Enhancing defect detection in photovoltaic cells through advanced cascade CNN with attention mechanisms on electroluminescence images

Ensuring the reliability of photovoltaic (PV) power plants hinges on effective and robust quality inspection methods during the manufacturing of solar modules. This paper presents the development of an advanced cascade Convolutional Neural Network (CNN) model that integrates Squeeze-and-Excitation (SE) blocks, attention layers and prediction fusion techniques for defect classification on electroluminescence images. The integration of SE blocks and attention layers mechanisms allows the model to selectively emphasize significant image regions and informative channels, significantly improving the precision of feature extraction. To further enhance classification accuracy, a combined prediction approach involving a secondary model is employed for cases where the primary model produces uncertain or incorrect classifications. This fusion technique boosts the robustness of the overall system by accounting for potential errors. The proposed advanced cascade CNN model achieved an accuracy of 96 %, substantially outperforming simpler cascade CNN architectures and demonstrating its effectiveness in identifying a wide range of PV cell defects.