Condition Monitoring and Fault Diagnosis Model of Pole Mounted Circuit Breaker Based on Multi‐Source Heterogeneous Cross Modal Information Fusion

ABSTRACT Aiming at the defects of current diagnosis methods for Primary and Secondary Integrated Pole Mounted Circuit Breaker in distribution networks, this paper proposes a fault diagnosis model based on multi‐source heterogeneous cross‐modal information fusion. First, Dynamic Time Warping (DTW) realizes time‐axis alignment of multi‐sensor data, combined with sliding window standardization to adapt to data non‐stationarity and eliminate interference from operating condition fluctuations. Second, a Spatio‐Temporal Graph Convolutional Network‐Transformer (ST‐GCN‐Transformer) with dynamic topology awareness is proposed, which dynamically captures spatio‐temporal correlation features of current and vibration signals via a load‐sensitive adjacency matrix. Cooperated with the two‐stream Transformer's gated cross‐attention, it achieves cross‐modal semantic alignment between electrical time‐series data and infrared image spatial data. Third, federated feature distillation and differential privacy protection mechanisms are introduced to ensure data privacy, realize cross‐regional knowledge transfer, and solve the inefficient transfer problem of parameter averaging in federated learning. Finally, a Deep Random Forest (DRF) with hierarchical feature learning and dynamic weight optimization is constructed to break the classification bottleneck of single‐layer forests. Experimental results show the model's Macro‐F1 values reach 0.9372 and 0.9358 on two datasets, with fault alarm accuracy of 93.88% and 94.25%, and false alarm rates of only 3.12% and 2.95%, demonstrating superior performance.