Research On Defect Detection Method Of Solar Cell Based On Domain Adaptation

New energy technology is gradually becoming a strategic development goal of various countries,and solar energy is one of the most important energy technologies.At present,the highest photoelectric conversion efficiency is monocrystalline silicon solar cells,which are prone to rupture,black spots,scratches,broken grids and other defects during the production process,which seriously affect the photoelectric conversion efficiency.Usually,the defect detection of monocrystalline silicon panels adopts the electroluminescent(EL)imaging principle.The traditional EL image based defect detection of monocrystalline silicon panels usually uses machine vision inspection technology,which cannot meet the requirements of high precision defect detection due to its large bottleneck of accuracy enhancement.Since the data-driven deep learning methods have greatly improved the accuracy and robustness of the model in recent years,this paper takes the EL image-based defect detection of monocrystalline silicon panels as the research object,uses the target detection algorithm in deep learning as the technical means,and incorporates the domain adaptive algorithm to reduce the labeling cost of defect data and achieve the goal of improving the detection accuracy and robustness of the model.The main research contents and results of this paper are as follows:(1)For the problem of defect detection accuracy of monocrystalline silicon cells based on EL images,a data enhancement method and feature extraction network structure based on defect features are proposed,and the loss function in target detection is optimized.The Poisson fusion-based Copy-Pastes and adversarial generative network-based Mosaic data enhancement algorithms are proposed for monocrystalline EL images,and the Inception-based feature fusion extraction network ELNet is proposed for slender defects such as broken grids and ruptures.The defect detection model based on Focal Loss and GIoU loss function training is proposed to address the category imbalance and border regression accuracy in the data.After a series of algorithm improvements,the detection accuracy reaches 91.4mAP.(2)A domain adaptation defect classification algorithm based on the consistency regularization method is proposed for the problem of small spacing of defect classes in EL images of monocrystalline silicon cells.In this paper,the sample distribution and feature extraction layer distribution are analyzed,and the sample distribution consistency regularization module and the feature consistency regularization module are designed,and the distribution distance measure loss function of K-L scatter and MMD distance are used respectively,which effectively reduces the distribution spacing between defects and supports end-to-end model training.By comparing the t-SNE dimensionality reduction visualization with algorithms such as the adversarial discriminative domain adaptation based on parameter sharing proposed in this paper,the domain adaptation algorithm based on the consistency regularization method improves the class spacing more significantly.(3)To address the problem of low robustness of defect detection in EL images of monocrystalline silicon cells,a pseudo-labeling domain adaptation learning-based defect detection algorithm for monocrystalline silicon cells is further proposed.The weakly supervised learning-based pseudo-labeling method reduces the cost of data labeling,incorporates a consistent regularization-based domain adaptation algorithm to close the distribution distance between the source and target domains,reduces the noise data brought by pseudo-labeling learning to reduce the detection accuracy of the model,effectively solves the problem of time-consuming and labor-intensive EL image labeling,and improves the defect detection accuracy of the model in the target domain.The method achieves 46.3mAP in domain migration of Cityscapes and Foggy Cityscapes public datasets,and 96.2mAP on the El image defects in this paper.(4)A FLASK-based defect detection platform for monocrystalline silicon panels was developed.In this paper,the defect detection model was quantified and compressed by FP16 model,and the TensorRT reasoning framework was deployed to the NVIDIA hardware platform,and the deployment scheme of the web platform based on Flask was built.The platform integrated model reasoning,defect statistics,model selection and other functions.In the actual application test,the reasoning speed of the model reached33 FPS.