Effective transfer learning of defect detection for photovoltaic module cells in electroluminescence images.

Due to the damage during production, transportation and installation, some defects inevitably occur in the solar cells, which will reduce the power generation efficiency. Benefiting from the development of deep learning, the performance of solar cell defect detection has been improved by a considerable margin. However, a problem persists that a model trained on one batch of data may not work on another batch following different production schemes. Moreover, compared with monocrystalline cells, more impurities on the surface of polycrystalline cells make it more difficult to inspect defects and obtain labels. To address this problem, we formulate it as a monocrystalline to polycrystalline task based on an unsupervised domain adaptation task in transfer learning. A standard network with CNN architecture is adopted as our classifier, which can tell whether the solar cell is cracked. Based on that, we embed adversarial learning in the network, where the domain discriminator is added to distinguish data domains, so that the feature extractor can learn domain-invariant features. Considering the scenario of defect detection in Electroluminescence (EL) images, we further propose attention-based transfer learning and class-aware domain discriminator to enhance the effect of knowledge transfer. When detecting a batch of new polycrystalline products, our method can improve F1-score by 0.2631, and the recall and precision can reach 84.70% and 90.15%, respectively. • Using transfer learning to adapt defect detection model to various production lines. • Attention mechanism is introduced in knowledge transfer to optimize features. • Align distribution of channel attention to reduce feature distribution gap. • Proposing class-aware domain classification to combat the impact of class imbalance. [ABSTRACT FROM AUTHOR]