제한적 데이터 환경에서 분포 인지형 Few-shot 프레임워크를 활용한 선박 방사소음 인식 = Distribution-Aware Few-Shot Framework for Ship Radiated Noise Recognition in Data-Constrained Environments

다국어 초록 (Multilingual Abstract)

Effective identification of ship radiated noise in underwater environments is a core technology required for maritime object recognition, marine safety management, and ecosystem preservation. Traditional signal identification methods demand large-scale labeled datasets for each ship type. However, in real sea condition scenarios, data collection is challenging, and particularly, acoustic signals of specific vessels are often inaccessible due to military or security reasons, resulting in data scarcity and imbalance. These limitations restrict the practical deployment of automated signal recognition technologies in marine applications. With the recent adoption of deep learning techniques in underwater acoustic analysis, classification accuracy has improved when large datasets are available, yet performance remains limited under data-constrained conditions. As an alternative, few-shot learning has emerged as a promising approach, enabling classification of new classes with only a few training samples. This study proposes a novel ship signal identification framework based on few-shot learning that maintains high performance even under limited labeled data conditions. The proposed method employs a CNN-based feature extractor to extract latent vectors that represent the intrinsic characteristic of ships. These latent vectors are used to measure inter-ship similarity through cosine similarity while incorporating probabilistic information. To enhance class separability in the embedding space, the model integrates Original Loss, Self-Calibration Loss, Inter-Calibration Loss, and Contrastive Loss. Such architectural design enables consistent classification performance and strong generalization ability under scarce data environments. Experiments were conducted using real underwater acoustic data collected in the southern seas of Jeju Island. Ship radiated noise from 21 ships were segmented into one-second intervals and converted into spectrograms for analysis. Under the 5-way, 5-shot learning condition, the proposed model achieved an average classification accuracy of 86.63%, outperforming representative few-shot learning methods such as ProtoNet, SPNet, MAML, MatchingNet, and RelationNet. Furthermore, backbone architecture comparison experiments revealed that the VGG-16 backbone achieved the highest performance, while also confirming that increasing the embedding dimensionality does not necessarily yield performance improvements. Additional evaluations under varying k-way and n-shot conditions further validated the model’s capability to generalize effectively with only a small number of samples. We analyzed the effect of the proposed Contrastive Loss weight (δ) and identified the optimal parameter value. This study confirms the practical potential of few-shot learning for effective ship signal recognition in data-constrained scenarios. By integrating Original Loss, Self-Calibration, Inter-Calibration, and Contrastive Loss into a Prototypical metric-learning framework, the proposed framework simultaneously enhances both accuracy and generalization performance in underwater acoustic signal recognition. The findings suggest that the proposed approach can be applied to a wide range of real-world applications, including real-time marine monitoring systems, maritime traffic management, military surveillance, and environmental monitoring, as well as underwater acoustic processing domains where data labeling is inherently difficult.

목차 (Table of Contents)

• 1. 서론 1
• 2. 관련 연구 4
• 2.1 메타 러닝과 Few-shot 학습 4
• 2.1.1 메타 러닝 4
• 2.1.2 대조 학습과 샴 네트워크 6
• 2.1.3 Few-shot 학습과 메트릭 학습 8
• 2.1.4 특징추출을 위한 백본 12
• 2.2 인공지능 기반의 선박신호 탐지 15
• 2.3 Few-shot 기반의 선박신호 탐지 19
• 2.4 관련 연구 요약 및 연구 목적 21
• 3. 연구 방법 23
• 3.1 데이터 23
• 3.1.1 수중 음향 데이터 23
• 3.1.2 자료 전처리 및 정규화 29
• 3.2 선박신호 분류를 위한 Few-shot 네트워크 33
• 3.2.1 제안된 모델 구조 33
• 3.2.2 백본 구조 36
• 3.2.3 거리 함수 38
• 3.2.4 손실 함수 40
• 3.2.5 Original Loss 41
• 3.2.6 Self-Calibration Loss 44
• 3.2.7 Inter-Calibration Loss 46
• 3.2.8 Contrastive Loss 49
• 3.2.9 실험 설정 52
• 4. Few-shot 기반의 분류 결과 및 분석 54
• 4.1 지도학습 기반의 선박 방사소음 분류 54
• 4.2 백본 및 임베딩 차원 비교 57
• 4.3 기존 Few-shot 기법과의 비교 60
• 4.4 k-way 및 n-shot 변화에 따른 성능 63
• 4.5 Contrastive Loss 가중치 민감도 분석 68
• 5. 결론 72
• 5.1 결론 72
• 5.2 연구의 한계 및 향후 연구 방향 75
• 참고문헌 76
• 부록 92