Dynamic Soft Sensor Method With The Time-delayed Variables Selected

Soft sensor can realize the real-time estimation of primary variable that are difficult to be measured on-line by using easy to be measured process variables(Secondary Variable)to construct inference model and has attracted significant attention in industrial processes.Present adopted static soft sensor methods ignore the dynamic characteristic of processes.Thus constructing dynamic soft sensor model which can effectively consider process dynamic information is quite important.Currently,the usual practice is to introduce the time-delayed variables of each secondary variable to reflect the dynamic information of process.However the key problem is which time-delayed variables should be selected.In this paper,the time-delayed variables selection and model building of dynamic soft sensor method are studied.Aim to appropriately select time-delayed variables to informatively obtain process' s dynamic information.Three dynamic soft sensor methods are improved.Furthermore,the problems of dynamic soft sensor methods application in debutanizer column are studied.The detailed content is arranged as follows:In order to deal with time delay as well as dynamic characteristics exist in industrial process systems,this paper puts forward a dynamic soft sensor method based on joint mutual information.In the proposed method,maximizing the joint mutual information is taken as the object to optimize the time delay and historical data length of each secondary variable by intelligent optimization algorithms.Naturally,the time-delayed variables which containing the process' s time delay information and dynamic information can be selected.After that,soft sensor models based on selected time-delayed variables are constructed.As above process,the determination of optimal parameters is independent of the following modeling algorithm.So the algorithm for constructing soft sensor models after selecting time-delayed variables can be freely chosen according to the nonlinear degree of the specific application.The simulation results on predicting butane concentration in the bottom of debutanizer column verify that this method can obtain good prediction accuracy.To cope with the issue of dynamic characteristic in industrial processes,this paper proposes a soft sensor method based on dynamic orthogonal forward regression(dynamic OFR).By introducing time-delayed variables of each auxiliary variables variable,the proposed method augments the input variable matrix of soft sensor model.Then OFR is applied to the augmented input variable matrix.The meaningful time-delayed variables which can well explain primary variables are selected automatically in the way of orthogonal forward selection.At the same time,a sparse soft sensor model is thus constructed based on these time-delayed variables.The proposed soft sensor is applied to butane concentration in the bottom of debutanizer column,and the simulation results reflect the superiority of the proposed method in terms of prediction accuracy and the computational complexity.Considering the nonlinear process widely in actual industrial process,a revised soft sensor method based on dynamic orthogonal forward regression is proposed.This make the soft sensor method based on dynamic orthogonal forward regression suitable for nonlinear process.Firstly,the forward selection method is adopted to select significant time-delayed variables in the augmented input variable matrix.Secondly,the selected time-delayed variables are projected to high-dimensional feature space by kernel method,and then the orthogonal forward regression method is applied to high dimensional space data.Finally,the useful time-delayed variables which consider the nonlinear information of process can be chosen and the dynamic soft sensor model which is sparse in both the original variable space and high dimensional variable space can be received.To use this method to real-life debutanizer column process,the simulation results show the advantages of applying this method to the nonlinear debutanizer column.