LLM4VC: Harnessing Large Language Models for Virtual Commissioning of IEC 61499 Automation Systems

This paper addresses a critical deficiency of virtual commissioning systems developed for industrial automation: they lack the ability to analyze the avalanche of generated events and recognize faulty situations in them efficiently. The solution, called LLM4VC, is a cloud-based pipeline that turns a large language model into an LLM-based Validation Agent (LVA). It is an autonomous program that ingests human-centred specifications and produces pass/fail verdicts for distributed IEC 61499 systems automatically during testing. The agent's knowledge is derived from user-approved artifacts: the Sequence of Operations (SOO) and the Cause-&-Effect matrix (C&E), and users' prompts. A lightweight parser plus GPT models prompts converts these materials into a Machine-Readable Expected Behavior (MREB), which is a JSON document enriched with timing requirements and interlock rules. A Python script streams normalized runtime logs from mul-tiple SoftPLC containers. At run time the LVA uses MREB and SoftPLC logs to generate a structured verdict together with a human-readable description. We demonstrate LLM4VC on a four-controller conveyor and jack system containing manually injected fault scenarios. The agent analyses about 12 000 log lines in around 38 s per test, identifies most of the injected faults and produces traceable explanations suitable for factory acceptance reports without manual intervention or hard-coded rules.