Research And Application Of Lightweight Object Detection Algorithm In Relay Protection Inspection Work

With the development of the smart grid,the structure of the grid becomes more and more complex,and the number of relay protection hard plates is increasing,which increases the workload of manual inspection and the risk of malfunction.How to improve the inspection efficiency and ensure the stability of the power system has become an urgent problem to be solved.To improve the current situation,this thesis applies deep learning technology and Android development technology to the task of relay protection inspection,and completes the inspection and verification of the protection pressure hard plate on the mobile terminal,including the following work:Firstly,a lightweight object detection algorithm is studied.On the basis of the YOLO algorithm,the Shuffle Net V2 network is used as the new backbone part,and the convolution operation with the channel number of 1024 and the pooling operation with the pooling kernel of 7×7 in the Shuffle Net V2 network are removed;then the 3×3 normal convolution in the enhanced feature extraction part is replaced by the depthwise separable convolutional structure in Mobile Net,and the use of C3 structure and high channel C3 structure is reduced;finally,appropriate channel pruning operations are performed on the whole network to further adjust and optimize the network structure.The experimental results show that the algorithm has less model parameters and faster inference speed under the premise of ensuring that the relay protection inspection work can be carried out normally.Secondly,a mobile terminal transplant algorithm of object detection model based on ncnn forward reasoning framework is studied,the 8bit data format quantification of the algorithm model in this thesis is realized,the dynamic size reasoning of the mobile terminal is completed,and it is aimed at the relay protection inspection work.According to the requirements of the ncnn forward reasoning framework,the original model calling method of the ncnn framework has been optimized.The experimental results show that the weight file of the algorithm in this thesis is only 1.6 MB,and the average inference speed on the Redmi Note10 Pro device reaches 14.78 ms.Finally,a mobile intelligent inspection application is designed and developed,which can complete the state identification of relay protection pressure plates by using the object detection model,and simultaneously complete the verification of the field pressure plates and the reference by parsing the inspection reference file designed in this thesis.In order to solve the problem of missed and mis-checked protection pressure plates,the application automatically completes the missed pressure plates and easily corrects the mis-checked pressure plates;the application supports incremental updates,which can reduce the waiting time for users to update the application in scenarios with poor network quality such as substations;in order to further optimize the user experience of the application,this thesis provides a targeted design of the application’s user interface and project architecture.After the field test in substation,the algorithm of this thesis can effectively solve the influence of environmental factors such as reflection and shading on object detection,and significantly improve the efficiency of inspection work.