Knowledge Extraction via Decentralized Knowledge Graph Aggregation
Richard Nordsieck, Michael Heider, Anton Winschel, Jörg Hähner · 2021
In many industrial manufacturing processes, human operators play a central role when it comes to parameterizing the involved machinery and dealing with errors in the process. However, large parts of the acquired process knowledge are tacit, leading to difficulties sharing the knowledge between operators. Therefore, knowledge extraction is a necessary but time and cost intensive process, requiring both specially trained personnel and experienced operators. In contrast, we propose that by gathering insights into what influenced operators' actual parameter choices, tacit process knowledge can be extracted during production in an example-based manner. This decentralized knowledge-decentralized in regards to who holds knowledge and where it was extracted-is then aggregated to a coherent knowledge graph. We showcase our methodology on a real-world dataset in the domain of fused deposition modeling (FDM), which is generated by operators providing their insights without additional assistance using extended human machine interfaces. Furthermore, we compare rules extracted from the aggregated knowledge graph against an established FDM knowledge base showing the viability of our approach even with limited amounts of data.