(268g) Utilizing Multi-Way Partial Least Squares Discriminant Analysis: An Industrial Case Study


This paper demonstrates the effectiveness of applying multi-way partial least squares-discriminant analysis (PLS-DA) to industrial processes.  Traditionally, multi-way partial least squares (PLS) can be used to generate regression models and to reduce the dimension of the variable space.  The models generated from this regression method are calculated based on the relationship of the dependent Y variables, also called quality variables, with the set of independent variables X.  The paper’s main contribution is focused on demonstrating how PLS-DA can be efficiently used in scenarios where the quality variables are not relevant or not available.  Therefore, these scenarios can be formulated as a classification problem.  An important outcome of this paper is to determine whether the conclusions are consistent, facilitating the analysis of the process chemometricians.  The methodology proposed is applied to an industrial case study, in which a prompt verification is required utilizing the traditional multi-way technique of principal component analysis (PCA).

The case study is based on a continuous reactor whose performance has deteriorated.  One key factor that influences the performance of the reactor is the way in which the start-up is performed.  Utilizing historical data from previous runs, the objective is to identify if there is abnormal behavior in the variables that are related to the reactor.  Only the start-up time frames for the entire set of runs available are grouped and organized.   Each “batch” is defined to be a start up run.  To solve this problem, a suitable methodology is defined that is comprised of the following three steps:  (1) a classification criterion is defined such that batches are classified as good or bad based on their performance in the case that quality variables are not obtainable or do not provide relevant information for the analysis; (2) the data set is unfolded batch-wise to two-way data for further analysis; (3) using PLS-DA, variables that are not relevant to the classification criterion are eliminated.  A posteriori validation is performed by utilizing a PCA model.  Finally, in applying the methodology based on PLS-DA, a reduced set of variables that is causing the differences between the good and bad start-ups is identified and analyzed for process troubleshooting.