The Control Systems group has a long tradition in modeling, system identification and control of dynamical systems. The group aims to be a nationally and internationally acknowledged center in mastering complexity of dynamic systems. Currently, the Control Systems group focuses its fundamental research around model approximation, networked systems, model predictive control and spatial-temporal systems.
The group covers a wide variety of applications in modeling and control system design in projects. These include power networks, electromechanical systems, automotive systems, chemical production processes, and various applications in the process industry. We disseminate our expertise and knowledge to students, the scientific community and to industry. Currently about 30 people work at the Control Systems group, including postdocs and PhD students.
Project description
Linear Parameter-Varying (LPV) systems are flexible models capable of representing nonlinear/time-varying dynamical systems in terms of a linear structure. Signal relations in this structure depend on a so-called scheduling variable, which embeds time-variance, nonlinear dynamical aspects, etc., into the behavior of the LPV model.
The LPV framework provides computationally efficient and robust control-synthesis approaches for nonlinear/time-varying systems — making it attractive to chemical-process and high-tech mechatronic applications. However, systematic LPV modeling based on measured data or first-principle knowledge is still unresolved — a widely recognized shortcoming of this promising theory. Currently it is not understood how (1) to achieve an exact and low-complexity embedding of a nonlinear behavior into an LPV structure, (2) to find optimally parameterized representations of LPV behaviors, and (3) to efficiently estimate them based on measured data.
This Phd project aims to surmount these challenges by establishing an innovative synergy between the Machine Learning (ML) and LPV frameworks. The aim is to develop computationally efficient model learning approaches capable of supporting control synthesis. The emerging ML framework provides powerful data-driven approaches to facilitate non-parametric learning of complicated data-relations. The flexibility of the ML framework in defining learning objectives (aim-relevant estimation) and its ability to facilitate optimal recovery of structural relationships (model structure selection) provide novel perspectives in terms of developing dedicated methods to solve the limiting problems (1)-(3) of the current LPV theory.
The results of the fundamental research will be applied to modeling problems in complex physical/chemical and/or electrical/mechatronic systems as e.g. high-purity distillation columns and high-performance positioning applications.
PhD position: Data-Driven Learning of Linear Parameter-Varying Models
