New imbalanced bearing fault diagnosis method based on Sample-characteristic Oversampling TechniquE (SCOTE) and multi-class LS-SVM

Abstract In actual industrial production, the historical data sets used for bearing fault diagnosis are generally limited and imbalanced and consist of multiple classes. These problems present challenges in the field of bearing fault diagnosis, for which traditional fault diagnosis methods (e.g., multi-class least squares support vector machine (multi-class LS-SVM)) are not very effective. Therefore, we propose a new multi-class imbalanced fault diagnosis method based on Sample-characteristic Oversampling Technique (SCOTE) and multi-class LS-SVM, where SCOTE is a new oversampling method proposed by us. SCOTE transforms multi-class imbalanced problems into multiple binary imbalanced problems. In each binary imbalanced problem, first, SCOTE uses the k-nearest neighbours (knn) noise processing method to filter out noisy points. Second, samples are trained by LS-SVM, and minority samples are sorted by importance according to the misclassification error of the minority classes in the training sets. Moreover, based on the importance sorting of minority samples, SCOTE performs a sample synthesis method based on the k* information nearest neighbours (k*inn) to address the binary imbalanced problems. Thus, when all the binary imbalance problems are addressed, the multi-class imbalanced problem will also be addressed. The 20 fault diagnosis examples represented by Case Western Reserve University (CWRU) bearing data and Intelligent Maintenance Systems (IMS) bearing data show that the proposed method has higher fault diagnosis recognition rates and algorithm robustness than 8 oversampling algorithms and 8 multi-class imbalanced algorithms.