Unsupervised pattern recognition methods for exploratory analysis of industrial process data

Abstract

The rapid growth of data storage capacities of process automation systems provides new possibilities to analyze behavior of industrial processes. As existence of large volumes of measurement data is a rather new issue in the process industry, long tradition of using data analysis techniques in that field does not yet exist. In this thesis, unsupervised pattern recognition methods are shown to represent one potential and computationally efficient approach in exploratory analysis of such data.

This thesis consists of an introduction and six publications. The introduction contains a survey on process monitoring and data analysis methods, exposing the research which has been carried out in the fields so far. The introduction also points out the tasks in the process management framework where the methods considered in this thesis - self-organizing maps and cluster analysis - can be benefited.

The main contribution of this thesis consists of two parts. The first one is the use of the existing and development of novel SOM-based methods for process monitoring and exploratory data analysis purposes. The second contribution is a concept where cluster analysis is used to extract and identify operational states of a process from measured data. In both cases the methods have been applied in exploratory analysis of real data from processes in the wood processing industry.