Image registration based fault localization in panoramas of mountain-mounted PV plants. (15th May 2023)

• Record Type: Journal Article
• Authors: Ying, Yuxiang; Ying, Pengfei; Men, Hong; Joo, Young Hoon
• Is Part Of: Solar energy. Volume 256(2023)
• Page Start: 16
• Page End: 31
• Publication Date: 2023-05-15
• Subjects: Mountain-mounted photovoltaic -- Fault localization -- Image registration -- YOLOv5s -- AKAZE; Solar energy -- Periodicals; Solar engines -- Periodicals; 621.47
• DOI: 10.1016/j.solener.2023.03.049
• Languages: English
• ISSNs: 0038-092X

Abstract

Abstract: Large-scale photovoltaic (PV) plants have been widely distributed in the mountains. However, the complex geographical environment and the irregular arrangement of PV strings bring great challenges to maintenance. In this paper, a novel scheme combining an improved YOLOv5s model and image registration algorithms is proposed to detect and locate module faults in mountain-mounted PV plants (MPVs) via aerial dual-light videos and panoramas. The improved YOLOv5s used to detect faults in IR images becomes more lightweight owing to the integration of MobileNetV3 and ECA-Net. To realize fault localization in panoramas of MPVs, image registration algorithms are adopted. IR and visual images are simultaneously extracted frame by frame from dual-light videos. Faults detected in the IR image are first mapped to the corresponding visual image by the homography ( H 1 ) matrix. To map faults from the visual image to the panorama, another registration between the visual image and panorama is achieved. Multi-scale matching between the visual image and the panorama is first performed to reduce the area of registration. The conversion of grayscale probability density function (GPDF) of the visual image and the adoption of AKAZE feature detection algorithm make the homography ( H 2 ) matrix used for registration more accurate. A MPV with a power generation of 212.85 kW is used for testing. The precision and recall of fault detection are 88.31% and 91.67%, respectively. The accuracy of locating faults to the corresponding modules is 85.3%, and the recall of locating is 94.12%. Experimental results show that the proposed scheme can effectively identify faults and accurately locate faults in MPVs. Highlights: A scheme that locates faults to corresponding modules in panoramas is proposed. A more lightweight YOLOv5s model is applied to detect faults in the entire IR image. Registration between high-resolution images becomes efficient and accurate. A new criterion is proposed to evaluate image matching. Registration between IR and visual images is achieved by manual calibration points.