| Due to the vigorous development of the photovoltaic power generation industry and the wide application of UAVs,UAV inspection of photovoltaic power plants has become a very popular and promising way of inspection of photovoltaic power plants.This method greatly improves the efficiency of inspection,and also reduces the labor cost of inspection,so many photovoltaic power plants use UAVs to inspect photovoltaic power plants.As the main work content of UAV inspection photovoltaic power station,there are still several problems to be solved in hotspot detection.First of all,this paper finds that the reflective interference on the photovoltaic array may not only cover the real hotspot,but it may also be detected as a rectangular hotspot,but so far,no one has proposed a way to eliminate reflective interference directly from infrared images containing reflective interference.Secondly,this paper finds that the classification of hotspots from the shape of hotspots is not conducive to further analysis of the causes of hotspots,and it is found that this classification method does not classify the severity of hotspot failures.Finally,this paper finds that the hotspot positioning method that transmits the coordinates of the component cannot map the detected hotspot to the specific position of the corresponding visible component.Therefore,in view of the above three problems,this paper conducts research on the infrared and visible video of photovoltaic array taken by UAVs in the process of inspecting photovoltaic power plants,and the specific research content is as follows:(1)In this paper,a two-stage reflective interference cancellation method based on deep learning is proposed.The first stage of the algorithm is the image registration stage,and the second stage is the image reconstruction stage.In the image registration stage,the Res Net-50 network is selected as the feature extraction and matching part of the supervised deep homography estimation model,and the homography matrix transformation module and the tensor perspective transformation module are proposed.In the image reconstruction stage,this paper proposes a position determination module for reflective interference areas,and selects Unet network as the image reconstruction network model to reconstruct the real scene in the reflective interference area.Finally,the image registration experiment and the image reconstruction experiment are carried out respectively,and the results are evaluated,and it is found that the gap between the predicted image and the real image is getting smaller and smaller,whether in the image registration stage or the image reconstruction stage,indicating that the algorithm can effectively eliminate the reflective interference in the infrared image of the photovoltaic array.(2)This paper proposes a hotspot classification algorithm based on YOLO V5,which divides hotspots into eight categories: small Circle,dim Circle,mid Circle,large Circle,cover,cover Circle,cover Rect and rect.Due to the different size and uneven number of hotspots of each category,this paper first uses data enhancement to expand the number of hotspots of each category,and then uses the YOLO V5 algorithm to identify hotspots.Finally,this paper conducts hotspot classification experiments and evaluates the results,and finds that the m AP value of all hotspot categories is as high as 92%,indicating that the algorithm can effectively identify each type of hotspot.(3)A hotspot localization algorithm mapped to visible light components is proposed.In this paper,the problem is first converted into infrared and visible image registration,and then the Res Net-50 network is used to estimate 8 degrees of freedom for describing the homography between infrared and visible images and three distortion parameters for describing the relative distortion relationship between infrared and visible light images,and according to these 8 degrees of freedom and 3 distortion parameters,infrared and visible image registration is realized.Finally,this paper conducts infrared and visible image registration experiments,and evaluates the results,and finds that the NMI value between infrared and visible images increases after image registration,indicating that the algorithm can align the texture of infrared image with the texture of visible image,and then map the hotspot detected on the infrared image of the photovoltaic array to the photovoltaic module on the visible image. |
