Lifecycle-Oriented Taxonomy and Evidence Synthesis of Physics-Integrated Reinforcement Learning and Hybrid Monitoring Approaches for Fluid Machinery

Abstract

Predictive monitoring of fluid-machinery assets is hindered by nonlinear dynamics, sparse fault data, and distribution shifts between simulation and field operation. This survey develops a lifecycle-oriented taxonomy and maps evidence across 127 screened records, including 77 studies retained for tiered synthesis. The tiering and integration-score analyses are used as descriptive evidence-mapping tools rather than as measures of empirical performance or deployment maturity. The taxonomy records physics provenance, integration mechanism, and stage-specific use of physics across design, training, and deployment. It also distinguishes physics-based validation from calibrated uncertainty quantification (UQ), treating them as complementary deployment-assurance roles rather than interchangeable safeguards. The synthesis shows a broad base of mechanism-level work, including physics-informed features, constraints, reward shaping, ROM/PINN surrogates, and simulator-based learning. However, evidence remains limited for lifecycle-spanning systems that combine deployment-oriented validation, calibrated UQ, and fallback logic. Adjacent transferable studies are used only to identify reusable mechanisms and are not treated as direct evidence of fluid-machinery deployment validation. The survey concludes with reporting recommendations for benchmark design, validator/UQ diagnostics, hardware-in-the-loop evaluation, and uncertainty-aware monitoring.