Multivariate Methods for Process Data Analysis - A Batch Process Application

摘要

A typical feature of modern developments in the chemical industry is the continuous effort to further reduce costs while saving natural resources. One aspect of this development is the implementation of prcess information systems which facilitate access to process data. Modern process information systems record and store chemical process data at virtually any rate and have now become standard for equipping new and retrofitting old plants. With the introduction of this technology, the question emerges of how to make use of the process data to better understand and improve existing processes with respect to quality and costs. One of the characteristics of real-world process data is the large number of highly correlated process variables, a fact which stresses the need for methods for handling these collinearities. Experience from the last decade has shown that methods from multivariate statistics can efficiently deal with this kind of data. Originally developed by Pearson [1] and Hotelling [2] to investigate sociological questions, these techniques have become a leading standard for analyzing high-dimensional, messy and ill-conditioned data. Common to all multivariate techniques is that they project the original high-dimensional variable space onto a lower dimensional subspace which can later be investigated with simple graphical techniques. This dimension reduction is one of the reasons why multivariate techniques such as Principal Component Analysis (PCA) [3] and Partial Least Squares Regression (PLS) [4-6] have become an essential part of the tools for analyzing process data.