Fault prediction methods of straddle-type monorail pantographs based on multi-domain features and deep learning

Abstract The straddle-type monorail system critically depends on the safe operation of pantographs, which are vulnerable to unstable contact, arc discharges, and structural impacts under complex conditions. To overcome the inefficiency of conventional maintenance and the limitations of traditional methods in handling non-stationary signals and small-sample scenarios, this study proposes a fault prediction method based on multi-domain feature extraction and deep learning. Contact force, stress, and acceleration signals were collected to build a multi-condition dataset. Features from time, frequency, and time-frequency domains were fused and reduced to enhance discriminability. An Attention-CNN-LSTM model was designed to capture critical patterns. Experiments show that the proposed method achieves 95.8% accuracy and an AUC of 0.98, demonstrating strong robustness and interpretability. The results confirm the effectiveness of the method in enabling high-precision pantograph fault prediction under limited data, providing solid technical support for predictive maintenance in straddle-type monorail systems.