Control Systems

Reports on control applications of tomography are rare and the proposed approaches were never physically demonstrated. Typical for these is an extraction of one or two global control variables from images and an almost exclusive use of PID controllers. These controllers are intended for single input single output systems. Taking full advantage of the richness of tomography data is impossible with PID controllers. On the contrary, state space models allow a much more detailed representation of the data from tomographic sensors (e.g. velocity or concentration profiles). State space models are often also a natural system description resulting from the finite element (or similar) discretization of the underlying distributed parameters of the controlled processes. State space models make it possible to design optimal controllers and to make full use of the arsenal of modern control theory in general. However, controllers based on state space models and using tomographic sensor data have hitherto been tested in simulations only. Moreover, they have mostly been designed using the LQG approach. This approach has serious drawbacks such as limitation to linear or linearized models, questionable robustness in the presence of non-Gaussian noise and inability to respect constraints of actuators and other process components.

TOMOCON objectives

TOMOCON will focus on developing a methodology to derive an adequate controller structure from advanced multi-parametric physical process models and thereby foster the development of new process models, which incorporate both state-of-the-art physical simulation techniques and intelligent soft-computing methods for situations, where an accurate process model is lacking. Process models are not only helpful for controller synthesis but can also give important clues what is really significant in the tomographic data. Accordingly, another focus will be on the suitable incorporation of tomographic data processing into the controller structure. As real-time implementation is the final objective, particular emphasis will be on methods that avoid the computationally demanding and ill-posed image reconstruction using a suitable process model parametrization or state estimation instead. Approaches based on state space models will be used, in particular the model predictive control (MPC) due to its ability to respect process constraints.

Many process features have significant nonlinearity. It can be time varying and difficult to describe by first principles model. This may call for modelling the nonlinearity, e. g. by neural network and/or fuzzy models. Similarly the behaviour of electrical tomographic sensors includes some strongly nonlinear relations (e. g. the relation between conductivity and the measured physical variable). If strong nonlinearity prevents using linearized models in MPC, special variants of MPC will be chosen, which are less computationally expensive than full nonlinear MPC, such as MPC based on Wiener–Hammerstein models or MPC with nonlinear prediction and linear optimization.

Another important challenge is handling safety-critical states (e.g. reactor runaway) and situations with component failures, e. g. due to fouling in a reactor or waxing in an inline separator. Since process models will be available, discrepancy between model output and reality can be used to detect these states and trigger adequate controller response. The developed techniques for controller design will be applied to selected demonstration plants and their control performance will be tested not only in simulation but in real pilot scale applications. Hence, a key issue of this work package is the derivation of multi-dimensional optimization strategies and suitable control variables, the derivation of process-specific optimal controller structures, the derivation of key parameters to be extracted from multi-dimensional tomography data and simulation as well as the demonstration of control-loop stability, the achievement of optimum-criteria and the efficacy for unconventional process states. R&D work will mainly be driven by rich experience and expertise at Technical University of Liberec (transferring their expertise in model predictive control in power plant applications to a wider range of process classes), University of Eastern Finland (tomography-based control for continuously operated reactors) and Lodz University of Technology Lodz (advanced feature extraction and fuzzy control). Chalmers University of Technology will bring in expertise in the design and analysis of advanced human-machine interfaces with treatment of 4D process data.