데이터-공정지식 융합형 다중관계 상호작용 그래프 기반 산업공정 고장진단 = Data-Knowledge Interaction-aware Multi-Relational Graphs for Industrial Fault Diagnosis

다국어 초록 (Multilingual Abstract)

Effectively representing relationships among process variables is essential for industrial process fault diagnosis. Although graph-based methods have been introduced to address the limited ability of conventional deep learning approaches to explicitly model such relationships, existing studies remain limited in their integrated use of data-driven and process-knowledge-based relations. Moreover, even when multiple relations are considered, interaction-based fusion that accounts for their redundancy and complementarity has not been sufficiently investigated. Most existing methods also focus on fusing all constructed relations, with limited attention to selecting and interpreting relation combinations that are effective for fault diagnosis.
To address these limitations, this study proposes a multi-relational graph-based fault diagnosis model that integrates data-driven and process-knowledge-based relations while explicitly considering interactions among them. Process variables are represented as nodes, and each type of relation is constructed as an individual adjacency matrix. A relation interaction map is then generated by measuring the structural similarity between the adjacency matrices and is used to select effective relation combinations. The selected relations are transformed into a unified process representation through a hierarchical multirelational embedding module. Specifically, relation-specific graph-level embeddings are first extracted, after which message passing is performed among the relation embeddings according to the relation interaction map. This enables the model to jointly capture information from individual relation graphs and structural interactions among different relations.
Experiments on the Tennessee Eastman Process dataset showed that the proposed model achieved performance comparable to or better than that of baseline models for most fault classes. Ablation studies further demonstrated the contributions of multirelational representation, the integration of data-driven and process-knowledge-based relations, relation interaction-based fusion, and effective relation selection. Analysis of the selected relation combinations also indicated that data-driven and process-knowledgebased relations were jointly selected and used in a complementary manner. These results demonstrate that the proposed method can identify effective relation combinations and learn process representations that incorporate interactions among multiple relations for industrial process fault diagnosis.

목차 (Table of Contents)

• I. 서론 1
• 1. 연구 배경 1
• 1.1 산업공정 고장과 고장 진단의 중요성 1
• 1.2 다변량 시계열 고장진단의 고려사항 2
• 2. 문헌조사 2
• 2.1 시계열 특징 학습 기반 고장진단 연구 3
• 2.2 그래프 기반 관계 구조 모델링 고장진단 연구 4
• 2.3 문헌 조사의 결론 및 본 연구의 필요성 7
• 3. 연구 목표 및 범위 9
• 3.1 연구 목표 9
• 3.2 연구 기여점 9
• 3.3 연구 추진방법 10
• II. 데이터-공정지식통합 다중관계 융합 공정 고장진단 모델 11
• 1. 전체 프레임워크 개요 11
• 2. 공정데이터 다중관계 그래프 구성 12
• 2.1 입력 데이터 및 다중관계 그래프 정의 13
• 2.2 데이터 기반 관계 그래프 구성 15
• 2.3 공정지식기반 관계 그래프 구성 17
• 3. 설명가능한 다중관계 상호작용 모듈 23
• 3.1 관계 상호작용 맵 생성 23
• 3.2 관계 상호작용 맵 기반 유효 관계 조합 선택 25
• 4. 계층적 다중관계 그래프 임베딩 모듈 27
• 4.1 Sensor Layer: 관계별 센서 그래프 임베딩 27
• 4.2 Relation Layer: 관계 상호작용 그래프 임베딩 31
• III. 제안 모델의 실험 검증 및 분석 35
• 1. 실험 세팅 및 방법 35
• 1.1 데이터셋 및 전처리 35
• 1.2 제안 모델 구현 세부사항 36
• 1.3 학습 설정 및 평가지표 37
• 1.4 실험 방법 38
• 2. 베이스라인 성능 비교 38
• 2.1 베이스라인 모델과 성능비교 38
• 2.2 클래스별 성능 및 오분류 분석 40
• 3. Ablation Study 41
• 3.1 다중관계 타당성 비교 41
• 3.2 공정지식 기반 관계 통합 영향 42
• 3.3 관계 상호작용 반영 융합 영향 43
• 3.4 Relation Selection 반영 영향 44
• 4. 선택된 유효 관계 조합 분석 46
• IV. 결론 48
• V. References 49