Research On Fault Detection Algorithm Of Photovoltaic Panel Based On Deep Learning

In recent years,with the development and growth of PV industry,the safe and stable operation of PV system is related to the power generation efficiency and economic benefits of PV power plants,and how to complete the daily operation and maintenance of PV power plants in an efficient and economic way has become a key issue.The traditional PV plant operation and maintenance is mainly based on the electrical characteristics of PV inverters or classical image processing methods.The rise and development of UAV inspection has greatly changed this status quo,as the infrared camera carried by UAVs can detect internal faults of PV panels,and combined with deep learning detection methods based on computer vision can efficiently and accurately detect a variety of common faults of PV panels,saving a lot of human and material resources.In this paper,the infrared photovoltaic panel fault detection task is investigated using deep learning based semantic segmentation and target detection algorithms with the following details:Firstly,the PV panel infrared images collected by UAV are manually annotated using Labelimg data annotation software after data augmentation and collated into two datasets,PV＿larege and PV＿roof,and then common data pre-processing and enhancement methods are introduced.Then,two deep learning-based semantic segmentation networks are designed to extract PV panel regions in order to eliminate the negative interference from irrelevant backgrounds in subsequent detection.Based on the improved Deep Labv3+ network,the Deep Labv3+ semantic segmentation network with fused edge features is designed for the PV＿large dataset with the problem of intermittent edge erosion,which further optimizes the edge details and achieves 99.96% segmentation accuracy;for the PV＿roof dataset with the problem of background mis-segmentation due to interference from roof color steel tiles,the A semantic segmentation model incorporating visual features is designed to extract the texture features of PV modules and colorful steel tiles under infrared light and selectively suppress them after clustering,which reduces the mis-segmentation of the roof background and the internal voids of PV modules,and the segmentation accuracy reaches 96.15% providing a good input for subsequent inspection.Finally,the fault detection model PV-YOLO for centrally distributed IR PV panels is designed by combining the Transformer-based PVT-v2 structure with the improved YOLOX.Experiments conducted on the PV＿large dataset demonstrate that the algorithm in this paper has a significant accuracy advantage at the same parameter magnitude,with a maximum m AP of 92.56%.For the characteristics of rooftop PV panel fault detection task,the network is further streamlined and the lightweight PV panel fault detection algorithm PV-YOLO-Tiny is designed,and the experiments on PV＿roof dataset are conducted to compare with the lightweight model of YOLO series to prove the superiority of the network.