Supervised modeling and monitoring for profiles

It is common in industrial and pharmaceutical industries that many process variables are measured successively over time and this generates multiple profiles of related variables. This research develops new techniques for extracting and utilizing knowledge from high-dimensional data in the form of (multiple) profiles. The profiles can be important to monitor and predict a corresponding response from a production cycle. Immediate practical applications are presented.;A reference method for analyzing this kind of data is Partial Least Squares Regression (PLS). PLS summarizes the profiles by extracting latent variables that capture the variation within the profiles as well as between the profiles and the response. PLS latent variables can be difficult to interpret and act upon. Two alternatives are developed for predicting a response from multiple high-dimensional profiles. The first method is Partitioned Partial Least Squares (PPLS) which is a modification of PLS that is especially useful when there exists a natural ordering of multivariate predictors. PPLS is an algorithm for selecting only the most relevant predictors, thereby building a more interpretable predictive model.;The second method is based on automated feature extraction and is motivated by the desire to extract features from profiles that have intuitive interpretation. Time alignment between profiles from different samples is an important challenge for the automated approach. A landmark based alignment algorithm is modified and profiles are approximated by piecewise linear segments allowing extraction of intercepts and slopes as intuitive features of the profiles. The least absolute shrinkage and selection operator (LASSO) is used to simultaneously select the most important features and build a predictive model for the response.