Advancements in AI-Driven detection and localisation of solar panel defects

Abstract

Renewable energy production has experienced rapid growth over the past three decades and is projected to triple its global capacity by 2030. Given that the utilisation of solar photovoltaic (PV) technology plays a vital role in generating renewable electricity, it is crucial to continuously monitor the condition of solar panels because a variety of defects can significantly reduce their power production. In this paper, we review the latest artificial intelligence (AI) algorithms developed for inspecting solar panels. We also discuss various low-resource hardware systems used to execute these algorithms. AI algorithms are trained using datasets and images, including optical, infrared, and electroluminescence images of solar panels. These images can be captured by unmanned aerial vehicles (UAVs), ground vehicles, and fixed cameras. In this paper, we compare the precision, accuracy, and recall rates of a selection of reviewed AI algorithms. To gain a deeper understanding of these AI algorithms, we introduce a generic framework of AI-driven systems that can autonomously detect and localise solar panel defects and we analyse the literature based on this framework. Some of the main AI and image processing algorithms reviewed are YOLO V5 BDL, weight imprinting, custom-designed CNN, modified edge detection, fuzzy-based edge detection, and the modified Canny algorithm. We also discuss the main hardware systems used to execute image processing algorithms to localise and detect defects in solar panels: the central processing unit (CPU), field programmable gate array (FPGA), and graphics processing unit (GPU). Finally, as a future direction, we suggest developing image processing algorithms specifically designed for hardware systems tailored for machine learning, such as tensor processing units (TPUs). This development would further enhance the capabilities of solar panel inspection and defect detection.