Research On Computer Vision Based Photovoltaic Module Defects Diagnosis

With the large-scale application of centralized photovoltaic(PV)power plants and distributed PV generations,the demand of PV operation and maintenance is arising.Nowadays,Unmanned Aerial Vehicle(UAV)automated inspection and intelligent module defects diagnosis are the two crucial technology to meet the demand.Utilizing the second technology to analyze aerial images of PV regions with complicated background,is an efficient way to solve the defects diagnosis problem.This paper proposed a Computer Vision(CV)based solution for visible module defects recognition,so as to achieve the goal of intelligent diagnosis.And the solution is combined tightly with the current situation of PV industry.The solution includes three parts: a progressive automatic labeling approach,an instance segmentation based module segmentation model and a transfer learning based module defects diagnosis model.Due to the lack of defect samples and currently high cost of manual annotation,it is difficult to directly apply deep learning technology to analyzing the aerial images of PV module.This paper proposed a progressive labeling method that is able to annotate PV module in a short time.Considering the fine-grained characteristic of PV defect,this paper employed a two-step method to construct the PV module defects diagnosis model.The first step is to segment modules from complex background and the second step is to sample defects in PV modules.Then it is possible to diagnose modules through a classification network.In view of the difference between the adopted dataset and other actual ones,this paper exploited a transfer learning based solution to deal with the deviation of data distribution and the frequently update of deep learning model.The methods proposed in this paper have been practically verified on an experimental platform and a rooftop PV generation.The solution is evaluated through extensive experiments and the numerical results show its effectiveness.