Hybrid deep learning approaches for photovoltaic fault detection: from supervised CNNs to self-supervised masked autoencoders

Photovoltaic (PV) fault diagnosis is challenged by severe class imbalance and the limited availability of labeled fault data. However, existing studies typically evaluate individual learning paradigms under different experimental protocols, making direct comparison difficult. This study presents a unified benchmark of four one-dimensional convolutional neural network-based strategies: supervised learning, semi-supervised pseudo-labeling, iterative self-training, and masked-autoencoder-based self-supervised pretraining. Each strategy was evaluated under three imbalance settings: no explicit mitigation, random oversampling, and class weighting. Experiments were conducted on the Lazzaretti PV fault dataset using stratified five-fold cross-validation. Performance was assessed using accuracy, macro-F1, ROC-AUC, class-wise metrics, and paired statistical tests on fold-level ROC-AUC values with Bonferroni correction. Without explicit imbalance mitigation, the supervised CNN and self-supervised model achieved near-perfect performance, with accuracies of 0.9949 and 0.9973 and macro-F1 scores of 0.9912 and 0.9942, respectively. Random oversampling substantially degraded minority-class discrimination across all four strategies. Under class weighting, CNN and self-supervised learning retained strong performance, whereas semi-supervised learning and self-training remained sensitive to biased or noisy pseudo-labels. The self-supervised model required substantially greater training time and memory than the supervised CNN but maintained low inference latency. These findings indicate that supervised CNN and masked-autoencoder-based self-supervised learning are the most reliable strategies under the evaluated protocol, whereas naïve pseudo-labeling and random oversampling are unsuitable for severe class imbalance. External validation on independent PV systems is required to assess generalizability.