Uncertainties Consideration in Empirical Frequency Response Function Data for Damage Identification Based On Artificial Neural Network

Abstract

The modern application of frequency response function (FRF) with artificial neural networks (ANN) has become one of the leading methods in vibration-based damage detection approach. However, since full-size empirically obtained FRF data is used as ANN input, a broad composition ANN input layer series would occur. Consequently, principal component analysis (PCA) is adopted to compress the FRF data magnitude. Despite this, PCA alone is unable to select the important FRF data features effectively, due to the exceedingly FRF data size in addition with existing uncertainties. Therefore, this study proposed the merger of a non-probabilistic analysis and ANN approach with PCA by considering the uncertainties effect and the inefficiency of using empirical FRF data. The empirical FRF data is obtained from a steel truss bridge structure. The results show that the PoDE values above 95% are measured at the particular executed damage locations and the DMI values show the damage severity at the actual damage locations. Overall, the results show that the proposed method is capable in considering the uncertainties effect on the empirical FRF data for structural damage identification.