AI-assisted auditing review of mechanical engineering drawings: a Perception–Cognition–Collaboration reference architecture and its implementation Perspectives

Mechanical drawing audits are essential to product design and manufacturing but remain labor-intensive and error-prone, particularly when checking tolerances, surface roughness, materials, treatments and assembly requirements. Although computer vision, document understanding, large language models (LLMs), knowledge representation and multi-agent systems (MAS) can support individual tasks, existing efforts remain fragmented. This paper proposes an engineering-oriented approach to AI-assisted mechanical drawing auditing and a three-layer Perception – Cognition – Collaboration (P – C – C) pre-validation reference architecture. Perception transforms carrier-specific evidence into provenance-bearing entities and relations. Cognition aligns them with a versioned Standards-and-Requirements Knowledge Graph and applies approved deterministic numerical and logical rules outside the LLM. Collaboration coordinates audit tasks, preserves conflicts, and routes critical or uncertain findings to authorized engineers. An audit of the Merge_places Computer Vision Dataset reveals limitations in schema consistency, lineage, class semantics and clause grounding. De-identified drawing regions illustrate clearance calculation, detection of a controlled tolerance mismatch, and the routing of GD&T and bill-of-materials (BOM) requests when evidence is incomplete. These bounded cases clarify interface behavior and define validation requirements for provenance, deterministic checking, interoperability, uncertainty, cybersecurity and human governance before engineering deployment. The framework integrates heterogeneous evidence, executable rules and accountable human decisions within a traceable and selectively automated audit workflow.