Extraction of optimal control patterns in industrial batch processes based on Support Vector Machines

In general industrial processes are complex systems that have to be optimized due to several performance criteria. As control optimization based on the development of physical models is in many cases very time consuming and cost intensive or not even feasible an alternative consists in analyzing historical process data, as lots of it is usually available. Hereby an approach is to use computational intelligent methods as these can identify characteristic control patterns and classify them according to their contribution for process optimization. As data coming from these processes is in general only available as time series a crucial question is how to generate and find these significant features from the measured data concerning a prior defined performance index. Traditional ways of finding these features need plenty of knowledge about the underlying process. In this paper we propose an automated feature extraction and ranking methodology based on Support Vector Machines for Regression (SVR). Furthermore the methodology is used to derive a model of the performance index in terms of the relevant features within the algorithm. The found model can finally be used to find a suitable control pattern of the complex system. The proposed concept is applied to a real world industrial batch process.