Research On Road Recognition Method For Cleaning Robot In Photovoltaic Plant

Solar energy,as the main source of photovoltaic power generation,is closely related to the efficiency of the entire photovoltaic industry.Photovoltaic modules are the core components of solar photovoltaic systems,but due to the strong sandstorms in the northwest region,their surface is prone to dust accumulation,seriously weakening the efficiency of photovoltaic power generation.Cleaning robots are needed to clean them in a timely manner.The ability of cleaning robots to accurately identify the road area of photovoltaic plants is a prerequisite for robots to perform photovoltaic power plant cleaning operations.The roads of photovoltaic plants in the northwest region of China are mostly covered with weeds and gravel on both sides,with uneven roads and obstacles from photovoltaic modules,which poses great difficulties for road recognition by cleaning robots.Therefore,this article focuses on the research of road recognition methods for cleaning robots in photovoltaic plants,mainly including the following aspects:(1)Aiming at the characteristics of photovoltaic plant’ s roads covered with weeds and gravel on both sides,fuzzy boundaries,and uneven pavement,a road recognition method for cleaning robots of photovoltaic plants based on the DeepLabv3+basic model of deep learning is proposed.Training and predicting are conducted on the established road scene dataset of photovoltaic plants.The experimental results show that the DeepLabv3+basic model can segment most roads relatively completely,and the road segmentation accuracy is significantly better than the Seg Net,UNet,and PSPNet models.However,there are also problems of missing segmentation and wrong segmentation,the segmentation of road edges is not clear enough,and the parameters of the model are large,making it take a long training time.(2)Aiming at the problems of applying the DeepLabv3+basic model to the segmentation of road scenes of photovoltaic plants,a method for road recognition of photovoltaic plants based on optimizing the DeepLabv3+model is proposed.Firstly,replace the backbone network of the original DeepLabv3+model with an optimized Mobile Netv2 network to reduce the complexity of the model;Secondly,the strategy of combining heterogeneous receptive field fusion and hole depth separable convolution is adopted to improve the hole spatial pyramid pooling(ASPP)structure,improving the information utilization rate and model training efficiency of ASPP;Finally,the attention mechanism CBAM is introduced to improve the accuracy of model recognition.(3)Establish an experimental platform,set training parameters,and use the optimized DeepLabv3+model and other models to train and test on road scene dataset of photovoltaic plants to verify the feasibility of the optimized DeepLabv3+model.The experimental results show that the average pixel accuracy of the optimized DeepLabv3+model is 98.06%,and the average intersection and merge ratio is 95.92%,which is 1.79% and 2.44% higher than the DeepLabv3+basic model,respectively,and higher than the Seg Net,UNet,and PSPNet models.At the same time,the optimized model has a small amount of parameters and good real-time performance,which can better achieve road recognition for cleaning robots in photovoltaic plants.