Efficient Cloud-Edge Fault Diagnosis in High-Speed Trains leveraging Dynamic Neural Inference

High-speed train fault diagnosis systems face significant challenges in achieving both rapid response and reliable detection under computational and communication constraints. While cloud computing offers substantial computational resources, the limited communication bandwidth and potential instability of train-to-ground networks make it impractical to continuously transmit all sensor data to remote servers. Conversely, edge computing alone is restricted by the computational resources available on-board. This paper proposes an efficient fault diagnosis method based on cloudedge collaboration leveraging dynamic neural inference. The proposed approach features a dynamic inference mechanism that intelligently allocates computational tasks between edge and cloud resources based on sample complexity, along with TernP Networks, a lightweight architecture specifically designed for multi-channel sensor signals through positional normalization, partial convolution, and point-wise convolution. Experimental results demonstrate that the proposed approach achieves improved accuracy and reduced latency while significantly reducing data transmission requirements, with the majority of cases being successfully processed at the edge. The dynamic framework demonstrates significant potential for realworld deployment in high-speed rail systems where reliable, real-time fault diagnosis is essential for operational safety.