Critical Component Overheating Monitoring Algorithm Integrating Temperature-sensing Patches and Computer Vision Trains

［Objective］ The present temperature sensor-based monitoring solutions for critical train component temperature entail high investment and operation-maintenance costs, hindering the widespread adoption of such systems in large-scale engineering projects. Consequently, a low-cost and low-maintenance temperature monitoring solution is urgently needed. ［Method］ A novel overheating monitoring algorithm integrating temperature-sensing patches and computer vision is proposed, which adheres to a ′locate, segment, and calculate′ identification logic. Precise localization of temperature-sensing patches in images is achieved by optimizing the YOLOV3 algorithm with a binary k-means clustering algorithm and an attention mechanism. A main-boundary separation module is embedded in the U-Net++ network architecture, and corresponding boundary supervision terms are added to the loss function, so that boundary segmentation performance could be enhanced and segmentation precision of temperature-sensing patches in the image be improved. The segmented images are calculated and overheating results are determined by analyzing the relative volumetric in the color changes of temperature-sensing patches. ［Result & Conclusion］ Comparative experiments are conducted on the positioning accuracy of five algorithms: SSD, Retina-Net, YOLOV3, YOLV4, and the improved YOLOV3, yielding their accuracies of 95.32%, 97.15%, 98.09%, 98.36%, and 99.21%, respectively, with the improved YOLOV3 approaching nearly 100%. For segmentation accuracy, comparative experiments are conducted among algorithms of DeepLabV3+, U-Net++, and the improved U-Net++, yielding their accuracies of 95.97%, 96.81%, and 98.36%, respectively, with the improved U-Net++ performing the best. In the test on a real dataset, the improved algorithm achieves an accuracy rate of 99.30%.