Leveraging Knowledge Graphs for explainable predictive maintenance strategies

Research output: Contribution to conferencePresentation/Speechpeer-review

8 Downloads (Pure)

Abstract

Many components within nuclear power plant systems, including heavy water (D2O) filters in CANDU reactors, have traditionally followed time-based maintenance strategies. Recent advancements in data capture and storage have enabled a shift toward condition-based maintenance, driven by models tracking filter degradation. To successfully deploy these models, it is crucial that they clearly articulate and justify the data analysis steps, including data selection and preprocessing. Knowledge graphs have the potential to enhance this explainability. Knowledge graphs offer a flexible structure for representing complex relationships, mirroring human understanding through interconnected nodes and edges. This paper applies knowledge graphs to model relationships between highlevel engineering concepts, real-world assets, and data analysis in industrial maintenance. We present a use study of knowledge graphs applied to D2O filters, demonstrating a dynamic system where nodes not only hold information but also encapsulate functionality by integrating software methods and feeding outputs back into the graph. This approach enables the graph to change and react to information over time and host data analysis pipelines within its structure. Additionally, reasoning can be conducted based on the nodes and relationships within. This approach offers two key benefits: it tracks system-wide events like outages by relating them to high-level concepts and applying reasoning to infer when they happen. It also enhances the explainability of the data analysis pipeline by linking its nodes to real-world assets and conceptual explanations, providing real-time explanations for decisions and broader inferences as data is processed. With the graph as a semantic interface that is both human-readable and machine-legible, this allows for more transparent predictions of filter Remaining Useful Life (RUL) while creating a framework to allow for reusable data preprocessing. A single, flexible structure that can capture, reuse, and explain knowledge across multiple domains
Original languageEnglish
Publication statusPublished - 29 Nov 2024
EventFifth Annual Conference on Disruptive, Innovative, and Emerging Technologies in the Nuclear Industry: DIET 2024 - Toronto, Canada
Duration: 27 Nov 202429 Nov 2024
https://www.cns-ai-nuclear.com/

Conference

ConferenceFifth Annual Conference on Disruptive, Innovative, and Emerging Technologies in the Nuclear Industry
Abbreviated titleDIET 2024
Country/TerritoryCanada
CityToronto
Period27/11/2429/11/24
Internet address

Keywords

  • knowledge graph
  • nuclear
  • predictive maintenance

Fingerprint

Dive into the research topics of 'Leveraging Knowledge Graphs for explainable predictive maintenance strategies'. Together they form a unique fingerprint.

Cite this