Self-Evolving Cognitive Digital Twin Architectures for Physics-Guided Fault Prognostics, Catalyst Deactivation Analysis, and Autonomous Optimization of Multiphase Catalytic Reactor Networks in Intelligent Refineries

The integration of physics-informed machine learning with high-fidelity process simulation is reshaping fault management and optimization in modern petroleum refining. This paper introduces a Self-Evolving Cognitive Digital Twin (SECDT) architecture specifically engineered for multiphase catalytic reactor networks the operational backbone of hydrocracking, fluid catalytic cracking (FCC), and hydrotreating units in intelligent refineries. The proposed framework couples ensemble physics-guided neural operators with probabilistic Bayesian state estimators, enabling real-time fault prognostics, quantitative catalyst deactivation tracking, and closed-loop autonomous optimization under uncertainty. Three methodological pillars underpin SECDT: (i) a Physics-Guided Neural Operator (PGNO) layer that enforces thermodynamic and kinetic consistency through embedded conservation laws; (ii) a Hierarchical Bayesian Fault Prognostic Engine (HBFPE) that fuses multimodal sensor streams with first-principles residuals to issue probabilistic remaining useful life (RUL) estimates; and (iii) a Reinforcement-Learning-Assisted Economic Model Predictive Control (RL-EMPC) module that autonomously adjusts reactor operating conditions to maximize long-run net present value while respecting safety envelopes and environmental constraints. Continual online learning via a catastrophic-forgetting-resistant Elastic Weight Consolidation (EWC) scheme allows the digital twin to self-evolve as catalyst activity declines and feedstock composition shifts. Industrial validation is reported for a ۱۲۰,۰۰۰ BPD hydrocracking complex operating over an ۱۸-month campaign. SECDT achieved a ۳۴.۷% reduction in unplanned shutdowns, a ۲۲.۱% improvement in catalyst cycle-length prediction accuracy over benchmark data-driven models, and a ۴.۸% gain in distillate yield through autonomous setpoint adjustment translating to an estimated annualized economic benefit of USD ۴۷.۳ million. The architecture generalizes to any multiphase fixed-bed, ebullated-bed, or slurry reactor and offers a principled pathway toward fully autonomous refinery operations.

کلیدواژه ها:

digital twin ، fault prognostics ، catalyst deactivation ، physics-informed machine learning ، autonomous optimization ، multiphase catalytic reactors ، intelligent refinery ، reinforcement learning ، remaining useful life ، Bayesian inference