A train bearing imbalanced fault diagnosis method based on extended CCR and multi-scale feature fusion network

The number of fault samples is much less than the normal samples in the actual operation of the train bearing, and the imbalanced characteristics of the fault data significantly decrease the performance of the diagnosis model. Therefore, a train bearing imbalanced fault diagnosis method (ECCR-MFFN) based on extended combined cleaning and resampling (ECCR) and the multi-scale feature fusion network (MFFN) is proposed. Firstly, the ECCR method is proposed, which adaptively determines the sampling area and provides rich fault information for the diagnostic model with high-quality synthesized samples. Then, MFFN is designed to obtain great feature extraction and classification results under imbalanced data conditions through feature extraction and fusion strategies of multi-branch different kernels. Finally, the superiority and effectiveness of the ECCR-MFFN under various data imbalance conditions are verified by comparative experiments on laboratory and public bearing datasets. The results demonstrate that the MFFN can effectively extract fault features under small imbalance rate (IBR) conditions and achieve ideal classification results. Compared with other data augmentation methods, the ECCR can synthesize samples with higher quality and has a more stable performance. Under the condition of IBR = 40:1, the accuracy of the ECCR-MFFN is 95.84% and 96.07%, which is significantly better than the comparison methods and offers a reliable method for dealing with data imbalance.