Fault location based on relevance vector machine and decision directed acyclic graph

Relevance Vector Machine (RVM) is one of the `state-of-the-art´ approaches for classification which exploits the probabilistic Bayesian learning frame work. Compared with the classical Support Vector Machine (SVM), RVM avoids the problem of parameter setting while learning and offers probabilistic outputs. These make RVM more suitable for real applications. In this paper, we have employed the DDAG approach to extend the RVM into a multi-classifier which enables its recognition of different faulty patterns, and further makes the fault location feasible. Compared with conventional methods, The proposed approach yields a smaller computing complexity whereas it maintains a higher diagnostic reliability. It has also been applied to the real problem of pulling motor fault isolation, and satisfactory results have been obtained in these experiments which has validated the effectiveness of proposed approach.