INTERPRETABLE MACHINE LEARNING FOR PHOTOVOLTAIC DEFECT SCREENING: A REPRODUCIBLE ELECTROLUMINESCENCE BENCHMARK

Electroluminescence (EL) imaging can reveal cracks, inactive regions, and other abnormalities in photovoltaic (PV) cells, but manual interpretation is labor intensive and depends on specialist judgment. In this study, I evaluate transparent image features as a reproducible screening baseline. I use the public ELPV dataset, which contains 2,624 normalized grayscale cell images from 44 modules with expert-assigned defect probabilities, and resize each image to 96 x 96 pixels. I compare three representations: 62 interpretable intensity and gradient features, 324 histogram-of-oriented-gradient features, and their 386-feature combination. Using stratified five-fold cross-validation, I assess class-balanced logistic models for two prespecified outcomes: any annotated defect and high-confidence defect. For any annotated defect, the combined representation achieves mean accuracy 0.721, F1 0.670, receiver-operating-characteristic area under the curve (ROC AUC) 0.778, and precision-recall area under the curve 0.751. For high-confidence defects, the interpretable representation produces the highest mean F1, 0.638, and ROC AUC, 0.816. On the primary outcome, out-of-fold F1 is 0.712 for monocrystalline and 0.639 for polycrystalline cells. I use standardized coefficients to connect predictions with observable intensity dispersion, gradient tails, and dark-pixel fractions. I present this benchmark as an auditable basis for triage and method comparison, not as autonomous disposition, electrical qualification, or field deployment. Module-grouped external validation and correlation with calibrated electrical measurements remain necessary.