딥러닝 기반 멀티 모달 분석을 통한 경량 LLM 돼지 건강 모니터링 = Deep Learning-based Multimodal Analytics for Pig Health Monitoring with a Lightweight LLM

다국어 초록 (Multilingual Abstract)

This dissertation investigates practical pig health monitoring under real-farm constraints through an Environment-Individual-Reasoning perspective. In commercial group-housed barns, health risks arise from coupled dynamics across housing environment, individual behavior, and growth, yet many precision livestock systems remain single-task and offer limited interpretability. As a result, farmers still face a semantic gap between heterogeneous numeric outputs (sensor streams, activity indices, and weight trajectories) and actionable daily management decisions. The core research problem is therefore formulated as producing reliable, low-cost, and auditable health assessments under noisy farm dynamics, limited annotations, and edge-deployment constraints.
To address this problem, an integrated pipeline is developed with validated module-level performance and a structured reporting interface. At the environmental layer, a closed-form continuous-time model is adopted to better capture control-induced fluctuations than discrete-time models, achieving an F1-best of 73.02% and an AUC-PR of 44.30% for anomaly detection on multi-sequence pig-house sensor streams. At the individual layer, identity-consistent biometric analytics in crowded pen videos are enabled via box-supervised instance segmentation with an explicit box-mask identity bridge, reaching 89.22% Box AP and 85.29% Mask AP while reducing annotation time by approximately 70% (26s vs. 94s per image). Building on stable identities, motion quantification integrates tracking with pixel-level optical flow to mitigate detection jitter and posture-induced confounds. For growth assessment, an AIGC-driven single-image 3D reconstruction pipeline with metric-scale prediction achieves accurate weight estimation (MAE of 1.48 kg, RMSE of 3.21 kg, and R2=97.88%), demonstrating that kilogram-level estimation is feasible without dedicated depth sensors when metric scale is explicitly recovered.
At the reasoning layer, daily assessment is formulated as evidence-grounded, template-constrained data-to-text generation. A unified JSON schema encodes pre-computed assessments, anomaly events, cohort baselines, and short histories to enforce evidence traceability, while a lightweight LLM serves as a constrained synthesis component to produce auditable four-section reports. Overall, the proposed framework bridges upstream analytics and operational decision support in a measurable and explainable manner, while remaining extensible to future system-level validation with fully synchronized farm datasets.

목차 (Table of Contents)

• 1 Introduction 1
• 1.1 Research Background and Significance 1
• 1.2 Conceptual Framework and Research Scope 2
• 1.3 Challenges and Problem Statements 5
• 1.4 Existing Methodologies and Limitations 6
• 1.4.1 Overview of Pig Health Monitoring Systems 6
• 1.4.2 Environmental Anomaly Detection 7
• 1.4.3 Individual -Level Biometric Monitoring 8
• 1.5 Dissertation Contributions and Organization 9
• 1.5.1 Dissertation Content and Contributions 9
• 1.5.2 Dissertation Organization 13
• 2 Continuous-Time Modeling for Macro-Environmental Anomaly Detection 15
• 2.1 Motivation 15
• 2.2 Dataset Characteristics and Problem Formulation 17
• 2.2.1 Data Acquisition and Nominal Dynamics 17
• 2.2.2 Anomaly Taxonomy and Synthetic Injection 19
• 2.2.3 Problem Formulation 21
• 2.3 Methodology 23
• 2.4 Experiments 26
• 2.4.1 Experimental Setups and Evaluation Metrics 26
• 2.4.2 Comparisons with SOTA Models 30
• 2.4.3 Ablation Studies 36
• 2.5 Chapter Summary 38
• 3 Micro-Individual Biometric Data Analytics 40
• 3.1 Motivation 40
• 3.2 Motion Estimation via Integrated Optical Flow and MOT 42
• 3.2.1 Problem Formulation 42
• 3.2.2 Methodology 43
• 3.2.3 Experiments 47
• 3.2.4 Discussion 56
• 3.3 Weakly Supervised Learning for Instance Segmentation 57
• 3.3.1 Problem Formulation 57
• 3.3.2 Methodology 59
• 3.3.3 Experiments 63
• 3.3.4 Discussion 71
• 3.4 AIGC-based Weight Estimation 71
• 3.4.1 Problem Formulation 71
• 3.4.2 Methodology 73
• 3.4.3 Experiments 81
• 3.4.4 Discussion 90
• 3.5 Chapter Summary 90
• 4 Reasoning-based Pig Health Assessment via LLM 92
• 4.1 Motivation and Problem Formulation 92
• 4.2 Data Foundations and System Schema Design 94
• 4.2.1 Domain-Adaptation Corpus Construction 95
• 4.2.2 Unified Data Representation Schema Design 99
• 4.3 Methodology 103
• 4.3.1 Model Architecture and Strategy Selection 104
• 4.3.2 Hybrid Reasoning Design and Evidence Binding 106
• 4.3.3 Model Adaptation via QLoRA 107
• 4.3.4 Template-Constrained Inference 108
• 4.4 Experiments 111
• 4.4.1 Experimental Results of Finetuning 111
• 4.4.2 Prototype Validation 114
• 4.4.3 Rule-based Report Auditing Metrics 122
• 4.4.4 LLM-based Peer Evaluation of Reports 124
• 4.5 Chapter Summary 126
• 5 Conclusions and Future Work 128
• 5.1 Conclusion 128
• 5.2 Future Work 131
• Bibliography 133
• 요약문 145
• Acknowledgements 147