Partial least squares regression applied to two chemical processe

Chemical processes often have many variables that are being monitored every minute or every second. Often times this can result in "data overload" and useful information that is buried within the collection of data is lost. Techniques that provide a quick method of extracting information from large sets of data can prove to be very beneficial.;Partial Least Squares (PLS) is a regression technique that utilizes principal components analysis (PCA) to reduce the dimension of regression problems. Sets of data can often be broken into two categories: process variables (predictor variables) and product quality variables (predicted variables). PLS attempts to find factors that capture variance in the predictor variables and achieve correlation between the predictor variables and the predicted variables. These factors are then used to regress the predicted variables onto the predictor variables.;PLS$sb{-}$Toolbox 2.01a (which runs in MATLAB$sp{rm TM}$) was used to analyze data from two chemical processes, Process 1 and Process 2. Process 1 consisted of two predicted variables and 33 predictor variables. Process 2 contained one predicted variable and 19 predictor variables. Three regression techniques (PLS, Principle Components Regression (PCR), and Multiple Linear Regression (MLR)) were used to build regression models for each process and the results from each were compared.;PLS provided the best regression model for Process 1. The two predictor variables, Y1 and Y2, were predicted with less than 2.9% error and 9.9% error respectively. None of the regression models were able to provide a good prediction of the predicted variable for Process 2. The models did not detect a substantial change within the predictor variables. It was concluded that the variable (or variables) contributing most heavily to the variation in the predicted variable was not being measured.