PERFORMANCE EVALUATION OF PRETRAINED CNN ARCHITECTURES FOR PHOTOVOLTAIC MODULE DEFECT CLASSIFICATION FROM IR AND EL IMAGES

The rapid deployment of photovoltaic (PV) systems has increased the demand for accurate, scalable, and automated defect detection, as manual interpretation of infrared (IR) and electroluminescence (EL) images is costly, subjective, and unsuitable for large-scale solar farms. Although deep learning has shown strong potential for PV inspection, many existing studies focus on single models or limited comparisons under inconsistent experimental settings, leaving a gap in standardized and fair benchmarking of modern pretrained convolutional neural networks. This study presents a comprehensive evaluation of six state-of-the-art pretrained CNN architectures, ResNet18, ResNet34, ResNet50, EfficientNet-B0, DenseNet121, and MobileNetV3-Large, for multi-class PV defect classification using IR and EL images. A curated dataset of 7,666 annotated PV images covering six defect categories (clean, dust, bird-related, electrical faults, cracks, and hotspots) was used. Images were resized to 224×224, augmented, normalized, and split into 80% training and 20% testing sets. All models were adapted for single-channel inputs and integrated with a unified classification. Training was performed using a weighted cross-entropy loss, the Adam optimizer, learning rate scheduling, and early stopping. Performance was assessed using accuracy, precision, recall, F1-score, and confusion matrices. Results show that EfficientNet-B0 achieved the highest accuracy (89.62%) and macro F1-score (0.8511), followed closely by DenseNet121 (89.25%), outperforming ResNet and MobileNet variants (86.89–87.61%). Although strong performance was observed, reliance on a single dataset limits the generalizability of the results. The findings identify EfficientNet-B0 as a strong candidate for real-time, large-scale PV monitoring and provide a standardized benchmarking framework to guide model selection in automated PV inspection systems.