Development of a CNN-based hierarchical inspection system for detecting defects on electroluminescence images of single-crystal silicon photovoltaic modules

Abstract

Electroluminescence (EL) image technology has long been a standard tool for detecting defects in the photovoltaic (PV) industry. In recent years, deep learning technology has been developed rapidly and applied to in-line inspection of EL images of PV modules, but not without challenges. In addition to imbalanced data, the recall rate and inspection speed for detecting defects of PV modules are unable to meet industrial expectations. Therefore, this study proposes a hierarchical inspection system to address these challenges. To alleviate the problem of imbalanced data, an EL image of sc-Si solar modules is cropped into images of single solar cells. This allows us to reduce the processing size of a large EL image without overlooking small defects and to increase the amount of data samples for learning fine-to-coarse defect features. To speed up detection, we design two light-weighted convolutional neural network (CNN) models according to feature map analysis. Experimental results show the precision and recall rate of our CNN models on the testing dataset reach 99.36% and 98.77% respectively, which are confirmed by t-distributed stochastic neighbor embedding visualization. With the same detection discriminability, our model needs only 49.63% of the detection time of ResNet-50. After alleviating the problem of imbalanced data, the performance of this hierarchical inspection system meets the requirement for the in-line inspection of the industry.