GD-TCN: Lightweight Data–Physics Integrated Fault Diagnosis for Heavy-Haul Railway Bogies

Timely and accurate diagnosis of bogie faults is essential for ensuring the safe and reliable operation of heavy-haul railways, yet real-time implementation remains challenging due to limited onboard computational resources. To address the aforementioned limitation, this paper proposes a lightweight fault diagnosis model (GD-TCN) that integrates data-driven techniques with physics-informed strategies. First, GD-TCN introduces a Ghost Depthwise-Separable Convolution (G-DSC) block that combines large-kernel depthwise convolutions with the efficient feature generation mechanism of GhostNet, significantly reducing both the parameter count and the computational complexity. Second, the model incorporates physical prior knowledge through an auxiliary regression task, thereby enhancing the robustness and interpretability of the learned representations. Experimental evaluations on bogie-fault datasets show that GD-TCN achieves a diagnostic accuracy of 97.67% with only 31K parameters and 14.68M FLOPs.