An Improved Metalearning Framework to Optimize Bearing Fault Diagnosis under Data Imbalance

Abstract

The intelligent diagnosis of rotating machinery with big data has been widely studied. However, due to the variability of working conditions and difficulty in marking fault samples, it is difficult to obtain enough high-quality fault marking data for training bearing fault diagnosis models in practical industrial application scenarios. Aiming at the problem of training data imbalance caused by lack of fault samples, a novel metalearning fault diagnosis method (MOFD) is proposed to get the bearing fault diagnosis solution under data imbalance. Firstly, in order to enhance the variety of fault samples, a Feature Space Density Adaptive Synthetic Minority Oversampling Technique (FSDA-SMOTE) is proposed in this paper, which takes the density difference of minority samples in the spatial domain within the class as the constraint of local neighbor similarity to generate new fault samples for data augmentation. In addition, in order to strengthen the model’s learning ability and diagnosis performance under limited fault samples, a residual-attention convolutional neural network (RA-CNN) was constructed to identify the deep features of fault signals, and a metalearning strategy based on parameter gradient optimization was applied to RA-CNN for refining the learning process of the diagnosis model. Finally, the reliability of the proposed method is verified through experimental analysis of public bearing dataset.