腦機介面(Brain-Computer Interface, BCI)創建了一條不依賴肌肉運動之通訊控制路徑，以促使大腦和外部機器之間的溝通，這是透過測量大腦電位差訊號所產生之波形進行觀察，將腦電圖（Electroencephalography, EEG）訊號轉換為控制命令，從而擴大使用者意念的應用，為意識清楚但肢體功能不完整的人提供新的溝通方式。想像運動(Motor Imagery)是一種對運動行為的心理意念的過程，而不具任何實際的運動表現，且經研究證實，想像運動和實際運動執行之間的大腦神經激活區域相互重疊，使得與想像運動有關之腦機介面開發成為一個新的研究方向。 基於上述所言，論文旨在開發一應用於想像運動之腦機介面的演算法架構，該架構以共同空間型樣法(Common Spatial Pattern)為基礎 進行特徵提取，並以集成學習(Ensemble Learning)理論做為依據，與傳統之單一分類演算法相比，集成策略具有更強的模型穩健性和泛化能力。論文中以BCI Competition III Dataset IVa與 BCI Competition IV Dataset IIa兩公開資料集進行演算法驗證，結果顯示其平均分類準確率分別可達 87.03%與80.18%，可有效提升與想像運動有關之腦電信號的分類性能 。;The brain-computer interface (BCI) creates a communication control path that does not rely on muscle movement to promote communication between the brain and external machines. It is observed by measuring the waveform issued from the electrocortical potential difference signal. Electroencephalography (EEG) signals are converted into control commands, thereby expanding the application of users′ thought and providing a new way of communication for people with normal thinking but incomplete motor functions. Motor imagery (MI) is a process of mental ideation of motor behavior without any actual motor performance. Recent research indicates that the brain nerve activation regions between imaginative movement and actual movement execution overlap each other, making it consistent with imaginary movement. Cause the development of the brain-computer interface to become a new research direction. As stated above, this paper aims to develop an algorithm framework for the brain-computer interface applied to MI. The framework is based on the common spatial pattern for feature extraction and ensemble learning theory. Compared with the traditional single classification algorithm, the ensemble strategy has stronger model robustness and generalization ability. In this paper, two public data sets, BCI Competition III Dataset IVa and BCI Competition IV Dataset IIa, are used for algorithm verification. The results show that the average classification accuracy can reach 87.03% and 80.18%, respectively, which can effectively improve the EEG signals classification performance related to MI.
