An incremental learning method based on dynamic-static collaborative prompts for fault diagnosis of critical components in bogies

As the demand for intelligent operation of high-speed railways increases, the non-stationarity and distributed drift of high-speed railway bogies under multi-coupled operating conditions become increasingly pronounced. With extended service duration, the failure severity of critical bogie components continuously evolves, necessitating ongoing updates to diagnostic models. Existing class-incremental fault diagnosis methods are prone to forgetting old classes or failing to adequately learn new ones when training on new classes, making it difficult to effectively balance stability and plasticity. Furthermore, the utilization of old knowledge during the incremental phase tends to be fixed, lacking effective interaction between the features of old and new classes. To address these issues, this paper proposes an incremental fault diagnosis method based on dynamic-static collaborative prompts (DSCP) to achieve a synergistic balance between stable memory and rapid adaptation. Firstly, a hybrid convolutional neural network-Transformer network is constructed as the fundamental framework. An extensible spherical distribution classifier is introduced to optimize inter-class decision boundaries through boundary scaling, enhancing the discrimination capability of the model. Subsequently, a feature replay module is designed, incorporating feature embeddings into the prompt vector to enhance the representation of replay sample relevance. Finally, a parameter constraint module is developed, employing a sensitivity elastic weight consolidation mechanism to statically impose constraints, thereby protecting critical parameters from historical tasks and enabling rapid adaptation. Comparative experiments conducted on three high-speed railway bogie datasets demonstrate that DSCP achieves superior diagnostic accuracy compared to existing incremental learning approaches.