Application Of Neural Network To Insulator Cracks Detection In The Power System

The patrol robot for HV Power Line inspection is a very important subject for power transportation of electricity engineering. The robot systems can observe the power lines and transmit the video signal to supervision station automatically. The style-working will greatly eases people's work condition and improves the patrol efficiency and reliability. The paper presents insulator inspection task for HV Power Line inspection.With the developing of image theory, digital image processing is widely used in every department of country. Based on image recognition, the paper detects insulato cracks and achieve the classifications of the cracks through simulating vision system of the robot. Center around the subject, the main work has been done as follows:Firstly, hardware design. The hardwares are completed, which include CCD, camera, control pan, image card and computer. The device can get the images of the object and be prepared for the computer.Secondly, image preprocessing. The image which is inputted is being preprocessed, including denoising, edge detection and binarization. Robert, Gauss, Prewitt, Canny methods are used to detect edges. In the processing of the image segment, the optimum threshold method is used. In addition, a new method that cuts the whole image into sub-image is introduced, which can identify the cracks and clear off the non-cracks in the processing of the image binarization. The experiment result will be shown.Thirdly, Features extraction. The paper extracts five features to form the feature vector, which include plane projection variance, vertical projection variance, number of the disrepair sub-image, autocorrelation coefficient and eulernumber, each of which will be compared among the five classifications, which will show the speciality.Lastly, image recognition. The feature vector are the input of artificial neural network and of Support Vector Machine respectively, which are used to recognize. In this paper, BP neural network, RBF neural network and SVM are proposed. The results are compared and analysed. In the BP neural network, the paper will study several improving methods in order to solve the conventional BP algorithm shortcomings. In the RBF neural network, the paper will study three arithmetics, which all have the merits and demerits, the compare will be done between the BP and RBF neural network. In the SVM, the paper will get the results. At last, the paper will compare the results of the three methods.The paper mainly studys the recognition of the crack classifications.