ASIAN JOURNAL OF CIVIL ENGINEERING (BUILDING AND HOUSING) VOL. 7, NO.

2 (2006) PAGES 155-168

TIME HISTORY ANALYSIS OF STRUCTURES FOR EARTHQUAKE LOADING BY WAVELET NETWORKS

A. Heidaria and E. Salajeghehb University of Shahrekord, Shahrekord, Iran b Bahonar University of Kerman, Kerman, Iran
a

ABSTRACT
Fast Wavelet Transforms (FWT) and Discrete Wavelet Neural Networks (DWN) are used for dynamic analysis of structures. A filter bank is used for the FWT and the number of points of the earthquake record is reduced. The DWN is used for approximating the dynamic responses of the structures. By such approximation, the dynamic analysis of the structure is not necessary if the structural properties are changed. Both feedforward neural networks and wavelet decompositions inspire this network. An algorithm of backpropagation type is proposed for training the network. In this network, the input is the damping ratio and the angular natural frequency of the structure and the output is the responses of the structures against these reduced points. After training the network, using the inverse wavelet transform, the results of the dynamic analysis are obtained for the original earthquake record from the output of the network. Some numerical examples are solved by the proposed method and the results are compared with those of the original record.

Keywords: time history analysis, signal processing, filter banks, fast wavelet transforms, wavelet neural network 1. INTRODUCTION
Time history dynamic analyses of structures for problems with large number of degrees of freedom, is time consuming. In order to overcome this difficulty, Wavelet Transforms (WT) are used. The WT are well known as useful tools for various signal-processing applications. In WT the use of a fully scalable window solves the signal-cutting problem. There are three kinds of WT, namely, continuous, discrete and fast wavelet transform [1-3]. During the last 10 years, the wavelet theory has been developed in mathematics and engineering. Wavelets are mathematical functions that cut up data into different frequency components, and then study each component with a resolution matched to its scale. They have advantages over traditional Fourier Transforms (FT) in analysing physical states where the signal contains discontinuities and sharp spikes. Unlike the FT, the wavelet transform has dual localization,

E-mail address of the corresponding author: [email protected]

156

A. Heidari and E. Salajegheh

both in frequency and in time [4]. These characteristics make wavelets an active subject with many exciting applications. The FT has been widely used in dynamic analysis of structures. It is a powerful tool to study the frequency content of signals but it has the drawback that it does not provide any localization in time. Recently, WT has been used for dynamic analysis of structures against earthquake record [5-8]. The results show that the WT is a powerful method for dynamic analysis and the time of analysis is reduced substantially. Discrete wavelet transform (DWT) is used for optimisation of structures against earthquake record [9]. Also the FWT combined with artificial neural networks are used for optimisation of structures against earthquakes [10-13]. The results show that the DWT, FWT and ANN are effective methods for optimisation of structures against seismic loads. In this paper, FWT and DWN are used to approximate the dynamic responses of the structures. By the FWT the main earthquake record is transformed into a record with very small number of points. Thus the time history dynamic analysis is carried out with fewer points. The decomposition starts from the original record, and produces two sets of new records and the process is repeated. The error is negligible, in particular in the first three stages of decomposition. After reducing the number of points of the earthquake record, the DWN is used to approximate the dynamic responses of the structure. For training the DWT, an algorithm of backpropagation type with mother wavelet used as the activation function is proposed. In this network, the input is the damping ratio and the angular natural frequency of the structure, and the output is the dynamic responses of a single degree of freedom (SDOF) structure against these reduced points. After training the network, using inverse wavelet transform (IWT), the results of the dynamic analysis are obtained for the original earthquake record from the output of the network. In the paper, the details of the approximation concepts, and some numerical examples of the dynamic analysis of structures will be presented. The details of the FWT, DWN and IWT will also be outlined. The numerical results of the dynamic analysis show that this approximation is a powerful technique and the required computational work can be greatly reduced.

2. FWT FOR DECOMPOSITION OF EARTHQUAKE RECORD

The wavelet transform can be simply achieved by a tree of digital filter banks. The main idea behind the filter banks is to divide a signal into two parts; the first is the low frequency part and the other is the high frequency part. This idea can be achieved by a set of filters. By applying a low pass filter to a signal s[t] of length N (number of points), the high frequency bands of the signal are removed and an approximate version of the original signal is obtained. A high pass filter removes the low frequency components of the original signal, and the result is a signal containing the details of the main signal [14]. A multi-level decomposition of the original signal is obtained by repeating the decomposition process. In the next stage, the low pass filtered output signal is used as input to the filter bank. If the computation of a wavelet transform is reduced to a FWT, then the resulting implementation is very efficient. Several FWT algorithms have been devised for

TIME HISTORY ANALYSIS OF STRUCTURES FOR EARTHQUAKE LOADING...

157

computation of wavelet transform coefficients. In this paper the Mallat algorithm [15] is used for dynamic analysis of the structures against earthquakes [6,7]. The original record is decomposed into two records. Detail record Dj and approximate record Aj for j=1,, J and k=1,,K, which are obtained as follows:
j D j = s[ t ]h j [t 2 k] t

(1) (2)

j A j = s[t]g j [t 2 k] t

where j and k are the integer numbers. The symbol * denotes complex conjugate, the value of 2 j is the dilation factor; the value of k is shifting factor, and [.] is used for discrete notation. The functions h j and g j are defined as [15]:
h j [t 2 j k] = 2 j/2 (2 j (t 2 j k))

(3) (4)

g j [t 2 j k] = 2 j/2 (2 j (t 2 j k))

The functions and are chosen as [14]

1 ( t ) = 1 1 ( t ) = 0.5 < t 1 0 0 t 0 .5 0 t 1 otherwise

(5)

The wavelets and the scaling functions must be deduced from one stage to the next as follows:
g 1[ t ] = g[ t ] h 1[t] = h[t]

(6a) (6b) (6c) (6d)

g j+1[t] = k g j [k]g[t 2k]

h j+1[t] = k h j [k]g[t 2k]

The number of points of the earthquake record is reduced by the FWT. The decomposition starts from the original record of the earthquake, and produces two sets of records, Dj, and Aj. Then from record A1, the two records A2 and D2 are evaluated and the process is continued until AJ and DJ are calculated for the Jth stage. For the earthquake record the approximation record (Aj) with low frequency components is the effective part. Therefore Aj is used for the dynamic analysis. In this paper, this process

158

A. Heidari and E. Salajegheh

is repeated in three stages, and the number of points of the original record is reduced to 0.125 of the primary points. The error is negligible, in particular in the first three stages of decomposition [16]. In this method the process can be inversed and the original record can be computed. This process is named as inverse wavelet transform (IWT). The original signal, can be achieved through the IWT process, by using Dj and Aj as:
J J ~ s[ t ] = D j h j [ t 2 j k ] + j=1 k j=1

g j[t 2 j k] A j~
k

(7)

~ where h j[ t 2 j k ] is called the synthesis wavelets and ~ g j[ t 2 j k ] is called synthesis scaling

functions, defined as [15]:

~ ~ (2 j ( t 2 j k )) h j[ t 2 j k ] = 2 j / 2
~ ~ g j[ t 2 j k ] = 2 j / 2 (2 j ( t 2 j k ))

(8) (9)

~ ~ are used for the high and low pass filters, respectively. The functions in which h and g ~ ~ and are mother wavelet and scaling function, respectively, that are used for reconstruction. These two filters are used for reconstruction of the IWT. In the present study, the Aj record is only used in the process of the IWT as the effect of Dj is negligible.

3. DYNAMIC RESPONSES OF STRUCTURE USING DWN

The DWN is used to approximate the dynamic responses of the structures. The DWN is an artificial neural network and consists of three layers. Artificial neural network (ANN) is a mathematical system that mimics the way in which the brain works. It consists of fully interconnected layers of processing units called neurons. There are always one input layer, one output layer and a number of hidden layers. Each layer of nodes receives its input from previous layer or from the network input. The output of each node feeds the next layer or the output of the network. Internal or hidden layers provide the interconnection between input and output layers. The output of neuron is related to the summed inputs by a linear or a nonlinear transfer function. Different ANN classes use different definition of activation functions and different training algorithms. The most common training network algorithm is the backpropagation algorithm, and a sigmoid function is a popular activation function. The ANN is capable of representing functional relations. To create such a representation, it suffices to train a neural network with a set of known input-output pairs. During the training, interconnection weights between neurons are set to produce the correct input-output relation. After this initial stage, upon presentation of a new input vector, an approximation to an output vector is returned.

TIME HISTORY ANALYSIS OF STRUCTURES FOR EARTHQUAKE LOADING...

159

Two neural networks that can be used in function approximation are backpropagation (BP) neural network and counterpropagation neural network. These two networks are not capable to approximating the output of the dynamic analysis. In this paper, to overcome this difficulty, a novel ANN is used. The DWN is inspired from the BP. The activation function in the DWN is mother wavelet. Based on discrete wavelet transform, the DWN has been proposed as a novel universal tool for functional approximation, which shows surprising effectiveness in solving the conventional problems of poor convergence or even divergence encountered in other kinds of neural networks. It can dramatically increase convergence speed. In spite of having great potential applicability of the DWN, there are only a few papers on the DWN theory. The topological structure of the DWN employed in this study is shown in Figure 1.

Figure 1. Wavelet neural network

The DWN consists of three layers: input layer, wavelet layer and output layer. The activation functions in wavelet layer are derived from a mother wavelet with varying translation and scale values. Each layer has one or more nodes. The input data vector Y is connected to the input nodes of the networks. The connections between input units and wavelet units, and between wavelet units and output units are called weights wti and Wt respectively. In the proposed DWN, the neural network simulator is described as follows: 1) Initialising the dilation parameter a, translation parameter b and node connection weights wi and W to some random values. All those random values are limited in the interval (0,1). 2) Inputting the data Yn(i) and the corresponding output values VnT , where i varies from 1 to U, representing the number of the input nodes, n indicates the nth data sample of the

160

A. Heidari and E. Salajegheh

3)

training set, and T denotes the target output state. The output value of the sample Vn is calculated with the following formula:
Vn =
=1 T

w i y n (i) b ) W ( =1
U

(10)

in which is taken as a Morlet mother wavelet

( ) = cos(1.75) exp(0.5 2 )

(11)

4)

To reduce the error, W, wi, a, b are adjusted using W , wi , a , b as:

W ( j + 1) = E + W ( j) W ( j)
E + w i ( j) w i ( j)

(12)

w ( j + 1) =

(13)

a ( j + 1) =

E + a ( j) a ( j)
E + b ( j) b ( j)

(14)

b ( j + 1) =

(15)

The error function E is taken as:

E= 1 N T (Vn Vn ) 2 n =1

(16)

where N stands for the data number of training set, and are the learning rate and the momentum term, respectively. In this paper, steepest descent method is used to minimize E. (5) Returning to step (3), the process is continued until the error function (E) satisfies the given error criteria, and the whole training of the DWN is completed. The method of evaluation of the errors in the DWN is similar to the BP. The results in DWN is better than the standard BP because the activation function in the standard BP is a simple function and this is not adequate for approximating the functions such as the earthquake records. The results of the training process are the approximate dynamic displacements of a structure with single degree of freedom at all the time intervals as explained further in the subsequent sections.

TIME HISTORY ANALYSIS OF STRUCTURES FOR EARTHQUAKE LOADING...

161

4. SIMULATION OF DYNAMIC RESPONSES BY DWN

One of the methods used for the evaluation of the dynamic responses of a linear analysis of the structures is the modal superposition. In modal superposition method it is shown that any multi degrees of freedom system (MDOF) can be converted into multi systems of single degree of freedom (SDOF). The solution of each SDOF system may be found by Duhamels integral. After calculating the response of each SDOF system, the response of MDOF system can be found by superposition method. The Duhamels integral is shown as [17]:
y a (t) = 1 D

t i +1
i

& &(t)e (t ) sin D (t )d y

(17)

where ya(t) is dynamic responses of SDOF, is the damping ratio, is the angular natural

dynamic analysis of the structure. The target vector (V nT ) in Eq. 16 is the responses of the SDOF system (ya(t)) for the A3 record points. In this method the responses of the SDOF are computed at all the time intervals of the earthquake record. The number of nodes of input layer, wavelet layer and output layer are chosen as 2, 12 and 336, respectively. If we use a=12, the value of 2a is equal to 4096. This number is greater than the number of points of the earthquake record. According to wavelet theory choosing at least 12 values can show the variant of the record suitability [15]. Total time intervals of the El Centro earthquake are 2688 and the number of reduced intervals by the FWT (3rd stage of decomposition) is evaluated as 336. It is to be mentioned that by the proposed method, a kind of response spectrum is created for time history dynamic analysis of a SDOF systems. In fact Eq. 10 provides the displacements for all possible SDOF structures at all the time intervals of the specified earthquake for a given and . Careful consideration of Eq. 10 together with the network given in Figure 1 indicate that by multiplication of some matrices, the displacements of a SDOF system is achieved at all the time intervals of the ground motion under consideration.

& is the acceleration record and D is equal to ( 1 2 )0.5 . In the Duhamels frequency, & y & , ti and ti+1 are constants. Therefore two y integral for a specific earthquake record, & independent variables exist and the input variables of the DWN are as and . In this & is chosen as the 3rd stage of decomposition of & & (i.e. A3), which is used for paper, & y y

5. MAIN STEPS OF DYNAMIC ANALYSIS WITH FWT AND DWN

The main steps in the dynamic analysis employing FWT and DWN against earthquake are as follows: (a) The functions and are defined as stated in Eq. 5. (b) The number of stages for decomposition of the record is chosen. In this paper three stages are used. (c) The FWT of the earthquake record is computed in three stages.

162

A. Heidari and E. Salajegheh

(d) The approximate version of the earthquake record in the 3rd stage (A3) is used for dynamic analysis. (e) The dynamic responses of a SDOF structure against A3 are calculated. This process is repeated for a number of SDOF structures with different and , for training the network. (f) The number of nodes in the input layer, wavelet layer, and output layer are chosen. In this paper 2, 12 and 336 nodes are used, respectively. (g) The activation function in wavelet layer is chosen by Eq. 11. In this paper, 12 activation functions are used. (h) The weights of the wavelet neural network are optimised. This process is repeated until the network is converged. (i) For each and , the trained network is used to calculate the dynamic responses of any SDOF structure against A3 record. (j) Equation 7 evaluates the responses of the SDOF structure against the original earthquake record. Now, the dynamic responses of the structure under investigation are determined by modal superposition method. It can be observed that during time history analysis, dynamic responses of a SDOF structure are not needed. In fact, the necessary responses are found by the trained DWN that is faster than the Duhamels integral. In addition, the time history analysis of the structure that was required for training is achieved for a record with less number of time intervals. For the El Centro earthquake, the best ratio of the reduced intervals with respect to the original record is chosen as 0.125. It is to be mentioned that the training process should be carried out once for the specified earthquake. It is to be noted that the procedure outlined in the present study is for structural responses under earthquake loads when the dynamic analysis is carried out by the modal superposition method. However, the investigation is under progress for other methods of dynamic analysis as the training process is different.

6. NUMERICAL EXAMPLES
Two examples are analysed against the El Centro earthquake record (S-E 1940). A personal Pentium 4 computer is used to evaluate the results. In each case, the execution time is measured. The number of points of the El Centro earthquake is 2688. The time interval for the record is 0.02 seconds. The number of points in A3 is 336. The analysis is carried out by the following methods: (a) Dynamic analysis with the original earthquake record (DAO) (b) Dynamic analysis with the reduced record (A3) (DAR) (c) Dynamic analysis using the FWT and the DWN (FDN). The training time for the El Centro earthquake for a SDOF system is 41 minutes that is carried out once for the earthquake record. The trained system is used for the time history dynamic analysis of any MDOF systems without repeating the training procedure.

TIME HISTORY ANALYSIS OF STRUCTURES FOR EARTHQUAKE LOADING...

163

6.1 Example 1: Plane Shear Building The plane shear building model of 7 storeys shown in Figure 2, is analysed, whose floor masses move only horizontally within a vertical plane. It is assumed that the mass of each rigid floor of the model includes the effect of masses of all the structural elements adjacent to the floor of the prototype building. The mass of each floor is 90 tons. The material properties are given as Youngs modulus 2104 kg/cm2, weight density 0.0078 kg/cm3 and damping ratio for all modes as 0.02. The moments of inertia for all columns are 2104 cm4. Maximum displacements of floors are given in Table 1. The displacement history of the top story, for DAO, DAR and FDN are shown in Figuers 3 to 5. Not only the maximum displacements of each floor are almost the same, but also displacement histories of the all cases are similar. The time of analysis for DAO, DAR and FDN are 3.16, 0.44 and 0.32 sec., respectively. As the number of the degrees of freedom of the structure is small, the difference between the DAR and FDN is not significant.

Figure 2. Shear building of 7 stories

6.2 Example 2: 72-Bar Space Truss A 72-bar space truss with 48 translation degrees of freedom is shown in Figure 6. The material properties are given as Youngs modulus 2.1106 kg/cm2, weight density, 0.008 kg/cm3, and damping ratio for all modes as 0.02. The mass of 2500 kg is lumped at each free node. The cross sections of all members are 20 cm2. The earthquake record is applied in X direction. Maximum displacement of joints 1, 3, 6, 8, 9, 12, 14 and 16 are given in Tables 2 to 4, respectively. The time of analysis for the DAO, DAR and FDN are 13.11, 2.01 and 1.54 sec., respectively.

164

A. Heidari and E. Salajegheh

Table 1. Maximum displacement of Example 1

Floor No. Maximum dynamic displacement (cm) DAO DAR FDN

1 2 3 4 5 6 7

1.952 3.758 5.459 7.027 8.353 9.340 9.876

1.943 3.724 5.441 6.987 8.291 9.301 9.823

1.941 3.729 5.448 6.988 8.322 9.313 9.858

10

20
Time

30

40

50

3.7

Figure 3. Displacement history of level 7 by DAO (cm)

7.5

15

22.5
Time

30

37.5

45

53.7

Figure 4. Displacement history of level 7 by DAR (cm)

TIME HISTORY ANALYSIS OF STRUCTURES FOR EARTHQUAKE LOADING...

165

7.5

15

22.5 Time

30

37.5

45

53.7

Figure 5. Displacement history of level 7 by FDN (cm)

Figure 6. 72 Bar space truss Table 2. Maximum displacement of Example 2 in X direction

Joint No. Maximum dynamic displacement (cm) DAO DAR FDN 12.565 12.576 12.561 12.599 12.587 12.563 8.690 8.701 8.713 8.700 8.711 8.709 5.069 5.101 5.098 4.824 4.867 4.834 1.502 1.503 1.501 1.703 1.711 1.708

1 3 6 8 9 12 14 16

166

A. Heidari and E. Salajegheh

Table 3. Maximum displacement of Example 2 in Y direction

Joint No. Maximum dynamic displacement (cm) DAO DAR FDN

1 3 6 8 9 12 14 16

12.497 12.502 8.882 8.874 5.234 5.067 2.192 1.703

12.501 12.586 8.902 8.997 5.278 5.103 2.214 1.706

12.499 12.531 9.899 9.101 5.229 5.109 2.209 1.709

Table 4. Maximum displacement of Example 2 in Z direction

Joint No. Maximum dynamic displacement (cm) DAO DAR FDN

1 3 6 8 9 12 14 16

5.445 7.554 2.778 3.616 4.358 2.833 1.128 1.703

5.419 7.568 2.786 3.624 4.378 2.856 1.146 1.695

5.451 7.549 2.781 3.620 4.401 2.860 1.133 1.691

7. CONCLUSIONS
An efficient method is presented for time history dynamic analysis of structures under earthquake loading. The main goal is to reduce the computational time of the analysis. The required time and efforts for existing dynamic structural analysis are excessive for largescale structures with many degrees of freedom. The idea of signal processing together with wavelets and filter banks are employed to overcome this problem. By the proposed method,

TIME HISTORY ANALYSIS OF STRUCTURES FOR EARTHQUAKE LOADING...

167

the acceleration record is filtered and the number of points of the record is reduced. The structure is analysed dynamically with less number of time intervals. In addition, a discrete wavelet neural network is presented to approximate the dynamic analysis of a SDOF structure for the cases in which the structural properties such as damping ratio and natural frequency are changed. The results are used for the analysis of any MDOF structure. In fact, the FWT is used to enhance the efficiency of the time history dynamic analysis of a SDOF system against a specified earthquake and the DWN is employed for the dynamic analysis of any MDOF system using modal superposition approach. The numerical results show that in the proposed method, the time of dynamic analyses is reduced to about 0.1 of the time required for the time history dynamic analysis using the original earthquake.

REFERENCES
Grossmann, A. and Morlet, J., Decomposition of Hardy function into square integrable wavelets of constant shape, SIAM J. Math. Anal., 15(1984) 723-736. 2. Farge, M., Wavelet transforms and their application to turbulence, Annu. Rev. Fluid Mech., 24(1992) 395-457. 3. Polikar, R., The wavelet tutorial, http://users.rowan.edu/~polikar/WAVELETS/WTtutorial. html, 2001. 4. Daubechies, I., The wavelet transform, time-frequency localization and signal analysis, IEEE Trans. Inf. Theory, 36(1990) 961-1005. 5. Salajegheh, E. and Heidari, A., Dynamic analysis of structures against earthquake by combined wavelet transform and fast Fourier transform. Asian J. Civil Engrg., 3(2002) 75-87. 6. Salajegheh, E. and Heidari, A., Time history dynamic analysis of structures using filter banks and wavelet transforms. Compuers Structures, 83(2005) 53-68. 7. Salajegheh, E., Heidari, A. and Saryazdi, S., Approximate dynamic analysis of structures against earthquake by fast wavelet transform, Iranian J. Sci. Tech., (Submitted), 2005. 8. Heidari, A. and Salajegheh, E., Dynamic analysis of structures against earthquake by an adaptive wavelet, SEE4 Conference, Tehran, Iran, 2003, Paper No. 36 on CD. 9. Salajegheh, E., Heidari, A. and Saryazdi, S., Optimum design of structures against earthquake by discrete wavelet transform, Int. J. Num. Meth. Engrg., 62(2005) 21782192. 10. Salajegheh, E. and Heidari, A., Optimum design of structures against earthquake by a modified genetic algorithm using wavelet networks. SEE4 Conference, Tehran, Iran, 2003, Paper No. 92 on CD. 11. Salajegheh, E. and Heidari, A., Optimum design of structures against earthquake by adaptive genetic algorithm using wavelet networks, Struc. Multidisc. Optim., 28(2004) 277-285. 12. Salajegheh, E. and Heidari, A., Optimum design of structures against earthquake by wavelet neural network and filter banks, Int. J. Earthquake Engng. Struct. Dyn., 34(2005) 67-82. 1.

168

A. Heidari and E. Salajegheh

13. Salajegheh, E. and Heidari, A., Optimum design of structures against earthquake by discrete wavelet neural network, CST Conference, Lisbon, Portugal, 2004, Paper No. 92 on CD. 14. Strang, G. and Nguyen, T., Wavelets and Filter Banks, Wellesley-Cambridge Press, New York, USA, 1996. 15. Mallat, S., A theory for multiresolution signal decomposition: the wavelet representation, IEEE Trans. Patt. Anal. Mach. Intel., 11(1989) 674-693. 16. Heidari A., Optimum Design of Structures Against Earthquake by Advanced Optimisation Methods, Ph.D. Thesis, University of Kerman, Kerman, Iran, 2004. 17. Paz, M., Structural Dynamics: Theory and Computation, McGraw Hill, New York, USA, 1997.