Data-driven linear parameter-varying control

To meet with the challenging performance demands of next-generation mechatronic systems, machines in these sectors must exhibit complex nonlinear behavior such as position-dependent or operating condition-dependent dynamics.

Developing mathematical models for control that accurately describe complex dynamics is difficult based on first principles knowledge, or even data, and it is often subject to high model uncertainty. Designing controllers directly from measured system data can circumvent the involved modeling steps, shifting the focus directly on the control objective.

For his PhD research, Tom Bloemers sought to develop new ways to enhance the performance of mechatronic systems in the system. He looked at ways to the develop effective control strategies for complex systems, which also simplified the amount of modelling needed by using data.

Challenges

Currently, the performance of the available methods that use measurement data to design operating-condition-unaware controllers can be severely limited in the face of complex dynamics, with several challenges present.

First, to adapt to varying conditions of the system, an operating-condition-aware controller is required. Second, to avoid the modeling step, the controller should be selected such that stability and performance can be analyzed and guaranteed during control design based on measurement data without knowing the underlying data generating system.

Key advancements

In his thesis, Bloemers outlines the development of a data-driven control design framework that provides performance improvements for complex mechatronic systems by incorporating knowledge about the operating conditions. And to achieve this, Bloemers’ thesis describes the key advancements that he made in his research.

First, a parameterization of the controller is introduced that ensures adaptability and awareness of the controller on the operating conditions of the system.

Second, the modeling step is circumvented by substituting mathematical models with local non-parametric estimates that result from measured data, for which algorithms are developed to analyze and design controllers even in case of multiple inputs and multiple outputs.

Third, it is shown that the proposed controller parameterization together with a special realization algorithm provides global guarantees during operation (even under varying scheduling trajectories) in terms of universal shifted stability and performance. These results allow for reliable data-driven design of operating condition-aware controllers that enable performance improvements and wide-range stability guarantees for complex mechatronic systems compared to traditional designs.

Applications

Performance of the developed methods is demonstrated on applications ranging from laboratory-scale setups to industrial applications.

Bloemers shows that designing controllers directly from available data, while incorporating knowledge about the position-dependent or operating condition-dependent behavior significantly improves the performance of next-generation complex mechatronic systems.

Title of PhD thesis: Data-Driven Linear Parameter-Varying Control. Supervisors: Rotland Tóth and Tom Oomen.