A Neuro-Symbolic Hybrid Framework for Automated P&ID Auditing: Integrating Graph Neural Networks with Knowledge-Grounded RAG for Human-in-the-Loop Safety Verification

Abstract Piping and Instrumentation Diagrams (P&IDs) represent the foundational safety records of chemical process facilities, yet their accuracy frequently degrades in brownfield facilities due to undocumented process modifications and Management of Change implementation. This documentation decay introduces latent topological hazards that traditional manual reviews often fail to detect. The aim of this paper is to automate topology-aware, standards-based safety auditing for P&IDs using a neuro-symbolic artificial intelligence framework. The architecture represents P&IDs as labeled property graphs across five core layers: (1) Graph Neural Networks (GNN)-based topological anomaly detection, (2) a standards-populated Knowledge Graph, (3) Graph-Retrieval-Augmented Generation (RAG) retrieval for audit finding generation, (4) an engineer review interface, and (5) human feedback-based calibration. The framework is examined conceptually across the three critical process safety scenarios: reversed check valve, incorrect valve fail-state positions, and missing thermal relief in blocked liquid lines. The flagged finding is classified into three severities: Critical, Warning, or Informational, to support safety auditing. An adjacent benchmark synthesis of graph-learning, engineering-drawing analysis, and knowledge-grounded AI studies suggests that the hybrid architecture addresses two key limitations of standalone methods: brittle rule coverage and insufficient audit traceability. Reported adjacent-domain benchmarks include evidence from previously published studies in adjacent domains, indicating that knowledge-graph-enhanced graph models can achieve high topological detection accuracy, that physics-informed graph architectures can attain strong classification performance, and that grounding retrieval-augmented generation in structured knowledge sources can substantially reduce unsupported generation. This human-in-the-loop system supports safety audits by grounding probabilistic AI anomaly detection in deterministic engineering codes.