Automatic Classification of Defects in Solar Photovoltaic Panels Using UV-Fluorescence: A Deep Learning Approach

作者：Sina Lotfian, Andrew Ballen, Dylan J. Colvin, Nadia Khan, B. J. Thompson, Kristopher O. Davis, Mengjie Li · 年份：2024 · DOI：10.1109/pvsc57443.2024.10749318 · 被引用次数：1 · 研究领域：Industrial Vision Systems and Defect Detection、Photovoltaic System Optimization Techniques

UV-Fluorescence (UVF) imaging has become increasingly popular as a non-contact, non-destructive inspection tool in recent years due to its high throughput capability. However, the constraints of human speed in visually processing images present a challenge in fully harnessing the technology's potential. Over the past decade, deep learning techniques have demonstrated human-level accuracy in various tasks. In this paper, we present a deep learning solution capable of accurately classifying defects in photovoltaic modules with high speed. The proposed pipeline consist of multiple stages. The input to the pipeline is a photo containing several modules. Each module is subsequently cropped and perspective-corrected to form rectangles in the next stage. Next, the module image is segmented into cells. Finally, the images of individual cells are inputted into a deep neural network classifier. Our leading model achieves an F1 score of 0.93 while processing an average of 240 images per second. Additionally, we provide a dataset of cell images consisting of 1447 labeled and 53264 unlabeled images, which can aid researchers in the field further.