Research On Time Series Modeling And Imputation Based On The Variational Auto-Encoders

As a common form of data storage,multivariate time series exists in complex systems in various fields of the real world and have been widely used in analysis.However,the problems of redundancy,irrelevant variables or incomplete data that may exist in the actually obtained data are often not conducive to further modeling,analysis and application.Therefore,it is of great practical significance to do feature extraction modeling for multivariate data and missing imputation for incomplete data.Taking multivariate time series data as the research object and variational auto-encoders(VAE)as the research method,this paper studies the problems of feature extraction modeling and missing imputation in practical application,so as to lay a foundation for further constructing high-precision analysis model.The research contents of this paper are as follows:In response to the "black box" problem that may arise from the traditional feature extraction methods,this paper considers that the latent space of VAE model has the ability of feature extraction and introduces mutual information(MI)theory.And a VAE model based on mutual information theory(MI-VAE)is proposed,which embeds the latent space as a feature extractor into the modeling process of multivariate time series.The mutual information theory introduced into the model is conducive to enhancing the ability of expressing the input data,and can effectively learn the essential features of multi-dimensional input.Finally,the experimental results show that the proposed model has better the ability of feature expression and higher reconstruction accuracy.In response to the problem of data missing of two types in nonlinear multivariate time series,this paper considers the ability of generation in the VAE model.It is taken as the basic imputation framework and two improved models are proposed.On the one hand,combined with shift correction(SC)method and extended it to the β-VAE,the model based on shift correction and β-VAE(SC-β-VAE)is proposed.It introduces hyperparameter λ through shift correction to correct the basic assumptions of the standard VAE model.At the same time,the hyperparameter β better balances the ability of generation and decoupling in the model,and can produce imputation results that are more consistent with the probability distribution of real data.Finally,the effectiveness of proposed model is verified on two real data set.On the other hand,combined with transfer learning(TL)and bidirectional long short-term memory(Bi-LSTM)network,a bidirectional VAE model based on transfer learning is proposed,which is called as transferred bidirectional VAE(T-Bi VAE).This model effectively alleviates the deviation of probability distribution caused by a number of data missing at random through transfer learning.The Bi-LSTM network introduced into the model is suitable for processing long time series and can make a more reasonable estimation for missing values.Finally,the effectiveness of proposed model is verified on one real data set.