To improve the efficiency and accuracy of fault identification for pantographs in straddle-type monorail transit systems, this paper proposes an intelligent diagnosis method that integrates structural priors with a self-supervised learning mechanism, addressing the challenges of signal non-stationarity, multimodality, structural coupling, and label scarcity. First, a synchronized multi-source dataset is constructed based on field-tested pantograph data from urban mainline operations, encompassing contact force, stress, and vibration signals, while representative fault types are defined via physical simulation. Then, a structure-driven feature encoding approach is proposed, combining wavelet packet decomposition and variational mode decomposition to achieve multi-scale modeling of non-stationary signals. A hybrid CNN-Transformer network is further designed to capture both local dynamics and global dependencies, and a modality attention mechanism is introduced to facilitate multi-signal fusion. To mitigate label deficiency, a self-supervised contrastive learning strategy based on modality perturbation and temporal augmentation is developed to enhance feature discriminability. Experimental results demonstrate that the proposed method achieves superior diagnostic performance across various fault scenarios and maintains high accuracy under limited data conditions, validating its effectiveness and robustness. This method exhibits strong potential for engineering applications and can be applied to intelligent sensing and fault diagnosis of key structural components in rail transit systems.
Sun et al. (Wed,) studied this question.