Detection of Hidden Faults in Electric Power Facilities Combining SAM and U-Net

Abstract. Fault detection in electric power facilities is a crucial component of power grid maintenance, with hidden faults posing greater challenges compared to overt faults. Notably, hidden faults often coincide with localized heating, making infrared imaging an effective detection modality. However, automatic identification of power equipment in infrared images remains challenging; traditional methods are often inefficient and lack accuracy, while deep learning approaches are hindered by limited sample availability and accuracy issues. Furthermore, the temperature-based criteria for diagnosing hidden faults lack robustness. To address these challenges, this study proposes a comprehensive approach: first, employing the Segment Anything Model (SAM) for rapid annotation of power facilities in infrared images; second, leveraging these annotations to iteratively optimize a U-Net model for automated power equipment identification; and third, integrating temperature information to identify abnormal regions using dynamic threshold segmentation, thereby locating potential fault components. Experimental validation was conducted on a transmission line in Jiaxing City, Zhejiang Province, demonstrating a detection success rate exceeding 90%. The results indicate high detection accuracy and efficiency, presenting a promising solution for intelligent inspection of electric power infrastructure.