腦機介面(Brain Computer Interface, BCI)提供大腦與外部設備之間一個 有效的溝通橋樑，透過腦電圖(Electroencephalogram, EEG)解碼並轉化為指 令，從而實現與外界交流及對外部設備的控制，進而協助肢體運動功能障礙 患者表達意念，並改善現有之生活品質，而想像運動(Motor Imagery, MI)也 已被證實是操作腦機介面的一種有效方式。然而，基於想像運動操作腦機介 面的研究中，常出現無法準確辨認使用者的操作指令以及演算法複雜度高 導致計算時間過長等問題。本論文旨在開發一基於想像運動之腦機介面分 類架構，該架構分別結合以聯合近似對角化為基礎之濾波器組共同空間型 樣法以及黎曼幾何之切線空間投影法以獲取多類別想像運動腦電圖訊號之 特徵，並透過特徵選取保留與類別相關性高之特徵，以降低特徵空間維度， 後續則藉由分類器進行解碼，藉此達到腦電圖訊號分類之目的。此方法不僅 透過聯合近似對角化之方法降低演算法於多類別分類上之計算複雜度，同 時有效提升想像運動之分類性能。最後，經由BCI Competition IV dataset 2a 及自行錄製之數據集進行測試，實驗結果成功地驗證本論文所提出演算法 之有效性；其中，在BCI Competition IV dataset 2a 的數據集測試下，9 位受 試者於四類別想像運動腦電圖訊號之平均分類準確率可達75.39%，而在自 行錄製的腦波數據集測試下，5 位受試者於三類別想像運動腦電圖訊號之平 均分類準確率可達72.26%。;The brain-computer interface (BCI) establishes an effective bridge between the human brain and external devices. BCI is a system capable of decoding electroencephalographic (EEG) signals into device commands to communicate with the external environment and control the devices, thereby assisting patients with executive dysfunction to express their intent and improve the quality of life. Nowadays, motor imagery (MI) has proved to be an effective way to operate BCI. However, BCI based on MI often fails to correctly recognize the user’s mental commands. Here we aim to develop a BCI classification architecture based on MI, which combines the filter bank common spatial pattern based on joint approximate diagonalization and Riemannian tangent space mapping to obtain features from multiclass MI EEG. To prevent over-fitting, we retain the features with a high correlation with the class through feature selection to reduce the dimensionality of the feature space. Finally, use the classifier to decode EEG signals. This architecture not only reduces the computational complexity of the algorithm for multiclass classification through joint approximate diagonalization but also effectively improves the classification performance of MI task. The architecture was validated by the BCI Competition IV dataset 2a and the in-house dataset. The results indicated that our proposed architecture had achieved 75.39% mean accuracy on the BCI Competition IV dataset 2a with four classes of MI tasks and had achieved 72.26% mean accuracy on the in-house dataset with three classes of MI tasks.
