Objective Traditional wheel slip detection methods based on acceleration thresholds suffer from limitations in threshold setting and poor adaptability to different lines, resulting in a relatively high false alarm rate and difficulty in meeting detection requirements under complex line environments. To improve detection accuracy, it is necessary to study a new train wheel slip detection method.

Method A wheel slip detection model composed of an AE （autoencoder） based on LSTM （long short-term memory） integrated with a temporal weighting layer is proposed. Deep temporal features of operation data are extracted using a two-layer cascaded LSTM, and the temporal weighting layer enhances the model's perception capability of slip-induced fluctuations. In addition, a slip section detection algorithm integrating sliding window statistical features with temporal continuity validation is designed. The algorithm builds a discrimination criterion based on reconstruction errors and utilizes a decision function to capture the start and end boundaries of slip events, thereby achieving precise localization of slip sections. Experimental data originate from real CBTC （communications-based train control） operation logs. The model performance is evaluated using four evaluation metrics of accuracy, precision, recall, and F₁ score. Comparative experiments and ablation studies are designed to verify the effectiveness of the method.

Result & Conclusion The improved LSTM-AE model achieves an accuracy of 99.71% in identifying specific wheel slip segments, with both precision and recall exceeding 99%. Compared with traditional acceleration threshold methods and ablation models, the proposed model can effectively identify wheel slip sections, improving the safety and reliability of rail transit systems.