Multi-class Intelligent Fault Diagnosis Approach Based on Modified Relevance Vector Machine

Abstract

In order to solve the problem of fault data with small sample and nonlinear in fault diagnosis and improve support vector machine, a fault diagnostic approach based on the multi-class classification method of One-Against-Rest (OAR) algorithm and decision tree is proposed combined with relevance vector machine. The above classification method modifies the current OAR algorithm using decision tree during the testing phase. To be specific, the K classifiers of OAR algorithm are arranged to form a decision tree in descending order and to reduce the average testing numbers of classifiers is the optimization object. Meanwhile, the threshold value of distance function is set and function value of each classifier is calculated in the sequence of decision tree. Once the function value of the i-th classifier exceeds the threshold, the testing sample will be assigned to the i-th class without any other evaluations. If none of values exceeds the threshold, the sample is classified to the class of the maximal decision function value as same as that of OAR. Theoretical analysis and experimental results both demonstrate that the presented approach performs better than traditional methods in terms of diagnosis time, diagnosis accuracy and diagnosis efficiency.