Structural damage detection using artificial neural networks and measured FRF data reduced via principal component protection

This paper deals with structural damage detection using measured frequency response functions (FRFs) as input data to artificial neural networks (ANNs). A major obstacle, the impracticality of using full-size FRF data with ANNs, was circumvented by applying a principal component analysis (PCA)-based data reduction technique to the measured FRFs. The compressed FRFs, represented by their projection onto the most significant principal components, were then used as the ANN input variables instead of the raw FRF data. The output is a prediction for the actual state of the specimen, i.e., healthy or damaged. A further advantage of this particular approach was found to be the ability to deal with relatively high measurement noise, which is of common occurrence when dealing with industrial structures. The methodology was applied to the measured FRFs of a railway wheel, each response function having 4096 spectral lines. The available FRF data were grouped into x, y and z direction FRFs and a compression ratio of about 400 was achieved for each direction. Three different networks, each corresponding to a co-ordinate direction, were trained and verified using 80 PCA-compressed FRFs. Twenty compressed FRFs, obtained from further measurements, were used for the actual damage detection tests. Half of the test FRFs were polluted further by adding 5% random noise in order to assess the robustness of the method in the presence of significant experimental noise. The results showed that, in all cases considered, it was possible to distinguish between the healthy and damaged states with very good accuracy and repeatability. (C) 2001 Academic Press.