Research On Intelligent Fault Diagnosis Method Based On Generative Adversarial Networks

With the automation development of industrial system,intelligent fault diagnosis technology is an important guarantee for its safe production.Meanwhile,with the improvement of industrial data acquisition technology,the data-driven artificial intelligence method is more suitable for the research of fault diagnosis technology.However,in the actual production operation,due to the complexity of fault working conditions and the need for a large number of marker samples for model training,these problems will lead to the imbalance between the acquired historical data,and the fault diagnosis model based on unbalanced data will lead to a large misdiagnosis rate among the few samples.Aiming at the problem of unbalanced data set in fault diagnosis,this paper develops research on fault diagnosis under unbalanced data based on methods of generative adversarial network,convolutional neural network and recurrent neural network.Aiming at the problem that the data quality of the boundary equilibrium generative network decreases with the increase of data diversity,the Kullback-Leibler divergence loss between the real data and the low-dimensional feature distribution of the generated data is added to the generator loss function.At the same time,the hyperparameter theat is introduced to dynamically adjust the KL loss,which improves the model's learning of low-dimensional features,enables generator training to obtain more potential information of real data,and improves the quality of generated data and the rate of model convergence.Aiming at the continuity characteristics of multivariate time series data,the recurrent neural network has a front-to-back dependency between processing time series data.Combining the gated recurrent neural network and the autoencoder model,a GRU-BEGAN model for generating multivariate time series data is proposed.On the input of the generator,additional time information is added to the input variables to generate time-matched time series data.In order to improve the quality of the generated data,the low-dimensional feature vector of the real data obtained by the discriminator encoder is introduced into the generator as a hidden variable,So that the generator contains information that can be explained to the real data.At the same time,in order to control the degree of similarity between the generated data and the real data,the hidden variables are randomly sampled with probability,so that the generated data has a certain diversity.On the open source data set MIT-BIH ECG timing signal data,the proposed GRU-BEGAN model has good performance in the evaluation index of generating time series data.Based on the unbalanced fault data set of the thermal and hydraulic system of nuclear power plant,a small number of samples are generated according to the proposed GRU-BEGAN model,and then the original unbalanced data set is expanded to reach a balanced state.The 1D-CNN and 2D-CNN fault diagnoisis models based on convolutional neural networks are established respectively,the experimental results show that the accuracy of the fault diagnosis model after data expansion has been significantly improved.