GENERATION OF DESIGN EARTHQUAKE AND DAMAGE DETECTION USING ANN

Abstract

Three ANN based models are developed for (i) prediction of peak ground acceleration, (ii) generation of spectrum compatible accelerograms and (iii) identifying location and extent ofdamage in structure utilizing artificially generated accelerograms. Apredictive model referred to as neural attenuation model in this work is developed to estimate the peak ground acceleration (PGA) at a given site using the learning capability ofartificial neural networks (ANN). The strong-motion data recorded during ten Indian earthquakes in the Himalayan region are used to train and test the model. Unlike the empirical attenuation relations, the neural attenuation model is evolutionary in nature and eliminates a priori assumption regarding the form of the model. The significant association of variables is directly learned by the network through supervised learning on the data set. The heuristic guidelines for training the model are derived by incrementing the size oftraining data set in phases and an optimal architecture of the network is determined in the final phase, where full data set is used for training. Aproblem dependent statistical cross-validation procedure is employed for the testing of the model trained with full data set. Asingle neural attenuation model is able to learn the difference in attenuation rate in different regional setting of central Himalayan and NE India region reasonably well. The model consistently gives better predictions than the existing empirical attenuation relations for ten Indian earthquakes. ANN based model is also developed to generate accelerograms from given response target spectra. The methodology adopted in this research work is a variation of the methodology proposed by Ghaboussi and Lin in 1998. In a two-stage approach to generate spectrum compatible accelerograms, a replicator neural network is used to compress the high dimensional Fourier spectra of accelerograms to much lower dimensional vectors. A modular compression was used to compress the Fourier spectra, where the optimal architecture of replicator neural network is determined to achieve a high speed of compression and reasonable compression ratio. In the second stage, accelerogram generator neural networks are trained to inversely map the compressed vector of Fourier spectra to their corresponding response spectra. Asample of thirty eight accelerograms are used to demonstrate the performance of the replicator neural network and accelerogram generator neural network. The methodology is extended to develop multiple accelerograms compatible with target spectra. More than hundred accelerograms recorded during ten Indian earthquakes, pre-classified based on their observed predominant frequencies, are used to train and test multiple neural networks. The networks are trained to generate accelerograms from response spectra in all the preclassified categories. The networks are able to generate spectrum compatible accelerograms from target response spectra in all the categories of different predominant frequency content. The networks are also tested with smoothed design spectra as inputs and are able to synthesize ensemble of realistic accelerograms with desired frequency content. The artificial accelerograms generated in this work are utilized to train an ANN based model for identifying the location and extent of damage in a building. The transient responses of the building due to artificially generated accelerograms were mapped to various simulated damage states. Transient responses due to recorded accelerograms were given as inputs to the model. An example is presented where accelerograms generated and recorded atwo site are used to train and test the model. The model is able to identify the location and extent of damage in the building. The study shows that the artificially generated accelerograms can be used to evaluate performance of structure including damage detection for future hypothesized earthquakes