Substation Equipment Recognition Based On Deep Convolution Neural Network

The real-time and accurate identification of substation equipment is the key to realize automatic and intelligent production of unattended substation automatic monitoring and robot inspection.Because of the wide variety of equipment and models and the similar color of equipment in the substation,the threshold segmentation method based on color space is difficult to detect simultaneously different devices.On the other hand,due to the non real-time nature of feature extraction operator,the traditional pattern recognition method is also difficult to be directly applied to real-time detection of substation equipment.Convolution neural network(CNN)can automatically extract low-level image features and high-level semantic features,and is real-time in GPU computing devices.In view of this,this paper proposes a fast identification method for substation equipment based on CNN based on Faster R-CNN and YOLOv2,and designs a corresponding identification network model.The model is made up of convolution neural network and the further processing of the output feature map.The recognition network can directly predict the target boundary and type of substation equipment with convolution feature map.Image and scene images of substation equipment are collected under various illumination conditions,and image data sets and scene data sets of substation equipment are constructed.In order to screen the basic network structure for the identification network,the performance and applicability of several typical classified CNN networks in the image processing of substation are analyzed based on the criterion of model size,characteristic statistical separability,classification performance and computing speed.Inspired by the advantages of these typical networks,a channel oriented grouping convolution module and a corresponding classification network CWGCNet are designed.In order to further improve the image feature extraction ability of classified CNN network,we propose to amplify training networks using sample amplification training.CWGCNet and selected Darknet-19,Inception V2 and SqueezeNet are used as the basic structure of the identification network.After the infrastructure,the channel oriented packet convolution module with dropout and 3 layers of full volume layer are used as the overall architecture of the identification network.Based on the Microsoft CNTK computing framework,all classification networks and identification networks are implemented by Python,and the substation target recognition software system is realized by Visual C++.Experiments were carried out on 1 computers equipped with Tesla K40 c GPU.The experimental results show that CWGCNet has high feature separability and real-time performance compared with several typical classified CNN networks.On the transformersubstation equipment image data set,Caltech256 can significantly improve the network feature extraction ability by using Caltech256 for the network,and the nonlinear scaling of Sigmoid forms relative to the number form.The factor makes the recognition network more easy to train;compared with the additional 3 layer full volume layer,the additional coiling layer using the CWGC module with dropout can significantly reduce the size of the model with a significant increase in network recall,recognition speed and accuracy;CWGC module with dropout after CWGCNet is used as a module.For the structure of the substation equipment identification network,the network can identify different types of substation equipment.The recognition accuracy of the transformer,switch and lightning arrester can reach 96.52%,97.85% and 82.62% respectively.The recall rate is 77.39%,69.33% and64.23% respectively.The recognition speed is 62 fps.Compared with YOLOv2,the recall rate of the network is identified.It is 14.03% higher and the precision is improved by 2.51%.