A diagnosis method for imbalanced bearing data based on improved SMOTE model combined with CNN-AM
Zhenya Wang, Tao Liu, Xing Wu, Chang Liu
Abstract
Abstract A boundary enhancement and Gaussian mixture model (G) optimized synthetic minority oversampling technique (SMOTE) algorithm (BE-G-SMOTE) is proposed to improve diagnostic accuracy under imbalanced bearing fault data conditions. It is designed to solve the problem that the diversity of samples generated by the original SMOTE model is limited, as well as the deep learning model is limited by the size of training samples and processing speed. Firstly, a few bearing fault data are clustered by G to achieve cluster division. Secondly, according to the cluster density distribution function designed in this paper, the weights of different clusters and sample weights to achieve intra-class balance are determined and data quality is improved. Then, to take full advantage of the limited fault data, based on the sensitivity of the support vector machine (SVM) to imbalanced data, the enhanced boundary is established between generated data and the SVM classifier under different penalty factor (PF) values. According to the accuracy, the optimal PF is determined, and fault datasets satisfying diversity are obtained. To improve the classification accuracy, a convolutional neural network with an attention mechanism is built. Finally, analysis using two practical cases shows the effectiveness of the proposed method.