Road Scene Recognition Of The Substation Inspection Robot Based On Deep Learning

Substation is an important part of the power system field.The safety and reliability of the substation can directly affect the normal operation of the whole power grid.However,traditional substation equipment patrol operations dominated by manual patrol can cause several disadvantages,for example,the stronger labor intensity,the greater environmental impact and the less objectivity for inspection.These drawbacks cannot adopt the increasing requirements of the maintenance management of substation.With the depth study of power grid construction in our country,inspection robots have been widely applied in different fields,which can be gradually replacing manual inspection.It can be seen that they play an extremely important role in maintaining the safety of substation equipment.In terms of substation inspection robots,a better road scene recognition technique is the key to implement autonomous navigation and normal inspection.So far,substation inspection robots usually use laser radars to acquire environmental information.Although these laser radars can detect the accurate distance,they cannot predict effectively the environment information because they lack the recognition and understanding for road scenes,which greatly affects the efficiency of substation inspection robots and their environmental adaptability.Based on this,a deep learning technique is applied to substation inspection robot in order to better recognize the road scene.Specifically,a full convolution road scene recognition network of the deep learning technique can be used to recognize the road scene,which can not only provide effective environmental information for substation inspection robot,but also enrich the theoretical research related to deep learning.The main works of this paper are summarized as follows:1.Summarize the current research status of the road scene recognition of substation inspection robot and deep learning techniques,introduce the specific steps of building deep learning framework on different hardware platforms for substation inspection robot road scene recognition,and construct the data set of road scene in substation,which provides support for subsequent algorithm research.2.According to the characteristics of the road scene of substation inspection robot,the paper designs a full convolution road recognition network for substation.This network can be composed of the encoder network and the decoder network that the former is applied to extract image features by building a relative shallow encoder network under the VGG16 model,and the latter is applied to restore the original size by choosing the global and local feature information under the mirror structure and the skip-level fusion structure.3.Test experiments of substation road scene recognition network.Firstly,verify the accuracy and the efficiency of the network by constantly testing on open source standard data sets and substation data sets.Secondly,verify the adaptability of the network in the real substation environment.Experimental results show that our proposed network has higher recognition accuracy and efficiency than the same type of network.Besides,in the real substation environment,the network can provide effective road environment information for substation inspection robots though its excellent scene recognition performance.