Research On Discharge Recognition Of Switch Cabinet Based On Deep Learning

With the rapid development of China's electric power industry,the insulation performance of the switchgear has also received a huge challenge.The real-time monitoring method of the insulation state has attracted people's attention and thought.For a long time,China's power sector has used regular maintenance,off-line maintenance and other methods to avoid insulation accidents,which is particularly inconvenient for enterprises that rely on electricity and areas where the city is steep.A method based on voice recognition has appeared in recent times.In this paper,the discharge sounds of several typical switchgear insulation components are collected on site,and the characteristics of short-time energy,wavelet energy,linear predictive cepstrum coefficient and Mel cepstral coefficient and their extraction methods are studied,and supervised learning through deep learning is studied.The discharge sound is identified.Firstly,this paper preprocesses the collected sound signal of the switchgear.It mainly includes emphasis,framing and windowing.Pre-emphasis can reduce the difference in the magnitude of the discharge sound itself during the acquisition process,eliminating the influence of the difference in volume and intensity on the quality of the discharge sound,and each segment can highlight its own characteristics.Framing and windowing can obtain a linear and smooth discharge sound short-time signal.Secondly,in order to improve the accuracy of recognition,it is necessary to increase the number of training samples.In this paper,the experiment has added the imitation of the impact sound,the background noise recorded on the scene as the interference item.At the same time,the real environment is simulated when the discharge occurs,and the background noise is incorporated.The accurate signal extraction is carried out by the aliasing sound processing method proposed in this paper,and the environmental adaptability and stability of the whole system are improved.Then,for the pre-processed discharge sound signal,the wavelet energy distribution of the discharge sound is extracted in the time domain and the frequency domain,respectively,in the feature extraction phase.According to the proportion of energy of each node,the characteristic parameter extraction of the wavelet energy is determined;the prediction coefficient is solved by making the mean square error of the error between the sampling value of the discharge sound signal and the result value of the linear prediction algorithm,and the prediction coefficient can accuratelyrepresent The actual discharge sound,which reflects the signal characteristics of the real sound;simulates the characteristics of the human ear's perception of the sound frequency,and uses the filter to map the spectrum of each frame of the sound signal from the linear frequency domain to the Mel frequency domain.Then,the Mel spectrum of the filter output is logarithmically transformed,and then mapped to the cepstrum domain by discrete cosine transform to complete the extraction of MFCC feature parameters.Finally,based on the characteristics of the discharge sound extraction,the construction and training of the discharge sound recognition model is completed based on the deep learning theory.In the training phase,a convolutional neural network model is constructed,and the appropriate kernel function type is selected considering the number of network layers and the size of the convolution kernel,so as to obtain the best training model and complete the determination of the category of the test sample.