표제지
ABSTRACT
목차
1. 서론 10
1.1. 연구 배경 10
1.2. 연구 내용 11
2. 관성자기센서 기반 자세 쿼터니언 추정을 위한 순환신경망 13
2.1. 서론 13
2.2. 방법 15
2.2.1. 제안 신경망 구조 15
2.2.2. 쿼터니언 기반 자세추정 평가 16
2.2.3. 모델 학습 과정 17
2.3. 실험 및 데이터처리 18
2.3.1. 실험 및 데이터 구성 18
2.3.2. 데이터 증강 21
2.4. 결과 및 고찰 22
2.5. 결론 29
3. 관성자기센서 기반 자세 행렬 추정을 위한 병렬 순환 신경망 31
3.1. 서론 31
3.2. 방법 33
3.2.1. DCM 기반 자세 추정 33
3.2.2. 병렬 신경망 구조 34
3.2.3. 학습 과정 36
3.2.4. 데이터 처리 38
3.3. 결과 및 고찰 39
3.3.1. 신경망 구조 분석 39
3.3.2. 기존 필터 알고리즘과의 비교 40
3.3.3. 쿼터니언 기반 RNN과의 비교 45
3.4. 결론 46
4. 순환신경망에서 데이터 증강이 관성자기센서 기반 자세추정 정확도에 미치는 영향 48
4.1. 서론 48
4.2. 방법 50
4.2.1. 신경망 구조 선정 50
4.2.2. 데이터 증강 기법 51
4.3. 실험 및 데이터 처리 53
4.4. 결과 및 고찰 56
4.4. 결론 61
5. 결론 62
참고문헌 64
국문요약 69
Table 2.1. Averaged RMSEs of orientation estimation for each disturbed condition. 23
Table 2.2. Comparison of orientation estimation performances according to the difference of training data. 27
Table 2.3. Averaged RMSEs of attitude estimation for two neural networks. 28
Table 3.1. Estimation performance(mean±SD of RMSEs) according to size and type of neural network 40
Table 3.2. The number of parameters according to size and type of the neural network. 40
Table 3.3. Estimation performance(mean±SD of RMSEs) for KF(Method 1), CF(Method 2), and proposed RNN(Method 3). 42
Table 4.1. Estimation performance(mean of RMSEs) for each network model on test data. 57
Table 4.2. Estimation performance(mean of RMSEs) for each network model on augmented test data. 58
Fig. 2.1. Architecture of the proposed RNN for 3D orientation estimation. 15
Fig. 2.2. Experimental setup 19
Fig. 2.3. RMS of external accelerations vs RMS of angular velocities 20
Fig. 2.4. Flow chart of the neural network training process. 21
Fig. 2.5. Example of estimation errors from Method 1(green), Method 2(red) and Method 3(blue) under accelerated condition. (a) magnitude of external acceleration, (b) attitude error, and (c) heading error. 25
Fig. 2.6. Example of estimation errors from Method 1(green), Method 2(red) and Method 3(blue) under both accelerated and magnetically disturbed condition. (a) magnitude of external acceleration, (b)... 26
Fig. 3.1. Architecture of the proposed parallel neural network. 35
Fig. 3.2. Profile of arccosine and its gradient 37
Fig. 3.3. Example of estimation errors from Method 1(green), Method 2(red) and Method 3(blue) under slow and magnetically disturbed condition. (a) magnitude of external acceleration, (b) magnitude of... 43
Fig. 3.4. Example of estimation errors from Method 1(green) and Method 2(red), Method 3(blue) under fast and magnetically disturbed condition. (a) magnitude of external acceleration, (b) magnitude of... 44
Fig. 3.5. Heading and attitude estimation performance of proposed and quaternion-based neural networks. 46
Fig. 4.1. Experimental setup 54
Fig. 4.2. Various data augmentation schemes. 55
Fig. 4.3. Overall flow chart of the neural network training and validation. 56
Fig. 4.4. Estimation performance(left: mean of RMSEs, right: rate of improvement) on rotated test data according to the number of rotation augmentation. 60