LiCPV: hybrid deep learning for PV fault detection

Solar photovoltaic (PV) systems are becoming more and more important for producing sustainable energy, but cell-level flaws—which are frequently detected by electroluminescence (EL) imaging—have a major impact on their dependability. Although deep learning models for PV defect classification have been used in earlier research, the majority of these studies rely on single architectures, small datasets, and traditional training techniques, which limits detection accuracy and generalization. This work introduces a novel hybrid learning framework that systematically combines transformer-based and convolution-based architectures—PiT-base and ConvNeXt—to exploit their complementary feature extraction capabilities for PV EL defect classification. Unlike prior approaches, the proposed method is further distinguished by the first comprehensive evaluation of the Lion optimizer in this application domain, demonstrating its superior stability and convergence behavior compared to traditional optimizers. The framework is evaluated under both offline and online learning scenarios, providing new insights into training and deployment performance that are largely unexplored in existing literature. Extensive experiments are conducted on two benchmark datasets—ELPV and PVELAD—thereby extending validation beyond the commonly used ELPV dataset. The proposed approach achieves accuracies of 82.45% (offline) and 81.94% (online) on ELPV, and 86.96% (offline) and 86.74% (online) on PVELAD, significantly outperforming previously reported state-of-the-art results.