Bearing fault diagnosis method based on multi-domain feature fusion and heterogeneous network under small sample conditions

To solve the problems of insufficient feature extraction of the current methods under small sample conditions and loss of information in the process of signal transformation from different domains, a bearing fault diagnosis method based on multi-domain feature fusion and heterogeneous networks under small sample conditions is proposed. The method firstly designs the data preprocessing module to transform and combine the raw vibration signals into multi-domain signals by Fast Fourier Transform (FFT) and Gram Angle Field (GAF), which provides rich feature conditions for the subsequent feature extraction. Then, heterogeneous branch networks are designed for different domain signals used in low-dimensional feature extraction in the high-dimensional nonlinear space of fault data. When the inputs or intermediate processes of one branching network is interfered by the outside world, another branching network will play the role of error correction, which enhances the fault-tolerance of the proposed method. Next, in order to enhance the critical feature extraction capability of the heterogeneous network, the Location-Aware Channel Enhancement Block (LACEB) is designed. The LACEB learns the unique weights for different channels and different locations in the feature map by adaptively adjusting the dynamic factors and feature location parameters. Further, the memory unit in the global feature extraction module is used to learn the context information of each time step, and the dependency between the global features and the local features is effectively established. Finally, in order to prevent the model from falling into local optimal, a learning rate adaptive optimization algorithm is designed to optimize the model training process. A variety of strictly comparative experiments were tested on the CWRU dataset and the MFS dataset, concluding that this method is capable of performing fault diagnosis tasks in different environments and devices.