Improving Structure Damage Identification by Using ICA-ANN Based Sensitivity Analysis

Abstract

Structure health monitoring detects the nature of structure damage in an early stage by the network sensors whose signals are normally highly correlated and mixed with noise. Feature reduction methods are applied in extracting attributes, that will be input into advanced classification models. The complicated data transformation and classification procedures make it difficult to identify direct relationship between sensors and abnormal dynamic structure characteristics, especially for complex buildings with large numbers of sensors. In this study, the sensor sensitivity analysis on a structure damage identifier is applied, which integrates independent component analysis (ICA) and artificial neural network (ANN) together. The approach is evaluated on a benchmark data from University of British Columbia. The experimental results show sensitivity analysis not only helps domain experts understand the mapping from different location and type of sensors to a damage class, but also significantly reduce noise and increase the accuracy of ICA-ANN classifier.