Rolling bearing intelligent fault diagnosis towards variable speed and imbalanced samples using multiscale dynamic supervised contrast learning

Abstract

Deep learning-based fault diagnosis methods have already attained remarkable achievements in this field. However, rolling bearing frequently operates under variable speed conditions, and the number of healthy samples collected is often significantly larger than that of failure samples. In this paper, a multiscale dynamic supervised contrast learning (MDSupCon) framework is proposed. First, a multiscale adaptive feature extraction network is designed as the backbone, which utilizes multiple convolutional kernels to enhance feature extraction capabilities under variable speed conditions, and the branch attention mechanism is incorporated to adaptively adjust the weights of various scale branches. Second, the joint channel-space attention mechanism is constructed to enhance the importance of critical features while reducing redundant information, thereby improving fault identification accuracy and interpretability. Third, the dynamic supervised contrast loss function is designed to assign dynamic compensation factors to samples of various categories according to the training results, which reduces the impact of easily classified samples and enhances the contribution of hard-to-classify samples in imbalanced scenarios. Additionally, a dynamic cross-entropy loss is designed to train the backbone and the classifiers. The MDSupCon has achieved superior results of 89.49% and 92.15% on two bearing datasets with an imbalance ratio of 20:1 and variable speeds.