Detection Of Surface Defects On Solar Cells Based On Active Learning Strategy

As a clean and pollution-free renewable energy source,the scale of solar power grid connection is increasing day by day.The crystal surface of solar panels is fragile and prone to various defects during production and use,leading to a decrease in power generation efficiency.Timely and accurate detection of these defects and proactive repair and replacement of panels with surface defects are crucial for maintaining the stable power generation efficiency of solar cell photovoltaic modules.In recent years,with the rapid development of deep learning technology,the surface defect detection technology of solar cells based on deep learning has gradually replaced traditional detection techniques.Deep learning based surface defect detection methods for solar cells often require obtaining a large number of datasets,and require a significant amount of time and labor costs to annotate the datasets for network training.In addition,the application environment for surface defect detection of solar cells has undergone significant changes.Traditional surface defect detection methods often have low generalization ability and severely degraded detection performance in detection environments with inter domain differences.This paper focuses on the idea of active learning and aims to train the designed detection network model with only a small number of labeled datasets,so as to obtain more accurate detection results of solar cell surface defects and strong generalization ability.The main research of the paper includes the following three aspects:(1)In order to reduce the cost of data annotation while ensuring the detection accuracy,this paper proposes a active learning method based on combined uncertainty measure(CUM)for solar cell surface defect detection.The proposed method redefines the uncertainty measurement scheme,fully considering the value of samples for model training from both classification and regression perspectives.By selecting higher quality training samples,When 25% of the training set is used,the test results are respectively 90.70%and 86.67 mAP values on Retina Net and SSD baselines.The experimental results show that the active learning method based on combined uncertainty measurement outperforms the advanced methods in other stages on the solar cell surface defect datasets.Our method is 4.72% higher in m AP on Retina Net baseline than advanced MI-AOD,and 1.97% higher in m AP on SSD baseline than advanced MDN.(2)The complex industrial environment can result in images containing a significant amount of background noise,and images collected in different application environments have significant inter domain differences.In order to reduce the labeling cost and deal with the detection difficulties caused by inter domain differences,this paper proposes a solar cell surface defect detection method based on active-meta learning,which uses the strategy of active learning to select more valuable samples for labeling,and introduces the meta learning idea of few-shot objection detection to reweight the model,so as to improve the generalization ability of the model to deal with different objection domains.The experimental results show that 87.0% and 84.8% m AP values were obtained on the night domain and exposure domain datasets,respectively.Not only the detection accuracy is guaranteed and the cost of data annotation is reduced,but also the generalization ability of the model is guaranteed.(3)Considering that there is currently no shared datasets available for detecting surface defects in solar cells,this paper collaborates with solar power generation companies to construct a solar surface defect datasets for network training.The datasets consists of over 4000 images containing surface defects in solar cells.In addition,through datasets augmentation technology,this article also constructed two datasets with inter domain differences,namely the nighttime domain and the exposure domain,for the training and testing of active meta learning methods.