Title Page
Abstract
초록
Contents
1. Introduction 14
2. Related Works 17
2.1. Challenges in EEG-Based Grasp Task Decoding for BCI 17
2.2. Contemporary Techniques for Developing Robust BCIs 18
3. Kinematic Representation-Based Deep Neural Network 20
3.1. Muscle Activity Detection and Prediction 20
3.2. Muscle Synergy Extraction via Non-negative Matrix Factorization 24
3.3. Classifying Synergy Activation Coefficients 25
4. Experiment 31
4.1. Participants 31
4.2. Experimental Setup 31
4.3. Experimental Protocol 32
4.4. Data Collection and Processing 32
5. Results 35
5.1. Evaluation Methods 35
5.1.1. Offline Model Benchmarking 35
5.1.2. Online Experiment Implementation 37
5.2. Decoding Performance 38
5.2.1. Offline Model Results 38
5.2.2. Multi-Class Confusion Matrix Analysis 40
5.2.3. Online Model Performance Analysis 42
5.2.4. Generated Images and Classification Insights 43
6. Discussion 47
7. Conclusion 55
Reference 56
Table 3.1. Specifications of the neural network: details of parameters and layer implementations 28
Table 3.2. Training procedure of KRDNN (Part 1) 29
Table 3.3. Training procedure of KRDNN (Part 2) 30
Table 5.1. Performance evaluation results on offline session. Grand-averaged classification accuracy of four grasp types using the proposed and comparable methods 39
Table 5.2. Performance evaluation results on offline session. Grand-averaged classification accuracy of two-class grasp categories using proposed and comparable methods 40
Table 5.3. Online session results across representative subjects. Trials were randomly given among the four classes. The number for each class was distributed equally 45
Table 5.4. Online session results across representative subjects: Reorganized results of four classes into binary 46
Figure 3.1. Flowchart representation of the training phase within the proposed dual-stage deep learning framework. The first stage creates synergy activation coefficient–based... 21
Figure 3.2. Flowchart representation of the test phase within the proposed dual-stage deep learning framework. This phase leverages the network trained during the preced-... 22
Figure 4.1. Experimental tasks consisted of four hand–grasp types for different objects such as a cup, ball, card, and bolt; cylindrical, spherical, lateral, and pincer grasps. The... 33
Figure 4.2. Experimental protocol of a single trial. The duration of a single trial was 10 s, wherein the subjects performed motor imagery (MI) during the designated stage for 4... 34
Figure 5.1. Experimental setup and environment for online session. The control PC decodes EEG signals in real–time and presents visual feedback of the classification result... 36
Figure 5.2. Experimental setup and environment for the online session mirror that of the calibration phase. A control PC decodes real-time EEG signals and provides visual... 37
Figure 5.3. Confusion matrices for each class across all subjects. These matrices encompass various grasp types, namely Cylindrical (Cyl), Spherical (Sph), Lateral (Lat), and... 41
Figure 5.4. Presenting confusion matrices for each class, exclusive to the high- performance subject group which has achieved an accuracy of 0.68 or above. 41
Figure 5.5. Presenting confusion matrices of each class, specific to the low-performance subject group with an accuracy of less than 0.68. 42
Figure 6.1. Visualization of true and predicted muscle SAC images by class. The score indicates the similarity with the true image on the left side of the predicted image. The proposed framework has the ability to visually identify the MI, which... 52
Figure 6.2. Visualization of true and predicted muscle SAC images for the 'Sph' class. The high similarity score is indicative of the KRDNN's ability to accurately distinguish... 53
Figure 6.3. Visualization of true and predicted muscle SAC images for the 'Lat' class. Distinct 'Lat'-specific muscle movements can be identified, and the image achieving the... 53
Figure 6.4. Visualization of true and predicted muscle SAC images for the 'Pin' class. Muscle movements unique to 'Pin' are clearly displayed, and the highest scoring image... 54