Designing learning control algorithms for mechanical systems with complex dynamics

Authors
park, Bumsoo
ORCID
https://orcid.org/0000-0003-0919-2507
Loading...
Thumbnail Image
Other Contributors
Kopsaftopoulos, Fotis
Samuel, Johnson
Paternain, Santiago
Mishra, Sandipan
Issue Date
2023-08
Keywords
Mechanical engineering
Degree
PhD
Terms of Use
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute (RPI), Troy, NY. Copyright of original work retained by author.
Full Citation
Abstract
Numerous mechanical processes are governed by complex multi-scale dynamics (e.g. manufacturing, robotic systems, and building thermal control), which often involve the interaction of various physical phenomena, making it difficult to derive suitable models for controller design purposes. In all of these systems however, the opportunity is that there is a massive increase in available data, with high-fidelity simulations, improved experimental measurement techniques, and increased computational power. Learning control techniques can leverage this data, offering the potential to overcome the challenges posed by complex dynamics in mechanical systems and optimize control strategies without relying solely on accuratephysics-based models. The work in this dissertation is a demonstration of the endeavors to develop intelligent control strategies for mechanical systems, particularly with complex dynamics, through the investigation of learning-based control strategies in two representative applications, namely (1) passive building thermal control and (2) metal-additive manufacturing (laser powderbed fusion). These applications have been chosen due to their inherent complexity, as the dynamic behavior of these systems is spatially and temporally distributed, governed by principles of mass and energy conservation, as well as phase changes, posing challenges in effective controller design. However, they also present a significant opportunity to improve control performance and system efficiency through the implementation of learning control approaches. For the first application, passive building thermal control, this thesis presents a robust control strategy that harnesses climatic resources to minimize the use of mechanical heating/cooling energy. The algorithm is developed with a focus on reducing training efforts and enhancing the adaptability of the learning control algorithm. To achieve this, first, an approach to incorporate domain knowledge in the form of an expert is introduced, followed by the design of the learning control algorithm. The expert is then used to assist the initialization of the learning control algorithm, such that the controller mimics the behavior of the expert to serve as baseline, accelerating the training process and enhancing the initial performance. The expert-assisted initialization also serves as a means to reduce undesirable behavior and enhance the performance in the final controller. Next, the baseline controllers are deployed in contrasting climates to further learn a representative control strategy for each climate. The developed learning control algorithms are tested in similar yet unforeseen climate and building conditions, in which we find that the control algorithms are able to adapt to climate and building-specific variations. The expert-assistance also proves to be an efficient means to reduce training efforts, and greatly enhance the performance of the developed controllers. In the second application, metal-additive manufacturing, a control strategy to effectively regulate the measurements from the process is presented. Methods to safely develop a learning control algorithm that is capable of robust control in the physical system are investigated. On pursuit of this goal, a sim-to-real (simulation-to-reality) learning control approach is proposed, involving three key steps. First, a physics-informed model is developed to replicate the system dynamics and serve as a basis for training the learning control algorithm. This model provides insights into anticipated measurement deviations and enables the controller to respond to prior deviations that can be compensated. Subsequently, the algorithm is trained using the developed model, reducing training efforts and ensuring the safe development of the control strategy. Finally, the trained control strategy is deployed in the physical system, and its performance is evaluated under various build conditions. Upon experimental deployment, we find that the algorithm is applicable to novel build geometries without further tuning or modification, showcasing the geometry-informed capabilities of the algorithm. Furthermore, the sim-to-real approach serves as an competent strategy for the mitigation of training time and safety issues.
Description
August2023
School of Engineering
Department
Dept. of Mechanical, Aerospace, and Nuclear Engineering
Publisher
Rensselaer Polytechnic Institute, Troy, NY
Relationships
Rensselaer Theses and Dissertations Online Collection
Access
Restricted to current Rensselaer faculty, staff and students in accordance with the Rensselaer Standard license. Access inquiries may be directed to the Rensselaer Libraries.
Collections