Comparative Analysis of Deep Learning Models for Photovoltaic Defect Classification Using Drone Imagery

Abstract

The effective production of solar power energy in terms of efficiency will take place due to the quick detection of the defects in the photovoltaic cells. Manual techniques used to detect defects in the photovoltaic panels are very costly and time-consuming. This paper therefore mainly refers to the method of designing an autonomous system that uses drones equipped with deep learning tools for detecting defects in photovoltaic panels. The focus of this research is to build an algorithm with six classes (clean; dusty; bird dropping; snow covered; electrical damage; and physical damage on this kind of image taken by UAVs. To avoid any leakage of information, the dataset, comprising 1575 pictures, was partitioned into three segments of 70%, 15%, and 15% before the data augmentation phase. The models were evaluated using the pipeline technique. We carried out testing of the following deep learning algorithms: VGG19, EfficientNet (B1, B6, B7), InceptionResNetV2, MobileNet (V3Large, V4), CNN, and ResNet (50, 101). From the results of the experiments, the best model among these models is EfficientNetB7, which outperformed all other models, achieving an accuracy of 96.21%, an F1-score of 96.21%, and a recall of 95.42%. The most efficient algorithm concerning accuracy and F1-score proved to be the EfficientNetB7, followed by the ResNet101 and ResNet50. As such, this technology ensures that there is no need for human involvement in the process, which makes it easy to conduct inspections several times without experiencing any trouble.