Incremental Fault Diagnosis of Train Bogies Through Temporal Prototype Evolution

Emerging fault conditions, dynamic operational regimes, and limited labeled data pose significant challenges for fault diagnosis in high-speed train bogies, particularly in multi-sensor monitoring environments. This paper proposes an incremental diagnostic framework that reformulates the principles of continually evolved classification to support the progressive modeling of multivariate sensor sequences. A Conformer-based backbone is adopted to capture both local dynamics and long-range dependencies across sensor channels, enabling a stable and expressive temporal feature space that is preserved throughout the learning process. To mitigate catastrophic forgetting and enhance class discrimination under class-incremental conditions, we introduce a prototype-centric classifier whose structure evolves over time through a graph attention mechanism, effectively capturing semantic dependencies among both existing and novel fault classes. Additionally, a pseudo-meta-incremental learning strategy is introduced, in which diverse temporal augmentations—such as stretching and masking—are applied to alleviate sample scarcity and enhance the classifier's adaptation during incremental phase. Experimental validations on a bogie dataset confirm that the proposed method maintains strong diagnostic performance under evolving fault distributions, demonstrating its practical utility for intelligent railway maintenance in non-stationary, safety-critical environments.