LLM-based agent for natural language interpretation of IIoT time-series data in a bakery digital twin system

Bauyrzhan Amirkhanov, Amirkhanova Gulshat, Yusubova Nazly, Alikhan Amirkhanov

2026

Closed access Metadata (CSL-JSON)
DOI 10.1117/12.3130206
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Industrial Internet of Things (IIoT) deployments in food manufacturing generate substantial volumes of time-series sensor data, yet operational personnel frequently lack the technical expertise required to query and interpret this data directly. This paper presents an edge-oriented architecture that bridges the gap between raw IIoT telemetry and human-comprehensible insights by integrating a locally deployed Large Language Model (LLM) agent with a digital twin infrastructure for bakery production monitoring. The proposed system introduces two principal contributions: (1) a read-only REST proxy that mediates between the LLM agent and an InfluxDB time-series database, enforcing a constrained API contract that eliminates the risks of direct database access and arbitrary query generation; and (2) a tool-augmented LLM agent equipped with domain-specific prompting that transforms natural language questions from plant operators into structured API calls, retrieves sensor data, and returns contextualized interpretations via a Telegram messaging interface. The architecture was validated on a production bakery facility monitoring temperature, humidity, weight, and energy consumption parameters. Experimental evaluation across 50 representative operator queries demonstrated an interpretation accuracy of 92%, with an average end-to-end response latency of 4.7 seconds. The results confirm that the proposed approach enables non-technical personnel to access and understand complex IIoT data streams without specialized training, while maintaining strict data security through architectural isolation.