Research On Transformer Fault Detection Method Based On Sound Signal Intelligent Processing Model

Transformer plays an important role in power system,its role includes: power transformation,transmission and distribution,etc.,if an accident will cause inestimable loss.At present,the development of computer and microelectronics technology makes fault diagnosis technology gradually improved,and many fault diagnosis methods have been applied to electrical equipment,therefore,the research of transformer fault diagnosis technology is of great significance.Under normal operation,the transformer produces a uniform buzzing sound with a frequency of100 Hz due to the magnetostriction of the iron core.When the transformer fails,the sound will also change.People who work in substations all the year round can hear whether there is a transformer fault by experience.Based on this feature,this paper mainly studies the 20Hz～20k Hz audible sound signal of transformer under partial discharge fault.Because transformers are mostly located in substations or in the field environment,it is inevitable that the noise of the surrounding environment will be mixed with the sound data collected in the field.This paper classifies the noise according to the characteristics of the mixed noise.For the classified unsteady noise and short-time steady noise,blind source separation algorithm and wavelet packet decomposition denoising are respectively used to reduce the noise,which better restores the sound of the transformer during normal operation.Secondly,this paper simulates partial discharge fault of transformer and designs three discharge models of pin plate discharge,along with the surface discharge and suspension discharge,carries out discharge experiment and takes the collected data as fault data.In order to extract the transformer partial discharge fault features effectively,this paper starts from the human ear model,draws on the feature extraction process of MFCC,and replaces the Mel filter with Gammatone filter to solve the problem of energy leakage.Finally,SPNCC is selected as the characteristic parameter and improved.Then,CNN and LSTM are fused according to the feature that transformer sound has local correlation in both time and frequency,and a dual input channel is set up to extract detailed features and overall features at the same time.Finally,this paper proposes a one-dimensional two-channel CNN-LSTM model based on deep learning.Experimental results show that the improved characteristic parameter SPNCC can better reflect the partial discharge fault characteristics of the transformer,and the average fault identification can reach 97% by combining with the one-dimensional two-channel CNN-LSTM model.Using SPNCC as the characteristic parameter and one dimensional two-channel CNN-LSTM model,different partial discharge faults can be distinguished well.The classification accuracy of needle plate discharge model,surface discharge model and suspension discharge model is 95.7%,98.1% and 96.9%,respectively.