Freight car roller bearing fault detection using artificial neural networks and support vector machines

Abstract

This paper deals with fault detection and diagnostics of freight rail tapered roller bearings using statistical features and auto-regressive model parameters extracted from bearing vibration signals. Two classifiers are used to distinguish cup, cone, and roller defects, namely, artificial neural networks (ANNs) and support vector machines (SVMs). Time domain vibration data is collected from normal bearings, and defective bearings with combinations of cup, cone, and roller defects. Features are used to train both ANNs and SVMs, and performance is compared for both classifiers. Mutual information between time domain features and target classes is computed in order to rank features with the highest relevance. Excellent performance was observed using laboratory data and the trained ANNs and SVMs, with testing accuracy as high as 100% in some cases.