Research On Defect Detection Technology Of Photovoltaic Modules Based On Infrared Thermal Imaging

The accumulation of dust and the destruction of photovoltaic modules will reduce the efficiency of photoelectric conversion,but also corrosion of the surface protective layer of photovoltaic modules,resulting in electrical short circuit of internal components,resulting in safety problems.Based on this,this topic is based on the actual needs of photovoltaic module defect detection,using UAV thermal infrared equipment for aerial photography and defect detection of solar photovoltaic modules.In view of the limitations of UAV storage and computing power,as well as the problems that the existing deep learning-based defect detection model has large model parameters and high computing cost in the complex environment of large photovoltaic power stations,an ultra-lightweight defect detection model based on deep learning is designed.The main contributions of this paper are as follows:1.First,to solve the problem of imbalance in the number of thermal infrared image samples taken by UAV,methods such as copy-paste operation or manual rendering samples are proposed to realize manual enhancement of data sets.In order to solve the problems of noise interference and image blur in photovoltaic module images,the median image filtering method is used to reduce noise interference,and the adaptive histogram balancing with limited contrast is performed by adapthisteq to improve color contrast.Then,aiming at the calibration of data sets with and without dust in centralized and low-height hot spot samples,experiments prove that low-height data sets adopt the calibration method without dust in hot spot as the detection object in Chapter 5.2.In view of the large number of parameters in the YOLOv5 model and the low detection accuracy of small targets,the backbone network of the YOLOv5 algorithm was replaced by Mobilenet V2 to redesign the network.Secondly,the CA module was inserted into different layers of the YOLOv5 model,and the experimental verification was carried out on the centralized and high-altitude data set,which proved that the network has significantly improved the detection of small target defects while reducing the number of parameters.3.In order to solve the problems of foreign body occlusion and missing detection in photovoltaic module defect detection,an ultra-lightweight defect detection model,YOLOv5 Lite X,was studied for centralized and low-height photovoltaic module data set.Firstly,the weighted bidirectional Bi FPN structure is used in the feature fusion stage to achieve effective cross-scale connection and weighted fusion of features.Secondly,the four-scale feature fusion structure is used to enhance the extraction ability of important features.focal-EIo U Loss was introduced to improve the prediction loss of the original boundary frame coordinates,so that the network could focus on the calculation of difficult samples.Under the same experimental conditions,the algorithm in this paper gave consideration to the advantages of precision and speed.