Data-driven optimization for dynamical systems
Wouter Weekers defended his PhD thesis at the Department of Mechanical Engineering on June 13th.
In an increasingly automated and interconnected world, the demand for fast, accurate, and reliable machines, from precision robots to medical equipment, is growing rapidly. But behind the scenes of these high-performing dynamical systems lies a complex challenge: how to select the right design parameters that will guarantee optimal performance. Traditionally, this has relied on building detailed mathematical models, which is a time-consuming task requiring expert knowledge and is prone to inaccuracies. In the PhD research of Wouter Weekers new model-free, data-driven optimization techniques were developed to address this problem directly at its core. The results are promising—and transformative. Read on to discover how these innovations can reduce design time, cut costs, and significantly improve system performance, particularly in high-speed industrial applications.
To fine-tune the performance of complex dynamical systems, like those used in manufacturing, logistics, and healthcare, engineers traditionally depend on mathematical models. These models describe how performance is influenced by design choices. However, real-world systems are rarely predictable. Variations in manufacturing, wear over time, and external disturbances introduce discrepancies between the modeled system and its actual behavior.
This mismatch means that even well-designed systems may underperform or behave unpredictably. Moreover, the development of accurate models requires extensive expertise and time, resources not always readily available. These limitations call for a new approach that goes beyond the traditional reliance on models.
From models to measurements: a data-driven approach
This research of Wouter Weekers introduces a paradigm shift: instead of depending on mathematical models, the new methods use only data collected during system operation. This ‘model-free’ optimization allows for more accurate and individualized parameter selection, reflecting the real behavior of each specific system rather than relying on theoretical approximations.
The focus is on optimizing control systems, the mechanisms responsible for ensuring systems operate accurately and efficiently. Choosing the right design parameters for these systems is critical for achieving high performance, but traditionally requires significant manual effort and domain knowledge. By automating this process through measured data, the approach not only reduces reliance on expertise but also tailors optimization to each system’s unique characteristics.
Tackling point-to-point motion challenges
A key application of these data-driven methods is the optimization of point-to-point motions, the fast and precise movements commonly required in industrial tasks, such as wire bonding in semiconductor packaging. In such processes, every millisecond counts.
However, optimizing the timing of these movements is particularly difficult. The duration of the motion is highly sensitive to small changes in design parameters, making the optimization problem both delicate and critical. The newly developed methods succeeded in tackling this sensitivity, resulting in significantly faster execution times without sacrificing accuracy.
These methods were validated on a real industrial wire-bonding system, and compared with two state-of-the-art techniques currently used in the semiconductor industry. The outcome? A substantial reduction in the time needed to perform point-to-point motions—demonstrating the clear advantage of the proposed data-driven approaches.
Extremum-seeking control meets machine learning
At the heart of these innovations lies a technique called Extremum-Seeking Control (ESC). ESC is a model-free method for performance optimization that iteratively adjusts parameters based on performance measurements. It’s flexible and broadly applicable, but there's a catch: ESC often requires a large number of measurements, which can be costly or time-intensive in practice.
To address this, the research of Weekers also presents novel ways to combine ESC with machine learning techniques, dramatically reducing the number of measurements needed. These enhancements retain the adaptability of ESC while improving efficiency, making the method more practical for real-world industrial use.
A new standard for performance optimization
This dissertation introduces a set of powerful, data-driven methods for optimizing control system design—without the need for detailed mathematical models. Special attention was given to improving the speed and accuracy of point-to-point motions in high-tech machines, which are vital to industrial processes. The methods proved robust, even when faced with extreme sensitivity in parameter-performance relationships, and delivered significant improvements when tested in real-world conditions.
Ultimately, by integrating Extremum-Seeking Control with machine learning, the research not only improves performance but also lowers the cost and effort of optimization. These methods represent a major step forward in the development of autonomous, efficient, and high-performing control systems—laying a solid foundation for future advances in robotics, manufacturing, and beyond.
Title of PhD thesis: . Supervisors: Prof. Nathan van de Wouw, and Dr. Alessandro Saccon.