Deep Learning-based Online Detection Of Defects In Silicon Photovoltaic Modules

Photovoltaic power generation inspired by daylight has become a good choice in today’s growing demand for new energy sources.As the use of photovoltaic power generation is increasing,its role is becoming more and more important.In photovoltaic power generation,defect detection and quality monitoring of PV modules are of great significance to guarantee the photoelectric conversion efficiency,safe and stable operation and avoid catastrophic accidents of PV power generation equipment.In industry,the defect detection of PV modules is generally accomplished by using electroluminescence(EL)detection imaging with manual visual inspection.In this thesis,we propose a method for intelligent detection of PV defects by fusion of EL detection images and electrothermal(ET)imaging detection data,and establish a defect detection model for PV modules using a combination of traditional image processing and deep learning based on the EL detection image dataset of PV cells.The identification and localization of defect categories and the evaluation of PV cell grades for a variety of common defects in PV cells were completed.The main research work of the thesis is as follows.(1)Characterization analysis of each defect type of PV cells.The causes,characteristics,and effects of dozens of different defects in the EL inspection of PV cells are analyzed,including large cracks,sub fissure,scratches,false welding,over weld,broken gates,black spots,pollution,chipped,failures and mixed grades.Through the analysis and combined with the actual situation,a standard on the evaluation of the EL inspection grade of PV was developed.The analysis process and this criterion provide the basis for the final evaluation of defect detection results.(2)For the shortcoming that EL inspection cannot detect the pre-breakdown defects inside the PV cell,a combined EL and ET inspection experiment was conducted,in which the defects such as cracks,broken gates,contamination,and pre-breakdown were mainly detected.For the ET detection image,after the processing of optical flow method and EL detection image fusion,it can better detect the defects.Finally,the deep learning method is also used to detect the fused PV cell images.This process provides a basis for using deep learning methods to detect industrial PV cell images.(3)Research on EL image segmentation method for PV cell modules.For the PV module EL inspection image size is very large and the defect is very small characteristics proposed first segmentation,then detection method.After comparing the segmentation by line feature-based Hough transform linear detection method and the segmentation by point feature-based template-matching corner detection method,the final segmentation method using point features completes a more accurate segmentation of the EL inspection image of PV module.Support is provided for the PV cell dataset used for deep learning.(4)A Deep Learning based target detection model for PV cell defects is investigated.Based on the PV cell dataset,a deep learning-based PV cell defect detection model is established,and the necessity of PV cell component segmentation is verified by comparing the detection effect of the model on different datasets.Finally,the category detection and localization of PV cell defects are achieved,and the conversion process from target detection to PV cell component grade evaluation is realized according to the PV cell EL inspection grade evaluation criteria.