Toward Autonomous Control for Automated Material Handling Systems via Spatio-Temporal Prediction and Traffic Impact-Aware Routing = 시공간 교통 예측과 영향도 인지 경로 최적화 기반 물류자동화시스템 자율제어 연구

다국어 초록 (Multilingual Abstract)

Semiconductor manufacturing represents one of the most complex and capital-intensive production processes in modern industry, requiring both high precision and large-scale automation. Within a modern fab, the Automated Material Handling System (AMHS), particularly the Overhead Hoist Transport (OHT) system, serves as critical infrastructure for Work-In-Process (WIP) transportation between tools. The efficiency of the AMHS plays a crucial role in determining factory throughput and equipment utilization. However, as production scale expands, traffic congestion and flow imbalance occur frequently. Localized bottlenecks can propagate through the network, ultimately degrading overall system performance. This dissertation addresses these challenges by proposing an autonomous control methodology for AMHS that integrates predictive intelligence with system-level decision-making. The proposed approach consists of three complementary components. First, a multi-scale spatio-temporal graph neural network-based traffic forecasting model captures complex transport dynamics by incorporating the structural characteristics of the OHT network and interactions among vehicles. Second, a horizon-aware vehicle arrival-time prediction model integrates route decomposition with multi-step traffic forecasts to maintain accuracy over long, multi-segment routes. Third, a traffic impact-aware routing algorithm is designed to dynamically incorporate predicted congestion effects into routing costs, enabling proactive and system-wide optimization. Extensive large-scale simulations using real operational data from a semiconductor fabrication validate the effectiveness of the proposed control scheme. The results demonstrate significant improvements in transport efficiency and delay reduction compared with heuristic-based methods, and further confirm the framework's scalability in large-scale manufacturing environments. By integrating spatio-temporal prediction with traffic impact-aware routing, this research provides a practical foundation for more predictive, coordinated, and autonomous material handling control in semiconductor manufacturing.

국문 초록 (Abstract)

현재의 반도체 제조는 가장 복잡하고 자본 집약적인 산업 중 하나이다. 그중 물류자동화시스템(AMHS), 특히 OHT 시스템은 설비 간 재공 이송을 담당하는 핵심으로서, 생산처리량과 장비 가동률을 결정짓는 중요한 요소이다. 라인 규모가 확장됨에 따라 물류 흐름의 불균형과 교통 혼잡이 빈번해지고, 국소적인 정체가 네트워크 전반으로 확산되어 제조생산성의 저하를 야기할 수 있다. 본 논문은 이러한 문제를 해결하기 위해 예측 기반 지능형 의사결정을 통합한 물류자동화시스템의 자율 제어 기법을 제안한다. 제안된 접근법은 세 가지 상호보완적인 구성 요소로 이루어진다. 첫째, 다중 스케일 시공간 그래프 신경망 기반의 구간 단위 교통 예측 모델을 통해 OHT 네트워크의 구조적 특성과 차량 간 상호작용을 반영하여 복잡한 교통 패턴을 정밀하게 학습한다. 둘째, 다계층 경로 네트워크와 장·단기 시계열 예측을 결합하여 경로 적응형 차량 도착시간 예측 모델을 제안하여 정확도를 향상시킨다. 셋째, 교통 영향도 개념을 정의하고 이를 동적으로 반영하는 경로 최적화 알고리즘을 설계하여, 예측 정보를 기반으로 한 시스템 수준의 능동적 경로 결정을 가능하게 한다. 실제 반도체 공장의 운용 데이터를 활용한 대규모 시뮬레이션 결과, 제안된 자율 제어 기법은 기존 휴리스틱 기반 방식 대비 운송 효율과 지연 시간 측면에서 유의미한 개선을 보였으며, 수백 대의 자율 차량이 운용되는 대규모 생산 환경에서도 우수한 확장성을 입증하였다. 본 연구는 시공간 예측과 지능형 경로 의사결정을 결합한 자율 제어 체계를 제시함으로써, 소프트웨어 중심의 자기 최적화 및 완전 자율형 제조 시스템으로 발전하기 위한 실질적 기반을 제안한다.

목차 (Table of Contents)

• Contents
• Abstract i
• Contents vi
• List of Tables vii
• List of Figures ix
• Chapter 1 Introduction 1
• 1.1 Semiconductor Manufacturing Landscape 1
• 1.2 Challenges in OHT Systems 4
• 1.3 Overview of Proposed Approach 5
• 1.4 Contributions 6
• 1.5 Dissertation Organization 8
• Chapter 2 Fundamentals of the OHT System 10
• 2.1 OHT Vehicles 10
• 2.2 Delivery Process 11
• 2.3 OHT System Architecture 13
• 2.4 Rail Layout and Topology 14
• Chapter 3 Literature Review 17
• 3.1 AMHS and OHT Operations 17
• 3.1.1 Evolution of AMHS Technologies 17
• 3.1.2 AMHS Layout Design and Optimization 19
• 3.1.3 Operational Control and Optimization 22
• 3.2 Traffic Prediction Approaches 26
• 3.2.1 Statistical and Deep Learning Models 27
• 3.2.2 Graph Neural Network Models 28
• 3.2.3 AMHS-Specific Prediction Studies 30
• 3.3 Research Gaps and Dissertation Justification 32
• 3.3.1 Research Gaps 32
• 3.3.2 Dissertation Justification 33
• Chapter 4 Vehicle Traffic Prediction Using Multi-Scale GNN 35
• 4.1 Background and Motivation 35
• 4.2 Problem Statement 37
• 4.3 Methods: Traffic Prediction Using Multi-Scale GNN 38
• 4.3.1 Data Processing 38
• 4.3.2 Edge-Level Traffic Features 39
• 4.3.3 Multi-Scale Time-Series Features 40
• 4.3.4 Multi-Scale Attention-Based ST-GNN 42
• 4.4 Experiments 45
• 4.4.1 Dataset and Experimental Setup 45
• 4.4.2 Comparative Models 45
• 4.4.3 Prediction Results 46
• 4.5 Chapter Summary 51
• Chapter 5 Horizon-Aware Arrival Time Prediction 53
• 5.1 Background and Motivation 53
• 5.2 Problem Statement 55
• 5.3 Methods: Horizon-Aware Arrival Time Prediction 56
• 5.3.1 Multi-Level Route Decomposition 56
• 5.3.2 Segment-Level Travel Time Prediction 58
• 5.3.3 Route-Level Arrival Time Prediction 59
• 5.3.4 Algorithm Pseudo Code 61
• 5.4 Experiments 61
• 5.4.1 Dataset and Experimental Setup 61
• 5.4.2 Comparative Models 62
• 5.4.3 Prediction Results 63
• 5.5 Chapter Summary 70
• Chapter 6 Harmonized Routing with Traffic Impact Prediction 71
• 6.1 Background and Motivation 71
• 6.2 Related Work: System-View Routing Approaches 73
• 6.3 Problem Statement 75
• 6.3.1 Operational Conditions and Scope 76
• 6.3.2 Concepts and Terminology 77
• 6.3.3 Route Selection Problem 77
• 6.4 Methods: Harmonized Routing With Traffic Impact 80
• 6.4.1 Practical Partitioned ST-GNN 81
• 6.4.2 Traffic Impact Prediction 82
• 6.4.3 Harmony Routing 87
• 6.4.4 Architecture Design for Deployment 87
• 6.5 Experiments 89
• 6.5.1 Simulation Setup 89
• 6.5.2 Simulation Results 91
• 6.6 Discussion 96
• 6.7 Chapter Summary 99
• Chapter 7 Conclusion 100
• 7.1 Summary of the Research 100
• 7.2 Contributions 100
• 7.3 Limitations and Future Research 104
• 7.4 Final Remarks 105
• Bibliography 106
• 국문 초록 125
• 감사의 글 126