Research On Insulator Detection In Power Patrol Based On Convolutional Neural Network

Power system patrol is an indispensable means to guarantee the safty of power grids.The burgeoning inspection with unmanned aerial vehicles(UAVs)that equipped with HD camera can obtain a large number of detailed inspection images in the transmission lines.If these patrol data are processed by manual analysis,the workload is enormous and inefficient,and the deviation caused by human experience and quality is inevitable.Simultaneously,due to the fact that insulators are widely used components in power systems and exposed in the wild all year around,as a fault-prone component,they seriously threat the grid security,thus computer technology is required for fault diagnosis.Combined with the demand of the project which is proposed by Sichuan Electric Power Company,the contents of this paper included the following aspects:(1)This paper accomplished the detection of insulator by building and improving a convolution neural network(CNN),aiming to solve the problem of traditional detection algorithms such as poor robustness,weak generalization ability and low accuracy.Firstly,by introducing as well as investiagating the characteristics of CNN and its wide applications,we designed each components of CNN with the engineering requirements and hardware support.Then,we use UAVs to collect glass and ceramic insulator photos in different lines and time as training samples with artificially expand.Next,Caffe was choosed as the tool in this paper,the network structure was improved combining the correlative parameter adjustment technologies for a neural network and the training process was optimized by contrast tests.With the improvement and optimizition of the network,it learned the nature and distributed expression of insulator characteristics automatically and successfully achieved the insulator detection in complex background,the accuracy for training and testing is 95% and 92% respectively.(2)This paper achieved the recognition of insulator explosion defect with a well-trained CNN to solve the problem like large workload and low efficiency by mannual analysis.First,by using the synthesis and abstraction of global and local features by hierarchy architecture of CNN,we used the well-trained network model as feature extraction tool and sent the feature maps into a self-organizing feature map network(SOM)for saliency detection.Then,based on the saliency detection,we can extract the insulators quickly and discard the complex background as well.Next we proposed a recognition algorithm for explosion defect by building the module via super pixel segmentation,contour detection and other image processing methods,and the accuracy is above 90%.The algorithm can help reduce labour,risks and the errors caused by manual analysis,guarantee the security and reliability of the power grid.(3)The verification and contrast tests of insulator detection and explosion recognition algorithm are implemented.Firstly,the insulator detection tests were carried out for different backgrounds,different categories and different quantities,and the proposed algorithm was compared with the traditional DPM and SVM algorithm based HoG as well.Then the performance of the network was analyzed by visualization of feature maps and kernels.Next,the explosion recognition algorithm was verified in different backgrounds.Based on the engineering project,this paper briefly introduced the framework of the insulator detection system finally.The test also demonstrated that the detection of the insulator and the recognition of explosion defect not only met the requirements of the project,but also refected the value of patrol data effectively and improved the efficiency of inspection and made the inspection more intelligent.