An approach for bearing fault diagnosis based on PCA and multiple classifier fusion

The purpose of this paper is to propose a new system, with both high efficiency and accuracy for fault diagnosis of rolling bearing. After pretreatment and choosing sensitive features of different working conditions of bearing from both time and frequency domain, principal component analysis(PCA) is conducted to compress the data dimension and eliminate the correlation among different statistical features. The first several principal components are sent to the classifier for recognition. However, recognition method with a single classifier usually has only a limited classification capability that is insufficient for real applications. An ongoing strategy is the decision fusion techniques. The system proposed in this paper develops a decision fusion algorithm for fault diagnosis, which integrates decisions of multiple classifiers. First, the front four principle components are chosen as input of individual classifier. A selection process of the classifiers is then operated on the basis of correlation measure for the purpose of finding an optimal sequence of them. Finally, classifier fusion algorithm based on Bayesian belief method is applied to generate the final decision. The result of experiments show that this new bearing fault diagnosis system recognize different working conditions of bearing more accurately and more stably than a single classifier does, which demonstrates the high efficiency of the proposed system.