Research On The Method Of Dimensionality Reduction And Pattern Recognition For The Characteristic Parameters Of Transformer Discharge Acoustic Signals

Oil immersed power transformers are important hub equipment in the power system,and their insulation and health status are crucial for the safe and stable operation of the power grid.Affected by factors such as production and manufacturing process,transportation and installation,and operating environment,insulation defects such as metal burrs,suspended particles,open welding air gap,etc.may appear inside the transformer,which will distort the distribution of local electric field,and in serious cases,partial discharge may occur at the insulation weak point.The cumulative effect of partial discharge will accelerate the deterioration of insulation materials,which may induce insulation breakdown fault to a certain extent,resulting in transformer shutdown and endangering the safe operation of power system.Partial discharge can reflect the fault status of equipment and the deterioration form of insulation.The discharge process is usually accompanied by acoustic,optical,thermal,electrical and other signal changes.Experienced frontline personnel can determine whether there is internal discharge by listening to the sound during transformer operation.This article transforms subjective experience judgment into a traceable recognition mode through operations such as sound signal acquisition,operation status pre judgment,signal blind source separation,multi-dimensional feature reduction,and pattern recognition,providing theoretical basis and effective reference for transformer operation status evaluation and real-time discharge monitoring.The specific research content of this article is as follows:(1)When a discharge fault occurs in a transformer,the actual collected mixed sound signal is the superposition of the operating sound and discharge sound sources of the transformer body.Due to the scarcity and difficulty in obtaining the actual discharge sound signal of the transformer,it is necessary to design mixing experiments to obtain mixed sound signals that can simulate the superposition of various sound sources during internal discharge of the transformer,in order to expand the fault sound sample library.This article first collects the sound signal of the transformer during normal operation on site at Wusheng 750 k V substation as background sound;Then,according to the form and characteristics of partial discharge,three electrode models of tip discharge,surface discharge and air gap discharge are designed,and the discharge acoustic signal in the oil medium is collected as the foreground sound;Finally,a linear instantaneous mixing model is used to obtain a mixed sound signal that can simulate the superposition of sound sources during transformer discharge faults to a certain extent.(2)Analyzing the characteristics of the above three types of acoustic signals,it was found that the frequency distribution of the normal operation of the transformer’s acoustic signals is mainly concentrated in the low-frequency part below about 1 k Hz,with 100 Hz and its multiples being the main frequency,while the high-frequency part above 1 k Hz has a very small proportion of frequency;The frequency distribution of the discharge sound signal is relatively obvious in both the low-frequency and high-frequency parts.Compared with the normal operation sound of the transformer,the proportion of the discharge sound in the low-frequency part is reduced,while the proportion in the high-frequency part is significantly increased;The frequency distribution characteristics of the mixed acoustic signal during simulating actual discharge faults are superimposed on the above two types of acoustic signals,and the characteristic difference between them and the normal operation sound of the transformer is mainly reflected in the frequency distribution of the high-frequency part.(3)By utilizing the difference in high-frequency characteristics between normal and mixed sound signals,a spectrogram of voiceprint time is drawn to pre determine whether there is discharge inside the transformer,and to achieve binary classification of normal or faulty transformer operation status.In order to further identify the discharge type during the discharge fault state,this article uses a fast independent component analysis algorithm to separate source signals such as discharge sound from the mixed sound signal;The Pearson correlation coefficient criterion and normalization method are used to solve the randomness problem of the amplitude symbol and magnitude of the separated signal,respectively.The kurtosis feature index is used to solve the randomness problem of the arrangement order of the separated signal,and waveform similarity and root mean square error are used to evaluate the separation effect.Extract the joint time-domain and frequency-domain feature parameters of the separated discharge acoustic signal,and construct a multidimensional feature parameter matrix.(4)In order to achieve effective dimensionality reduction of multidimensional features,A feature selection and dimensionality reduction method based on correlation and inter class differences is proposed.Firstly,the Pearson correlation coefficient matrix is used to analyze the correlation between features,and the inter class difference and intra class dispersion of each feature sample are combined to select the optimal feature quantity for recognition effectiveness testing;Then,the accuracy of support vector machine recognition is used as the criterion for feature dimension selection,and the final feature dimension and corresponding features are determined from low to high dimensions based on the change in accuracy;Finally,a comparative analysis was conducted between the recognition results of our method and traditional dimensionality reduction algorithms.The results showed that our method retained the original feature attributes compared to traditional dimensionality reduction algorithms,and the recognition accuracy of the selected features exceeded 95%,providing an effective criterion for feature dimensionality reduction.This article further verifies the effectiveness and necessity of separation,and explores the impact of different noise interferences on the separation effect.The research in this article can be extended to the final selection of information domain features,multi domain joint features,etc.in the future,providing theoretical basis and reference for the construction of transformer voiceprint recognition systems.