Studies on Techniques to Improve YOLO in Fault Detection Using RGB Images of Solar Panels

Solar panels can have short lifespan and significantly lower efficiency with the occurrence of various faults. Currently, most fault detection methods suffer from cost-ineffectiveness due to either requiring expensive equipment, such as a thermal camera or lack of information on the source of the faults. Most of the research of using You Only Look Once (YOLO) to detect faults in solar panels revolves around thermal images too. This paper aims to propose an alternative method to detect faults (primarily cracks and soiling) in solar panels by applying YOLOv7 on Red, Green, Blue (RGB) images captured by an Unmanned Aerial Vehicle (UAV). To further improve the performance of this scheme, this paper will focus on applying various techniques to improve the accuracy and training time of the model, which include gray scale conversion, data augmentation and change of Gradient Descent (GD) optimizer. Based on preliminary results, it was found that the combination of gray scale conversion and data augmentation leads to an accuracy of 95.91% while Nadam stands out as the best optimizer because it reduces required training time for convergence by 75 epochs.