Recognition And Segmentation Of Catenary Insulator Based On Deep Learning

In recent years,china has made remarkable achievements in the construction of electrified railway,and the railway mileage has been in the leading position in the world.The insulator is a key component in the railway transportation power grid.However,the insulator is exposed in the outdoor environment all the year round,so it is easy to accumulate dust and other impurities on its surface.Conductive solution will be formed in wet weather,resulting in partial discharge of insulator and frequent thermal faults,which will seriously affect the running safety of the train.At present,the insulator inspection is still mainly manual,but this method can no longer meet the needs of maintaining the development of China’s railway.Therefore,how to realize automatic inspection has important practical significance.The main research of this paper is as follows:(1)In order to locate the position of the insulator in the image,the traditional Histogram of Oriented Gradient(HOG)and Support Vector Machines(SVM)algorithms are used to detect insulators.The algorithm has strong robustness and fast training speed.It can successfully identify and locate the insulators in the image using a small amount of data and without complex parameter settings.However,this algorithm can only extract the shallow features in the image,which is better for the simple applicable background,and the detection efficiency is low.It has certain limitations for the realization of intelligent detection of insulators.(2)For the shortcomings of HOG and SVM algorithms.An improved Faster-RCNN detection algorithm was proposed to improve the detection precision of insulators.The original Faster-RCNN algorithm uses Vgg16 as the feature extraction network,which has the disadvantages of large amount of parameters and slow convergence.Therefore,this paper uses Inceptionv2 to replace Vgg16 as the feature extraction network.The Inceptionv2 network structure consists of by multiple Inception＿modules,which improves the network depth and thus has higher recognition accuracy.A normalization layer(Batch Normalization,BN)is added before input layer to normalize the data and improve the network,convergence speed.The experimental results show that the mean average precision of the Faster-RCNN algorithm based on the Inceptionv2 network on the test set is improved by about 3.5%compared with the original algorithm.(3)When using infrared technology to detect thermal failure of insulators.In order to reduce the influence of irrelevant background in the thermal image and improve the efficiency of insulator thermal fault detection,this paper uses the Mask-RCNN algorithm based on deep learning to segment the insulator in the infrared image and obtain the specific structural area of the insulator.Due to the complex and diverse background of insulator collection,the feature extraction network adopts the residual network Res Net101 to replace the Res Net50 network in the original algorithm,thereby improving the network feature extraction ability.The experimental results show that the mean average precision of the Mask-RCNN algorithm based on Res Net101 on the test set is improved by about 3.2% compared with the original algorithm.(4)Consider the practical application of the algorithm.Finally,the detection and segmentation algorithm model based on deep learning is exported and deployed,and the image is transmitted to the backend through the browser to complete the detection and segmentation tasks.The back-end then transmits the detection and segmentation results to the front-end,realizing the separation of the front-end and the back-end.It further proves the feasibility of the detection and segmentation algorithm based on deep learning.