Enhancing robustness of deep learning for autonomous ships : a case study on ship fuel consumption prediction

다국어 초록 (Multilingual Abstract)

Data-driven models, such as deep learning models, are increasingly adopted within Maritime Autonomous Surface Ship (MASS) components for perception, prediction, and decision-making tasks. However, most existing approaches implicitly assume that operational data are independently and identically distributed (i.i.d.), overlooking the pronounced distributional shifts induced by changing routes, environmental conditions, and operational settings across voyages. Using ship fuel consumption prediction as a representative data-driven component within MASS, this thesis demonstrates that such voyage-level distributional shifts lead to a failure of generalization in components of MASS relying on data-driven models. As a result, these components exhibit severe performance degradation when deployed under unseen operational conditions. Through systematic in-distribution and out-of-distribution experiments on real-world voyage data, we show that high predictive accuracy under idealized i.i.d. evaluation settings does not guarantee robustness under realistic deployment scenarios. To address this limitation, we propose the VaRNN, a Voyage-aware Recurrent Neural Network that disentangles voyage-invariant and task-specific representations and selectively enforces alignment across voyages using a commonality loss applied to the voyage-invariant subspace. Extensive voyage-based cross-validation experiments confirm that the proposed VaRNN consistently improves predictive performance compared to conventional machine learning and deep learning baselines, while representation analysis and ablation studies highlight the necessity of its disentangled alignment for stable learning. These findings suggest that accounting for voyage-level distributional shifts is essential for ensuring reliable behavior of data-driven components within MASS, underscoring the broader value of distribution-aware modeling under real-world maritime conditions.

목차 (Table of Contents)

• List of Tables iv
• List of Figures vi
• List of Abbreviations vii
• Abstract xi
• 1 Introduction 1
• 2 Related Work 8
• 2.1 Maritime autonomous surface ship 8
• 2.1.1 Decision-making and control systems 10
• 2.1.2 Situational awareness 12
• 2.2 Ship fuel consumption prediction 16
• 2.3 Out-of-distribution generalization 20
• 3 Problem Definition 25
• 3.1 Categories of ship fuel consumption prediction 25
• 3.1.1 Categorization by model deployment stage 25
• 3.1.2 Categorization by operational characteristics 27
• 3.1.3 Categories of SFC prediction and i.i.d. compliance 29
• 3.2 Post-departure prediction for a tramp ship: ID approach 31
• 3.2.1 Problem assumptions 31
• 3.2.2 Formal definitions 32
• 3.3 Pre-departure prediction for a tramp ship: OOD approach 34
• 3.3.1 Problem assumptions 34
• 3.3.2 Formal definitions 34
• 4 Methodology 37
• 4.1 Data collection and processing 37
• 4.1.1 Data collection 37
• 4.1.2 Data preprocessing 38
• 4.1.3 Feature selection 43
• 4.1.4 Segmentation 43
• 4.1.5 Train–validation–test split 45
• 4.1.6 Normalization 47
• 4.2 Evaluation metric 48
• 4.3 In-distribution approach 48
• 4.3.1 Machine learning models 49
• 4.3.2 Deep learning models 49
• 4.4 Out-of-distribution approach 51
• 4.4.1 Representation modeling for pre-departure SFC prediction 51
• 4.4.2 The architecture of VaRNN 52
• 4.4.3 Training objective of VaRNN 53
• 5 Experiments 57
• 5.1 Dataset description and computational environment 57
• 5.2 Results of data preprocessing 57
• 5.2.1 Removal of erroneous samples 57
• 5.2.2 Attribute construction 58
• 5.2.3 Moving median filtering 58
• 5.2.4 Feature selection 58
• 5.2.5 Segmentation 62
• 5.3 ID approach 63
• 5.3.1 Experimental setups 63
• 5.3.2 Comparison of prediction models 68
• 5.4 OOD approach 70
• 5.4.1 Limitations of ID approach 71
• 5.4.2 Experimental setups 72
• 5.4.3 Verification of VCS in preprocessed raw data 74
• 5.4.4 Comparison of prediction models 79
• 5.4.5 Analysis on learned representations 82
• 5.4.6 Ablation studies 84
• 6 Discussion 88
• 6.1 Contributions to SFC prediction 88
• 6.1.1 Interpretation of findings under distributional shifts 88
• 6.1.2 Methodological implications 89
• 6.2 Application of VaRNN to pre-departure decision support 89
• 6.3 Limitations 91
• 6.4 Future work 92
• 6.5 Implications for MASS 92
• 6.6 Overall findings 93
• 7 Conclusion 94
• References 96
• Appendix A Detailed Sensor List of Raw Data 109
• Appendix B Normal Ranges of Sensors 115