Analysis of building seismic records using artificial neural networks

Abstract

This chapter uses an artificial neural network to analyze the building acceleration records obtained during earthquakes. The neural network contained a buffer layer with six delay data points. The hidden layer either contained a nonlinear activating function or a simple linear activation function. The neural network was first tested on a model ten-storey building frame using real earthquake ground acceleration and simulated response at each floor based on a second-order linear mathematical model. The training of the network was successful and the fundamental vibration frequency and damping ratio were identified by exciting the network with harmonic acceleration. The neural network was then applied to a set of real data during the 1989 Loma Prieta earthquake from a four-story building. The data exhibited pronounced time-variation in the building characteristics. The neural network was not able to simulate this time-variation behavior over the long duration of thirty seconds, but was accurate in characterizing behavior within short duration windows of two to four seconds.