Fault diagnosis for motor bearing has a great economic and social significance, the patents on motor bearing
fault detection and diagnosis method have been presented in the past years. However, there are some shortcomings including
local extremum or over-fitting in the fault diagnosis method of these patents. Relevance vector machine (RVM), requires
much fewer kernel functions compared with support vector machine (SVM), and provides the posterior distributions
rather than just a point estimate. Thus, fault pattern recognition for motor bearing based on multi-layer RVM classifier
was proposed in the paper, and two multi-class combination patterns of RVM classifiers including ‘binary tree’ and
‘one-against-one’ were used, among which ‘binary tree’ pattern is used to distinguish the fault state from the normal state,
‘one-against-one’ pattern was used to recognize the fault types. The fault pattern recognition results for motor bearing
among RVM classifier, SVM classifier and RBF neural network (ANN) classifier were analyzed, and the experimental results
indicate that RVM classifier has the better fault pattern recognition ability than SVM classifier or RBFNN classifier.
fault diagnosis, multi-class combination pattern, motor bearing, pattern recognition, RVM.
School of Mechanical Engineering, Donghua University, Shanghai 201620, China.