Comparison of Damage Classification Between Recursive Bayesian Model Selection and Support Vector Machine

Abstract

All damage identification activities inevitably involve uncertainties, and the resulting classification ambiguity in contaminated structural health monitoring (SHM) features can dramatically degrade the damage assessment capability. Probabilistic uncertainty quantification (UQ) models characterize the distribution of SHM features as random variables, and the UQ models facilitate making decisions on the occurrence, location, and type of the damages. A Bayesian framework will be adopted and the damage classification is transformed into a model selection process, in which the most plausible structural condition is determined by means of the recursively updated posterior confidence. In contrast to the probabilistic approach, machine learning is another candidate approach, which employs training data and extracts features from the recorded measurements. A support vector machine (SVM) is employed to classify the frequency response function data obtained from rotary machine under different damaged conditions. With different size of feature and different kernel functions, the classification of ball bearing damages are studied. Comparison between the Bayesian model selection approach and SVM is concluded in this paper.