Multi-System Learning Control - MultiSysLeCo

01 April 2018 → 30 June 2022
Regional and community funding: various
Research disciplines
  • Engineering and technology
    • Computers and logic systems
    • Motion planning and control
Learning Control
Project description

Machine controllers have to evolve to keep up with current mechatronic trends. First, machines are getting more complex, involving more complex or simply more subsystems, requiring their controllers to evolve accordingly. Second, the number of tasks and conditions under which machines operate is increasing, for example due to mass adjustment, making it more difficult for a single controller to deal with this variation, as well as making tuning this controller more difficult. Finally, more and more additional requirements need to be considered, such as energy-optimized controllers instead of satisfactory lightweight structures, ease of maintenance and tuning, ... all of which require a more complex controller and more difficult control design. With classic control, these challenges are tackled by experienced control technicians. However, this is a difficult job and can result in sub-optimal solutions with poor performance or in very long and costly tuning procedures. Learning control can be used as a more systematic approach to meeting these needs. This learning is currently done on a system-by-system basis, so it is repeated independently for each system, with each controller attempting to independently cope with changes in conditions, without taking advantage of the similarity between the systems to speed up this process.

The new trend to interconnect mechatronic systems (directly or via the cloud) offers new ways to improve these learning algorithms: instead of learning by machine, we propose to learn for multiple systems in parallel by sharing information and research load, resulting in an overall learning algorithm that is more efficient (shorter convergence periods) and more effective (better performance across all systems, which is achieved through a monotonically improving learning process). Since there are no learning control techniques for multiple systems to date, we will start from the existing learning control with one system and expand it.

The development of multi-system learning controls will lead to improvements for a wide range of companies, such as faster commissioning of a new fleet of systems, faster commissioning of a single system when added to an existing fleet, an adaptation of a fleet after a software update ... There will also be improvements for systems operating in variable but similar conditions and for systems operating in the same environment, allowing them to adapt faster by sharing information. Furthermore, there will also be an advantage for facilitating technology providers and technical service providers who can make use of these techniques.

In addition to the direct impact for a wide range of companies, the proposed research is a first innovative step in a new field, as there is currently very little activity in this field, even at a global level, but one that we believe is an important research topic will be in the years to come.