Вы находитесь на странице: 1из 8

Key Engineering Materials Vols. 245-246 (2003) pp.

141-148
online at http://www.scientific.net
Journal
© 2003 Citation
Trans Tech(to Publications,
be inserted bySwitzerland
the publisher)
Copyright by Trans Tech Publications

Structural Health Monitoring and Damage Assessment Using Measured


FRFs from Multiple Sensors, Part II: Decision Making with RBF Network

C. Zang1, M.I. Friswell1 and M. Imregun2


1
Department of Areospace Engineering, University of Bristol
Queen’s Building, Bristol BS8 1TR, UK,
Email: c.zang@bristol.ac.uk m.i.friswell@bristol.ac.uk
2
Department of Mechanical Engineering, Imperial College,
Exhibition Road, London SW7 2BX, UK, Email: m.imregun@imperial.ac.uk

Keywords: structural health monitoring, damage assessment, RBF neural network, GSC, GAC

Abstract. This paper is the second part of structural health monitoring and damage assessment
using measured FRFs from multiple sensors, discussing the decision making technique with radial
basis function (RBF) neural networks. In PART 1 of the paper, the correlation criteria showed their
capability to indicate various changes to the structure’s state. PART 2, presented here, develops the
methodology of decision theory to identify precisely all of the structure states. Although, the
statistical approach can be used for classification, interpreting the information is difficult. Neural
network techniques have been proven to possess many advantages for classification due to their
learning ability and good generalization. In this paper, the radial basis function neural network is
applied for function approximation and recognition. The key idea is to partition the input space (the
indicators of the correlation criteria) into a number of subspaces that are in the form of hyper
spheres. Then, the widely used k-mean clustering algorithm was selected as a logical approach to
detecting the structure states. A bookshelf structure with measured frequency responses from 24
accelerometers was used to demonstrate the effectiveness of the method. The results show the
successful classification of all structure states, for instance, the undamaged and damage states,
damage locations and damage levels, and the environmental variability.

Introduction
In PART 1 of the paper, it has been shown that the global shape and amplitude correlation criteria
(GSC and GAC) can be used to quantify the closeness between two sets of measured vibration data
and the averaged integration of the GSC and GAC (AIGSC, AIGAC) can be used as damage
indicators to recognize various structure states [1]. This paper develops an approach that can be
used to successfully identify and assess structural states and damage (existence, location, and
extent) using the proposed damage indicators.
Generally speaking, the methodology frequently used for structural pattern recognition is based
on the statistical approach, whereby decision theory derived from statistics of input patterns is used
to design a classifier [2]. However, this approach requires much heuristic information and has
difficulty in expressing structural information unless an appropriate selection of features is possible.
Recently, neural network methods have been proven to possess many advantages for classification
due to their learning ability and good generalization [3]. Reviews of applications in structure health
monitoring and damage detection, based on neural network techniques, were given by Doebling et
al. (1996 [4]) and Zeng (1998 [5]). The most widely used neural networks are multiplayer networks,
usually coupled with the back-propagation (BP) algorithm. For instance, Zang and Imregun (2001
[6,7]) used back-propagation neural networks to successfully detect the damage in the case of two
representative structures: a railway wheel and space antenna. However, the training computation is

Licensed to University of Bristol - Bristol - UK


All rights reserved. No part of the contents of this paper may be reproduced or transmitted in any form or by any means without the
written permission of the publisher: Trans Tech Publications Ltd, Switzerland, www.ttp.net. (ID: 137.222.10.57-06/01/06,10:32:07)
142 Damage Assessment of Structures V
2 Title of Publication (to be inserted by the publisher)

expensive due to the slow convergence speed and there is no guarantee that absolute minima can be
achieved. On the other hand, radial basis function (RBF) neural networks have attracted extensive
interest because of certain advantages over other types of ANNs, such as better approximation
capabilities, simple network structures and fast learning algorithms. Such features motivate us to
apply this method to structural damage detection and assessment.

Radial Basis Function (RBF) Neural Networks


Radial basis function (RBF) neural networks were introduced into the neural network literature by
Broomhead and Lowe (1988 [8]). They implement an input-output mapping using a linear
combination of radially symmetric functions. The basic structure of the radial basis function
network is shown in Fig. 1.

Fig. 1 Radial basis function (RBF) neural network structure

It has three layers, where there are feed-forward connections between the nodes. In the first
layer, each input neuron corresponds to a component of an input vector x and weightless
connections pass inputs to the hidden layer (the second layer). The hidden nodes, each of which
computes a kernel function (also called activation function), are the radial basis function units. The
most used radial functions are Gaussian functions. They are characterized by two parameters: the
centroids represented by cj and the width represented by s j . The output from the jth Gaussian
neuron from an input object xi can be calculated by the following equation:

2
- xi - c j
Out j = exp (1)
s 2j

where xi - c j is the calculated Euclidean distance and s j determines the portion of the input space
where the jth RBF will have a non-significant zero response.
Key Engineering Materials Vols. 245-246 143

The output layer consists of k neurons which correspond to the possible classes of the problem.
Each output neuron is fully connected to the hidden layer and computes a linear weighted sum of
the outputs of the hidden neurons:

yk = wkjOut j (2)
j

RBF neural networks are trained in three stages (Moody and Darken 1989 [9]). The first stage is
to optimize the radial hidden layer centers cj by the k-means clustering algorithm [10], an
unsupervised technique that places unit centers centrally among clusters of training points. Once
the unit centers are established, the next is to determine the receptive ‘width’ of each unit using a
nearest neighbor heuristic. Finally, the weights of connections between the hidden and the output
layer are optimized by linear regression.

Case Study: Bookshelf Structure

Testing configuration
The structure of PART 1, a three-story bookshelf structure model [11], will be used to demonstrate
the method. Configuration of all test cases, including the healthy and various damage states is listed
in Table 1. Environmental variability was simulated by varying the shaker input level to 2, 5, and 8
volts respectively. In total, 6480 FRFs with 24x270 measurements were obtained for correlation
analysis. Damage indicators (AIGSC and AIGAC), each calculated for 24 FRFs, are plotted in Fig.
2.

Fig. 2 Damage indicators (AIGSC, AIGAC) over 270 measurements of 24 FRFs

The following features can be noted from Fig. 2:


· Damage indicator AIGSC is sensitive to the bookshelf structure state. Various damage
states result in the significant reduction of the AIGSC values while the same states of the
144 Damage Assessment of Structures V
4 Title of Publication (to be inserted by the publisher)

bookshelf structure keep values of the AIGSC more or less unchanged. It seems that a
threshold may be found to distinguish the healthy states from the damaged states. But,
further classification of damage levels and locations is unlikely to be straightforward. On
the other hand, the AIGSC is insensitive to environmental variability. Changes in the
excitation level for the same structure state hardly affect the shape correlations.
· Unlike the AIGSC, the damage indicator AIGAC is sensitive to both the environmental
variability and the structure states. Any change of either the structure state or the shaker
excitation level leads to enormous variations of the AIGAC values. There is difficulty in
building a threshold directly to classify the structure states.
A comparison of the damage indicators for various healthy states in Fig. 2, shows that a relative
gap exists between the original healthy state (HS1) and the other healthy states (HS2, HS3, HS4,
and HS5), especially for the AIGSC indicator. Furthermore, values of the AIGSC go down
gradually with continuing measurements in the initial state. This is probably because the dynamic
system of the initially built bookshelf structure was not stable. Structure states from HS2 to HS5 are
stable and consistent. FRFs for all of these healthy states are overlaid in Fig. 3. Clearly, FRFs of
HS2 to HS5 are in good agreement.

Fig. 3 Comparison of healthy-state FRFs

In summary, damage indicators (AIGSC and AIGAC) provide the information of the various
states of the bookshelf structure. Using both of them simultaneously provides the possibility to
devise an intelligent approach to classifying the healthy and damaged states of the structure and
estimating the damage locations and extents at the same time.

RBF neural networks for classification


The RBF neural network for structure damage detection and assessment was built with 2 input
nodes, 44 hidden nodes and 6 output nodes. The output definition is displayed in Table 2.
Key Engineering Materials Vols. 245-246 145

Table 2. Output definition of a RBF neural network


RBFNN
Value Definition
Output
1 The structure is in healthy state
Node 1
0 The structure is not in healthy state
1 The structure damage occurs at 1C location
Node 2
0 The structure damage does not occur at 1C location
1 The structure damage occurs at 3A location
Node 3
0 The structure damage does not occur at 3A location
1 The damage extent occurs at level 1 (The bolts were removed)
Node 4
0 The damage extent does not occur at level 1
1 The damage extent occurs at level 2 (The brackets were removed)
Node 5
0 The damage extent does not occur at level 2
1 The damage extent occurs at level 3 (The torque was applied to the bolts)
Node 6
0 The damage extent does not occur at level 3

Damage indicators (AIGSC, AIGAC) were employed as two input nodes of the RBF network. A
total of 180 damage indicators were randomly selected as the training set and another 90 as the
testing set. The output of the network, including both the training and testing sets, is plotted in Fig.
4. Clearly, all of the structure states such as the healthy or damage state, damage location, and
damage extent can be classified accurately although there are some fluctuations in the outputs. For
instance, values of the outputs at cross section 8 were 0.0015, 0.99, 0.92, 1.028, -0.005, 0.0625
respectively. Multiple damage existed in the structure at both 1C and 3A locations and the damage
extent was at Level 1, that is, the bolts were removed between the bracket and the plate.

Fig. 4 Output of RBF neural network with the training and the testing sets
146 Damage Assessment of Structures V
6 Title of Publication (to be inserted by the publisher)

To investigate the ability to detect damage with imposed variability in the environmental
conditions, the excitation levels were considered in the output of the RBF network. Another RBF
network was defined with nine output nodes. The output of the first 6 nodes had the same
definitions as the previous network. The output of nodes 7, 8, and 9 were used to describe the three
excitation levels as 2, 5, 8 volts respectively. The same number of training and testing sets were
randomly selected and the output results of the network are plotted in Fig. 5. To observe the result
more clearly, values of output neurons were set to zero if the practical outputs were less than 0.5. It
can be seen from Fig. 5 that most of 270 measurements can be successfully trained and tested except
the failure of 3 damage indicator samples. The first exception is the output of node 5 for which there
is a failure to identify the damage location 3A. The other two exceptions are the outputs of node 8,
and 9. The environmental conditions of EV2 and EV3 were mistaken for EV3 and EV2
respectively.

Fig. 5 Output of RBF neural network with the training and the testing sets

Concluding Remarks
Damage indicators, the averaged integration of the GSC and GAC (AIGSC, AIGAC), have the
capability to recognize various structure states and environmental variability. Using both of them
Key Engineering Materials Vols. 245-246 147

simultaneously provides the possibility to identify various structure states and assess the damaged
locations and extents.
Radial basis function (RBF) neural networks are effectively employed for devising health
monitoring and damage assessment rules using damage indicators. Results from the demonstration
of various damage states of a bookshelf structure, including damage location and level, show the
feasibility of this approach. It is also capable to identify the change of environmental conditions.
Both shape and amplitude functions have the capability to cope with incomplete FRF data
obtained from a few sensors. Such a feature will make the method well suited to on-line industrial
applications.

REFERENCES
[1] C. Zang, M. I. Friswell, and M. Imregun: Structural Health Monitoring and Damage Assessment
Using Measured FRFs from Multiple Sensors, Part I: The Indicator of Correlation Criteria,
The 5th International Conference on Damage Assessment of Structures, University of
Southampton, UK (2003).
[2] K. Fukunaga: Introduction to Statistical Pattern Recognition, 2nd edition, San Diego, CA:
Academic Press, (1990).
[3] C. M. Bishop: Neural Networks for Pattern Recognition, New York: Oxford University Press
(1995).
[4] S. W. Doebling, C. R. Farrar, M. B. Prime, & D. W. Shevitz: Damage identification and health
monitoring of structural and mechanical systems from changes in their vibration
characteristics: a literature review, Report No. LA-13070-MS, Los Alamos National
Laboratory, (1996).
[5] P. Zeng: Neural computing in mechanics, Applied Mechanics Reviews, Vol. 51, No. 2, (1998)
pp. 173-197.
[6] C. Zang, and M. Imregun: Structural damage detection using artificial neural networks and
measured FRF data reduced via principal component projection, Journal of Sound and
Vibration, Vol. 242, No. 5, (2001), p. 813-827.
[7] C. Zang, and M. Imregun: Combined neural network and reduced FRF techniques for slight
damage detection, Journal of the Archive of Applied Mechanics, Vol. 71, No. 8, (2001), p. 525-
536.
[8] D. Broomhead, and D. Lowe: Multivariable functional interpolation and adaptive neural
networks. Complex System, Vol. 2, (1988) p. 321-355
[9] T. J. Moody and C. J. Darken: Fast learning in networks of locally tuned process units. Neural
Computation, Vol. 1 (1989) p. 151-160
[10] C. J. Darken and T. J. Moody: Fast adaptive k-mean clustering: some empirical results.
Proceedings of International Joint Conference on Neural Networks, IEEE press, (1990) p. 238.
[11] Los Alamos National Laboratory Website: http:// www.lanl.gov
148 Damage Assessment of Structures V
8 Title of Publication (to be inserted by the publisher)

Table 1: The testing configuration of the healthy and the damaged states

Damage Damage Exciation Measured FRFs


State Definition
location extent level (V) times Samples
Original 2 10 240
The original state of the
Healthy 5 10 240
structure
(HS1) 8 10 240
The bolts were removed 2 5 120
between the bracket and the 1C DL1 5 5 120
Damaged
plate at 1C location 8 5 120
State
The bracket were 2 5 120
(DS1)
completed removed at 1C 1C DL2 5 5 120
location 8 5 120
Recoverd The recovered healthy state 2 10 240
Healthy of the structure from 5 10 240
(HS2) damaged state (DS1) 8 10 240
The bolts were removed 2 5 120
between the bracket and the 3A DL1 5 5 120
Damaged
plate at 3A location 8 5 120
State
The bracket were 2 5 120
(DS2)
completed removed at 1C 3A DL2 5 5 120
location 8 5 120
Recoverd The recovered healthy state 2 10 120
Healthy of the structure from 5 10 120
(HS3) damaged state (DS2) 8 10 120
The bolts were removed 2 5 120
between the bracket and the 1C&3A DL1 5 5 120
Damaged
plate at 1C &3A location 8 5 120
State
The bracket were 2 5 120
(DS3)
completed removed at 1C 1C&3A DL2 5 5 120
and 3A location 8 5 120
Recoverd The recovered healthy state 2 10 120
Healthy of the structure from 5 10 120
(HS4) damaged state (DS3) 8 10 120
A torque value of 5ft.lbs was
8 10 240
applied to the bolts
Damaged
A torque value of 5ft.lbs was
State 1C DL3 8 10 240
applied to the bolts
(DS4)
A hand tight torque was
8 10 240
applied to the bolts
Recoverd The recovered healthy state 2 10 240
Healthy of the structure from 5 10 240
(HS5) damaged state (DS4) 8 10 240
In total: 270 6480

Вам также может понравиться