RRDNet: Robust Fault Diagnosis of Railway Switch Machines Under Imperfect Sensing

Railway Switch Machines (RSMs) are the actuators that directly control the train route and thus require fault diagnosis for condition-based maintenance. Most existing fault diagnosis research for RSMs assumes clean, fully perfect sensing, which rarely holds for trackside equipment in harsh field environments. In practice, environmental interference and sensor faults jointly lead to imperfect sensing, where multichannel vibration signals are corrupted, partially missing, or even invalid on certain channels. To bridge the gap, this paper first proposes an end-to-end Robust RSM Diagnosis Network (RRDNet) for the RSM diagnosis task under imperfect sensing. First, a Single Channel Signal Transformer (SCST) tokenizes each vibration sequence into patches and learns long-range temporal representations while confining disturbances within each channel. Second, a Dynamic Graph ATtention (DyGAT) fusion layer models channels as nodes, constructs a sample-specific topology, and learns to down-weight perturbed channels while enhancing informative ones to obtain a fused descriptor for the final classification. Experiments on a multichannel vibration dataset collected from three triaxial accelerometers, with 16 RSM states and eight imperfect sensing modes, show that RRDNet achieves a Macro-F1 of 94.56% with balanced Macro-Prec and Macro-Rec, and consistently outperforms other models under a unified protocol. Ablation and comparative studies further confirm that the isolate-then-fuse design of SCST and DyGAT is critical for accurate RSM diagnosis under imperfect sensing.