Research On Partial Discharge Recognition Method Of High Voltage Switchgear Based On Ultrasonic Signal

The widely used high-voltage switchgear plays a vital role in the electric power system.Only timely judgment and identification of the switchgear partial discharge status can ensure the reliability of system operation,for this reason,this paper researches the identification method of high-voltage switchgear partial discharge based on ultrasonic signals.In this paper,the experimental platform of partial discharge of high-voltage switchgear is built,the ultrasonic sensing module and data acquisition module as well as the typical discharge model are designed,the experimental plan is formulated,and the ultrasonic signals of partial discharge under different working conditions are obtained by using different discharge electrodes,changing the space of discharge electrodes and the voltage of industrial frequency test in the experiment.In order to avoid the interference of noise on signal feature extraction,this paper proposes a denoising algorithm for partial discharge ultrasonic signal based on adaptive filtering,wavelet packet decomposition and genetic algorithm,which does not need to anticipate the a priori statistical knowledge of the signal,and the processing of time domain waveform is more refined than other noise reduction methods such as frequency domain filtering.Experimental results show that this denoising algorithm has good results even when the signal is drowned by noise.In order to obtain the signal recognition feature,this paper preprocesses the denoised partial discharge ultrasonic signal,extracts the feature vector and time-frequency spectrogram,and performs dimensionality reduction on the mixed feature vector.In order to realize the partial discharge recognition of ultrasonic signal of high-voltage switchgear,three partial discharge recognition models combining signal feature vector with support vector machine,convolutional neural network and long-short time memory network are given in this paper,and the performance of each model is investigated by using confusion matrix and accuracy comparison,and the influence of feature vector on different discharge types and recognition models is analyzed.The results show that the inverted Mel-log spectrogram and the deep learning model are more suitable for partial discharge recognition in terms of accuracy and recognition speed.