Supplementary Results: Contextualizing Timed Automata via Knowledge Graphs for Agentic LLM-Based Fault Diagnosis

This repository contains the supplementary results for the paper:

Contextualizing Timed Automata via Knowledge Graphs for Agentic LLM-Based Fault Diagnosis
Tom Westermann, Felix Gehlhoff, Alexander Fay

Overview

The paper proposes a two-stage agentic diagnosis pipeline for fault diagnosis in Cyber-Physical Production Systems (CPPS). A learned timed automaton and its detected anomalies are integrated with the physical plant model into a semantically enriched knowledge graph. Two LLM-based agents operate on this knowledge graph:

1. The State Description Agent generates natural-language descriptions for each automaton state and for the overall production process of each module.
2. The Syndrome Diagnosis Agent identifies root causes of detected anomaly syndromes by reasoning over the augmented knowledge graph.

The approach is evaluated on the HaiCPPS benchmark across ten system configurations (DS1–DS10) of increasing complexity, comprising one to four production modules (Mixing, Distillation, Filter, Bottling). All agents use Claude Opus 4.6 at default temperature without task-specific fine-tuning.

Directory Structure

Published_Results/
├── KnowledgeBase/          # Knowledge graph, ontologies, and plant model
├── ProcessDescriptions/    # Module-level process descriptions (State Description Agent)
├── StateDescriptions/      # Per-state natural-language descriptions (State Description Agent)
└── SyndromeDiagnosis/      # Diagnosis reports and aggregated results (Diagnosis Agent)

Each subfolder contains its own README with detailed descriptions of the files it contains.

HaiCPPS Configurations

Dataset	Modules	# Variables	# GT States	Flow Type
DS1	Mixing	17	8	convergent
DS2	Distillation	17	6	divergent
DS3	Filtering	8	4	linear
DS4	Bottling	10	5	linear
DS5	Mixing + Bottling	27	13	convergent
DS6	Distillation + Bottling + Bottling	37	16	divergent
DS7	Filter + Bottling	18	9	linear
DS8	Mixing + Filter + Bottling	35	17	convergent
DS9	Filter + Distillation + Bottling + Bottling	45	20	divergent
DS10	Mixing + Distillation + Mixing + Bottling	61	27	conv./div./conv.

Contact

Tom Westermann — tom.westermann@hsu.hamburg
Institute of Automation Technology, Helmut Schmidt University, Hamburg, Germany