Research Of Insulator State Identification Based On CNN

The reliability and operation of power transmission equipment play an important role in the stability and safety of the power system.As one of the main components of the high voltage overhead transmission line,it is important to discover the insulator fault to ensure the safe and reliable of power supply.Therefore,is of great significance for identifying the fault insulator.Neural network mainly through the simulation of the human brain learning process,implement the abstract expression of the data,has gained widespread attention and focus in many fields.As a kind of neural network,CNN has a wide application in target detection and image recognition,and it has practical significance to study the deep research of CNN.Based on deep theoretical research of CNN,this thesis applies it to the identification of fault insulator.Based on the description of the characteristics by CNN,a new algorithm for image segmentation of insulators was proposed.Firstly,the pre-training CNN model was used to extract the region of interest,and then the filter and the Hough transform were used to extract the insulator information.The results show that this method can effectively extract the insulators from complex background of aerial images.From the point of kernel functions correlation,through BP algorithm error propagation analysis and correlation analysis for space vector theory,the independence kernel functions could extract more comprehensive features.Used wavelet reconstruction to remove the correlation between kernel functions can not only obtain higher recognition rate,but also shorten the training time.Studied the generalization of neural network,affined transformation of the first kernel function,thus formed the kernel function with more generalization ability.The results show this method can guarantee the correct mapping of valid samples and improve the generalization of the network.Accorded to the recognition of small samples,the confidence judgment function was set according to the Neyman-Pearson criterion.The samples which were difficult to be identified in the trained single-layer network were re-extracted in the next layer,thus form the multi-layer CNN structure.The experimental results show that this model can achieve better recognition of small samples when the sample size is small.