Rolling bearing fault diagnosis based on imbalanced sample characteristics oversampling algorithm and SVM

Aiming at the shortcomings of the standard support vector machine (SVM) in the field of rolling bearing fault diagnosis, such as poor performance on imbalanced datasets, sensitivity to noise, and heavy dependence on its own parameters, an oversampling algorithm based on sample characteristics (OABSC) was proposed.First,improved agglomeration hierarchical clustering was used to divide the failure samples into multiple clusters. Then, the sample distance and the neighborhood density in each cluster were comprehensively considered to identify and remove "suspected noisy points", and sort the remaining samples according to the amount of information. Further, K★ -information nearest neighbors (K★ INN) oversampling algorithm in each cluster was utilized to synthesize new samples to balance the dataset. Finally, bearing failures at three different imbalance ratios were simulated and the parameters of the SVM classifiers were optimized by using particle swarm optimization (PSO). The experiments show that, compared with the existing algorithms, the proposed OABSC algorithm is better applicable to the field of bearing fault diagnosis where the data is distributed in multiple clusters and is imbalanced.It has higher G-mean value and AUC value, and stronger algorithm robustness. © 2020, Editorial Office of Journal of Vibration and Shock. All right reserved.