An Incremental Learning Method With Feature-Attention Distillation and Logit Adjustment for Rotating Machinery Fault Diagnosis

Deep learning-based diagnosis methods can accurately identify the fault mode, which have attracted widespread attention from researchers. For mechanical systems, in actual industrial environments, data from different fault modes will continue to emerge, which requires the model to be updated in a timely manner and maintain high diagnosis accuracy. Class incremental learning (CIL) is proposed for classification problems with a continuously increasing number of modes, which meets the requirement of industrial diagnosis. However, directly incorporating new class data into the training set to optimize the network will cause the model to forget old class knowledge, resulting in irreversible performance degradation, known as catastrophic forgetting. To solve this problem, this paper constructs an Incremental Learning method with Feature-Attention Distillation and Logit Adjustment (ILD-FADLA) for fault diagnosis. Specifically, a residual network composed of convolutional blocks is utilized as the feature extractor, and channel attention and spatial attention are superimposed in each residual block to enhance the feature extraction capability. To alleviate catastrophic forgetting, the proposed ILD-FADLA distills the attention weights at each layer and feature relation information before classifier. In addition, logit adjustment cross-entropy loss is employed to mitigate the bias of the classifier towards new classes. Experimental results on two private datasets show that the proposed ILD-FADLA improves the average accuracy of the incremental phase by 17.76% and 22.42% over the baseline method.