Experimental investigation of seismic damage identification using PCA-compressed frequency response functions and neural networks

This paper presents an experimental investigation of seismic damage identification of a 38-storey tall building model using measured frequency response functions (FRFs) and neural networks (NNs). The 1:20 scale reinforced concrete structure is tested on a shaking table by exerting successively enhanced ground earthquake excitation to generate trifling, moderate, serious and complete (nearly collapsed) damage, respectively. After incurring the earthquake excitations at each level, a 20-min white-noise random excitation of low intensity is applied to the structure to produce ambient vibration response, from which FRFs are measured for post-earthquake damage detection by means of the NN technology. Principal component analysis (PCA) is pursued to the measured FRFs for dimensionality reduction and noise elimination, and then the PCA-compressed FRF data are used as input to NNs for damage identification. After a study on tolerance of PCA-reconstructed FRFs to measurement noise, different PCA configurations are designed for overall damage evaluation and damage location (distribution) identification, respectively. It is shown that the identification results by means of the FRF projections on a few principal components are much better than those directly using the measured FRF data, and agree fairly well with the visual inspection results of seismic damage during tests.