Traffic Data Imputation Method Based On Low-rank Tensor Completion And Tensor Decomposition

With the development and popularization of sensing technology and monitoring equipment,more and more traffic data can be obtained with high-resolution spatio-temporal characteristics which are widely used in Intelligent Transportation System(ITS),such as traffic flow prediction,trip planning and so on.However,with the massive collection of traffic data,the loss of traffic data due to the damage to collection equipment and weather is an inevitable common problem.Accurate imputation of missing traffic data is very important for many applications in intelligent transportation systems.We aim to impute the missing traffic data by transforming the problem of imputation the missing traffic data into a tensor completion or tensor decomposition problem.Therefore,this thesis mainly focuses on the task of imputation the missing traffic speed data and traffic flow data.Based on low rank tensor completion(LRTC)and Tucker decomposition technology,this thesis studies two main problems in traffic data imputation:(1)The data initialization problem of traffic data imputation,and proposes a high-precision initialization strategy on this basis;(2)A high-precision imputation model applicable to all kinds of missing situations in the real world.The main contributions of this thesis are as follows:(1)Aiming at the problem of data initialization in traffic data imputation,we propose a low-rank tensor completion model based on truncated nuclear norm minimization,and use an efficient Alternating Direction Method of Multipliers(ADMM)optimization algorithm to solve the optimization problem of the model.By setting flexible rank to truncate different dimensions of tensor,the hidden patterns in spatio-temporal traffic data can be better explored.On the real-world traffic data set,through extensive comparative experiments under five different missing conditions,it is proved that the proposed model has high imputation accuracy and good initialization performance.(2)A tensor decomposition model based on spatio-temporal regularization is proposed.The model firstly uses high-precision initialization method to pre-fill the missing original data,then carries out high-order truncated singular value decomposition(HOSVD)by setting the threshold of singular value ratio,and adds the regularization term related to spatio-temporal into the optimization model.Finally,the model is optimized by gradient descent optimization method.Extensive experiments are carried out on public data sets(Seattle,USA),and it is proved that the proposed model is superior to the most advanced deep learning method and tensor completion method in five cases similar to the real world.