基于深度学习的光伏组件红外、电致发光和红绿蓝图像自动缺陷检测

Abstract This study presents an automated defect detection system for photovoltaic modules that combines image processing techniques with deep learning models. The system identifies 21 types of defects using three imaging methods: infrared imaging, red–green–blue imaging, and electroluminescence imaging. Infrared imaging, captured using a thermal imager mounted on a drone, detects thermal anomalies such as hotspots and open circuits. red–green–blue imaging identifies surface-level defects, including glass breakage, soiling, and vegetation shading. Electroluminescence imaging, obtained with a charge-coupled device camera in a controlled darkroom environment, reveals internal defects such as microcracks, cell degradation, and busbar corrosion. Cross-referencing results from infrared and red–green–blue images facilitates the identification of defect causes, while electroluminescence imaging confirms internal issues and provides targeted recommendations for improvements. The system development was divided into four main stages. First, electroluminescence images were preprocessed and segmented using median filtering, thresholding, edge detection, and perspective correction techniques. Second, color features in infrared and red–green–blue images were enhanced through color space transformation to improve detection accuracy. Third, image datasets were augmented using rotation, flipping, and generative techniques to ensure model reliability. Finally, deep learning models were trained using transfer learning methods and optimized through cross-validation and hyperparameter tuning to achieve optimal performance for each imaging type. The proposed system demonstrates exceptional accuracy: 99.06% for thermal defect detection in infrared images, 100% for surface defect detection in red–green–blue images, and 99.2% for internal defect detection in electroluminescence images. The average processing time per image is less than 0.1 s, making the system suitable for real-time and large-scale photovoltaic module inspections.