Self-supervised learning for train bearing fault diagnosis based on time–frequency dual domain prediction

Train traction motor bearings (TTMBs) are critical for ensuring stable and safe train operation. However, most intelligent fault diagnosis methods face significant challenges due to the complexity and variability of TTMB fault data, as well as the high cost of obtaining labeled data. Accordingly, this paper introduces a self-supervised learning model that leverages time–frequency dual domain prediction (TFDDP). First, a data augmentation module is designed to generate diverse training data, thereby enhancing the model’s robustness and generalization ability under different conditions. Secondly, an encoder combining multiple attention mechanisms and residual networks is proposed to extract critical time–frequency features efficiently and improve fault mode recognition capabilities. A cross-correlation smoothing matrix loss function is then designed, employing a smoothing matrix to effectively regulate the interplay between frequency-domain predictions and time-domain features, thereby strengthening the alignment between related features. Additionally, by introducing adaptive moment estimation with a weight-decay optimizer and the one-cycle learning rate scheduler, the model’s risk of overfitting is reduced, and convergence speed is accelerated. Finally, comparisons with existing optimal models indicate that TFDDP achieves competitive diagnostic performance across various conditions and performs excellently in TTMB fault classification even with limited labeled samples.