Fault Detection Of UAV Photovoltaic Module Based On Infrared Image Processing

In recent years,China is paying more attention to the development and utilization of solar energy resources,photovoltaic installations continue to climb.However,the long-term operation of the photovoltaic system and bad weather conditions are easy to cause module failure,affect its normal operation,and even create security risks.Therefore,regular inspection of photovoltaic power stations is particularly important.The unmanned aerial vehicle equipped with a camera can effectively meet the operation and maintenance requirements of photovoltaic power stations.This paper takes unmanned aerial vehicle inspection of photovoltaic power station as the background and aims to improve the efficiency of infrared image fault detection of photovoltaic modules.The fault detection technology of infrared photovoltaic module image is studied.Combined with the actual demand,the infrared image fault detection scheme suitable for different scenes is designed.Firstly,the characteristics of infrared photovoltaic modules are analyzed and the UAV inspection scheme is designed.The characteristics of photovoltaic module under infrared imaging and the causes of photovoltaic module failure are analyzed,and the technical difficulties and interference factors in the infrared image fault detection of photovoltaic module are pointed out.According to the characteristics of large-scale centralized photovoltaic power station,the infrared image inspection scheme of UAV photovoltaic module is constructed.Secondly,the traditional image fault detection scheme is designed.In order to avoid the influence of background interference information such as weeds and puddles in infrared images on detection results,a photovoltaic module fault detection scheme based on image segmentation algorithm was designed,including the improved watershed algorithm and Hough line detection algorithm to identify photovoltaic modules,and the application of edge detection algorithm in detection module faults was studied.The temperature information of every pixel in the infrared image can be extracted.Based on this feature,the temperature threshold detection algorithm is designed.This algorithm can effectively screen out the temperature anomalies in the module,so as to complete the detection of photovoltaic module heating fault.Infrared images can draw gray histogram according to gray information.In order to eliminate interference information in the histogram and select appropriate gray threshold,a detection scheme fitting gray curve setting threshold is designed.By analyzing the characteristics of gray curve and image gray level,the threshold is selected,and the median filter is applied to remove the remaining interference,and the component fault detection is completed.The experiment of traditional image processing scheme is carried out,and the application scenario of traditional image processing scheme is analyzed.Finally,a deep learning fault detection scheme is designed.Two deep learning object detection algorithms,YOLOv4 and YOLOv7-tiny,are improved.In view of the problem of less original data,off-line data enhancement is carried out to expand the data set.Because the YOLOv4 model has many parameters and a large model,Dense Net121 network is used to replace the original Dark Net53 backbone network to achieve model lightweight and reduce calculation parameters,improve the detection speed;A new path aggregation network is designed to further improve the fusion of shallow features,and YOLO detection head is added to the new path to achieve the purpose of improving the detection accuracy of small targets;After the backbone network,the attention mechanism module is added to further improve the focus of the model and improve the overall detection accuracy.Based on the YOLOv7-tiny model,two path aggregation networks were designed to study the influence of different path aggregation networks on the feature fusion effect.By adding different types of attention mechanisms,we study the effect of adding attention mechanisms at the back of the backbone network and the front of the path aggregation network on the accuracy of the model.The feasibility and effectiveness of the improved algorithm are proved by experimental evaluation of all algorithms.