Study On Identification Of Discharge Status Of 12kV Ring Grid Cabinet Based On Audio Signal

12kV Ring cabinets are a very important part of the power system,play a connecting and regulating role in the distribution network,and are an important part in the realization of digital construction.In the process of its manufacture and operation,partial discharge will be caused by various defects,which will threaten the stable operation of the power grid,so on-line monitoring of the ring network cabinets is particularly necessary.There are many methods of monitoring,mainly aiming at the phenomenon produced in the process of discharge,select a certain value to monitor.The traditional monitoring cost is high and the phenomenon is not easy to monitor.In order to fully reflect the insulation state of the ring net cabinet,this paper puts forward the sound-based insulation on-line monitoring combined with the structural characteristics of the ring net cabinet Testing.In order to simulate the discharge condition of ring net cabinet and facilitate the experiment,this experiment simulates the discharge condition of insulating tube,insulating sleeve and insulating porcelain bottle in ring net cabinet.and compared the discharge of each insulation device under wet environment.By using the pickup,the collected sound is transmitted to the sound card,and the sound signal is stored and processed by Labview the upper computer platform,and the discharge sound of each insulator and the non-discharge sound in the environment are collected.The pre-processing of sound is introduced,including:two-gate endpoint detection,adaptive filtering and bandpass filtering.Comparison of Mel cepstrum coefficient,linear predicted cepstrum coefficient and wavelet packet energy.The feature extraction method of quantitative analysis is used to extract the sound signal after filtering noise.Fisher feature dimension reduction is proposed to solve the problem of data redundancy.After feature extraction,the sample has a feature label,which facilitates the establishment of the model.The main learning models include hidden Markov model,Gaussian mixture model,neural network model and so on.This paper presents a Gaussian naive Bayes algorithm and a single classification support vector machine to train and learn the samples separately.The cross-validation and grid search methods are used to train and test the feature samples under the condition of collecting limited data samples.Compare the recognition rate to find a moresuitable algorithm for noise recognition of ring cabinets.On-line monitoring of discharge sound signals by establishing identification model can effectively prevent insulation failure.