Photovoltaic panel failure detection using class-conditioned generative adversarial networks

Detection of defects and failure in photovoltaic (PV) modules increasingly relies on multi-modal data. Artificial intelligencemodels trained on imaging data can help identifying PV defects and failures. However, the effectiveness of these imaging-baseddiagnostic models is frequently limited by severe class imbalance in real-world datasets, where safety-critical fault categoriesare significantly underrepresented. This data scarcity leads to unreliable classification performance and poor probabilisticcalibration. This study proposes an Auxiliary Classifier Generative Adversarial Network (AC-GAN) framework tailored forthe class-conditioned synthesis of high-fidelity thermal images of PV defects. By integrating self-attention mechanisms andspectral normalization, the model achieves stable training and preserves the subtle structural signatures characteristic ofdiverse thermal anomalies. Evaluated on the Infrared Solar Modules dataset, the proposed model achieves a FID of 35.4and outperforms class-weighted and oversampling baselines across multiple downstream metrics. While augmentation yieldsmarginally higher classification accuracy, the AC-GAN framework achieves substantially better probabilistic calibration, withan Expected Calibration Error (ECE) of 1.5%, underscoring its advantage in uncertainty-aware fault diagnosis. These resultsconfirm that fault-conditioned synthetic image generation provides high-quality, functionally informative training samples thateffectively address minority class under-representation in PV inspection datasets. This methodology offers a scalable solutionfor enhancing the reliability of autonomous solar operations and maintenance (O&M), ensuring that diagnostic models remainrobust even when field-acquired data for rare failure modes is unavailable.