Digital Twin and Smart Automation for Bitumen Extraction Process

• Author / Creator

  Soesanto, Jansen

• The advent of Industry 4.0 integrates advanced digital technologies and Artificial Intelligence (AI) into system engineering. This research explores the potential of AI in smart automation for industries, bridging it with physics-informed approaches, particularly through Explainable Artificial Intelligence (XAI) and transfer learning from physics to AI. The study unfolds across three interconnected phases, each targeting a specific aspect of industrial automation, with a focus on the bitumen extraction process from oil sands. The solid form of oil sands presents a complex challenge in producing Synthetic Crude Oil (SCO), a process characterized by natural disturbances from ore quality and plant scheduling capacity in upstream mining. The Primary Separation Vessel (PSV), central to the extraction process, is interconnected with the secondary separation unit, with both impacting each other's optimization and control. Our focus is on the digital twin development and autonomous operation of the PSV, including autonomous Real-Time Optimization (RTO) and advanced control. The multi-input multi-output, nonlinear, high-dimensional state-action spaces, and constrained processes present additional challenges that we aim to tackle.
  
  The first phase develops a high-fidelity digital twin for an industrial-scale bitumen extraction facility, incorporating multiparticle settling under non-ideal environments. Modifying the PSV model and integrating it with adjacent units, this plant-wide model accurately captures process dynamics, bitumen quality, and potential losses. High-dimensional parameters in the first-principle model are addressed using systematic parameterization techniques, Bayesian optimization, and sensitivity analysis to fully utilize industrial data. High-fidelity modeling proves crucial for automation validation and significantly contributes to the field of Explainable Reinforcement Learning (XRL).
  
  The second work focuses on developing autonomous control strategies, introducing a Model Predictive Control (MPC) for multimode operation with disturbances. Plant-model mismatches causing fluctuations in MPC motivate the integration of Reinforcement Learning (RL) for model scheduling and multitasking in the third work. Alongside MPC, this work showcases the capabilities of Reinforcement Learning-based Controller (RLC), achieving performance comparable to MPC with less controller effort. Notably, RLC speeds up computation 10 times faster than MPC. This work extensively tests the continual learning of RLC in multi-mode operations, ensuring their adaptability to changing environments.
  
  To enhance the feasibility of RL in real-world training, this study employs transfer learning approaches such as imitation learning and Simulation-to-Real (Sim2Real) pretraining. This strategy significantly reduces process trips during online training. Generative Adversarial Imitation Learning (GAIL) and Sim2Real pretraining decrease trip count by factors of 8 and 27, respectively, compared to direct agent training. GAIL opens new training pathways for agents in startup and shutdown tasks. The proposed ``MPC Safeguarded Exploration" approach strategically uses the alarm system and existing MPC controllers to further decrease trips during online training while maintaining agent explorability and adaptability.
  
  The third phase shifts to supervisory RTO in bitumen extraction, tackling the complexities of multivariable decision making and the interconnected extraction process under disturbances. This phase pioneers a novel framework that uses RL for setpoints optimization and multi-MPC scheduling. It combines the robustness of MPC with the adaptive optimization capabilities of RL to outperform existing operational strategies. First principle analysis elucidates and verifies the RL ability to manage trade-offs in microscale particle settling and balancing workload distribution across each unit to optimize the overall recovery rate. A key finding is that the RL agent anticipate MPC control policy and optimize its strategies accordingly. This ability to foresee and integrate decisions across control layers enhances collaboration among decision-making layers and optimizes operations in the context of plant-wide connectivity. Furthermore, the agent manages a second objective in control performance by scheduling MPC models based on operational changes. The RL policy reveals that operational modes depend on factors beyond ore grades, such as tailings density. These insights underscore the significance of Explainable Reinforcement Learning (XRL) in enhancing the acceptability of RL in complex industrial applications. The exploratory power and explainability of RL policies open new avenues for real-world implementation, transitioning RL from a learning agent to a teaching agent approach in industrial automation.
  

• Subjects / Keywords

  • autonomous control
  • digital twin
  • real time optimization
  • reinforcement learning
  • fused multi model predictive control

• Graduation date

  Spring 2024

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-wgex-2h61

• License

  This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.