Georgia Institute of Technology
Domain Knowledge Powered Graph Neural Network (GNN) And Retrieval-Augmented Generation (RAG) Decision Support For Failure Detection And Recovery (FDR) In Manufacturing System Operations (MSO)
Abstract
dc:description.abstractModern manufacturing systems are increasingly characterized by high structural complexity and data-driven operations, creating growing demand for efficient failure detection and recovery mechanisms. However, conventional approaches often rely on computationally expensive simulation-based evaluation for performance assessment and manual, experience-driven processes for failure detection and recovery in manufacturing process, resulting in limited scalability and delayed decision-making. To address these challenges, this thesis proposes a domain knowledge–driven framework that integrates graph neural networks (GNN) and retrieval-augmented generation (RAG) to support failure detection and recovery in manufacturing system operations. The central technical rationale of the framework is to use knowledge graphs as a relational knowledge backbone that transforms heterogeneous engineering artifacts into machine-interpretable structures. Compared with flat parameter tables, isolated CAD attributes, or unstructured document repositories, knowledge graphs explicitly represent entities, attributes, and relationships among design components, geometric entities, failure events, machine contexts, and recovery actions. On the learning side, CAD-derived design knowledge graphs provide graph-structured inputs for GNN-based performance prediction and requirement-deviation-based failure detection. On the reasoning side, ontology-based failure knowledge graphs organize historical failure records and recovery knowledge into structured evidence that can be retrieved and used by RAG-enabled language models for context-aware recovery recommendation. The proposed framework therefore establishes a common graph-based abstraction rather than a single merged knowledge schema. Under this abstraction, design knowledge and operational failure knowledge retain their domain-specific semantics while being converted into task-specific computational representations for learning and retrieval-based reasoning. This KG-centered technical route enables the framework to connect design-stage structure-aware prediction with operation-stage evidence-grounded recovery decision support.
Degree
thesis:*- Name thesis:degree_name
- Mechanical Engineering, MS
- Department thesis:degree_department
- Mechanical Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Shan, Liuyang
- Advisor dc:contributor.advisor
-
- Jiao, Jianxin (Roger)
- Committee members dc:contributor.committeemember
-
- Choi, Seung-Kyum
- Jariwala, Amit
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82201