Smart data adapter: A hybrid pattern-LLM approach

Abstract

Adopting digital twin technologies in small and medium-sized enterprises (SMEs) is often hindered by heterogeneous, poorly documented production data. Current semantic type detection approaches require massive labeled datasets making them impractical for resource-constrained SMEs. We propose a zero-shot hybrid framework combining pattern-based classification with selective large language model (LLM) reasoning for manufacturing-specific data types. The two-stage architecture uses rule-based patterns for high- confidence cases, forwarding ambiguous columns to a multi-step LLM reasoner. Evaluation on four manufacturing datasets shows the hybrid approach achieves weighted F1 within 7 24 percentage points of pure LLM classification performance while reducing LLM invocations by 39% on average. Processing time decreased by up to 2.8×. Our framework addresses a critical gap: automated, computationally efficient data type recognition for manufacturing SMEs without requiring training data, contributing to automated simulation and digital twin construction.