System Identification of Linear and Nonlinear Systems
The availability of high accuracy, low complexity and validated dynamic models is of paramount importance in model-based operational strategies such as monitoring, control, estimation and optimization. This research focuses on experiment design, identification in open-loop, closed-loop, and in dynamic networks, as well as on parameter estimation in large scale physical models. Particular attention is given to network and communication aspects, and nonlinear and linear parameter-varying (LPV) models, including the relationship with machine learning.
Identification of Linear-Parameter Varying (LPV) models
Operating industrial process systems (distillation columns, pulp digesters, etc.) under varying operating conditions and the need to handle position dependent and nonlinear dynamics in motion control systems (e.g., wafer scanners) requires the development of modeling and control approaches capable of efficiently handling the system behavior in terms of an intermediate: still linear, but with varying dynamical representation. The objective of this research is to develop data-driven LPV approaches from experiment design to model validation that are capable of capturing the dynamical behavior of physical systems using measured data.
- Data-Driven Learning of Linear Parameter-Varying Models
Fusion of System Identification and Machine Learning
To systematically explore (learn) nonlinear, spatial or time-varying structural and dynamical signal relations in data-driven modeling of process and electromechanical systems, an innovative synergy of the state-of-the art of the data-driven modeling framework and techniques presented in image processing and statistical learning theory are aimed at. The objective is to develop computationally efficient model learning approaches capable of non-parametric learning of complicated data-relations, support of control design objectives and handling of uncertain prior knowledge of the dynamics.
- Learning and Adaption Based Control
- Data-driven modeling using Symbolic methods
Identification in dynamic networks
The problem of identifying dynamical models of particular modules that appear in a complex network structure is a new problem that has been formulated and now is being analyzed and treated in detail. The problem occurs in many fields of application, including power grids, systems biology, sensor networks etc. The identification on the basis of measurement data of a particular link in a dynamic network shows closely resembles the classical closed-loop identification problem. However in its structure it is richer as it also involves the selection of sensing and actuation locations, and possibly the detection of the cause-effect relationships.