Answers are grounded in this paper and cite their sources.
The field reliability of solar cells is of critical importance in renewable energy systems. A key task in field reliability studies is the automatic detection of defects in photovoltaic (PV) modules, as such defects are often associated with degradation and failure-related conditions of solar cells. Accurate defect detection is essential for reliability assessment, but its performance is frequently hindered by the severe class imbalance inherent in electroluminescence (EL) image datasets. Standard deep learning models tend to be biased toward majority classes, resulting in poor detection of minority defect classes. To address this challenge, this paper proposes a robust imbalance-aware learning framework that integrates a class-weighted loss formulation with a stratified mini-batch sampling strategy to improve minority-class prediction. The proposed framework incorporates both classlevel weighting and sample-level adjustments during training to mitigate bias toward majority classes while preserving overall classification performance. The framework is evaluated using an imbalanced dataset derived from Fashion-MNIST, a real EL image dataset, and an additional credit card default dataset. To assess its performance, experiments are conducted across multiple train-test splits, and results are evaluated using per-class metrics and variability measures. The results indicate that the proposed framework can improve minority-class sensitivity and reduce performance variability compared with standard training approaches, while maintaining competitive performance for majority classes. These findings suggest improved defect identification and provide a useful tool for reliability assessment and failure analysis of PV systems.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: