Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Civil and Environmental Engineering | en_US |
| dc.contributor.advisor | Zhu, Songye (CEE) | en_US |
| dc.creator | Zhang, Yiang | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/14505 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Deep reinforcement learning empowered active vibration control strategies | en_US |
| dcterms.abstract | Although classical active control techniques demonstrate superior performance in controlling structural vibrations, deriving the optimal control strategy often faces challenges due to the difficulty or complexity in accurately modeling dynamic systems in practical scenarios. In contrast, machine learning based (data driven) control methods eliminate such a requirement by directly learning control strategies from structural behavior. However, past studies on learning-based control algorithms primarily focused on their applicability without guaranteeing optimal control performance, resulting in a notable performance gap between learning-based model-free vibration control and model-based optimal control. This raises a challenging question: | en_US |
| dcterms.abstract | How can optimal control performance be achieved without accurate knowledge of structural dynamics? | en_US |
| dcterms.abstract | To answer this question, this thesis established and developed a novel model-free approach step-by-step, from linear systems to nonlinear systems, and validated it through numerical and experimental examples. The initially proposed framework was based on neural networks (NNs) and deep reinforcement learning (DRL) to control both linear and nonlinear systems effectively. The proposed learning algorithm could determine the control policy through interaction with the environment (i.e., the actual structures) without knowing structural dynamic models. First, the proposed method was demonstrated on linear systems, including a single-degree-of-freedom (SDOF) model, and then extended to multi-degree-of-freedom (MDOF) systems (such as a shear building model). The DRL-based strategy achieved control performance comparable to that of classical model-based controllers, such as the linear quadratic regulator (LQR), even under conditions of partial observation and measurement noise. Therefore, the DRL-based strategy could provide optimal control without the need for detailed structural models, which has never been reported in the literature. Following that, the research extended to a nonlinear system, i.e., a shallow, simply-supported buckled beam. The findings illustrate the adaptability and efficacy of the DRL-based approach in preventing buckling under high loading levels without additional energy requirements. This work advances the current understanding of the DRL applications in structural control. | en_US |
| dcterms.abstract | Though the exploration of pure DRL methods in structural control reveals their potential, the inherent limitations of DRL, such as extensive training data requirements and instability in real-world applications, limit their practical applicability. A pretraining stage was added by utilizing Imitation learning (IL). This IL-DRL framework initially involved training an NN controller through behavior cloning of an extremely simple but stable control strategy, followed by fine-tuning using model-free DRL. The framework's feasibility and effectiveness were systematically examined in simulation by using a verified numerical model of a full-scale stay cable system with an active damper. The control performance of the proposed model-free method closely approached that of the model-based full-state feedback controller, even under conditions of a reduced number of observed states or measurement noises. Moreover, it demonstrated superior sample efficiency compared with other state-of-the-art model-free DRL algorithms. The combined IL-DRL framework was validated through a shaking table experiment using an active mass damper to control an SDOF structure. It was demonstrated that the proposed approach could successfully train an optimal NN controller in the experiment without using any structural information. The response of the controlled structure under different excitations was compared to a classical optimal linear quadratic Gaussian (LQG) controller. The proposed controller achieved more displacement and acceleration reduction with a similar control power requirement. The results show that combining IL and DRL can significantly improve the sampling efficiency and performance of NN controllers and is suitable for real-world structural vibration control applications. | en_US |
| dcterms.abstract | Furthermore, this thesis introduced a multi-objective control strategy based on the IL-DRL framework to enhance the control performance, enabling the controller to optimize multiple objectives concurrently. This approach was validated experimentally on a three-story frame and through numerical simulations of the Canton Tower under wind-induced vibrations. The results demonstrate that the IL-DRL method achieves rapid convergence and surpasses the performance of classical model-based controllers (like LQR and LQG) with direct optimization to the target control performance indices (like the inter-story drift). These results affirm the method's efficacy as a model-free, multi-objective control strategy suitable for real-world applications. | en_US |
| dcterms.abstract | In summary, a series of comprehensive investigations was conducted in this thesis to evaluate the efficacy and applicability of the model-free learning-based control strategies for structural vibration mitigation. For the first time in structural control literature, DRL was rigorously demonstrated to achieve optimal control performance comparable to conventional model-based methods. Building upon this foundation, a novel IL-DRL framework was proposed, representing the first successful integration of IL with DRL for vibration control applications. The developed strategy significantly enhanced performance, efficiency, and adaptability, showing considerable promise for future real-world applications, particularly in tackling complex, high-degree-of-freedom structures that have not been reported in the current study. Furthermore, the framework's inherent multi-objective control capability was rigorously examined and substantially strengthened within the IL-DRL architecture, achieving performance that surpasses model-based methods. This research contributes substantively to the advancement of intelligent structural control technologies while establishing a robust methodological foundation for future investigations in adaptive vibration mitigation. | en_US |
| dcterms.extent | xxiv, 214 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2026 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.accessRights | open access | en_US |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14505