Research On Acoustic Recognition Method Of Partial Discharge Ultrasonic Signal Of Power Transformer

Power transformers are key equipment in the power grid,and it is of great significance to perform state assessment and fault diagnosis on transformers.When the internal insulation of the transformer is defective,partial discharge will occur,and the detection of partial discharge can characterize the insulation performance.When there is a partial discharge in the transformer,an abnormal ultrasonic wave will be emitted.The current ultrasonic diagnosis methods of partial discharge include phase diagram,time-of-flight diagram and frequency correlation method.Due to the long duration of ultrasound,the pattern recognition of existing methods is relatively rough,and only some defect types can be realized,and it is not possible to finely distinguish all discharge types.The ultrasonic signal of partial discharge is also a "sound" in a broad sense.Research on recognition based on the "sound" attribute of the ultrasonic signal only stops at the introduction of acoustic features.Whether the acoustic recognition system can effectively recognize this "sound"is a question worthy of exploring.To solve this question,this paper designs and produces four typical oil-paper insulated partial discharge defect models,builds a partial discharge ultrasonic testing experimental system,collects ultrasonic signals of different types of partial discharges,and combines the partial discharge ultrasonic signals collected during the factory test of the transformer,and passes these signals through the Sound Activity Detection in voiceprint analysis The length of time is compared,the ultrasonic samples are screened according to the results,the partial discharge ultrasonic sample library of transformers is established,the acoustic recognition technology is introduced,and the partial discharge ultrasonic signals of different modes are analyzed.Based on the acoustic recognition technology,the spectrograms of different types of partial discharge ultrasonic signals are drawn,the differences of the spectrograms of partial discharge ultrasonic signals of different modes are analyzed,and the Mel Frequency Cepstral Coefficient and the Gammatone Frequency Cepstral Coefficient of ultrasonic signals are extracted.The Gaussian mixture model of partial discharge ultrasonic signals of different modes is obtained based on the feature coefficient training.It is found that the spectrogram,feature coefficient and model of different types of partial discharge ultrasonic signals are quite different.In order to highlight the "sound" attribute of ultrasound,the concept of frequency reduction coefficient is proposed,and the influence of frequency reduction coefficient on the recognition effect is analyzed.The key parameters in the algorithm includes frame length,feature coefficient dimension and model component number.At the same time,the feature coefficients are further simplified based on the recognition contribution rate,which improves the recognition efficiency and accuracy.The results show that the recognition effect based on MFCC can reach to 97.78%when the frequency reduction coefficient is set as 10,the frame length is set as 5ms,the feature coefficient dimension is set as 8,and the number of model components is set as 4,the recognition effect based on GFCC can reach to 87.36%when the frequency reduction coefficient is set as 10,the frame length is set as 10ms,the feature coefficient dimension is set as 16,and the number of model components is set as 8.