A Surrogate-Assisted Online Adaptive Reinforcement Learning and Approximate Bayesian Computation (OARL-ABC) Method for Calibration of Digital Twins
Author(s)
Wei, Xiao (Olin)
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Digital twin technology has become a cornerstone in modern industry and research, providing virtual replicas of physical systems that enable real-time condition monitoring, future state simulation, and design optimization capabilities over a product's entire life-cycle. The accuracy and reliability of digital twins depends critically on the calibration process, which aligns the digital model with real-world data. As physical systems evolve over time, digital twins must be continuously recalibrated to remain accurate representations of their connected physical counterparts. Recent trends in digital twin applications have demanded more complex model forms and more stringent recalibration time-frames, creating an urgent need for improved calibration methods that can handle this increased complexity while maintaining computational efficiency.
This thesis addresses these challenges by developing and validating a new surrogate model assisted online adaptive reinforcement learning and approximate Bayesian computation (OARL-ABC) method for the calibration and validation of digital twins. The approach builds upon a reinforcement learning-based calibration framework that performs model selection via reinforcement learning techniques and parameter calibration via Bayesian inference methods, specifically approximate Bayesian computation (ABC). The method employs a hybrid Bayesian reinforcement learning calibration framework that combines the adaptability and efficiency of reinforcement learning for optimizing complex, dynamic systems with the uncertainty quantification and updating capability of Bayesian inference methods. This combination integrates the principled uncertainty handling of Bayesian inference with the adaptive learning capabilities of reinforcement learning, making it particularly well-suited for the calibration of highly complex digital twins.
To enhance the efficiency of Bayesian reinforcement learning, this thesis integrated surrogate modeling to provide computationally efficient approximations of complex models for more rapid evaluation in the learning process. This surrogate-assisted OARL-ABC method successfully reduced the computational intensity of the sampling requirements by utilizing multiple surrogate model representations of real systems. Through investigation of various surrogate modeling techniques, Bayesian network models were identified and implemented as the optimal choice in this context, as these models maintain the Bayesian framework's ability to manage uncertainty while significantly reducing computational demands, thereby accelerating the calibration process without compromising accuracy.
The resulting OARL-ABC method demonstrated increased adaptability to different model forms, improved efficiency in complex applications, and robust capability to quantify inherent problem uncertainties. The method's ability to capture complex system behaviors, account for uncertainties, and perform continuous learning in a single integrated loop achieved substantial reductions in computational cost compared to conventional approaches for higher-order models.
The effectiveness of this calibration method was validated through application to a rotordynamic system with both physical and digital counterparts that could be scaled in complexity and manipulated to represent different stages of a product's life-cycle. A machinery fault simulator (MFS) test rig provided the physical platform, designed as a flexible simulator for various types of rotating machinery faults that could be easily reconfigured to represent different systems of increasing complexity and degradation throughout the product life-cycle. The MFS uses an electric motor attached to variable disks, weights, bearings, and a shaft to simulate different fault conditions found in real-world rotating machinery such as turbines or compressors. By systematically varying the distribution of weights around the axle, multiple imbalance fault conditions were generated and matched to their respective total output vibration signatures, measured by accelerometers mounted to the bearing housings. These physical experiments were paired with a high-fidelity finite element rotor dynamics models created in Python using Rotordynamic Open-Source Software (ROSS), an open-source rotordynamics library serving as the digital representation of the simulated system to capture the complex dynamics of rotating systems.
The validation process constructed surrogate model representations of various experimental configurations, from which the reinforcement learning algorithm selected the best representation for each state of the physical system while Bayesian inference calibrated the model parameters. The performance of the proposed method was then systematically compared in this rotordynamic context against benchmark performances using conventional ABC calibration, conventional OARL-ABC without surrogate model assistance, and OARL-ABC using Bayesian network surrogate models across increasingly complex rotating machine simulations.
Through this comprehensive rotordynamic testing campaign, this research successfully demonstrated the improved performance and scalability of the surrogate assisted OARL-ABC calibration method for real-world digital twin systems. The approach delivered substantial improvements in calibration efficiency while enhancing robustness and reliability when applied to increasingly complex digital twin models with demanding recalibration requirements. These results establish a foundation for more effective and widespread application of digital twin technology in modern industry and research, enabling continuous model adaptation and uncertainty quantification at computational costs that make real-time calibration practically feasible for complex systems.
Sponsor
Date
2025-12
Extent
Resource Type
Text
Resource Subtype
Dissertation (PhD)