A novel fault diagnosis method based on deep stable learning for bearings with imbalanced data samples

To resolve the fault diagnosis of bearing with imbalanced data samples, data augmentation method and the diagnosis algorithm-level methods have poor generalisation and adaptive ability for complex industrial process tasks without any expert knowledge. But the intrinsic fault features play a very important role in the fault diagnosis which can be hardly affected by the imbalanced data samples. A novel deep stable Resnet34 model is proposed to obtain the intrinsic fault features to diagnose the imbalanced bearing fault in this paper. The Resnet34 is utilized to extract the features from the 2D time frequency maps transformed from the vibration signal by the continuous wavelet transform. After that, the extracted features are mapped into high dimension feature space by the random Fourier Feature (RFF), and the intrinsic fault features are obtained by the learning sample weighting and decorrelation method, which feed to the classifier of Resnet34 to diagnose the imbalanced data samples. The proposed method is tested on the imbalanced bearing under constant and variable working conditions. The diagnosis results demonstrate that the proposed deep stable Resnet34 can diagnose the bearing fault with imbalanced data samples effectively and has higher diagnosis accuracy and stronger generalization than other diagnosis methods, such as individual Resnet34, ensemble Resnet34 and Resnet34 with resampling.