ResPTFE: Self Supervised Learning-Based Train Bearing Fault Diagnosis in Labelled Data-Starved Scenarios

Abstract Fault diagnosis of train traction motor bearings (TTMB) is a key component to ensure the safe operation of trains. However, due to the high cost of acquiring labelled data in different fault modes of TTMB, the lack of labelled data leads to the poor performance of most existing intelligent diagnosis methods. To this end, a self-supervised TTMB fault diagnosis model based on residual parallel time-frequency encoder (ResPTFE) is proposed in this paper. Firstly, a time-domain data enhancement method is proposed to generate diverse fault samples. Secondly, a time-frequency encoder combining residual network with squeezing and excitation attention mechanism is constructed, which can effectively extract key time-frequency features. Subsequently, a cross-correlation regularised loss function is designed to enhance the correlation between corresponding features and ensure the independence between different features by constraining the interaction between frequency domain and time domain features. Finally, the experimental results show that ResPTFE is more suitable for TTMB fault diagnosis in the case of scarce labelled data.