An Adaptive Symmetric Loss in Dynamic Wide-Kernel ResNet for Rotating Machinery Fault Diagnosis Under Noisy Labels

In engineering applications, the effectiveness of rotating machinery fault diagnosis is often disturbed by both vibration noise and labeling noise. Vibration noise makes it more difficult for the model to extract a potential fault impact feature, while labeling noise makes the feature extraction go in the wrong direction through error backpropagation. Both of these factors lead to a decrease in fault diagnosis accuracy. Hence, an adaptive symmetric loss in dynamic wide-kernel ResNet (AS-DWResNet) is proposed to solve the rotating machinery fault diagnosis under noisy labels. Firstly, wider convolution kernels are designed for branch convolution to enhance the extraction of vibration features. Secondly, a parallel network structure is constructed to fully and accurately extract the fault features by dynamically weighting different branches to improve the basic diagnostic performance of the network in the background of vibration noise. Thirdly, an improved adaptive symmetric cross-entropy (ASCE) loss function is adapted to reduce the effect of label noise in training datasets and hyperparameters in loss functions. The experimental results show that in the background of vibration noise and severe noisy labels, the proposed method still keeps more than 99.08% accuracy and has stronger noise robustness and stability than other networks and loss functions.