A long-short-term feature extraction network based on soft-parameter-sharing for high-speed train bogies multi-object fault diagnosis under long-tailed distribution

To guarantee the safe and efficient operation of high-speed trains, it is crucial to recognize faults in the essential components of the bogie, which is the main load-bearing unit of high-speed trains. Current methods primarily concentrate on recognizing faults within single components, yet they are lacking in the concurrent diagnosis of multiple rotating components, such as gears and bearings. Furthermore, acquiring adequate fault signals in engineering practice is challenging, and the likelihood of faults in various components differs rendering fault diagnosis under long-tailed distribution a recurring problem. To address these issues, this study proposes a long-short-term feature extraction network with soft-parameter-sharing for simultaneous fault diagnosis of bogie gears and bearings under long-tailed distribution. The method integrates a CNN-based shared backbone network for short-term feature extraction and a Transformer-based multi-task network for long-term feature extraction. By combining short-term and long-term features, the proposed method effectively captures fault features representing the health status of different components. A cross-attention mechanism facilitates information transfer between sub-task networks for soft-parameter-sharing. Additionally, the inclusion of a cost-sensitive function mitigates the impact of class-imbalanced datasets on model performance, while a Pareto-efficient algorithm adaptively assigns loss function weights for each task. The experimental results on the two datasets demonstrate that the proposed method outperforms the comparison method, achieving over 96% accuracy, thereby validating its efficacy and potential for multi-objective fault diagnosis under long-tailed distribution conditions. Our code is available at https://github.com/lyj05121308/LSFEN.