Multi-scale self-supervised ordinal learning for severity grading of photovoltaic defects in electroluminescence images

Detection and severity assessment of defects in photovoltaic (PV) cells are significant for ensuring efficient energy harvesting at reduced cost. Electroluminescence (EL) imaging is significant among the techniques employed for PV defect detection. Defects appear as subtle patterns in low contrast EL images, making severity grading difficult. This difficulty is further affected by an imbalance across classes in input datasets. The present work addresses these research gaps by proposing a novel Multi-scale, Ordinal, Self-supervised Learning Network (MS-OSSLNet) for PV cell surface defect severity grading in EL images. The proposed framework extracts multi-scale features from global and local-patch-level information, uses DINO-based self-supervised representation learning and ordinal regression via the CORAL framework to model the progression of PV cell defect severity. Further, class imbalance in the dataset is addressed using a smoothed class-weighted loss function. The proposed framework, when tested on the ELPV dataset, demonstrated a test accuracy of 60.91% with a weighted F1-score of 0.6417 and a balanced accuracy of 53.27%. However, an Earth Mover’s Distance metric of 0.2234 reveals distribution similarity, while a mean absolute error of 0.6497 and a Quadratic Weighted Kappa of 0.5479 indicate that most misclassifications occurred between adjacent severity levels, thereby preserving ordinal consistency. Further, Spearman’s Rank and Kendall’s Tau values of 0.5169 and 0.4770 reveal the presence of moderate, positive monotonic ordinal association. The model demonstrated an average processing time of 11.28 ms with a throughput of 88 frames per second. Ablation analysis supports the influence of multi-scale feature extraction, while GradCAM interpretations revealed the model’s ability to capture features across spatial scales and ordinal consistency. Model robustness is supported by minimum standard deviation across performance metrics for various random initializations. The results obtained support the capability of integrating self-supervised representation learning techniques with multi-scale representation and imbalance-aware ordinal optimization for PV defect analysis. Inherent to its efficacy in grading defect severity, the proposed framework could be enhanced to enable an intelligent inspection system in renewable energy applications.