Application of Computational Intelligence Techniques to Process Industry Problems.

Abstract

In the last two decades there has been a large progress in the computational intelligence research field. The fruits of the effort spent on the research in the discussed field are powerful techniques for pattern recognition, data mining, data modelling, etc. These techniques achieve high performance on traditional data sets like the UCI machine learning database. Unfortunately, this kind of data sources usually represent clean data without any problems like data outliers, missing values, feature co-linearity, etc. common to real-life industrial data. The presence of faulty data samples can have very harmful effects on the models, for example if presented during the training of the models, it can either cause sub-optimal performance of the trained model or in the worst case destroy the so far learnt knowledge of the model. For these reasons the application of present modelling techniques to industrial problems has developed into a research field on its own. Based on the discussion of the properties and issues of the data and the state-of-the-art modelling techniques in the process industry, in this paper a novel unified approach to the development of predictive models in the process industry is presented.