Parameter identification of rail-to-ground transition resistance based on hybrid network of improved BiLSTM-Transformer in urban rail transit

Abstract Due to long-term train operation, tunnel humidity, and the accumulation of metallic dust, the rail-to-ground transition resistance in urban rail transit (URT) may vary, leading to severe stray current. To achieve accurate identification of rail-to-ground transition resistance, this paper proposes an improved BiLSTM-Transformer model. A simulation platform for the multi-train dynamic operation with insulation deterioration is established to obtain rail potential data under different positions and rail-to-ground transition resistance. In the feature extraction stage, a denoising autoencoder is incorporated to enhance the robustness of feature representation, and a multi-task learning framework is adopted to jointly optimize regression prediction and domain adaptation. Experimental results demonstrate that the proposed method achieves high accuracy and strong generalization ability in rail-to-ground transition resistance identification, providing an effective tool for safety evaluation in URT.