Research On Fault Diagnosis And Data Filling Algorithm Based On Deep Learning

There are a lot of multivariate time series data in industrial systems,and the appearance of missing values hinders advanced analysis such as mining,identification,and forecasting of multiple time series.The interpolation method of the existing data is requested to solve the missing value by the deletion method,statistical interpolation method,machine learning based interpolation method and generative interpolation method.However,these methods can not handle time information and multi-stage information well.In this paper,through the improvement and combining the generation model of the Bi-directional Gated Recurrent Units and the Boundary Equilibrium Generative Adversarial Networks,A Bi GRU-BEGAN network model is constructed to realize the effective interpolation of the missing values in the multivariate time series,and it is also obtained in the downstream fault classification task.Firstly,for the time characteristics of time series data,this paper uses the Bi-directional Gated Recurrent Units（Bi GRU）to effectively generate time series data.By deriving and analyzing the parameter update process of the forward and backward reverse gradient derivation of the Bi GRU network,the Bi GRU-BEGAN network model can effectively deal with the gradient disappearance problems in the training process,and the effectiveness of the model in the task of generating data is determined by calculating some statistical indicators and regression indicators.Secondly,in view of the time characteristics of time data,this article analyzes the specific optimization goals of the loss function for the process of self-encoding and reconstruction of the data by the Boundary Equilibrium Generative Adversarial Networks,furthermore,it is proposed to add the ₁L norm between the generated data and the original complete data to the loss function of the generator,and detailed analysis of the proposed Bi GRU-BEGAN network model and step flow.Through the derivation of the parameter update of the Wasserstein distance loss error in the gradient descent process,the advantages of the model are explained.Finally,conduct experiments on the Bi GRU-BEGAN network model and several other types of baseline models on the nuclear power plant thermal and hydraulic system simulation data set,the experiments show that the model is effective both in interpolation accuracy and in subsequent fault diagnosis tasks.We also conducted experiments on filling and fault classification tasks on data with different missing rates.Experiments show that the Bi GRU-BEGAN network model can maintain a certain degree of adaptability when the missing data accounts for a relatively high proportion.