Leveraging ResNet50 Architecture for Solar Panel Defect Classification

Abstract

The rapid growth of photovoltaic (PV) energy systems has increased the need for efficient, automated methods to monitor solar panel conditions and identify operational defects. Various aspects, including dust accumulation, bird droppings, snow coverage, and structural failures, can negatively affect power generation efficiency and reduce the operational lifespan of PV modules. As a result, intelligent image-based diagnostic systems have gained considerable attention as an alternative to conventional manual inspection approaches. This study presented a deep learning frame-work for the classification of solar panel conditions using the ResNet50 architecture. The proposed approach categorized solar panel images into three major groups: Normal Condition (NC), Surface Coverage Conditions (SCC), and Structural Damage (SD). The dataset images were subjected to several preprocessing operations prior to training, including image resizing and augmentation techniques to improve data variability and enhance model generalization capability. To address the imbalance among dataset categories, class-weighting techniques were incorporated during the training stage to ensure balanced learning across all classes. The model was trained and evaluated by means of commonly adopted performance indicators. Experimental findings demonstrated satisfactory classification capability, achieving an accuracy of 88.42%, precision of 88.58%, recall of 88.23%, and F1 _ score of 88.32%. These results reinforce the effectiveness of the proposed framework in distinguishing between normal and defective PV panel conditions, ensuring a practical and reliable approach for smart PV monitoring, fault diagnosis, and predictive maintenance systems.