LLM Hallucinations Suppression in Safety-Critical Engineering Drawing Conversion
Academic Journal of Computing & Information Science · 2026 · 9 (5)
DOI: 10.25236/ajcis.2026.090506 ↗Kurz gefasst
Fachartikel in der Zeitschrift Academic Journal of Computing & Information Science, 2026.
Abstract (Originalsprache)
: Converting engineering drawings from Chinese GB to German DIN standards demands near-perfect accuracy: even a single hallucinated value or a misplaced annotation can render a drawing unsafe for manufacturing. Large language models (LLMs) are fluent, but they lack spatial reasoning and often generate uncontrolled outputs — two fatal flaws in this domain. We therefore design a neuro-symbolic framework that couples an LLM with a knowledge-based engineering shell. The core components are a Trident Router that assigns each annotation to a risk-appropriate path, an ontology-driven retrieval-augmented generation (RAG) module with entity masking, a hard symbolic validation loop, and a spatial-aware CAD layout adapter. Before generation, structural constraints protect critical entities; after generation, symbolic checks reject any output that violates standard or format rules. Together, these mechanisms prevent technical hallucinations (e.g., altered tolerances or standard codes) and layout collisions when the translated text is reinserted into the CAD drawing. By detailing the logic and theoretical operation of each module, we show how deep neuro-symbolic integration turns an LLM from an unconstrained text generator into a reliable, auditable, and standard-compliant industrial assistant.
Zitieren
APA
(2026). LLM Hallucinations Suppression in Safety-Critical Engineering Drawing Conversion. Academic Journal of Computing & Information Science, 9(5). https://doi.org/10.25236/ajcis.2026.090506
Vancouver
LLM Hallucinations Suppression in Safety-Critical Engineering Drawing Conversion. Academic Journal of Computing & Information Science. 2026;9(5). doi:10.25236/ajcis.2026.090506
BibTeX
@article{euvimed2026,
title = {LLM Hallucinations Suppression in Safety-Critical Engineering Drawing Conversion},
journal = {Academic Journal of Computing & Information Science},
year = {2026},
volume = {9},
number = {5},
doi = {10.25236/ajcis.2026.090506}
}
EUVIMED