Dynamic normalization supervised contrastive network with multiscale compound attention mechanism for gearbox imbalanced fault diagnosis.

Deep learning has gained significant success in fault diagnosis. However, the number of gearbox health samples is inevitably much larger than that of fault samples in real-world engineering, which severely limits the diagnostic performance of such methods. Thus, this paper put forward a dynamic normalization supervised contrastive network (DNSCN) with a multiscale compound attention mechanism to recognize imbalanced gearbox faults. First, a multiscale adaptive feature extractor (MAFE) possessing branch weight adjustment capability has been devised to serve as a contrastive learning backbone to effectively mine signal features. Second, a multiscale compound attention mechanism is designed to reweight the features from the MAFE, thus improving the accuracy and confidence of fault recognition. Third, a dynamic normalized supervised contrastive loss function for imbalanced scenarios is presented. It balances the contributions of minority and hard-to-classify samples in the loss function using class normalization and dynamic adjustment based on the training accuracy, respectively. DNSCN achieved accuracies of 91.58% and 90.96% on two gearbox datasets with extreme imbalance ratios, which proved the superior performance of this approach. • The supervised contrastive learning is modified to deal with the gearbox class imbalance problem from a new perspective • A multiscale adaptive feature extractor is designed as the backbone of contrastive learning to efficiently mine features in imbalanced signals. • A multiscale attention compound mechanism is presented to enhance further the importance of critical features extracted from the contrastive learning backbone, thereby improving fault identification accuracy. • A dynamic normalized supervised contrastive learning method is devised to cope with the effect of class-imbalanced samples in fault diagnosis. • Two gearbox experiments have been implemented to demonstrate that our method outperforms the most current advanced intelligent methods under severe data type imbalance. [ABSTRACT FROM AUTHOR]