. The IASC–ASCE SHM Task Group developed a series of Benchmark
SHM problems and established a set of specic Benchmark results for a specially designed
test structure in the Earthquake Engineering Research Laboratory at the University of British
Columbia [1]. However, the Benchmark Problem structure is largely linear in its response,
aside from pre-determined damage. Hence, methods that work well for this largely linear
system may not work as well for more non-linear pre-cast or reinforced concrete structures.
Current vibration-based SHM methods are based on the idea of modal or frequency domain
damage detection where changes in modal parameters, such as frequencies, mode shapes and
modal damping, are a result of changes in the physical mass, damping and stiness of the
structure [2]. SHM can simplify typical visual or localized experimental methods, such as
acoustic or ultrasonic methods, magnetic eld methods, radiography, eddy-current methods or
thermal eld methods [3], as it does not require visual inspection of the structure and its
connections or components. Doebling et al. [2] give an excellent review of the numerous
dierent approaches for vibration-based damage detection methods.
For civil applications, current methods are more applicable to steel frame or bridge struc-
tures where vibration response may be more linear. These problems typically also have
known, or reasonably estimated, input loads. However, the insensitivity of modal parameters to
(localized) damage can be a major limitation for the many methods that rely on identifying
these parameters to assess and locate damage.
A common method for identication of civil structural model parameters is the eigensystem
realization algorithm (ERA). The ERA method is based on knowledge of the time domain
free response data. In ERA, a discrete Hankel matrix is formed, and the state and output
matrices for the resulting discrete matrix are determined. The resulting modal parameters are
found by determining the eigenvalues of this continuous time system. Dyke et al. [4] use
cross-correlation functions in conjunction with the ERA method for identication of modal
parameters that are used to identify frequency and damping parameters. Caicedo et al. [5]
introduces SHM methods based on changes in the component transfer functions of the struc-
ture, or transfer functions between the oors of a structure, and use the ERA to identify the
natural frequencies of each component transfer function. Lus and Betti [6] also proposed a
damage identication method based on ERA with a Data Correlation and Observer=Kalman
Identication algorithm. Bernal and Gunes [7] also used the ERA with Observer=Kalman Iden-
tication for identifying modal characteristics when the input is known, and used a Subspace
Identication algorithm when the input cannot be measured.
Wavelet analysis approaches for SHM are found in Corbin et al. [8] and Hou et al. [9].
Damage, and the moment it occurs, can be detected by a spike or impulse in the plots

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 911

of higher resolution details from wavelet decomposition of the acceleration response data.
Wavelets oer the advantage of determining not only the extent of the damage but also the
time of its occurrence.
A major drawback of these current approaches is their inability to be implemented in real-
time, as the event occurs, and the diculty with highly non-linear systems. The wavelet and
ERA methods require the entire measured response to process and identify damage. Further,
their reliance on modal properties, which can be subject to noise, has potential problems.
Modal properties have also been shown to be non-robust in the presence of strong noise and
insensitive to small amounts of damage [9]. Finally, modal approaches are intractable for
highly non-linear structures like the pre-stressed concrete rocking structure examined in this
research.
It should be noted that Verboven et al. [10, 11] did develop a frequency domain maxi-
mum likelihood estimator for modal identication that was more robust to the typical noise
environment. Adaptive identication methods have also been used to identify modal param-
eters by Sato and Qi [12] and Loh et al. [13]. Loh et al. [13] used the adaptive fading
Kalman lter technique, and Sato and Qi [12] used an adaptive H∞ lter, to achieve real-time
capable or near real-time capable results. What these approaches provide in potential real-time
identication of modal parameters comes with signicant computational complexity.
This paper presents the development of a much simpler and ecient algorithm for contin-
uously monitoring structural status. Adaptive LMS ltering is employed for its computational
simplicity to develop a SHM method that takes advantage of this lter’s ability to adaptively
model noisy signals to identify changes in structural parameters in comparison to a base struc-
tural model, using a series of coupled adaptive LMS lters. In contrast to most of the prior
work in civil SHM, this research uses adaptive ltering to directly assess non-linear, dynamic
changes in stiness and is applied to a highly non-linear structure based on common pre-cast
reinforced concrete structural elements.

SHM PROBLEM DEFINITION

A seismically excited structure can be modelled using standard linear equations of motion, or
a more complex computational model if desired:
M · {v} + C · {v̇} + K · {v} = −M · xg (1)
where M, C and K are the mass, damping and stiness matrices of the model, respectively,
{v}, {v̇} and {v} are the displacement, velocity and acceleration, respectively, and xg is
the ground motion acceleration. This system represents a linear, undamaged baseline model.
A more complex model can be used, based on greater knowledge of the structural dynamics,
without loss of generality.
If damage occurs, structural properties such as natural frequency and stiness will change.
These changes can be time varying or result from simple modelling error. For the damaged,
or mis-modelled structure the equations of motion can be re-dened:
M · {v} + C · {v˙} + (K + K(t)) · {v} = −M · xg (2)
where v,
 v˙ and v are the measured responses of the damaged structure, and K(t) contains
time-varying changes in the stiness. Identifying the K term enables the structure’s condition

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
912 J. G. CHASE ET AL.

to be directly monitored without using modal parameters. Damping changes, C, could also be
identied and can occur due to hysteresis. Change in the mass matrix, M, is not likely to be
signicant and is ignored. For the rocking structure analysed in this research, the time-varying
change in stiness as the structure rocks is the only quantity identied.
To determine K using adaptive LMS, a new form of K is dened with time varying
scalar parameters,  i , to be identied using the LMS lter. For example, using a linear three
storey shear building baseline model, K can be sub-divided into three matrices with entries
of 1, −1 and 0 to allow independent identication of changes in storey stiness k1 , k2 and k3
by putting a 1 or −1 wherever it would appear in a detailed form of Equation (1).
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
1 0 0 1 −1 0 0 0 0 1 + 2 −2 0
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
K = 1 ⎢
⎣0 0 0⎥ ⎢
⎦ + 2 ⎣ −1 1 0⎥ ⎢
⎦ + 3 ⎣ 0 1 −1 ⎥ ⎢
⎦ = ⎣ −2 2 +  3 −3 ⎥

0 0 0 0 0 0 0 −1 1 0 −3 3
(3)

where

1 = k1 ; 2 = k2 ; 3 = k3 (4)

Hence, the change in stiness is dened:


⎡ ⎤
k1 + k2 −k2 0
⎢ ⎥  3
K = ⎢ ⎣ −k2 k2 + k3 −k3 ⎥
⎦ =  i Ki (5)
i=1
0 −k3 k3

The stiness of the damaged structure, or eective stiness changes due to non-linear be-
haviour such as rocking, yielding or hysteresis, can then be determined by identifying the K
matrix at every time step. Rewriting Equation (2) using Equations (3)–(5) yields:

n
M · {v} + C · {v˙} + K · {v} +  i Ki v = F (6)
i=1

where n is the number of degrees-of-freedom (DOFs) of the model and F is a known, or


estimated, input load vector. Note that n is the maximum number of coecients to identify
changes in each storey stiness. A lesser number can be used if some storeys are assumed
not to suer damage, or a greater number could be used for a more complex structural model
with more DOF per storey, to obtain greater resolution in structural changes.
Note that the baseline stiness model, K, need not be a shear building model as illustrated
by example in Equations (3)–(5), where a shear model is used for simplicity and as a basic
standard. The baseline model can be more or less complex, or even zero. The latter choice
(K = 0) makes the identication of K the same as the complete (time varying) stiness
model identication problem.
The varying stiness term is therefore simply the error between, in this case, the linear
baseline model and real measurements when actual measured values (v,  v˙ and v)
 are used in

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 913

Equation (6). Hence, Kv is the error in the baseline model.



n
 i Ki v = F − Mv − Cv˙ − Kv (7)
i=1

where v, v˙ and v are measured values of the structural displacement, velocity and acceleration
that are obtained either directly and=or from a dynamic state estimator. Equation (7) is only
valid at any point in time if  i has the correct values. At any discrete time, k, the dierence
between the linear model and actual measurements can be dened:

n
yk = Fk − Mvk − Cv˙k − Kvk =  i Ki v (8)
i=1

where Fk is the input at time k, and vk , v˙k and vk are the measured displacement, velocity and
acceleration at time k. The elements of the vector signal yk can be readily modelled in real-
time using an adaptive LMS lter [14] so that the coecients  i can be readily determined
from the reduced noise modelled signal at each time step.

ADAPTIVE LMS FILTERING

Adaptive lters are digital lters with coecients that can change over time based on how
well the existing coecients are performing in modelling a noisy signal. Because of their
self-adjusting performance and built-in exibility, adaptive lters have found use of mod-
elling signals in many real-time applications, particularly in advanced telecommunications.
The least mean squares (LMS) algorithm is one of the most widely used and computationally
simplest adaptive ltering algorithms. The algorithm approximates steepest descent method
optimization, using an estimate of the gradient instead of its actual value to considerably
simplify the calculations for real-time applications. The goal in this case is to model the
individual, scalar elements of the signal yk in Equation (8).
In adaptive LMS ltering, the coecients are adjusted from sample-to-sample to minimize
the mean square error (MSE), between a measured noisy scalar signal and its modelled value

m−1
ek = ŷk − WkT X k = ŷk − wk (i)xk−i = ŷk − n̂k (9)
i=o

where Wk is the adjustable lter coecient vector or weight vector at time k, ŷk is the noisy
measured signal to be modelled or approximated, X k is the vector input to the lter model of
current and previous lter outputs, xk−i , so WkT X k is the vector dot product output from the
lter to model a scalar signal ŷk , and m is the number of prior time steps or taps considered.
Note that more taps makes the algorithm converge faster and improves accuracy, however,
it takes more computation. The Widrow–Hopf LMS algorithm for updating the weights to
minimize the error, ek , is dened [14]:
Wk+1 = Wk + 2 · ek · X k (10)
where  is a positive scalar that controls the stability and rate of convergence. The overall
procedure is outlined in detail by Ifeachor and Jervis [14] and other signal processing texts.

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
914 J. G. CHASE ET AL.

ADAPTIVE LMS FOR SHM

Two-step method
The elements of a noisy signal vector, yk , as dened in Equation (8), can be modelled using
multiple, individual adaptive LMS lters from Equations (9) and (10).
⎡ ⎤ ⎡ ⎤
(ŷk )1 (WkT X k )1
⎢ ⎥ ⎢ ⎥
⎢ . ⎥ ⎢ . ⎥

yk = ⎢ . ⎥ = ⎢
. ⎥ ⎢ . ⎥ (11)
. ⎥
⎣ ⎦ ⎣ ⎦
(ŷk )n (WkT X k )n
where each WkT is updated individually for n dierent input signals and (WkT X k ) i is the output
for the i-th individual adaptive LMS lter. In the two-step method, adaptive LMS lters
approximate the noisy signal and, from Equation (8), the lter approximation is dened as:

n
yk =  i Ki vk = [K1 vk · · · Kn vk ] k = A k (12)
i=1

where dimensions of matrix A are n × n and  k is an n × 1 vector of coecients  i at time k.


Therefore, the  i values can be determined analytically by solving Equation (12) at each time
step as long as the matrix A is full rank.
Note that the  i values will be correct as long as the lter has converged to zero error.
During convergence these values will be only an approximation. Finally, this two-step method
is fast and robust to noise because of its use of LMS lters. However, the computation required
is more intense than desired due to the matrix solutions required for each time step, particularly
as the number of DOFs rises [15].

One-step method
The one-step method uses a further approximation of the gradient to combine the two steps
of noisy signal modelling and coecient solution in the two-step method. The linear model
error between the measured noisy signal and its modelled value from the lter dened in
Equation (9) can be expressed as:
 
m−1 n
ek = yk − ij Ki vk = yk − Q k (13)
j=0 i=1

where vk and yk are noisy signals, Q k is an n × 1 vector,  ij are weights where i = 1; : : : ; n
and j = 1; : : : ; m. Hence, the change in the i-th storey stiness, ki , will be the sum over j
of  ij . This averaged approach low pass lters the signal vk and further reduces the impact
of noise. An exact unltered solution would use m = 1. Note that there are no prior time
steps involved when estimating the error at time k, because yk is not stationary and adaptive
LMS based algorithms are not as eective for non-stationary signals [14]. More directly, the
error, ek , in Equation (14) is the error at this time step and is also a function of the response at
time k only.
Hence, the mean square error (MSE) can be dened as:
ekT ek = ykT yk + QTk Q k − 2ykT Q k (14)

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 915

To minimize the MSE with respect to the weights  ij , the optimum solution occurs when the
gradient of MSE is zero. The gradient of MSE is dened as:
  @Q k @Q k
∇MSEij = ∇ij ykT yk + QTk Q k − 2ykT Q k = 2 QTk − ykT = −2ekT (15)
@ij @ij
where ∇MSEij is one element of an n × m matrix ∇MSE, and @Q @ij is an n × 1 vector dened
k

as:


@Q k @  
m−1 n
= ij Ki vk = Ki vk (16)
@ij @ij j=0 i=1
Therefore,

∇MSE = −2 ekT Ki vk (17)
where ∇MSE is an n × m matrix and [ekT Ki vk ] is the same across an entire row for all
i = 1; : : : ; n rows. The weight matrix of dimension n × m can then be updated.

wk+1 = wk − ∇MSE = wk + 2 ekT Ki vk (18)
where the n × m matrix [ekT Ki vk ] is the same for all m elements in the i-th row, for which
the expression is shown.
However, Chase et al. [15] showed that decoupling the gradient estimation by approxi-
mating Ki as a zero matrix with a single 1.0 valued element in the (i; i) location provides
better results for the Benchmark Problem by signicantly reducing convergence time at near-
zero gradient values. This matrix therefore is all zeros with a single 1.0 value in the i-th
element of the matrix diagonal. Therefore, the following weight update formula is employed:

wk+1 = wk + 2 ekT (i)vk (i) n×m (19)
where the term [ekT (i)vk (i)] is the same for all m elements in row i. Note that the error calcu-
lation in Equation (13) still uses Ki as originally dened with coupling, and Equation (19)
only changes the means by which weights are updated.

ROCKING STRUCTURE DESIGN

Recent earthquakes, such as Northridge 1994 and Kobe 1995, have conrmed the adequacy
of conventional seismic design techniques at preventing structural collapse but have also high-
lighted the extensive damage levels that can be expected in the process. A desire now exists
to seismically design structures to not only preserve life safety but to do so with a minimum
of damage accumulation. Previous experimental work by Priestley and MacRae [16], Cheok
and Lew [17], Stone et al. [18], Nakaki et al. [19], and Priestley et al. [20], has shown that
pre-cast concrete frames constructed from post-tensioned segmental elements perform notably
better than conventional frames when subjected to lateral loading. Structural deformations are
accommodated by gap openings at the pre-cast component interfaces. Similar strength and
inter-storey drift capacities are achievable but damage to the structural elements is signi-
cantly reduced and the frames possess a re-centring capacity that results in negligible residual
displacements.

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
916 J. G. CHASE ET AL.

Figure 1. Model test structure in the experimental set-up.

The post-tensioned segmental element concept as developed previously was used by apply-
ing a damage avoidance design (DAD) philosophy [21] in which the only permissible damage
is that which occurs to replaceable components. An experimental investigation of a shake table
test of a quarter-scale two storey one bay post-tensioned pre-cast reinforced concrete frame
structure has been performed at the University of Canterbury [22]. A photo of the model test
structure is shown in Figure 1.
A computer model of the test structure was developed using the nite element program
RUAUMOKO [23], and validated using the experimental dynamic frame test results. A multi-
spring modelling approach, based on subassembly investigations, was used [24, 25]. To in-
crease the rocking of the modelled structure using the El Centro North–South component
ground motion record used in the experimental tests, the seismic mass of the structure was dou-
bled compared to the experimental investigations. A schematic drawing of the RUAUMOKO
model and a drawing of the test set-up are shown in Figure 2.

RESULTS

The adaptive LMS-based SHM method was tested using acceleration data generated using
the veried rocking structure model [24] for the El Centro North–South component ground
motion input with PGA = 0:348g. The data was generated using a veried model to reduce
the dynamic testing of the test article and allow uniform testing of dierent sample rates and
ltering approaches. Each simulation covered 15 s. Lateral accelerations were taken for the

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 917

(a) (b)

Figure 2. Test set-up and RUAUMOKO model of shake table frame test: (a) test
set-up and; (b) RUAUMOKO model.

base and each of the two storeys in Figure 2, sensor noise added, and numerically integrated
to obtain displacement and velocity data.
The numerical integration included oset and drift correction. In practice, additional sensors
and=or more advanced lters would be required to obtain reasonable estimations of the ve-
locity and displacement, particularly if permanent deformation occurs. For rocking structures
designed to avoid permanent deformation through rocking connections it is less of an issue.
Both advanced ltering based estimators and numerical integration provide an added level
of low-pass ltering. In this research, estimation of displacement and velocity is assumed,
in this case with errors due to the acceleration sensor noise modelled and simple numerical
integration employed, to focus primarily on the methods developed.
This model-based testing approach is not equivalent to explicit experimental testing. It
is undertaken to enable multiple methods and sampling rates to be tested without explicit
interpolation or decimation of the data, and without re-testing the structural artefact. To ensure
that the simulations are as real as possible, a calibrated, non-linear nite element model of
several hundred degrees of freedom is used to ensure all the non-linear dynamics are captured.
In addition, the noise added is modelled based on: (a) the heavy 50 Hz power source noises
evident in this lab and (b) the measured sensor noise variation with a conservative uniform
distribution. Hence, the modelled noise is very similar to what is encountered in this laboratory
setting and uses conservative choices, versus a simple random addition. However, errors due
to poor calibration of sensors, sensor failure or other systematic errors, are lost with this
approach. Finally, the use of the nite element model can act as a pre-lter of the data, with
any unmodelled behaviour not being captured. Therefore, the approach taken represents an
ideal experimental system and may not fully capture experimental reality.
To determine the optimal lter parameter step size, , and number of taps, m, several
dierent combinations were tested. To compare results, the size of the lter error,
e = yout − yout-lter , was computed and normalized for each of the two storeys.
e 2
(20)
yout 2

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
918 J. G. CHASE ET AL.

Table I. Two-step method results using bi-linear model for


dierent values of , with m = 10 taps.
f = 200 Hz f = 1000 Hz
 Taps = 10 Taps = 10

1st 0.977 0.883


0.1 2nd 0.941 0.750
1st 0.769 0.353
1 2nd 0.603 0.280
1st 0.264 0.168
10 2nd 0.211 0.123
1st 0.169 0.177
50 2nd 0.126 0.151
1st 0.155 0.490
100 2nd 0.117 0.535
1st 0.155 1.721
150 2nd 0.120 1.993
1st 0.162 60.804
200 2nd 0.132 21.776
1st 6:3 × 103 3:9 × 1031
300 2nd 2:9 × 103 1:9 × 1031

where e2 is the two-norm of the error between the true baseline model error and the
one estimated by the LMS lter, and yout 2 is the two-norm of the true baseline model
error. The smaller the value of this ratio, the less error energy, in the L2 sense, in the lter
approximation, and therefore the lower the error in the system identication as the lter
converges.
Both the two-step and one-step methods were tested for baseline models using both tan-
gential or eective stiness, and a bi-linear stiness model. Both of these models are far
simpler than the non-linear nite element model used to generate the analysis data. The
eective stiness model provides a simple linear baseline model with almost no knowledge
of the non-linear structural dynamics that represents a linear approximation to the non-linear
dynamics [26–28]. The bi-linear model is a better approximation incorporating more knowl-
edge of the non-linear structural dynamics in the baseline model. Sampling rates of 200 Hz
and 1000 Hz were also investigated. For each case a variety of values for  and m were
tested and the error metric recorded for both the rst and second storeys to determine the
best combination.

Two-step method results


Tables I and II show the results for the two-step method with bi-linear stiness model for
dierent values of  and m = 10 taps in Table I, and dierent numbers of taps in Table II,
which uses the values of  = 10 and 100 for sampling rates of 200 Hz and 1000 Hz based on
the results in Table I, where these values resulted in the best error metric results of 10–16%.
Note that in Table I larger values of  eventually result in unstable lter results, characterized
by sharp unstable spikes in the estimated values over short periods of time, which occur during

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 919

Table II. Two-step method results using bi-linear model for


dierent of taps, m.
f = 200 Hz f = 1000 Hz
Taps  = 100  = 10

1st 0.264 0.353


1 2nd 0.211 0.280
1st 0.169 0.195
5 2nd 0.126 0.148
1st 0.155 0.168
10 2nd 0.117 0.123
1st 0.155 0.159
15 2nd 0.120 0.115
1st 0.162 0.155
20 2nd 0.132 0.113
1st 7:7 × 1012 0.177
50
2nd 4:9 × 1012 0.151

rocking motion of the structure when the system is rapidly changing state between the stier
initial state and more compliant rocking state. Similarly, Table II shows that too many taps
can be unstable as the system cannot track fast enough, while too few do not average out
noise and high-frequency variation well enough.
At a sampling rate of 200 Hz,  = 100 and m = 10 are the best combination. At a sampling
rate of 1000 Hz, the values are  = 10 and m = 20. Note that a sampling rate of 1000 Hz and
m = 50 would result in the same low-pass lter corner frequency as m = 10 at 200 Hz, and
the results for m = 50 at 100 Hz is very nearly equal to the results with m = 20. Hence, the
results are conformable in terms of the implicit low-pass ltering of the signal noise using
this approach.
Tables III and IV show the same results for the two-step method using the eective stiness
baseline model, which is a less accurate representation of the actual behaviour. Hence, it
represents a case where the SHM method has less knowledge of the structural dynamics,
creating a more dicult test case. The results show error metric values of 0.85–1.13, which
are signicant levels of error in comparison to the bi-linear baseline model in Tables I and II,
as well as in an absolute sense. However, it should be noted that the error metric is a measure
of total error energy and may not be a complete description.
To determine the cause of this larger error, the lter error, e, is shown for both baseline
models and sampling frequencies in Figure 3. It is clear that the error, while larger, is not
signicantly dierent in magnitude. The dierence is that the eective stiness model is always
converging to a stiness value that is either higher (not rocking) or lower (rocking) than the
baseline model value. As a result, it is always trying to converge as the structure switches
states. As a result, the error is larger.
This same result is evident in Figures 4–7 which show the calculated baseline model error,
yout , and lter estimated value, yout-lter , for both models and both storeys, using a sampling
rate of 200Hz. In the ideal case, given instant lter convergence, these plots would be identical.
It is clear that the lters track better and more rapidly with the more accurate bi-linear baseline

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
920 J. G. CHASE ET AL.

Table III. Two-step eective stiness model results for


dierent values of , with m = 10 taps.
f = 200 Hz f = 1000 Hz
 Taps = 10 Taps = 10

1st 1.011 1.050


0.1 2nd 0:986 0.939
1st 1.070 1.128
1 2nd 0.918 0.839
1st 1.118 1.277
10 2nd 0.844 0.872
1st 1.251 1.970
50 2nd 0.883 1.632
1st 1.349 5.595
100 2nd 0.979 5.542
1st 1.444 18.662
150 2nd 1.091 19.539
1st 1.582 853.336
200 2nd 1.243 271.312
1st 7:6 × 104 5:26 × 1032
300
2nd 3:2 × 104 2:34 × 1032

Table IV. Two-step method results using eective stiness


model for dierent numbers of taps, m.
f = 200 Hz f = 1000 Hz
Taps  = 10 =1

1st 1.070 1.050


1 2nd 0.918 0.939
1st 1.116 1.115
5 2nd 0.859 0.864
1st 1.118 1.128
10 2nd 0.844 0.839
1st 1.135 1.129
15 2nd 0.841 0.827
1st 1.156 1.134
20 2nd 0.843 0.822
1st 1.251 1.197
50 2nd 0.883 0.828

model, and as a result the error is lower. However, from a qualitative standpoint the lters
appear to work well for most of the time in both cases. Note that the two-step method was
used in all of these cases.

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 921

Figure 3. Two-step method error plots: (a) bilinear model, 200 Hz; (b) bilinear model,
1000 Hz; (c) eective stiness model, 200 Hz; and (d) eective stiness model, 1000 Hz.

One-step method results


The same series of tests were run using the one-step method. The results are summarized in
Table V, which shows the best combination of parameters for both the bi-linear and eective

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
922 J. G. CHASE ET AL.

Figure 4. Bi-linear stiness model error, yout , and lter estimated values for rst storey.

stiness baseline models. In each case, the results are very similar to the best results in
Tables I–IV. These results show the similar functionality of these two methods for this highly
non-linear structure. The slightly larger error metrics for the one-step method, particularly in

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 923

Figure 5. Bi-linear stiness model error, yout , and lter estimated values for second storey.

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
924 J. G. CHASE ET AL.

Figure 6. Eective stiness model error, yout , and lter estimated values for rst storey.

estimating changes in second storey stiness, are the trade-o for reduced computational eort
and a greater level of gradient approximation by using Equation (19).
Overall, the computational dierence between the one-step and two-step methods does not
appear to be justied by signicantly better SHM performance using the more intense two-step
method. In addition, there is very little dierence in results based on sampling rates of 200 Hz
and 1000 Hz, where the higher sampling rate is not justied by the results presented, although
this dierence was much more pronounced for the more linear Benchmark Problem [15].
However, the baseline model used does signicantly impact SHM performance, which is

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 925

Figure 7. Eective stiness model error, yout , and lter estimated values for second storey.

expected given the easier identication problem using the bi-linear stiness baseline model
that better models the actual behaviour of the non-linear structure studied.
Finally, the changes in the rst and second storey stiness are presented in Figures 8 and 9
for both the bi-linear and eective stiness models at both sampling rates, and using the
two-step method. Figure 8 shows the rst storey results and Figure 9 shows the second storey
results with the top pane in each showing the storey acceleration response to help identify
dierent response regimes. Each plot is labelled with ve dierent response regimes that

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
926 J. G. CHASE ET AL.

Table V. Summary of one-step method results for both baseline models.


e2 =yout 2
Model Frequency  Number of taps 1st storey 2nd storey

Bilinear 200 Hz 300 10 0.163 0.106


1000 Hz 100 10 0.155 0.103
Eective stiness 200 Hz 10 20 1.135 0.834
1000 Hz 10 10 1.180 0.803

are dened:
1 = the building is rocking

2 = the building is moving in a hybrid fashion

3 = the building is rocking

4 = the building is moving in a hybrid fashion

5 = the building is moving in a linear, elastic range.

The ‘hybrid fashion’ regime occurs when the building is out of the elastic range but not fully
rocking. Note that when the building rocks during the strong motion in region ‘1’ there are
sharp acceleration spikes due to the rocking mechanism impacts. In contrast, these spikes are
lower in the hybrid motion regime where the ground motion is not as strong, but the response
is not fully linear either. Finally, note that in region ‘3’ the rocking is smaller in this lower
strength, second strong motion, portion of the El Centro record.
For both cases, the bi-linear stiness model has a negative steady state value at the end
of the simulation, representing a model error that over-estimates the actual nal stiness. In
contrast, the eective stiness under-estimates the nal, linear stiness as expected. In each
labelled response regime the changes in stiness as determined by the lter are modied
accordingly. During rocking the stiness drops and uctuates as the lter converges. During
hybrid motion the stiness increases gradually towards the steady state value, representing a
partially reduced stiness.
The changes in stiness are much larger for the eective stiness baseline model than
the bi-linear stiness baseline model. The worst case changes in stiness as a percentage
for the bilinear model are 6.5% for the rst oor, and 10.0% for the second oor. These
results capture both the bi-linear model periods of non-convergence as well as its potential
over-estimation of the structural stiness in the elastic range [24]. For the eective stiness
model the results are 21.0% for the rst oor, and 40.0% for the second oor, indicating
under-estimation in this model.
In both cases, there is a clear illustration that the baseline models used could be improved.
However, the overall trends in Figures 8 and 9 also clearly show that the SHM methods
presented did capture the proper changes and trends in the time-varying stiness, despite
model error, for each of the response regimes. In particular, it should be remembered that the

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 927

Figure 8. 1st storey stiness variations using the two-step method: (a) 1st oor acceleration record;
(b) bilinear model; and (c) eective stiness model.

pre-cast rocking structure is highly non-linear and represents a form of worst case test for
such a SHM system. As a result, this approach is at least adequate for indicating the general
behaviour and response of the structure to major ground motion, as well as any potential
signicant damage and when it occurred.

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
928 J. G. CHASE ET AL.

Figure 9. 2nd storey stiness variations using the two-step method: (a) 1st oor acceleration record;
(b) bilinear model; and (c) eective stiness model.

CONCLUSIONS

This research has demonstrated the potential for using adaptive LMS-based digital ltering
SHM methods for highly non-linear structures. The particular structure examined is a pre-
stressed concrete rocking structure designed to ameliorate damage to the concrete structural

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
SHM OF A NON-LINEAR ROCKING STRUCTURE 929

elements during large motions. A calibrated nite element model was used to generate ac-
celeration data using noise modelled on the existing laboratory environment and sensors,
representing an ideal experiment in which signicant modelling error and=or systematic sen-
sor errors are not accounted for. The SHM method was tested using two dierent baseline
models, a fully linear eective stiness model and a bi-linear stiness model that assumed a
little more knowledge of the structure’s dynamic behaviour. The approach was developed and
tested for sampling rates of 200 Hz and 1000 Hz and optimal lter parameters determined for
each baseline model and sampling rate. In addition, two dierent SHM ltering approaches,
the one-step and two-step methods, were developed and compared.
The overall results show that this approach is capable of capturing the trends in dynamic
stiness changes for this non-linear structure during a large magnitude seismic event in which
the pre-cast rocking structure changes response regimes rapidly. It is also seen to be capable
of simultaneously identifying model error. There is little dierence found in increasing the
sampling rate for this case or in using the more computationally intense two-step method.
However, there is a signicant reduction in lter error when the bi-linear stiness baseline
model is used. This result indicates that greater knowledge of the structural dynamics enabled
more eective overall damage identication, as might be expected. Overall, while there re-
mains a great deal of work to be done in this area, the real-time capability of this simple
ltering approach to SHM oers great promise.

ACKNOWLEDGEMENTS
This work has been generously supported by the German Scientic Society (Deutsche Forschung-
sgemeinschaft DFG) and the EQC Research Foundation of New Zealand.

REFERENCES
1. Johnson EA, Lam HF, Katafygiotis LS, Beck JL. A benchmark problem for structural health monitoring and
damage detection. Proceedings of the 14th ASCE Engineering Mechanics Conference, Austin, Texas, 21–24
May, 2000.
2. Doebling SW, Farrar CR, Prime MB, Shevitz DW. Damage Identication and Health Monitoring of Structural
and Mechanical Systems from Changes in Their Vibration Characteristics: a Literature Review. Los Alamos
National Laboratory, Report LA-13070-MS, 1996.
3. Doherty JE. Non-destructive evaluation, Chapter 12 in Handbook on Experimental Mechanics, Kobayashi AS
(ed.). Society for Experimental Mechanics, Inc.: Bethel, CT, USA, 1997.
4. Dyke SJ, Caicedo JM, Johnson EA. Monitoring of a benchmark structure for damage identication. Proceedings
of the 14th ASCE Engineering Mechanics Conference, Austin, Texas, 21–24 May 2000.
5. Caicedo JM, Dyke SJ, Johnson EA. Health monitoring based on component transfer functions. Proceedings of
the 2000 International Conference on Advances in Structural Dynamics, Hong Kong, 13 –15 December 2000.
6. Lus H, Betti R. Damage identication in linear structural systems. Proceedings of the 14th ASCE Engineering
Mechanics Conference, Austin, Texas, 21–24 May, 2000.
7. Bernal D, Gunes B. Observer=Kalman and subspace identication of the UBC benchmark structural model.
Proceedings of the 14th ASCE Engineering Mechanics Conference, Austin, Texas, 21–24 May 2000.
8. Corbin M, Hera A, Hou Z. Locating damage regions using wavelet approach. Proceedings of the 14th ASCE
Engineering Mechanics Conference, Austin, Texas, 21–24 May 2000.
9. Hou Z, Noori M, Amand R. Wavelet-based approach for structural damage detection. Journal of Engineering
Mechanics (ASCE) 2000; 126(7):677– 683.
10. Verboven P, Parloo E, Guillaume P, Van Overmeire M. Autonomous structural health monitoring—Part I: modal
parameter estimation and tracking. Mechanical Systems and Signal Processing 2002; 16(4):637– 657.
11. Verboven P, Parloo E, Guillaume P, Van Overmeire M. Autonomous structural health monitoring—Part II:
vibration-based in-operation damage assessment. Mechanical Systems and Signal Processing 2002; 16(4):
659 – 675.

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930
930 J. G. CHASE ET AL.

12. Sato T, Qi K. Adaptive H∞ lter: its application to structural identication. Journal of Engineering Mechanics
(ASCE) 1998; 124(11):1233 –1240.
13. Loh C-H, Lin C-Y, Huang C-C. Time domain identication of frames under earthquake loadings. Journal of
Engineering Mechanics (ASCE) 2000; 126(7):693 –703.
14. Ifeachor EC, Jervis BW. Digital Signal Processing: A Practical Approach. Addison-Wesley: Boston, MA,
1993.
15. Chase JG, Hwang KL. Adaptive LMS methods for real-time structural health monitoring of steel framed
structures. Proceedings of the 13th World Conference on Earthquake Engineering (13WCEE), Vancouver,
Canada, 1 August 2004.
16. Priestley MJN, MacRae GA. Seismic tests of pre-cast beam-to-column joint sub-assemblages with un-bonded
tendons. PCI Journal 1996; 41(1):64 –80.
17. Cheok GS, Lew HS. Model precast concrete beam-to-column connections subject to cyclic loading. PCI Journal
1993; 38(4):80–92.
18. Stone WC, Cheok GS, Stanton JF. Performance of hybrid moment-resisting pre-cast beam-column concrete
connections subjected to cyclic loading. ACI Journal 1995; 91(2):229.
19. Nakaki SD, Stanton JF, Sritharan S. An overview of the PRESSS ve-storey pre-cast test building. PCI Journal
1999; 44(2):26–39.
20. Priestley MJN, Sritharan S, Conley JR, Pampanin S. Preliminary results and conclusions from the PRESSS
ve-storey pre-cast concrete test building. PCI Journal 1999; 44(6):43 – 67.
21. Mander JB, Cheng C-T. Seismic resistance of bridges based on damage avoidance design. NCEER Technical
Report 97-0014, 1997.
22. Murahidy AG, Carr AJ, Spieth HA, Mander JB, Bull DK. Design, construction and dynamic testing of post-
tensioned precast reinforced concrete frame building with rocking beam column connections and ADAS elements.
Proceedings of the NZSEE Conference, Rotorua, New Zealand, 19 –21 March 2004.
23. Carr AJ. RUAUMOKO Program for Inelastic Dynamic Analysis: Users Manual. Department of Civil
Engineering, University of Canterbury, Christchurch, New Zealand, 2004.
24. Spieth HA, Carr AJ, Murahidy AG, Arnolds D, Davies M, Mander JB. Modelling of post-tensioned pre-cast
reinforced concrete frame structures with rocking column connections. Proceedings of the NZSEE Conference,
Rotorua, New Zealand, 19 –21 March 2004.
25. Spieth HA, Arnold D, Davies M, Mander JB, Carr AJ. Seismic performance of post-tensioned pre-cast concrete
beam to column connections with supplementary energy dissipation. Proceedings of the NZSEE Conference,
Rotorua, New Zealand, 19 –21 March 2004.
26. Calvi GM, Pampanin S, Fajfar P, Dolsek M. New methods for assessment and design of structures in seismic
zones: Present state and research needs. Proceedings of the International Workshop Mitigation of Seismic
Risk—Support to Recently Aected European Countries, Belgirate (VB), Italy, 2000.
27. Priestley MJN. Myths and Fallacies in Earthquakes Engineering, Revisited. IUSS Press: Pavia, Italy, Ch. 5,
pp 93 –101, 2003.
28. Pampanin S, Priestley MJ, Sritharan S. Analytical modelling of the seismic behaviour of pre-cast concrete frames
designed with ductile connection. Journal of Earthquake Engineering 2001; 5(3):329 –367.

Copyright ? 2005 John Wiley & Sons, Ltd. Earthquake Engng Struct. Dyn. 2005; 34:909–930

Vous aimerez peut-être aussi