Online Adaptive Reinforcement Learning and Approximate Bayesian Computation (OARL-ABC) for real-time long-term calibration of rotordynamic models

Efficient, timely and accurate calibration of digital twins is critical for their application in real-time decision-making and condition monitoring. Digital twins, as dynamic virtual representations of connected physical systems, require continuous recalibration to adapt to evolving system states, environmental changes, and degradation over time. Without timely and effective recalibration, discrepancies between the digital and physical counterparts accumulate, potentially leading to compromised safety and operational inefficiencies. This study addresses these challenges by introducing the Online Adaptive Reinforcement Learning and Approximate Bayesian Computation (OARL-ABC) framework, a novel and model-agnostic calibration algorithm. Building upon the foundational RL-ABC framework, OARL-ABC offers enhanced adaptability, scalability, and robustness, facilitating real-time adjustment of model parameters across diverse systems. By integrating reinforcement learning strategies with approximate Bayesian computation, the proposed framework effectively accounts for uncertainties in model parameters and environmental variables, ensuring reliable predictions in complex and dynamic scenarios. To validate its efficacy, the OARL-ABC framework is applied to the calibration of a rotordynamic system for rotating machine fault diagnosis, demonstrating its capability to maintain alignment between the digital and physical domains in real-time. However, while the results demonstrate improved real-time calibration performance in this context compared to conventional calibration methods, OARL-ABC remains limited by its use of sampling techniques for repeated full-scale model runs. Further improvement should prioritize applying the method to a broader range of use cases, exploring simplified model forms, or enhancing the sampling scheme to run more efficiently.