Research On GAN-based High-voltage Reactor Fault Vibration Signal Generation Technology

With the development of social economy,the power demand capacity continues to increase,and high-voltage reactors have been widely used in power grids.The application of data-driven deep learning in fault signal recognition of electrical equipment,to overcome the shortcomings of traditional fault signal recognition methods,is of great significance for improving the safety of power systems.Due to the low data value density of the high-voltage reactor,that is,the amount of data under the fault state is small,so deep learning is used to enhance the data of small sample data,which solves the problem of low-quality fault signal recognition of high-voltage reactors,and becomes the key of deep learning in the deep research and development of fault signal recognition of high-voltage reactors.In this paper,in order to study the BKD2-140000 / 800-110 reactor fault vibration signal recognition and solve the problem of small data amount under the fault state,a convolution neural network fault vibration signal recognition model is established;a one-dimensional DCGAN is used to generate fault vibration signal to achieve data enhancement;a one-dimensional CNN is proposed to realize fault vibration signal recognition of the reactor.After experimental verification,the method proposed in this paper can effectively solve the problem of BKD2-140000 / 800-110 reactor fault vibration signal data augmentation and recognition.This article mainly completed the following two aspects of work:1.It is proposed to use one-dimensional DCGAN to generate high-voltage reactor fault vibration signals.DCGAN is difficult to converge in the actual training process.From the perspective of the loss function,the reason for the instability of the training process is analyzed,and measures for stable training are given.It is proposed to use one-dimensional CNN to identify fault vibration signals of high-voltage reactors.Simulation experiments verify that the one-dimensional CNN can effectively identify the fault vibration signal of the high-voltage reactor,and the one-dimensional signal DCGAN enhanced data method,compared with the traditional method,can effectively improve the high-voltage reactor fault recognition rate.2.According to the BKD2-140000 / 800-110 reactor actual fault vibration signal data,use one-dimensional DCGAN and one-dimensional CNN to complete the generation and identification of the reactor fault vibration signal,and compared with the traditional method to improve fault vibration signal recognition rate by about 3%,which verifies the effectiveness of the method of generating and identifying the fault signal of the high-voltage reactor proposed in this paper.