Multi-Agent System for Smart Factory Management

Abstract

Manual retrieval of engineering knowledge from Piping and Instrumentation Diagrams (P&IDs), diagnostic manuals, and technical procedures remains a critical bottleneck in industrial plant operations: every minute of diagnostic delay carries direct costs in pro- duction and safety. Existing approaches either require extensive manual digitisation or rely on general-purpose AI systems that produce plausible but plant-unverifiable outputs. This work presents maestro, a governed Multi-Agent System that transforms P&ID drawings and industrial documents into structured, audit-traceable plant intelligence through a seven-layer architecture spanning OCR-led extraction, ISA-5.1 semantic en- richment, human-governed knowledge promotion, and agentic diagnostic reasoning. Gov- ernance is enforced architecturally: the agent’s tool contracts structurally prevent control- action execution, independent of any prompt or instruction. Validated on a refrigeration-loop P&ID, maestro extracted complete instrument and equipment candidates with zero false positives – eliminating engineering time spent on phantom candidates – and executed multiple diagnostic scenarios with response times between 8 and 18 seconds, combining topology traversal, live telemetry retrieval, and 85-chunk expert knowledge retrieval within a single governed pipeline. These results demonstrate that plant-aware agentic reasoning is feasible at Technol- ogy Readiness Level 4–5 without sacrificing governance or auditability, establishing a reproducible pathway toward deployable industrial AI.