Insulator Detection Based On Feature Learning

With the development of intelligent grid and power system automation,computer vision techniques are increasingly applied on the intelligent inspection and online monitoring for equipment on power system.Insulators are indispensable insulation components in power transmission lines,and its operating conditions directly affect the reliability and safety of power grids.At the same time,insulators play the role of electrical insulation and supporting in the transmission lines.And the contamination,cracks,and aging on the surface of the insulators seriously threaten the safe operation of transmission lines.According to statistics,the highest percentage of accidents in the current power system failure is caused by insulator defects.Therefore,it is very important to monitor the condition of the insulator.How to detect insulators quickly and accurately in complex background images collected patrol robots has become a hot topic of research.Therefore,combined the demand of the project which is proposed by Jiangsu Provincial Electric Power Research Institute's science and technology,the contents of this paper included the following aspects:(1)To accurately recognize and locate the insulator,an insulator detection method based on shared convolutional neural network is adopted.First,the region proposal network is used to extract and abstract the feature of insulator images.After the feature of the next layer of convolution layer,the anchor mechanism of different aspect ratios is used on the last convolutional feature to obtain a series of candidate regions.Then the candidate regions are sent to sub-network Fast R-CNN which can achieve the identification and positioning of insulators.Throughout the entire process,the two networks share the features of the convolutional layer.This can reduce redundant calculations when the insulators are detected and can reduce the amount of network computation.By conducting experiments on the insulator data set,this method can detect most of the insulators in the image,but when the insulators occupy a small proportion in the image,there is a problem of insulator missed detection.(2)To solve the problem of missed detection of small insulators,an insulator detection method based on cross-connected convolutional neural network is proposed.This method utilizes the high-level features of convolutional neural networks that can embody stronger semantic information to facilitate target recognition.The low-level features facilitate the target positioning due to its high resolution.By connecting the convolutional layers of the last three layers of the network with the fully connected layer in the regional proposal network,the three-layer convolution features are simultaneously sent to the classification layer and the regression layer,which can obtain a series of high quality insulator candidate regions.Next,the region proposalsare input into the insulator detection sub-network that takes the deconvolution operation,then the ROI features of candidate regions are sent to the cascaded Adaboost classifier to detect insulators.Evaluations on the region proposal generation method were performed and comparative experiments were conducted.The method can effectively recognize and localize insulators of different sizes with complex background.(3)To achieve the saliency detection and self-explosion identification of insulators,an insulator saliency detection method based on multi-scale reconstruction error fusion is proposed.By considering the consistency of the insulators from the morphological and the difference from the image background,the simple linear iterative clustering is used to generate a series of image segments.The image boundaries are extracted as background templates.Next,we obtain the dense and sparse reconstruction errors for each image.The context-based propagation is used to propagating the reconstruction errors in each cluster.The condition random field is used to fusion the two reconstruction errors.Experiment results show that the precision of the proposed method is higher than other methods.The method can clearly and accurately extract the insulator saliency regions.In this paper,insulator detection problem is studied from two aspects: target detection and saliency detection.For each specific problem,we designed the corresponding network structure,and proposed specific and effective methods.Through experiments,it is found that the method of this paper can replace manual analysis and reduce the risks and errors due to the patrol staff's experience and judgement.It also can make the power system more intelligent.It lays a good foundation for subsequent insulator fault detection.