WF-SwinT: A Wavelet Fusion Method for Fault Diagnosis of Variable-Speed Rolling Bearings

Metro train rolling bearings operate under variable-speed conditions due to continuous acceleration and deceleration, leading to time-varying characteristics in vibration signals. These conditions complicate fault diagnosis, as time-domain features evolve dynamically and frequency-domain features become obscured by fluctuations in rotational frequency. To address these challenges, this study proposes a novel fault diagnosis method named WF-SwinT that integrates continuous wavelet transform (CWT) with a deep learning feature extraction network. The proposed method utilizes two different wavelet bases to construct a dual-wavelet time–frequency characterization approach, which converts the vibration signals captured by the accelerometer into 2-D time–frequency maps that capture complementary fault-related information. These maps are processed by a dual-branch Swin Transformer network, where an attention-guided feature fusion module dynamically integrates multiscale features from both branches. Furthermore, feature fusion is achieved through spatially invariant convolutional filters, and multilayer perceptrons (MLPs) are followed by classification. The experimental results show that the proposed method outperforms the comparison method in terms of accuracy, precision, recall, and ${F}1$ score. Especially, the proposed method achieves an accuracy of 99.56% and 99.17% on two bearing vibration signal datasets, respectively. It provides an effective fault diagnosis method for variable-speed conditions.