Digital twin fault diagnosis method for high speed train Bogies based on HGKN-SINDy and residual graph sequences

Abstract

To address the difficulty of unified modelling for heterogeneous signals, the limited representation of high-order collaborative relationships and the ambiguity of fault boundaries in multimodal monitoring data from high-speed train bogies, this article proposes a digital twin fault diagnosis method based on a hypergraph kernel neural operator integrated with sparse identification of nonlinear dynamics (HGKN-SINDy) and residual graph sequences. First, temperature, vibration and current signals are represented as a unified node channel state matrix. An HGKN-SINDy digital twin model is then constructed by combining hypergraph kernel neural operators with sparse nonlinear dynamic identification to characterise high-order coupling relationships among sensors and the dynamic evolution of the system. Second, digital twin residuals are generated from the discrepancy between observed and predicted derivatives, and residual graph sequences are constructed through dynamic graph generation and sliding window sampling. Finally, a multi-scale relation GraphSAGE model is developed to extract complementary fault features from instantaneous local, two-hop propagation and window-level dynamic co-evolution relationships for bogie fault classification. Experiments on multimodal data from a CR400BF high-speed train bogie show that the proposed method outperforms the comparative methods in modelling accuracy, classification performance and result stability, while the three relation views exhibit good complementarity. Further sensitivity and disturbance experiments support the adopted derivative scheme and demonstrate a certain tolerance to limited channel loss. The proposed method also reduces confusion between fault types with similar residual patterns, providing an interpretable digital-twin diagnostic approach for intelligent operation and maintenance of high-speed train bogies.