Research On On-line Correction Methods Of Model Based On Statistical Theory

In the continuous process of industrial production, it is difficult or even unable to set up strictly accurate mathematical model in order to represent the real industrial process. Thus, there must be errors of the model. Besides, even if model can describe changes of process accurately, when model put into operation, due to the time-varying and nondeterminacy of system, the error of model maybe increase as time goes by. Model correction is a normal method to solve the problems mentioned above, model correction can always make the model follow the latest process changes. Thus model correction has important significance to systems running correctly.Based on the analysis of process modeling and model correction, the author take advantage of the statistical method to model, model monitoring, model on-line correct and do simulation experiment.This thesis introduces modeling methods and monitoring methods based on data at beginning and focus on the research of partial least squares algorithm (PLS) and principal component analysis algorithm (PCA). Select a combined index that combines the squared prediction error (SPE) and T2 to monitor the model, the combined index is concise in practical application. A new improved PLS algorithm based on the combined index is proposed, named combined index based block-wise RPLS with forgetting factor. Firstly, the algorithm update the PLS model with a certain size of data. Secondly, in order to prevent data saturation, keep a certain size of data by controlling the number of data-block. Thirdly, forgetting factors maintain influences of new data and reduce the confidence coefficient of old data. Finally, use the combined index to calculate forgetting factors, it can adjust the forgetting factors according to the model status. With the dynamic adjustable capacity of the improved PLS algorithm, an on-line model correction strategy is proposed.Select the Tennessee-Eastman process as the simulation object. With the simulation experiment, the superiority of the new improved PLS algorithm and on-line model correction strategy which are proposed in this thesis is proved.