Study On Noise Separation And Recognition Technology Of Shunt Reactor Based On Blind Source Separation And Depth Learning

As China’s power system enters the new stage of large-scale system,high-voltage,long-distance transmission and inter-regional network,the capacitance effect of transmission lines becomes particularly obvious with high-power and long-distance transmission,in order to solve this problem,UHV shunt reactors are widely used in UHV AC power transmission projects,which cause power frequency voltage rise and transient surge over-voltage.In the process of suppressing over-voltage,UHV shunt reactor gradually exposes some problems of its own,among which the problem of noise exceeding standard is one of the more serious problems in its operation.The application environment of UHV shunt reactor is very complex.The noise of other electric equipments around it will cause serious influence.In this paper,an acoustic signal acquisition method based on the microphone array is proposed,which is based on the arrangement of the electrical equipment in the substation.Based on the analysis of the advantages and disadvantages of the existing beamforming algorithms,a masking-based beamforming algorithm is proposed,it lays a foundation for the follow-up under-determined blind source separation.In addition,since the type and amount of mixed noise in substations are unknown,the advantages and disadvantages of the existing under-determined blind source separation algorithms are analyzed,in this paper,a density peak clustering algorithm in k neighborhood for non-sparse signal processing is proposed,the mean square error and deflection angle of the proposed algorithm are smaller than those of the common clustering algorithm.Next,through experiments,the noise around the reactor is separated,and the hybrid estimation matrix is calculated based on the radial basis function minimum mean square error method to recover the source signal,by comparing with single sound source,the main noise sources of the reactor body and its surroundings are obtained.Finally,a method to obtain a large number of time-frequency spectra of mixed noise by using independent source signal library is proposed.The mixed estimation matrix obtained by under-determined blind source separation matrix is adopted,and the appropriate step size is set up,a large number of mixed time-frequency Spectra composed of different kinds of noises are obtained,and a sound pattern library of reactor noise is established.Training and test samples of mixed time-frequency spectra are constructed by using the existing deep learning model,and the training set is preprocessed,through the training model,the test error data under different CNN models are obtained,and the hybrid time-frequency spectrum is classified,which provides the training samples for the classification of the voiceprints collected by the single sensor.