Train Bearing Fault Diagnosis Based on Time–Frequency Signal Contrastive Domain Share CNN

Trackside acoustic detection using the fault diagnostic model is an important part of intelligent operation and maintenance in railroad transportation systems. However, the Doppler effect caused by motion can bring obstacles to the fault diagnosis of train bearings. Signal correction is the main way to solve the Doppler effect at present, but in practical applications, there are problems of unknown kinematic parameters, modeling difficulty, complicated calculation processes, and the need to set diagnostic criteria manually. The time-frequency distortion of signals caused by the Doppler effect poses a challenge to traditional deep learning methods and concat fusion of time-frequency signal features has the problem of more redundant information and missing complementary information. Therefore, this study proposes a time-frequency signal contrastive domain adaption learning method to transfer the existing knowledge of bearing fault diagnosis to train bearing. First, the method divides the bearing signal into time and frequency domains and constructs a long short-term memory network-convolutional neural network (LSTM-CNN) feature extractor. Second, an unsupervised domain share CNN (DSCN) is constructed to extract domain-invariant features of the time and frequency signals in the source and target domains using the multi-kernel maximum mean discrepancy (MKMMD) loss function. Finally, the common and complementary features are learned from the time and frequency domains using the contrastive learning loss, and then the bearing fault classification is completed. The experimental results demonstrate the effectiveness and potential application of the proposed method in train bearing fault maintenance.