Electroluminescence image-based multi-class defect classification of solar photovoltaic cells using a lightweight CNN: An analysis of input image resolution

Abstract

Due to the rapid growth in large-scale solar farm photovoltaic (PV) installations, existing asset management activities, such as maintenance mechanisms, have become increasingly demanding and inefficient. Developing an automated system to monitor PV cell conditions is now essential for significantly reducing this maintenance burden. Inspired by this, we proposed a lightweight (simple and computationally efficient) convolutional neural network (CNN) to extract complex features and classify defects from electroluminescence (EL) images of solar PV cells. Based on the prediction, this multi-classification model categorizes EL cell images into three main classes: good, cracked, and corroded. To address dataset scarcity and class imbalance, a hybrid augmentation strategy combining geometric transformations and Gaussian noise injection was employed to enhance data diversity and improve model robustness. However, to investigate the influence of input image resolution on defect classification performance, a sensitivity analysis was conducted to determine the optimal image resolution. The EL images were resized to five spatial resolutions ranging from 50 × 50 to 250 × 250 pixels, and the CNN configuration was systematically evaluated using Adam and RMSprop optimization algorithms. The results demonstrate that image resolution plays a critical role in defect classification, with a small number of misclassifications observed at resolutions of 50 × 50 and 100 × 100 pixels. In contrast, all evaluation metrics and confusion matrices achieved perfect classification performance at resolutions of 150 × 150 pixels and above, indicating that this resolution is sufficient to preserve the defect-related features required for reliable EL image classification. The proposed lightweight CNN architecture was quantitatively compared with ResNet18 and VGG16 on the same dataset and training conditions, demonstrating an effective balance between predictive accuracy and computational efficiency. This balance enables efficient learning and robust generalization without the challenges associated with higher resolutions, making the proposed model a practical and scalable solution for real-world photovoltaic defect classification.