Application Research On Solar Cell Defect Detection Based On Convolutional Neural Network

Solar cells are highly susceptible to cracking during transport and installation due to the fragility of the crystal structure.Cracks not only seriously reduce power generation efficiency and service life,but even local overheating leads to fire accidents,causing direct economic losses.In addition,grain pseudo-defects make crack detection very difficult,so achieving automatic detection of cracks and break defects in solar cells has profound practical significance.However,manual inspection is time-consuming,labour-intensive and not very accurate;image processing techniques are complex to operate and have poor real-time performance.In view of the strong real-time processing ability,easy deployment and ability of the YOLOv5-s model to take into account the accuracy of model detection,the YOLOv5-s algorithm is based on the YOLOv5-s algorithm to realise the need for real-time,high-precision detection of solar cell defects.An improved YOLOv5-s model that fuses deformable convolution version 2(DCNv2)with coordinate attention(CA)is proposed for crack and break defects.The main research elements are as follows.(1)The YOLOv5-s model,which combines speed and accuracy,is applied to the field of solar cell defect detection,and can efficiently detect both crack and break defects.To address the problem that the defect dataset provided by real power stations is not standardised and the number of defects is small,image pre-processing and data augmentation are used to improve the visual effect of the images,while increasing the number of samples in the dataset to form a self-made dataset,and the augmented dataset can improve the stability of the model.(2)To address the problem of difficult multi-scale crack detection,a convolutional optimisation strategy is proposed to optimise the feature extraction network using DCNv2 module.The convolution module optimised using DCNv2 can achieve adaptive sampling,expand the perceptual field of small target defects and reduce the loss of effective feature information.The experimental results show that the mean average precision(m AP)of the model reaches 93.3%,which can effectively reduce crack leakage and false detection cases,while meeting the real-time requirements.(3)In order to solve the problem of inaccurate regression frame prediction in defect detection,CA mechanism is introduced to improve the path aggregation network(PANet)structure.CA and cross-cascade are embedded in the network,and CA-PANet is proposed to shorten the path of feature information exchange between deep networks and shallow networks.The experimental results show that the model test m AP reaches 93.9%,which can effectively improve the accuracy of the prediction frame while meeting the real-time requirements.(4)Design a fusion-improved YOLOv5-s model based on DCNv2 and CA mechanism for defect detection of solar cells.The fusion-improved model was used to test the defective solar cell images in real scenes,and its defect recognition,classification and localisation effect was good,with m AP reaching 95.4%.The model was able to detect 51 solar cell images per second,meeting the industrial real-time detection requirements.