A guide to independent component analysis : theory and practice

Abstract

<div lang="en" class="abs"><p class="abs_header">Abstract</p><p>Independent Component Analysis, or ICA, is a new technique for visualizing measured data. Appropriate data representation is necessary for the extraction of important information that is hidden in the measurements. Feature extraction and data reduction are two of the most important results of ICA that enable the application of the technique in many different fields of engineering. This report presents the theory and practical application of ICA in a way that is easy to understand for engineers from various kinds of expertise and works towards a step-by-step guideline for practical usage of ICA.</p><p>ICA assumes a data model that states that a collection of measurements originate from a number of independent sources that are linearly mixed. Based on higher order statistical properties of the measurements, ICA is able to estimate these independent sources or Independent Components (ICs) very accurately. It turns out that ICA is able to separate linearly mixed data from any distribution, but the Gaussian. This being nongaussian is what enables ICA to find the ICs. </p><p>In this report, ICA is applied to two cases. The first case is about prediction of failures in a screw insertion process. The second one is about the prediction of an upcoming paper break in a paper factory. </p><p>This work is the result of the three months internship of Jelmer van Ast at the Control Engineering Laboratory of the Department of Process and Environmental Engineering at the University of Oulu. He is an MSc student at the Delft Center for Systems and Control of Delft University of Technology under the supervision of Prof. Robert Babuška. The research resulting in this report has been done under the supervision of Mika Ruusunen. </p></div>