使用最大對比光束集成濾波器之腦活動多類動作想像腦機介面系統

Abstract

近年來，腦機介面被廣泛應用在各種領域，像是義肢控制、復健、拼字系統等等，它包含了訊號擷取、訊號前處理、特徵選擇、分類等程序，為了改善效能，提升每一筆資料的分類準確率就是必須的。在本論文中，我們嘗試使用最大對比光束集成濾波器(MCB)來增進分類準確率。 光束集成技術是一種透過各個感測器收集的資料來計算訊號源的方法，這項技術也被廣泛應用於聲納、雷達、腦電波訊號定位等領域。最大對比光束集成濾波器則是一種能夠有效抑制雜訊干擾，並最大化運動想像時活躍時期和控制時期對比的光束集成技術。在本論文中，我們利用實驗資料去估算最大對比光束集成濾波器和相關的大腦訊號源資料。 和使用穩態視覺誘發電位為基底的腦機介面相比，動作想像基底腦機介面有著資訊傳遞速率高和可透過持續訓練增進效能的優點，它依靠使用者進行動作想像時產生的特定腦電波，如事件相關非同步/同步現象，來進行指令判定。事件相關非同步/同步現象是指在動作想像相關事件下，特定頻段中產生的大腦活動能量變化，我們從大腦訊號源資料計算事件相關同步/非同步現象當做我們動作想像任務分類的輸入，並使用序列浮動前向選擇演算法(SFFS)來找尋近似最佳的特徵集合，再來使用支持向量機(SVM)做分類。 總結來說，我們成功地利用最大對比光束集成濾波器改善了動作想像任務的分類準確率，其表現可和傳統的共同空間模式(CSP)方法相比。同時也確認了以單一受測者資料算出的最大對比光束集成濾波器可應用在不同時間點的同一受測者實驗中，並可獲得不錯的動作想像分類準確率。

Recently, brain-computer interface (BCI) based on electroencephalography (EEG) is widely used in many fields like prosthesis control, rehabilitation and spelling systems. Typically BCI systems consist of signal acquisition, signal preprocessing, feature selection, and classification. To improve performance, it is essential to increase accuracy of each trial of EEG data. In this work, we tried to use maximum contrast beamformer (MCB) technique to improve the classification accuracy. Beamforming technique is a source localization method using spatial filter estimated from distributed sensors. The techniques are widely applied to many fields such as sonar, radar, and EEG source localization. MCB is a kind of beamforming technique which can suppress the noises while maximizing the contrast between the active state and the control state. In this work, we used experimental data to estimate MCB filter and used the source data obtained by MCB to perform classification. Compared to BCI using steady-state visual evoked potential, motor imagery based BCI has advantages like high information transition rate and performance improvement through constant training. Motor imagery based BCI depends on identifying the specified EEG change such as event-related desynchronization and event-related synchronization (ERD/ERS) when users perform motor imagery tasks. ERD/ERS are the desynchronization and synchronization of brain activity power in the given frequency band which are related to motor imagery events. We calculated ERD/ERS from brain source data as classification input and adopted sequential floating forward selection to find the approximate optimal feature set using support vector machine. We further obtained final test accuracy with the optimal feature set. In summary, we successfully used MCB to improve classification performance of motor imagery tasks. The performance is comparable to common spatial pattern. Single subject MCB filter can be applied to data acquired on different day and lead to good classification results of motor imagery tasks.