Research Information System
International Conference on Cognitive Computing, Intelligence and Data Science Applications, İstanbul, Turkey, 27 - 28 December 2024, pp.1-14, (Full Text)
In this study, we aimed to detect defects on solar panels using a specially designed 2D CNN architecture to classify images from the publicly available ELPV dataset. The images were organized into four classes based on the crystalline type and defect probability values specified in the dataset’s CSV file. An initial analysis showed class imbalance, which was addressed through oversampling to match the largest class, resulting in both balanced and unbalanced datasets for comparative analysis. Our 2D CNN model was trained for 30 epochs on both datasets, yielding F1 scores of 83.72% for the unbalanced data and 95.45% for the balanced data, indicating that data balancing significantly enhanced defect detection. To further demonstrate the model’s efficacy, accuracy-loss plots and confusion matrices were generated for both datasets. Additionally, heatmaps from the final CNN layer highlighted areas of focus, revealing that the model trained on balanced data targeted defect regions more accurately. These findings underscore the importance of data balancing in improving solar panel defect classification.