A Hybrid IoT-Based Cognitive Digital Twin Framework for Hydraulic Excavators: Real-Time Monitoring, Synchronization, and Validation = 유압 굴착기를 위한 하이브리드 IoT 기반 인지 디지털 트윈 프레임워크: 실시간 모니터링, 동기화 및 검증

국문 초록 (Abstract)

유압 굴착기의 복잡도가 증가하고 지능적이고 데이터 기반의 의사결정에 대한 수요가 커짐에 따라, 중공업 기계의 디지털 트윈 기술에 대한 관심이 가속화되고 있다. 그러나 기존의 많은 디지털 트윈 구현은 모델링, 데이터 수집, 동기화, 시각화 및 의사결정 지원 계층이 서로 분리된 단편적 형태로 남아 있어, 비선형적이고 일시적인 작동 조건에서 신뢰성 있는 실시간 모니터링, 예측 유지보수 및 인지적 해석을 지원하는 데 한계가 있다. 유압 굴착기에서는 유압 서브시스템이 강한 비선형성, 부하 의존적 거동, 매개변수 불확실성, 압력 과도 현상 및 결합된 액추에이터 동역학을 나타내기 때문에 이러한 과제가 특히 중요하다. 본 논문은 유압 굴착기의 실시간 모델링, 동기화 및 검증을 위한 3단계 하이브리드 IoT 기반 인지 디지털 트윈(Cognitive Digital Twin, CDT) 프레임워크를 제안한다. 1단계에서는 Simulink 기반 굴착기 모델과 최적화 기반 매개변수 추정, 그리고 머신러닝 기반 잔차 보정을 통합하여 하이브리드 물리-AI 모델링 기반을 구축한다. 물리 기반 모델은 비선형 유압 및 마찰 동역학을 통해 해석 가능성을 제공하며, 데이터 기반 계층은 현장에서 획득한 CAN 버스 데이터를 활용하여 시뮬레이션과 실제 거동 간의 불일치를 보완한다. 실험적 검증을 통해 시뮬레이션 결과와 측정된 기계 거동이 강하게 일치함을 확인하였으며, 이는 비선형 굴착기 시스템에 대한 하이브리드 모델링 전략의 유효성을 입증한다.
2단계에서는 물리적 굴착기를 디지털 트윈 환경에 연결하기 위한 배포 가능한 엣지-클라우드 원격 측정 아키텍처를 개발한다. CAN 버스 데이터는 엣지 로거를 통해 수집되어 클라우드 스토리지에 업로드되며, MATLAB에서 처리되어 Simulink 호환 데이터 구조로 변환된다. 본 아키텍처는 현장 데이터 전송, 원격 측정 지연 보정, 동기화 평가 및 배포 지향적 모니터링을 지원한다. 이 단계는 디지털 트윈의 성능이 모델 정확도뿐 아니라 신뢰할 수 있는 데이터 파이프라인, 시간 정렬 및 동기화 준비도에 의해 좌우됨을 보여준다. 3단계에서는 실제 작동 신호와 대표적인 버킷 굴착 부하 주기를 활용하여 검증 기반 인지 평가(validation-aware cognitive assessment)를 수행하는 프레임워크로 확장한다. 유압 보정을 통해 압력 불일치를 104.5 bar에서 94.5 bar로 감소시켰으며, 이는 9.6\%의 RMSE 개선에 해당한다. 시나리오 수준의 검증에서는 액추에이터 압력, 액추에이터 위치, 토양-버킷 상호작용 및 서브시스템 에너지 흐름 거동을 추가로 평가한다. 또한 머신러닝 기반 유압 예측은 결정계수 R2=0.99, 평가 데이터셋에서 RMSE 1.10 bar를 기록하여 측정 압력 값과 강한 일치를 보였으며, 이상 탐지 계층은 고압 피크, 급격한 전이 및 저압 편차 주변에 집중된 21건의 이상 관련 압력 이벤트를 식별할 수 있다.

다국어 초록 (Multilingual Abstract)

The increasing complexity of hydraulic excavators and the demand for intelligent, data-driven decision-making have accelerated interest in digital twin technologies for heavy industrial machinery. However, many existing digital twin implementations remain fragmented across separate modeling, data acquisition, synchronization, visualization, and decision-support layers. This limits their ability to support reliable real-time monitoring, predictive maintenance, and cognitive interpretation under nonlinear and transient operating conditions. In hydraulic excavators, this challenge is particularly significant because hydraulic subsystems exhibit strong nonlinearities, load-dependent behavior, parameter uncertainty, pressure transients, and coupled actuator dynamics. This thesis proposes a three-phase hybrid IoT-based Cognitive Digital Twin (CDT) framework for real-time modeling, synchronization, and validation of hydraulic excavators. Phase~1 establishes the hybrid physics--AI modeling foundation by integrating a Simulink-based excavator model with optimization-driven parameter estimation and machine-learning-based residual correction. The physics-based model provides interpretability through nonlinear hydraulic and friction dynamics, while the data-driven layer compensates for simulation--reality mismatch using field-acquired CAN bus data. Experimental validation demonstrates strong agreement between simulated and measured machine behavior, confirming the effectiveness of the hybrid modeling strategy for nonlinear excavator systems. Phase~2 develops a deployable edge--cloud telemetry architecture for connecting the physical excavator to the digital twin environment. CAN bus data are acquired through an edge logger, uploaded to cloud storage, processed in MATLAB, and converted into Simulink-compatible data structures. The architecture supports field-to-model data transfer, delayed telemetry reconciliation, synchronization assessment, and deployment-oriented monitoring. This phase demonstrates that digital twin performance depends not only on model accuracy but also on reliable data pipelines, temporal alignment, and readiness for synchronization. Phase~3 extends the framework toward verification-aware cognitive assessment using real operational signals and a representative loaded-bucket digging cycle. Hydraulic pressure calibration reduces the pressure mismatch from 104.5~bar to 94.5~bar, corresponding to an RMSE reduction of 9.6\%. Scenario-level validation further evaluates actuator pressures, actuator positions, soil--bucket interaction, and subsystem energy-flow behavior. In addition, machine-learning-based hydraulic pressure prediction shows strong agreement with measured pressure values, with $R^{2}=0.9999$ and an RMSE of 1.10~bar on the evaluated dataset. An anomaly-assessment layer identifies 21 anomaly-relevant pressure events concentrated around high-pressure peaks, rapid transitions, and low-pressure deviations. The results demonstrate that the proposed framework provides a practical pathway from hybrid modeling to deployable telemetry integration and scenario-level cognitive interpretation. Rather than claiming a fully autonomous diagnostic system, this thesis establishes a verification-aware CDT foundation capable of supporting pressure prediction, model calibration, synchronization analysis, anomaly-relevant event identification, and operational assessment of excavator behavior. The framework contributes to the development of intelligent, data-connected, and physically interpretable digital twins for hydraulic excavators and related heavy industrial cyber-physical systems.

목차 (Table of Contents)

• Chapter 1: Introduction 1
• 1.1 Background 1
• 1.2 Problem Statement 4
• 1.3 Purpose of the Study 5
• 1.4 Research Objectives and Research Questions 5
• 1.5 Significance of the Study 9
• 1.6 Scope and Delimitations 9
• 1.7 Novelty of the Thesis 11
• 1.8 Contributions 12
• 1.9 Definition of Key Terms 12
• 1.10 Thesis Organization 12
• Chapter 2: Literature Review 17
• 2.1 Introduction 17
• 2.2 IIoT-Enabled Cyber-Physical Systems 19
• 2.3 Physics-Based Modeling of Nonlinear Industrial Systems 22
• 2.4 Data-Driven and Machine Learning Approaches for PHM 24
• 2.5 Hybrid PhysicsAI Modeling Approaches 26
• 2.5.1 Physics-Guided Machine Learning 27
• 2.5.2 Data-Augmented Physics-Based Models 27
• 2.5.3 Fully Coupled Hybrid Frameworks 27
• 2.6 Digital Twin Technologies in Industrial Systems 30
• 2.6.1 Digital Twins for Heavy Equipment and Hydraulic Machinery 31
• 2.7 Cognitive Digital Twins 32
• 2.8 Co-Simulation and Synchronization in Real-Time Digital Twins 33
• 2.8.1 Synchronization Fidelity in Real-Time Digital Twins 35
• 2.9 Critical Analysis and Research Gaps 35
• 2.10 Chapter Summary 38
• Chapter 3: Phase 1: Hybrid PhysicsAI Modeling and Simulink-Based Verification of Hydraulic Excavator Digital Twins 40
• 3.1 Chapter Overview 40
• 3.2 System-Level Problem Definition 41
• 3.2.1 Hydraulic Excavator as a Nonlinear CyberPhysical System 41
• 3.2.2 Modeling Challenges 42
• 3.3 Hybrid PhysicsAI Modeling Framework 43
• 3.3.1 Framework Rationale 43
• 3.3.2 Hybrid Model Structure 44
• 3.3.3 Field Data Interface 46
• 3.4 Parameter Estimation Formulation 46
• 3.4.1 Estimation Objective 46
• 3.4.2 Gradient-Descent Update 47
• 3.4.3 Nonlinear Least-Squares Estimation 48
• 3.4.4 Parameter-Estimation Performance 48
• 3.4.5 Role of Parameter Estimation in the Digital Twin 49
• 3.5 Residual Learning and Hybrid Correction 49
• 3.5.1 Residual-Learning Objective 49
• 3.5.2 Candidate Learning Models 50
• 3.5.3 Model Selection Logic 51
• 3.6 Simulink-Based Verification Environment 52
• 3.6.1 Simulink Physics Twin Core 52
• 3.6.2 Verification Metrics 53
• 3.6.3 Verification Evidence 54
• 3.7 Implementation Workflow 59
• 3.8 Chapter Summary 61
• Chapter 4: Phase 2: Deployable EdgeCloud Architecture for Synchronization-Aware Excavator Digital Twins 63
• 4.1 Chapter Overview 63
• 4.2 Experimental Testbed and Physical Excavator Platform 65
• 4.3 Phase 2 EdgeCloud DT Architecture 67
• 4.3.1 System-Level Design Rationale 67
• 4.3.2 Layered Architecture 69
• 4.3.3 Relationship to the Three-Phase CDT Framework 70
• 4.4 Field Data Acquisition and CAN Telemetry Processing 71
• 4.4.1 CANedge3 Logging and MF4 Storage 71
• 4.4.2 Cloud Upload and Storage 72
• 4.4.3 MATLAB-Based Processing Workflow 73
• 4.4.4 Simulink-Compatible Data Generation 73
• 4.5 Latency Analysis and Synchronization Evaluation 74
• 4.5.1 Latency Measurement Points 74
• 4.5.2 Stale-State Horizon 76
• 4.5.3 Synchronization Jitter 76
• 4.5.4 Packet Continuity and Data Integrity 77
• 4.6 Predictive Buffering and Delayed Telemetry Reconciliation 78
• 4.6.1 Motivation for Predictive Synchronization 78
• 4.6.2 Predictive State Propagation 79
• 4.6.3 Residual-Based Verification 79
• 4.6.4 Correction Logic 80
• 4.6.5 Prediction Error Reduction 81
• 4.7 Performance Evaluation 82
• 4.7.1 Dataset Characteristics 82
• 4.7.2 Upload Throughput 82
• 4.7.3 Overall Pipeline Performance 83
• 4.7.4 Comparison with Baseline Deployment Strategies 84
• 4.8 Deployment Challenges and Lessons Learned 85
• 4.9 Chapter Summary 87
• Chapter 5: Phase 3: Scenario-Level Cognitive Assessment of Hydraulic Excavator Digital Twin Operation 91
• 5.1 Chapter Overview 91
• 5.2 Phase 3 Scenario-Level Validation Workflow 93
• 5.2.1 From Deployable Telemetry to Cognitive Assessment 93
• 5.2.2 Validation Structure 94
• 5.3 Real Operational Data and Signal Characteristics 95
• 5.3.1 Field-Derived Signal Overview 95
• 5.3.2 Pump and Valve Command Behavior 97
• 5.3.3 Measured Hydraulic Pressure and Boom Angle 99
• 5.4 Hydraulic Pressure Calibration 101
• 5.4.1 Calibration Objective 101
• 5.4.2 Calibration Response Before and After Parameter Update 102
• 5.4.3 Error Reduction and Parameter Sensitivity 104
• 5.5 Loaded-Bucket Digging-Cycle Simulation 106
• 5.5.1 Operator Command Sequence 106
• 5.5.2 Actuator Pressure and Position Response 107
• 5.5.3 SoilBucket Interaction 109
• 5.5.4 Energy-Flow Analysis 110
• 5.6 Machine Learning Pressure Prediction and Anomaly-Relevant Assessment 112
• 5.6.1 Pressure Prediction Accuracy 112
• 5.6.2 Prediction Error Distribution 114
• 5.6.3 Model Performance Comparison 115
• 5.6.4 Anomaly-Relevant Event Detection 116
• 5.7 Integrated Discussion 118
• 5.7.1 Implications for CDT Development 118
• 5.7.2 Relationship to the Three-Phase CDT Framework 119
• 5.7.3 Anomaly-Relevant Observations 120
• 5.8 Limitations and Threats to Validity 120
• 5.9 Chapter Summary 122
• Chapter 6: Conclusion 125
• 6.1 Summary of the Thesis 125
• 6.2 Major Findings 126
• 6.3 Response to Research Questions 128
• 6.3.1 Response to RQ1 128
• 6.3.2 Response to RQ2 129
• 6.3.3 Response to RQ3 129
• 6.3.4 Response to RQ4 130
• 6.3.5 Response to RQ5 130
• 6.4 Main Contributions 131
• 6.5 Limitations 131
• 6.6 Recommendations for Future Work 134
• 6.7 Final Remarks 135
• References 136