Study On Defect Model Of Photovoltaic Cell Based On Improved YOLOv5

Production process defects and human error caused by the surface of the photovoltaic cells,such as cracks,cracks,welding and other minor defects,will reduce its work efficiency,and even affect the service life.The characteristics of surface defects are subtle and difficult to observe,so there are many difficulties in the detection of surface defects of photovoltaic cells.Therefore,it is of great significance to design a target detection model which has strong detection performance and can identify defects of various photovoltaic cells.In this paper,different types of defect image data in the actual inspection of photovoltaic power station are taken as the research object.By making a series of improvements on the YOLOv5s model,the photovoltaic panel single sheet is extracted and the defect types are analyzed in detail to predict the defect degree,so as to improve the operation and maintenance efficiency of photovoltaic power station and reduce the operation and maintenance cost.The main research contents are as follows:(1)Focusing on the defect detection of photovoltaic cells,this paper collected the images of photovoltaic power stations by collecting on-site and historical data,and constructed the data set of substation defect detection by using labelImg software to manually label the corresponding labels of images.The data set covers a total of 1218 images across eight defect detection categories.(2)Starting from the actual requirements of photovoltaic power station inspection,this paper takes YOLOv5s as the basic model,improves the backbone network,loss function and neck network of the basic model respectively,and proposes a lightweight model.First,in order to better balance the conflict between network parameters and feature fusion,a CSP-Faster module was designed based on the lightweight idea of FasterNet,and further lightweight improvement was carried out for the model backbone out for the model backbone network.Based on this,the basic network loss function is improved.By integrating the Normalized Gaussian Wasserstein Distance(NWD)with CIoU,a new loss function Nwd-Ciou is proposed.The deficiency of the original model loss function in detecting small target defects of photovoltaic cells is improved.Finally,the CARAFE lightweight operator is introduced into the sampling module of the model neck network to improve the detection performance of the model.The improved model was tested on the data set of substation defect detection,and the results showed that the average accuracy reached 67.5%,which was 5.8%higher than that of the original YOLOv5s(60.7%),and the reference amount reached 59.5M,which was 15.5%lower than that of the original YOLOv5s(70.41M),which could better meet the requirements of photovoltaic cell defect detection.(3)Based on YOLOv5s model,a high-performance model is proposed to improve the design of the backbone network,loss function and neck network of the basic model.First,in order to make full use of the advantages of full-dimensional dynamic convolution,this paper improves the original model backbone by replacing the ordinary convolution of the standard convolution layer in the backbone with fulldimensional dynamic convolution.On this basis,by using the advantages of coordinate convolution,the neck network of the original model is improved by replacing the common convolution in the CSP module before up-sampling operation with coordinate convolution.Finally,the Biformer attention mechanism was introduced between the neck network and the detection head to improve the original model.Compared with the original YOLOv5s(60.7%),the average accuracy of the improved high-performance model is increased by 10.7%,and the number of parameters is increased by 71.4%,which can better meet the requirements of photovoltaic cell defect detection.