混合腦機接口在康復機器人上的應用
發(fā)布時間：2018-11-05 10:46
【摘要】：腦機接口(Brain-computer interface,BCI)通過借助計算機或其他外部電子設備,旨在建立一種不依賴人體神經和肌肉組織等正常傳輸通道,而直接進行人腦與外界之間信息交流的新途徑。在助殘康復、智能生活、娛樂等領域有著廣泛的應用前景。本文從實現對康復訓練機器人的控制來進行康復訓練出發(fā),以運動想象和P300腦電信號為切入點,并結合他們各自的優(yōu)勢構建混合BCI系統(tǒng)。本文主要做了以下工作:(1)經典的共同空間模式(Common spatial pattern,CSP)用于兩類運動想象的特征提取,通過對CSP進行擴展將其用于多類問題上。本文首先對多類CSP方法一對多CSP(One versus rest CSP,OVR-CSP)進行了研究。由于OVR-CSP濾波器的性能依賴于其選擇的頻帶,當在不合適的頻率段進行濾波的特征上執(zhí)行分類時,其分類精度一般很差。在此基礎上進一步的研究了對頻帶進行固定劃分的Filter bank共同空間模式方法,通過頻帶的劃分雖然能夠進一步提高分類正確率,但卻還是遠低于兩類問題。(2)針對常用多類CSP算法在BCI信號處理方面存在識別率較低的問題,通過引入堆疊降噪自動編碼器(Stacked denoising autoencoders,SDA),提出了一種多類變頻帶運動想象腦電信號的兩級特征提取方法。首先將原始信號通過變化頻率段帶通濾波器得到不同頻段的信號,其次利用OVR-CSP將不同頻段信號變換到使信號方差區(qū)別最大的低維空間,然后通過SDA網絡提取其中可以更好表達類別屬性的高層抽象特征,接著將獲得的特征使用Relief F方法進行特征選擇,選擇出最大權值所對應頻帶的特征,最后使用Softmax分類器進行分類。在對BCI競賽IV中Datasets 2a的4類運動想象任務進行的分類實驗中,平均Kappa系數達到0.70,表明了所提出的特征提取方法的有效性和魯棒性。(3)通過對現有P300范式的研究,提出了一種基于變概率的刺激范式(Variable probability paradigm,VPP)。在該范式中,字符呈現不均勻分布,其密度從中間向兩邊依次減小。字符識別分為兩步進行,先進行隨機行閃爍確定字符所在行,然后所選行中的字符再進行隨機閃爍以確定目標字符。使用該范式和基于區(qū)域的范式進行數據采集及處理,結果表明VPP的信息傳輸率比基于區(qū)域的范式提高約10%,證明了該范式的可行性。(4)為了實現對康復機器人的多維控制,本文設計了一種基于運動想象(Motor imagery,MI)和P300信號的混合BCI控制策略。使用P300信號作為兩種信號間切換的“開關”,選擇以游戲圖標組成的VPP作為游戲菜單的控制面板,MI作為機器人的控制信號來實現患者康復訓練。通過離線數據采集實驗進行模擬控制,結果表明了該系統(tǒng)的可行性。
[Abstract]:The brain-computer interface (Brain-computer interface,BCI) aims to establish a new way to directly communicate information between human brain and the outside world by means of computer or other external electronic devices, which is independent of normal transmission channels such as human nerve and muscle tissue. In the disability rehabilitation, intelligent life, entertainment and other fields have a wide range of applications. In order to control the rehabilitation training robot, this paper starts from the exercise imagination and P300 EEG signal as the starting point, and combines their respective advantages to construct a hybrid BCI system. The main work of this paper is as follows: (1) the classical common space model (Common spatial pattern,CSP) is used to extract the feature of two kinds of motion imagination, and the CSP is extended to multi-class problems. In this paper, one-to-many CSP (One versus rest CSP,OVR-CSP of multi-class CSP methods are studied. Because the performance of OVR-CSP filter depends on its selected frequency band, the classification accuracy is generally very poor when the classification is performed on the characteristics of the filter in an inappropriate frequency band. On this basis, we further study the Filter bank common space pattern method for the fixed division of the frequency band. Though the classification of the frequency band can further improve the classification accuracy, But it is still far lower than two kinds of problems. (2) aiming at the problem of low recognition rate in BCI signal processing of common multi-class CSP algorithm, the stack noise reduction automatic encoder (Stacked denoising autoencoders,SDA is introduced. A two-stage feature extraction method for multi-frequency band motion imagination EEG signals is proposed. First, the original signal is obtained by the bandpass filter in the variable frequency band, and then the signal in different frequency bands is transformed into a low-dimensional space in which the variance of the signal is the most different by using OVR-CSP. Then the high-level abstract features which can better express the category attributes are extracted by SDA network, and then the features obtained are selected by Relief F method, and the features corresponding to the frequency band corresponding to the maximum weights are selected. Finally, Softmax classifier is used to classify. In the classification experiment of four kinds of motion imagination tasks of Datasets 2a in BCI competition IV, the average Kappa coefficient is 0.70, which indicates the effectiveness and robustness of the proposed feature extraction method. (3) by studying the existing P300 normal form, In this paper, we propose a variable probabilistic stimulus normal form (Variable probability paradigm,VPP). In this paradigm, characters are distributed unevenly, and their density decreases from middle to both sides. Character recognition is divided into two steps: the random line flashes to determine the line of the character, and then the character in the selected line is flashed randomly to determine the target character. The results show that the information transfer rate of VPP is about 10% higher than that of region based paradigm. The feasibility of this paradigm is proved. (4) in order to realize multidimensional control of rehabilitation robot, a hybrid BCI control strategy based on motion imagination (Motor imagery,MI) and P300 signal is designed in this paper. The P300 signal is used as the switch between the two signals. The VPP composed of game icons is chosen as the control panel of the game menu and MI is used as the control signal of the robot to realize the rehabilitation training of patients. The simulation control is carried out by off-line data acquisition experiment, and the results show that the system is feasible.
【學位授予單位】：杭州電子科技大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TP242
[Abstract]:The brain-computer interface (Brain-computer interface,BCI) aims to establish a new way to directly communicate information between human brain and the outside world by means of computer or other external electronic devices, which is independent of normal transmission channels such as human nerve and muscle tissue. In the disability rehabilitation, intelligent life, entertainment and other fields have a wide range of applications. In order to control the rehabilitation training robot, this paper starts from the exercise imagination and P300 EEG signal as the starting point, and combines their respective advantages to construct a hybrid BCI system. The main work of this paper is as follows: (1) the classical common space model (Common spatial pattern,CSP) is used to extract the feature of two kinds of motion imagination, and the CSP is extended to multi-class problems. In this paper, one-to-many CSP (One versus rest CSP,OVR-CSP of multi-class CSP methods are studied. Because the performance of OVR-CSP filter depends on its selected frequency band, the classification accuracy is generally very poor when the classification is performed on the characteristics of the filter in an inappropriate frequency band. On this basis, we further study the Filter bank common space pattern method for the fixed division of the frequency band. Though the classification of the frequency band can further improve the classification accuracy, But it is still far lower than two kinds of problems. (2) aiming at the problem of low recognition rate in BCI signal processing of common multi-class CSP algorithm, the stack noise reduction automatic encoder (Stacked denoising autoencoders,SDA is introduced. A two-stage feature extraction method for multi-frequency band motion imagination EEG signals is proposed. First, the original signal is obtained by the bandpass filter in the variable frequency band, and then the signal in different frequency bands is transformed into a low-dimensional space in which the variance of the signal is the most different by using OVR-CSP. Then the high-level abstract features which can better express the category attributes are extracted by SDA network, and then the features obtained are selected by Relief F method, and the features corresponding to the frequency band corresponding to the maximum weights are selected. Finally, Softmax classifier is used to classify. In the classification experiment of four kinds of motion imagination tasks of Datasets 2a in BCI competition IV, the average Kappa coefficient is 0.70, which indicates the effectiveness and robustness of the proposed feature extraction method. (3) by studying the existing P300 normal form, In this paper, we propose a variable probabilistic stimulus normal form (Variable probability paradigm,VPP). In this paradigm, characters are distributed unevenly, and their density decreases from middle to both sides. Character recognition is divided into two steps: the random line flashes to determine the line of the character, and then the character in the selected line is flashed randomly to determine the target character. The results show that the information transfer rate of VPP is about 10% higher than that of region based paradigm. The feasibility of this paradigm is proved. (4) in order to realize multidimensional control of rehabilitation robot, a hybrid BCI control strategy based on motion imagination (Motor imagery,MI) and P300 signal is designed in this paper. The P300 signal is used as the switch between the two signals. The VPP composed of game icons is chosen as the control panel of the game menu and MI is used as the control signal of the robot to realize the rehabilitation training of patients. The simulation control is carried out by off-line data acquisition experiment, and the results show that the system is feasible.
【學位授予單位】：杭州電子科技大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TP242
【參考文獻】
相關期刊論文 前10條
1 汪彩霞;魏雪云;王彪;;基于堆棧降噪自動編碼模型的動態(tài)紋理分類方法[J];現代電子技術;2015年06期
2 馬玉良;許明珍;佘青山;高云園;孫曜;楊家強;;基于自適應閾值的腦電信號去噪方法[J];傳感技術學報;2014年10期
3 徐守晶;韓立新;曾曉勤;;基于改進型SDA的自然圖像分類與檢索[J];模式識別與人工智能;2014年08期
4 劉斌;魏夢然;羅聰;;基于腦電BCI的研究綜述[J];電腦知識與技術;2014年07期
5 劉沖;顏世玉;趙海濱;王宏;;多類運動想象任務腦電信號的KNN分類研究[J];儀器儀表學報;2012年08期
6 劉廣權;黃淦;朱向陽;;共空域模式方法在多類別分類中的應用[J];中國生物醫(yī)學工程學報;2009年06期
7 李明愛;劉凈瑜;郝冬梅;;基于改進CSP算法的運動想象腦電信號識別方法[J];中國生物醫(yī)學工程學報;2009年02期
8 馬，
本文編號：2311821
本文鏈接：http://www.sikaile.net/kejilunwen/zidonghuakongzhilunwen/2311821.html
