A Switching Cabinet Contact Temperature Anomaly Classification Warning Method Based on Dual-Stream Residual Network and YOLOv10 Enhancement

Abnormal contact temperature of high-voltage switchgear poses a threat to the safety of the power system, and traditional detection methods have limitations in accuracy and efficiency. This study proposes a dual-stream residual-YOLOv10 hierarchical warning method by integrating infrared thermal imaging and visible light image data. The method uses cross-attention mechanisms for spatiotemporal alignment and improves the C2f structure of YOLOv10, combining spatial refinement units and channel refinement units to enhance small target localization accuracy. A hybrid local channel attention module is introduced to improve feature capture. The method combines absolute temperature values and rate of change for comprehensive scoring to achieve hierarchical warnings, and optimizes the model through online learning and knowledge distillation. Experimental results show that in performance tests, the final training accuracy of the model approaches 98%, with loss values stabilizing around 0.3. Compared with other algorithms, the model performs well across various abnormal category recognition. Application analysis indicates that after applying this model, the average response time for abnormal detection has been reduced from over 15 minutes to 4-8 minutes, and the false alarm rate has decreased from 7%-8% to 2%-3%. In different scenarios, the contact detection accuracy exceeds 95%, and the false alarm rate is controlled at 1.23%-2.15%, with the three-level warning response time being less than 19ms. In summary, this method effectively addresses the challenge of hierarchical warning for abnormal contact temperatures in switchgear, offering high reliability and strong practicality, performing excellently across multiple scenarios, and demonstrating promising application prospects.