Please use this identifier to cite or link to this item:
|Title:||The application of partial least squares to problems in chemical engineering|
|Authors:||Kresta, Vaclav James|
|Advisor:||MacGregor, John F.|
Marlin, Thomas E.
|Keywords:||Chemical Engineering;Chemical Engineering|
|Abstract:||<p>The objectives of this work were to investigate the applicability of using Partial Least Squares (PLS) in three areas of chemical engineering: (1) for performance monitoring, (2) for inferential model development, and (3) in dynamic model identification. One further goal of this work was to gain understanding of the properties of the PLS method in the context of chemical engineering. Properties investigated included: (1) effects of scaling, (2) the objective function in the algorithm, and (3) the configuration of the X and Y data blocks. The first application of PLS was in the area of performance monitoring of a multivariate processes. A multivariate Statistical Process Control (SPC) procedure has been proposed for handling large numbers of process and quality variables. PLS is used to reduce the dimensionality of these large and highly correlated data sets down to a few latent variables which contain most of the information about the process behaviour under normal operating conditions. By compressing all the information on the process down to low dimensional spaces, and using simple plots of the data in these spaces, together with meaningful control limits, the essential idea and philosophy of Shewart's (1931) SPC methods have been preserved and extended to handle the large number of variables collected in most process industries today. PLS has also been shown to be a very powerful approach to inferential model building when large numbers of highly correlated measured variables are available. By retaining all the measurements without overfitting the data PLS is able to utilize all the information obtained from the process measurements. Further, PLS models are extremely robust to missing data and sensor failures, an important feature of any inferential control scheme. The importance of obtaining a representative reference data set, (i.e., one that represents the process relationships for the expected range of process conditions, and the structure of the process and control configuration under which the model is to be used), when using empirical models has also been demonstrated. The comparison of PLS, Ridge Regression (RR) and Least Squares identification of non-parsimonious (Finite Impulse Response (FIR) and Autoregressive exogeneous (ARX)) models for multiple input single output (MISO) systems shows that the biased estimates obtained from PLS and RR provide smoother estimates of the impulse and step weights. The difficulty in characterizing structured noise when using these non-parsimonious models is discussed. Finally, the use of PLS for identification of multiple input multiple output (MIMO) processes was investigated.</p>|
|Appears in Collections:||Open Access Dissertations and Theses|
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.