Insulator Fault Detection Based On Improved YOLO Network And Extreme Learning Machine

At present,China’s electric power system is rapidly developing,this makes the security of the contact network put forward more stringent requirements.To ensure the safe operation of the power grid.As the main equipment of catenary.Accurate and efficient detection of insulator faults is an important part of ensuring safe operation of power system.This paper is based on the summary of relevant research at home and abroad.In this paper,the OCS insulator detection algorithm is divided into two stages: target detection and defect identification,an insulator target detection algorithm and an insulator fault identification algorithm based on deep learning are designed.The main research results of this paper are as follows:Firstly,it is based on the YOLOV3 target detection algorithm,combined with bottleneck layer,a lightweight multi-scale network is constructed to extract spatial context information,its purpose is to further improve the reasoning speed of the network,integrating the advantages of lightweight networks is proposed for a streamlined network,this method optimizes the traditional complex network,in addition,the network parameters are minimized,the operating speed is improved,and the problem of insulator dichotomy in this paper can be solved.Then there is the performance loss caused by parameter reduction,K-means++clustering algorithm is introduced to fit insulator prior knowledge,the introduction of this algorithm improves the performance of the location algorithm and makes up for the lack of precision caused by the reduction of network parameters to a certain extent,insulator fault classification module is a "on" module,which outputs the insulator target image of the insulator positioning module and then carries out fault classification.The core algorithm is the convolutional neural network image classification model.As for the model positioning algorithm,according to the needs of insulator dichotomization,a lightweight network model is proposed based on the predecessors.By calculating the accuracy,recall rate and other indicators,it is proved that Light-15 can greatly improve the operating speed of the network in insulator positioning field through the sacrifice of small precision.Secondly,an insulator fault detection method based on integrated convolutional wavelet limit learning neural network is proposed to solve the problem of insufficient application of activation function in classification model algorithm.First through the industrial camera installed on the unmanned aerial vehicle(uav)the scene of the insulator image data preprocessing,and then the convolutional neural network,the automatic encoder,extreme learning machine is combined with advantages of wavelet function,construct integrated convolution wavelet neural network learning limit,and set up multiple deep neural network step by step a stack;finally,the insulator image samples are input to several deep neural networks for automatic feature learning,and the prediction results are integrated and the final fault detection results are output.In this method,the wavelet function is mostly the activation function of the hidden layer of the model of extreme learning machine,which solves the deficiency of the traditional activation function in the performance of local feature extraction of time and frequency,thus limiting the improvement of the recognition rate.Finally,the experimental results are summarized systematically,through TOP-1,the comprehensive comparison of accuracy,recall rate,average precision value and operation element was made.It is verified that the proposed method has more advantages in image feature extraction and fault detection accuracy compared with other methods,and is suitable for insulator fault automatic identification.This method overcomes the complicated conditions of the traditional model.It creates a feasible condition for the application of embedded micro-computing system in power grid inspection.For the classification model,due to the introduction of wavelet function,The results show that this method can suppress noise interference and wild value samples greatly and has excellent performance in fault diagnosis.