DINO-Proto: DINO-Augmented Prototypical Network for Few-Shot Photovoltaic Fault Classification

Photovoltaic (PV) anomaly detection plays a critical role in ensuring the reliability and efficiency of solar energy systems. However, existing diagnostic approaches typically rely on large amounts of annotated thermal imagery, which is expensive and labor-intensive in real-world inspection scenarios. Although few-shot learning provides a promising paradigm for addressing data scarcity by enabling effective recognition from limited labeled samples, its application to PV thermal inspection remains underexplored, particularly due to the complex and subtle textural variations present in infrared imagery. To address these challenges, we propose DINO-Proto, a novel fewshot learning framework built upon a Prototypical Network architecture with a DINO (Self-Distillation with No Labels)inspired data augmentation strategy. Within the proposed DINOProto framework, a lightweight convolutional feature encoder is trained using diverse augmented data that reflect environmental conditions commonly encountered in real-world photovoltaic (PV) inspections, thereby improving class separability under limited supervision. We evaluate DINO-Proto on the Infrared Solar Modules dataset with 11 PV fault categories under various few-shot settings. Extensive experimental results demonstrate that the proposed approach consistently outperforms state-of-theart few-shot methods, achieving superior classification accuracy as a result of more compact intra-class distributions under datalimited conditions.