딥러닝 기반 안전도 및 뇌파 분석을 통한 사용자 상태 인식 = User State Recognition using Deep Learning-based EOG and EEG Analysis

다국어 초록 (Multilingual Abstract)

Biosignal analysis technology, which infers human internal and external states non-invasively, is gaining attention as a core technology in Brain-Computer Interface (BCI) and Human-Computer Interaction (HCI). In this study, we propose deep learning-based modeling techniques to precisely recognize user's eye movement intentions and emotional states using Electrooculography (EOG) and Electroencephalography (EEG), and verify their effectiveness.
First, for EOG-based eye movement classification, we constructed a high-resolution dataset in a controlled laboratory environment and a Polysomnography (PSG) dataset in a real clinical environment. We proposed a hybrid model combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) to simultaneously learn spatial features and temporal contexts of time-series signals. Experimental results showed that the proposed model achieved a high accuracy of 99.63% in the laboratory environment, confirming the feasibility of precise interface implementation. However, performance degradation was observed in cross-domain evaluation on clinical data, experimentally demonstrating that domain adaptation techniques are essential for future real-world applications.
Second, for EEG-based emotion recognition, we proposed an attention-based neural network model that effectively reflects the characteristics of each EEG frequency band. To overcome the limitation of existing models that consider the importance of frequency bands equally, we introduced a single-head self-attention mechanism that self-learns band-wise interactions. Experimental results on benchmark datasets SEED and SEED-IV showed that the proposed model achieved accuracies of 93.48% and 82.22%, respectively, outperforming state-of-the-art methods. In particular, by demonstrating excellent classification performance even on the SEED-IV dataset with fine-grained emotion classes, we confirmed that the proposed attention technique effectively captures EEG patterns of complex emotional states.
This study demonstrated that deep learning technology can significantly improve the accuracy of biosignal recognition, while also presenting the domain gap problem that may occur in real-world applications and directions for solving it. The results of this study are expected to be utilized as core foundational technologies for the implementation of high-precision BCI systems and intelligent healthcare services in the future.

목차 (Table of Contents)

• Ⅰ.서론 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 1
• 1.연구 배경 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙1
• 2.생체신호기반인간상태인식기술의 요구 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 2
• 3.EOG기반안구방향분류의 필요성∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 2
• 4.EEG기반감정인식연구의중요성 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 3
• 5.연구의목적∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙4
• Ⅱ. 관련 연구 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 5
• 1.EOG기반안구방향분류연구∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 5
• 2.EEG기반감정인식연구∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙7
• 3.EEG그래프신경망 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 9
• 4.EEG어텐션메커니즘∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙10
• Ⅲ. 연구 방법 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙12
• 1.데이터셋∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙12
• 1) Curry 기반 고해상도 데이터셋 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙12
• 2) PSG 기반 임상 환경 데이터셋 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙13
• 3) SEED 데이터셋 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙14
• 4) SEED-IV 데이터셋 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙15
• 2. 신호 전처리 및 특징 추출 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙16
• 3. EOG 기반 안구 방향 분류 모델링 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙18
• 1) 머신러닝 모델 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙19
• 2) 딥러닝 모델 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙19
• 4. EEG 기반 감정 인식 모델링 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙21
• 1)스칼라밴드가중치 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙23
• 2) 밴드 어텐션 및 특징 융합 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙24
• 3) 단일 헤드 셀프 어텐션 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙24
• 4) 다중 헤드 셀프 어텐션 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙25
• Ⅳ. 실험 및 결과 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙27
• 1.실험 환경 및설정∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙27
• 2.EOG기반안구방향분류실험결과∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙28
• 1) 머신러닝 모델 성능 비교 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙28
• 2) 딥러닝 모델 성능 비교 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙29
• 3) 교차 도메인 평가 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙31
• 3.EEG기반감정인식실험결과∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 32
• 4.결과 요약 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙34
• Ⅴ. 고찰 및 결론 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙35
• 1.고찰 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 36
• 2.한계점및향후 연구 방향 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 36
• 1) 이산적 움직임 평가의 제약과 실시간 처리 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙36
• 2) 도메인 간 격차(Domain Shift) 해소 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙37
• 3)고정된특징추출방식의 의존성∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙37
• 4) 감정의 복잡성과 멀티모달 융합 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙38
• 3.결론 ∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 38
• 참고문헌∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙ 40
• ABSTRACT∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙∙47