Smarter Industrial Fault Diagnosis with Knowledge Graphs and Large Language Models
summary
Prototype called GraphRAG converts 15 years of Dutch maintenance Excel logs and English OEM PDFs into a Neo4j graph using GPT-4/GPT-4o multimodal extraction, merging 759 entities to 715, then serving a retrieval-augmented chatbot. Extraction yields ~51.5% exact matches and ~70% overall accuracy with ~20% potential hallucinations (some extractions outperform human labels). A nine‑month build includes entity resolution with Levenshtein, a GraphRAG index+retrieve+generate pipeline, and a user evaluation (six evaluators) reporting strong functional and traceability scores.
AI-written summary · about
chapters (15)
- 0:00Introduction and presenter
- 1:26About the presenter
- 3:11Problem statement in fault diagnosis
- 5:35Theoretical background of KG and LLMs
- 7:22Research questions and methodology
- 8:48Case study with Mres Be company
- 11:17Knowledge graph ontology design
- 13:25Data annotation and evaluation
- 15:57LLM extraction with schema prompting
- 18:04Knowledge fusion and entity resolution
- 20:33Extracted knowledge graph visualization
- 21:37Graph-based chatbot prototype
- 24:30User evaluation results
- 25:52Conclusion and key findings
- 26:34Q&A session
related talks
21mAgentic GraphRAG: Autonomous Knowledge Graph Construction and Adaptive Retrieval
53mPractical GraphRAG - Making LLMs smarter with Knowledge Graphs
15mGraph-Driven Knowledge Retrieval: Neo4j for Healthcare Building Codes
20mPractical GraphRAG: Making LLMs smarter with Knowledge Graphs
27mLies, Damn Lies, and AIs
25mAdaptive GraphRAG: A Framework for Knowledge Graph Quality, Consistency, Evolution
24mEmpowering Lifelong Learning: Transforming Data into Insights With GraphRAG
12mAgentic GraphRAG: Autonomous Knowledge Graph Construction and Adaptive Retrieval