Photovoltaic (PV) Module Defect Image Classification Analysis Using EfficientNetV2 Architectures

This paper presents a study of Photovoltaic (PV) module defect image classification using EfficientNetV2, the newest convolutional neural network (CNN) architecture. Five classes of defects were considered in the study: Broken Glass, Cracked Solar Cell, Gridline Corrosion, Normal, and Yellowing. The images were pre-processed and augmented to increase the diversity of the data. Current methods often take a long time since they manually capture and analyze thousands of images to look for errors and patterns in the modules. Furthermore, there is a lack of research that implemented EfficientNetV2 architectures in classifying defect images in PV modules. Therefore, the objective of this study is to classify and analyze the performance of EfficientNetV2 with two models; EfficientNetV2-S and EfficientNetV2-L, and identify which model gives the highest performance. The best performance model of EfficientNetV2 will be compared with other CNN architecture namely VGG16. The finding demonstrates that EfficientNetV2-S gives the highest performance with achieved 93.69% of accuracy, precision (93.75%), recall (93.36%), and F1-score (93.69%) when compared to VGG16 architecture.