Research On Variable Weighting Algorithm Based On Just In Time Learning

Due to the complexity,strong non-linearity,and time variability in the actual industrial process,the application of the global soft sensor model is greatly limited.Justin-time learning algorithm as a local soft sensor modeling method can well deal with such problems.The similarity measurement is the most critical step in the just-in-time learning algorithm,which largely determines the prediction accuracy of the model.Based on the locally weighted partial least squares regression,this work studies the Justin-time learning algorithm that weights variables with diversity of methods.The specific research work is as follows:First,aiming at the problem that the commonly-used similarity measure in the justin time learning algorithms only considers input variables,an output-relevant just-intime learning algorithm that weights variables is proposed,and development and comparison of just-in-time learning algorithms that weight variables based on correlation coefficients and regression coefficients are conducted.Numerical examples and the industrial process of debutane tower are used to verify that the just-in-time learning algorithm based on output-relevanted weighting can effectively improve the performance and robustness of the model.It is found that different orders of the weight coefficients have different influences on the model.For a given data set,there is a specific order that enables the model to obtain the best prediction results.Second,based on the real-time learning algorithm that weights variables according to output correlation,the effects of even-order exponent(or absolute values)of correlation coefficients or regression coefficients on model prediction performance are studied.Numerical simulation examples for time-varying processes and prediction of the hydrogen sulfide content in the sulfur recovery unit reveal the fact that increase of the order of the weight coefficient does not necessarily result in better prediction.In the case of more abnormal data,the absolute value of the weight coefficient is recommended and its prediction result appears to be the best.Then four commonly used weight functions are selected for locally weighted just-in-time learning.Experimental results show that for different industrial processes,different weight functions can be used to improve the model prediction performance.Therefore,in the actual industrial process,according to different data characteristics,selecting the appropriate order and weight function can maximize the prediction performance of the model,thereby improving the industrial production efficiency.Third,aiming at the problems of high-dimensionality,strong non-linearity and noise interference in industrial process data,a soft-sensor algorithm based on incomplete and high-dimensional data for nonlinear process is proposed.First,the probabilistic principal component analysis method is used to estimate the missing data,then the partial least square method is used to perform supervised dimensionality reduction on the sampled data,and finally the locally weighted partial least squares regression is performed in the latent space.Compared with the traditional locally weighted partial least squares regression method after dimensionality reduction by principal component analysis,through numerical examples and application on the debutane tower process,it is shown that the proposed algorithm can effectively improve the prediction performance of the model.Finally,weighted partial least squares regression algorithm based on diversity weights is proposed which can effectively reduce the dimensionality.At the same time,the effect of different orders of correlation coefficient on model performance is further studied.The effectiveness and precision of the algorithm is verified through numerical examples and the industrial process of debutane towers.It provides a guidance for further improving the robustness of modeling for industrial processes.