Seismic damage evaluation of a 38-storey building model using measured FrF data reduced via principal component analysis

Abstract

This chapter investigates the seismic damage of a 38-storey tall building model using measured frequency response function (FRF) data and neural network technique. The 1:20 scale structural model was tested on a shaking table by exerting successively enhanced ground earthquake excitations to generate trifling, moderate, severe, and complete (nearly collapsed) damage, respectively. A total of 27 accelerometers were instrumented on the structure for the measurement of FRFs in healthy state and after incurring each of the damage scenarios. A multilayer neural network with FRF data as input variables was configured and trained for damage occurrence, and extent detection in the structure. In order to circumvent the difficulty of huge dimension of input vector when using full-size FRF data with neural network, principal component analysis (PCA) was introduced to compress the size of FRFs, and the projection of measured FRFs onto the most significant principal components was finally used instead of raw FRF data as neural network input for damage identification. The results showed that this approach could indicate the structural damage condition with high fidelity.