AlphaTensor-Enhanced Deep Learning Framework for Accurate Photovoltaic Defect Detection and Classification

Abstract

The increasing deployment of photovoltaic (PV) systems has created a growing demand for intelligent and reliable defect detection methods capable of supporting efficient maintenance and maximizing energy production. However, existing deep learning approaches often suffer from high computational complexity, sensitivity to environmental variability, limited capability in detecting subtle defect patterns, and performance degradation caused by class imbalance. To address these challenges, this study proposes an Enhanced AlphaTensor-Optimized Convolutional Neural Network (EAO-CNN) for automated photovoltaic defect classification. The proposed framework combines AlphaTensor-based matrix optimization with depthwise separable convolutions, squeeze-and-excitation blocks, adaptive attention mechanisms, mixed-precision training, focal loss, and advanced data augmentation to improve computational efficiency, feature representation, and classification robustness. The framework was evaluated using a photovoltaic dataset containing annotated images representing six defect categories: dusty, electrically damaged, bird-dropped, clean, snow-covered, and physically damaged panels. Experimental results show an overall test accuracy of 74.73% and a Macro-F1 score of 0.7476, with distinct class-wise performance across the six defect categories. The Snow-covered category achieved the highest precision, recall, and F1-score of 93.75%, together with a ROC-AUC of 0.9958 and PR-AUC of 0.9851, indicating strong discriminative capability for this class under imbalanced data conditions. The proposed framework provides an efficient and robust solution for intelligent photovoltaic inspection by integrating AlphaTensor-based optimization with lightweight attention-enhanced deep learning, offering improved computational efficiency, reliable defect recognition, and enhanced suitability for real-world solar monitoring applications.