PV Module Defect Detection Based On Improved Lightweight Network

Under the background of ’carbon neutrality’ and ’carbon peaking’,countries have vigorously developed clean energy such as photovoltaic,wind power,nuclear power,and hydropower to gradually replace fossil energy power generation,while photovoltaic power generation is favored by countries due to its cost advantages.However,as PV energy systems reach terawatts,even a slight degradation in solar cell performance could lead to a loss of thousands of GWh of annual global power generation.PV modules will inevitably produce certain defects in the process of production,transportation and installation,resulting in great economic losses and even fires.Therefore,the defect detection of photovoltaic modules is highly valued.The purpose of this thesis research is to classify whether there are defects in solar cells in photovoltaic modules and solar cell defect localization and detection based on improved lightweight networks,in order to solve the problem that the current neural network model parameters are too large and unfavorable for real-time fault diagnosis,the main research content of this paper is as follows:Firstly,the composition,power generation principle,classification,defect types and detection means of PV modules are studied.Secondly,the principles of deep learning and the basic structure of convolutional neural networks and their roles are explored,and the evaluation metrics commonly used for deep learning are analyzed,and finally,classical classification algorithms as well as target detection algorithms are investigated.Secondly,a method of solar cell defect classification based on the improved Mobile Net V3 network is proposed for the problem that the complex background of solar cells leads to a large classification model with many parameters,which is unfavorable for operation on mobile and embedded devices.Firstly,the principle of the lightweight network Mobile Net V3 is studied,and the loss function of the Mobile Net V3 network is improved considering the problem of extremely uneven sample distribution of solar cell data,and the key parameters of the loss function are tuned in order to obtain the best network model.The improved model shows excellent performance for defect detection,and the convergence speed of the model is improved,which in turn leads to a significant improvement in defect detection efficiency.In this paper,the performance of the classifier of Mobile Net V3 network is compared with the other three classifiers,and the results show that the performance of the chosen random forest classifier is slightly better than the other three classifiers.The improved model not only has improved feature extraction capability,but also its generalization capability is slightly improvedFinally,for the recognition accuracy of solar cell detection of small defects such as broken grids,hidden cracks and shadows can be further improved,a solar cell defect localization detection algorithm based on improved G-SSD network is proposed;firstly,for the problem of difficult detection of small targets,the size of the input image and the size of the shallow prior frame used to detect small targets are improved;secondly,to achieve the network model’s lightweight,the feature extraction module of the SSD network is improved;again,the three defect images of solar cells are labeled using the labeling tool Label Image to build a dataset;then,the model is trained on the built dataset separately with the model before improvement for the purpose of comparison experiments,and it is verified that the detection accuracy of the improved model is improved;finally,the model is compared with the Faster R-CNN and YOLOv5 two target detection algorithms for comparison experiments,also achieved a relatively high detection accuracy.The experimental results show that the method has good feasibility and effectiveness in the detection of defects in PV modules.