Health-Aware Joint Learning of Scale Distribution and Compact Representation for Unsupervised Anomaly Detection in Photovoltaic Systems

Accurate and efficient anomaly detection is crucial to ensure the reliability and optimal performance of photovoltaic (PV) systems. However, traditional PV anomaly detection methods struggle to fully capture the interdependencies among health-related features in electroluminescence (EL) images and show limited robustness against contaminated training data. To address these limitations, an unsupervised joint learning-based health-aware model is proposed, termed improved scale learning anomaly detection (ISLAD). The scale learning mechanism provides supervisory signals to effectively model the intrinsic data patterns and structural features among EL image during alignment training. Subsequently, compact representation learning is integrated with aggregated scale distributions to enhance the model’s discriminative ability for anomalous states, enabling accurate detection of subtle anomalies and improved robustness. The experimental results on monocrystalline and polycrystalline EL image datasets validate the effectiveness of ISLAD. Compared to several representative anomaly detection methods, ISLAD achieves an average $F1$ -score improvement of 10.32%, with a further gain of 16.09% under data contamination scenarios.