A novel fault diagnosis of high-speed train axle box bearings with adaptive curriculum self-paced learning under noisy labels

Due to uncontrollable or objective reasons, mismarked labels (i.e., noise labels) are inevitably generated in the supervised fault diagnosis dataset of train axle box bearings. Self-paced learning is an effective method to deal with noise, but it needs to set additional hyperparameters, which limits its application. Therefore, this paper proposes a novel fault diagnosis method based on adaptive curriculum self-paced resistance learning under noisy labels. First, this method uses discrete wavelet packet transform to perform localized analysis of one-dimensional signals in both the time and frequency domains. Second, the memory effect of convolutional neural networks is exploited to learn predefined curriculum. In the self-paced learning framework, an adaptive method based on epoch statistics is proposed to select confidence samples to update the curriculum, which provides meaningful supervised information for the model to be trained next. Third, the resistance loss combined with the original loss function is used to optimize the model with the selected confidence samples. The proposed method was validated using the fault-bearing dataset from Paderborn University and the high-speed train bogie fault simulation test bench. Even under uniform noise labels ranging from 0% to 50%, the method achieved over 90% fault diagnosis accuracy in both datasets.