Multi-Condition Soft-Sensing Regression Modeling Based On Domain Adaptation

Due to limited conditions or physical measurement cost,it Is hard to directly measure some key parameters in industrial processes.Using auxiliary variables to build a data-driven soft measurement model for key parameters is an effective means.The traditional soft measurement method requires that modeling data and test data belong to the same distribution.However,this hypothesis in the actual process does not hold.Thus,the prediction of key quality variables under the changing distribution of process data becomes the focus of this paper.With the operation of the process,the history database stores a large number of data sets.The real-time learning method is used to search in the historical database for a similar sample of the current sample for building a local prediction model.However,the differences in operating conditions between the datasets in the database will reduce the modeling accuracy.Targeting this issue,the first part of this paper introduces a multi-source integrated real-time learning soft measurement model based on the measurement of operation condition differences.The method of public difference information extraction(JIVE)is used to extract the special process information,and the current processes used as the query domain to calculate the relative entropy characterization distribution difference between the special information in the current and previous processes and obtain the nearest neighbor domain.Finally,a multi-source integrated real-time learning model is established based on the nearest neighborhood and corresponding measurement information,and the weighted prediction results are generated.The above integrated method can obtain accurate and robust prediction results,without reducing the distribution differences between operating conditions.Hence,it is not applicable to the situation where the current process data and all historical data have great differences in operating conditions.In the second part of this paper,the domain adaptive regression method is adopted,and a C-time nearest neighbor structure preserving domain adaptive regression(TNN-RL-DAR)suitable for process data is proposed.Given the large sample size and the continuous and sequential nature of the process data,the domain adaptive regular term and C-time nearest neighbor Laplacian regular term are introduced to the nonlinear iterative partial least squares(NIPALS)method,which can significantly reduce the time of model training while effectively improving the results of domain adaptive regression.If operating data have different marginal distributions but the same conditional distribution,the above unsupervised domain adaptive model is satisfactory.However,when there are differences between conditional distribution,this model cannot achieve ideal results.In this paper,the error upper bound theory is introduced to analyze this phenomenon and a semi-supervised domain adaptation method is adopted to solve the problem of multiple conditions with different conditional distributions.Based on the TNN-RL-DAR algorithm,a C-time nearest neighbor structure preserved semi-supervised domain adaptive regression(TL-Semidar)for process data is proposed in this paper.In the experimental part,the effectiveness of the proposed method is firstly verified,and then the number of labeled samples in the target condition is discussed.Finally,the adaptive adjustment strategy for the penalty coefficient is put forward and verified.