運動想像腦電波之連續式辨別

Abstract

腦機介面是一種利用腦電波訊息作為人類與機器溝通的一種新媒介，隨著腦 機介面近年來技術的發展，使得所謂的以意念控制機器的想法逐漸有了實現的空 間。此技術若能開發成功，除了對於那些患有神經肌肉損傷的病患將有莫大的幫 助外，相信也可以廣泛應用在工業控制上。我們可以藉由腦電波儀取得一個受試 者在執行某些心智動作時所引發出的腦電波，經由精密分析，進一步地將腦電波 轉換為機器控制指令。這樣的溝通系統可以分成同步與非同步兩大類，分類的依 據在於受試者執行心智動作的時間點是否決定於自我意願，一個非同步的腦機介 面意味著使用者可以在任何時間點下達命令；相反地，一個同步的腦機介面則意 味著使用者必須在特定的時間點才可下達命令。目前以非同步的腦機介面較具有 實用價值。 在此論文中，受試者將藉由左右手動作的想像來控制系統，我們的問題在於 如何較精確且穩定地分辨受試者在某個時間點所下達的命令（或心智狀態）。爲 了建立一個較穩定且具有較精確度的非同步腦機介面，我們需要藉由訊號處理與 樣型辨識的技術，包含了訊號前處理、特徵擷取、特徵選取、與辨認。細部來說， 收到的腦電波必須先經由濾波器濾除雜訊和去除生理雜訊，例如眼動。之後，我 們會套用時頻分析(time-frequency analysis)來分析腦電波，由於這些時頻特 徵元素數量龐大，因此，我們還必須應用特徵選取的技術，包含有t 測試（t-test） 和向前特徵選取法（forward feature selection）來降低特徵點的維度和選取 出具有分辨力的特徵點。最後，我們採用單群分辨法（one-class classification）來分辨這些特徵點。 我們分析的資料是來自腦機介面國際型比賽（BCI competition III）所提 供的資料。在這三筆資料中，我們取出C3、Cz、C4三個位置的訊號加以分析，經 由向前特徵選取法處理後，僅取兩個特徵點來分辨受試者的命令。我們的分析結 果得到（1）左（右）邊動作想像和其休息狀態FN和FP平均為20%和30%。（2）左 （右）邊動作想像和非左（右）邊動作想像FN和FP平均為30%和25%。

Brain-computer interface (BCI) provides a communication channel for patients with sever neuromuscular disorders to signal their intentions directly with their brain activities, instead of the normal output pathways between the brain and muscles. When a subject is performing specific tasks, the electroencephalographic (EEG) signals induced by her/his neuronal activities are recorded, analyzed, and translated to the corresponding commands for computers or other devices. A BCI system can be synchronous or asynchronous depending on whether the task is cue-triggered or self-paced. Asynchronous BCI systems are more practical yet more complicated due to the requirement of continuous analysis of ongoing EEG without any timing information about the mental status of the subject. Toward building an asynchronous BCI system, we develop signal preprocessing and classification techniques, including signal preprocessing, feature extraction, feature selection, and classification, that can be used to continuously discriminate between EEG recordings when the subject is resting or performing left-hand/right-hand imagery tasks. The acquired EEG signals are first filtered for artifact removal. Then we use Morlet wavelet to extract time-frequency components. During the training stage, these abundant components are examined through t statistic and forward feature selection and the components with large discernment capability can be determined. During the classification stage, the EEG signals go through the same preprocessing procedure and the discriminative wavelet components are calculated. Then, two one-class classifiers are applied to discriminate the resting state from the left-hand motor imagery and from the right-hand motor imagery, respectively. The one-class classifier focuses only on the feature distribution of one of the motor imagery task on which the subject concentrates. In this way, we do not need to model the widespreading distribution of the resting state, which may comprise slight but fickle mental task. We obtained three datasets from the website of the BCI competition III. From these EEG recordings, three channels (C3, C4, and CZ) are employed and two most discriminative wavelet components are selected by using the proposed techniques. According to our experiments, the false negative (FN) and false positive (FP) recognition rates for both left- and right- hand motor imagery tasks are about ?20% and ?35%, respectively. We i also applied the proposed techniques to discriminate the left-hand (right-hand) motor imagery class from the non-left-hand (non-right-hand) motor imagery class which comprises of resting and right-hand (left-hand) motor imagery EEG data. In this case, the FN and FP recognition rates are about ?30% and ?30%, respectively. These experiments clearly demonstrate the stability and accuracy of the proposed techniques.