Many chemical systems and processes in industry and research laboratories exhibit to some extent non-linear and non-stationary characteristics (like non-linear drift or irregular baseline shifts encountered in routine control procedures). Usually, for on-line quality control, such systems and processes are monitored, identified and controlled by standard statistical tools (like Shewart control charts) and (quasi-)linear modelling techniques (e.g., Auto Regressive Moving Average models and Kalman filters), respectively. The development of multivariate non-linear monitoring, identification and control techniques, which provide in potency a much better process or system description, is however still in its infancy: several problems remain unsolved (for example, the appropriate number of principal components to be taken into account for constructing so-called multivariate control charts for monitoring processes characterized by multiple variables) and hence, none of these novel techniques is applied on a wide scale yet in routine laboratories and (industrial) research environments. Moreover, the majority of both standard and multivariate non-linear techniques explicity assume stationarity of the considered dynamical system or process.

Due to their adaptive nature and multivariate non-linear modelling abilities, (recurrent) neural networks might serve as promising alternatives, suitable for monitoring, identification and control of multivariate non-linear non-stationary chemical processes. The proposed STW project aims to investigate the applicability of regular and recurrent neural networks in each particular subfield. Main target of the proposed project is to develop and implement a validated system of neural network models for monitoring, identification and control of a real-world chemical system or process.

This project will be performed in close collaboration with the Foundation of Neural Networks (SNN) at Nijmegen, the Netherlands.

Resultaten van het onderzoek