|
Generative artificial intelligence (GenAI), particularly large language models (LLMs), is increasingly considered for supporting production and technology management. In manufacturing quality engineering, LLMs may reduce the cognitive cost of defect analysis, accelerate root-cause diagnostics, and support quality documentation. Still, they can also generate plausible recommendations that are not grounded in evidence. This creates risks of erroneous process interventions, loss of traceability, and non-compliance with artificial intelligence governance requirements. This paper proposes an auditable LLM-augmented Digital Quality Twin (LDQT) architecture that combines an LLM, Retrieval-Augmented Generation (RAG), a defect-centric digital quality twin, and human-in-the-loop approval. The architecture is evaluated in a discrete-event simulation of a production stream with stochastic process-drift events and three intervention policies: manual, rule-based, and LDQT. The revised computational study includes Monte Carlo analysis, explicit hypothesis testing, and sensitivity analysis. Under the adopted assumptions of 400 replications and 4000 items per replication, the LDQT scenario indicates an approximately 20% lower mean unit cost of quality and downtime than the manual baseline, as well as a higher first-pass yield. These findings should be interpreted as simulation-based evidence of architectural potential, not as proof of practical effectiveness in a real factory. The contribution is methodological: the study defines the Digital Quality Twin concept and provides metrics for balancing efficiency, auditability, and decision safety in industrial GenAI deployments.
|
