Supervision of Industrial Processes using Self Organizing Maps

Background

In the last decade, there has been an increasing interest in visualization techniques for multidimensional data (Card et al., 1999; Vesanto, 1999; Keim, 2002; Ferreira de Oliveira et al.; 2003; Kreuseler et al., 2002). Information visualization approaches have been successfully used in scenarios similar to that described in the introduction, allowing to integrate vast amounts of data with available prior knowledge in a user centric fashion, by transforming data and information into a visual representation to exploit the pattern recognition capabilities of the human visual system.

One approach for multidimensional data visualization is the so-called dimensionality reduction (DR) approach. The DR approach consists of obtaining a dimension reduction mapping which projects, in a continuous way, points from the data space onto a 2D or 3D space, which can be visualized, without losing significant information. Once all information is presented in a visual way, the human ability for detection, reasoning and inference with complex patterns can be exploited for process understanding and discovery of new knowledge. These mappings serve as a bridge between the high dimensional space of process data and 2D visualizations –actually maps of the process–, that allow the human to operate with reasoning tools valid in the data space, such as rules or theoretical models –e.g. mechanical or electrical equations– in an insightful way, boosting his ability to find new relationships and, in sum, to discover new knowledge.