Research on Fault Diagnosis of Pantograph Based on Multimodal Data Fusion

The pantograph is a core component for high-speed trains to obtain electrical energy, and its failure will directly threaten the safety of train operation. Aiming at the problems of high missed detection rate and insufficient positioning accuracy of traditional single-modal diagnosis methods, this paper proposes a pantograph fault diagnosis method based on a Multimodal Temporal-Spatial Attention Fusion Network (MTSA-Net). By fusing three types of modal data, namely vibration signals, visual images, and maintenance text records, the accuracy of fault prediction is significantly improved. Firstly, the vibration signal is processed by the Ensemble Empirical Mode Decomposition (EEMD) with adaptive parameter optimization to extract multi-scale sample entropy features; meanwhile, a Convolutional Neural Network (CNN) is used to extract visual defect features; and Natural Language Processing (NLP) technology is employed to analyze the semantic information of maintenance texts. A cross-modal graph convolution module and a temporal attention mechanism are introduced to realize feature-level fusion, which effectively solves the problems of large distribution differences of multimodal data and weak cross-modal correlation. Experimental results show that the average accuracy of MTSA-Net in the diagnosis of common pantograph faults reaches 98.2%, which is 12.6 percentage points higher than the time-frequency entropy method based on a single vibration signal and 7.3 percentage points higher than the image-text bimodal fusion method. This provides a new idea for the intelligent maintenance of key rail transit equipment.