Research On Substation Slope Hidden Danger Identification Method Based On Semi-supervised Generation Adversarial Network

The safety of the substation slope is related to the normal operation of the power system.It is of great significance to monitor the health status of the substation slope and identify the hidden danger.In order to complete the task of substation slope hidden danger identification,the diagnostic algorithm is required to have both signal classification and abnormal diagnosis functions.After finding that the semi-supervised generated adversarial network meets this diagnostic requirement,this paper improves the traditional semi-supervised generated adversarial network to improve the ability of sample generation and the ability to correctly classify the time-frequency graph samples of vibration signals.In this paper,a framework is designed to identify the hidden danger of soil vibration of substation slope by using data self-expanding mechanism.When the signal is sent to the diagnosis algorithm,the diagnosis algorithm can determine the corresponding category of events that cause the current vibration signal.If the current vibration signal does not belong to the category included in any of the training sets,the diagnostic algorithm can tell the extent of the anomaly.Specific work from the following aspects.The selection of vibration sensor and the location design of measuring point were carried out,and the vibration data of slope soil were collected.Data sets are made based on multiple time-frequency feature extraction methods.The traditional deep learning method was used to compare the classification accuracy of slope vibration signals under four different feature extraction methods,and the optimal time-frequency feature extraction method was selected.In the same simultaneous frequency feature extraction method,different numbers of data sets are produced,which verifies the fact that expanding data sets can improve the classification accuracy,and highlights the necessity of data expansion.A conditional semi-supervised generation adversarial network is proposed to expand the data set.Firstly,the input noise and the generated sample are deeply bound.Secondly,a new generator loss function is proposed.These two improvements improve the ability of targeted expansion of data sets.It is found that the sample quality of conditional generated adversarial network is higher than that of traditional semisupervised generated adversarial network.This conclusion is verified by three methods:intuitive perception,quantitative calculation of image similarity and visualization of dimensionality reduction.It is proved that the classification accuracy of slope vibration signals can be improved by generating sample and expanding data set with high quality.This paper explores the reason why the accuracy of discriminator decreases due to the strong generating ability and studies the data self-expanding mechanism.The explanation given in this paper for the declining accuracy of the discriminator is that the realistic sample generation will interfere with the discriminator,so as to lower the confidence probability of the discriminator to the real sample.This explanation is verified by the distribution of error cases in the classification results.A data self-expanding mechanism based on Wasserstein distance is proposed to solve the problem of realistic sample interference discriminator in the training process of semi-supervised generation adversial nets,which improves the accuracy.