Mechanical fault diagnosis under long-tailed data: an improved federated learning method integrating attention, asymmetric convolution, and margin calibration

Abstract Due to the inability to collect sufficient fault samples for specific types of gear and bearing faults, the data presents a long-tailed distribution, which hinders the effective construction of neural network diagnostic models. Furthermore, when Federated Learning (FL) methods are introduced to address this long-tail problem, they struggle to effectively extract feature information from samples belonging to the tail fault types. To address these issues, this paper proposes an improved FL method. Firstly, federated features are utilized to retrain the diagnostic model, enhancing the fault feature extraction capability for tail samples. Secondly, the Convolutional Block Attention Module (CBAM) mechanism is introduced to improve the Residual Network (ResNet) model within the FL framework, strengthening the capability and efficiency of extracting key local feature information from both channel and spatial dimensions. Thirdly, traditional convolution is replaced with asymmetric convolution to enhance the capability and efficiency of extracting asymmetric feature information from samples. Finally, a margin calibration algorithm is employed to optimize the network model's classification margin, thereby achieving higher diagnostic accuracy and efficiency. Experimental analysis based on measured fault samples from gears and bearings demonstrates that the proposed improved FL method effectively increases the average and maximum accuracy by 8.78% and 3.40%, respectively.