基于卷积神经网络的脑电信号检测与脑机接口实现

发布时间：2018-07-18 21:23

【摘要】：随着人工智能算法的快速发展以及医疗系统的应用需求,许多国家和地区开始建立以大脑为研究核心的脑计划项目,脑-计算机接口(Brain-Computer Interface, BCI)技术研究正是脑计划项目中的研究热点之一。BCI系统通过全新的方式建立大脑与外界设备的连接。根据获取脑电诱发电位的方式不同可将BCI系统分为两类：基于运动想象的脑-机接口系统和基于诱发电位的脑-机接口系统。基于运动想象的脑-机接口系统不需要通过与外界交流来获取诱发电位,只需要受试者自身的单边肢体运动或想象运动,从而在大脑对侧产生相对应的事件诱发电位；基于诱发电位的脑-机接口系统主要通过与外界的交流来获取诱发电位,一般基于视觉刺激,通过特定频率的刺激信号诱发大脑皮层产生相关电位,典型的诱发电位主要有P300信号与稳态视觉诱发电位(Steady-State Visually Evoked Potentials, SSVEP)。脑-机接口系统的关键在于有效诱发信号的获取以及信号的特征提取,传统的BCI分类算法主要包括线性判别分析算法(Linear Discriminant Analysis, LDA)及其规则化算法：逐步线性判别分析算法(Stepwise Linear Discriminant Analysis, SWLDA)、收缩线性判别分析算法(Shrinkage Linear Discriminant Analysis, SKLDA)。 SWLDA和SKLDA算法主要解决LDA算法存在的小样本问题,虽然其能够在一定程度上改善LDA算法的小样本问题,但其在充足样本下的算法识别率无法有效保持。近年来,随着深度学习算法的发展,其在图像处理领域的应用已经展现出巨大的潜力,对于图像识别的性能有了质的提高。在其他领域,如语音识别、生物医学信号处理等,关于深度学习的研究与应用也在不断深入。本论文基于深度学习算法典型模型,主要针对脑电信号中的P300信号进行研究；同时基于SSVEP信号,设计并实现基于OMAP3530的脑-机接口系统。论文研究成果如下：1)基于LDA、SWLDA、SKLDA算法,以BCI Ⅲ竞赛数据为基础,测试85个目标字符在三种算法下的分类准确率,可以验证SWLDA算法、SKLDA算法能够在一定程度上改善小样本问题,但无法保证充足样本下的算法分类准确率。2)设计基于卷积神经网络(Convolutional Neural Netwoks, CNN)算法模型,借鉴CNN算法在图像特征学习中的“局部感受野”与“权值共享”思想,以P300实验数据为基础,不断进行特征学习、特征表示以及各隐藏层权重参数的优化。以185个目标字符作为实验数据,得出不同训练字符、不同样本叠加平均下的分类识别率,并与传统算法进行详细对比,分类准确率有了进一步提高,并且随着样本数目的增多算法会有更好的性能。与传统算法相对比,算法具有更好的泛化能力,对于深度学习算法模型与脑-机接口系统的融合具有重要参考价值。3)基于稳态视觉诱发电位,设计并实现基于OMAP3530的脑-机接口系统。通过LED视觉刺激器的设计以及诱发电位的有效采集与传输,搭建完整的脑-机接口系统。受试者通过简单的训练学习即可应用,且系统通过无线桥接技术以及滑动窗口机制,使得系统识别率在90%以上,系统实时信息交互时间小于2秒,完全满足实际需求,为运动障碍病人带来福音。
[Abstract]:With the rapid development of artificial intelligence algorithm and the application requirement of medical system , brain - computer interface ( BCI ) technology is one of the hot topics in brain planning project . BCI system can be divided into two categories : brain - computer interface system based on motion picture and brain - computer interface system based on evoked potential .
The brain - computer interface system based on evoked potential is mainly used to acquire evoked potential through communication with the outside , which is generally based on visual stimulus . The evoked potential is induced by stimulation signals at a specific frequency . The typical evoked potential is mainly composed of 300 P signal and steady state visual evoked potential ( SSVEP ) . The key of the brain - computer interface system is the acquisition of effective evoked signals and the feature extraction of signals . The traditional BCI classification algorithm mainly includes linear discriminant analysis ( LDA ) and its regularization algorithm : stepwise linear discriminant analysis ( SWLDA ) and contraction linear discriminant analysis algorithm ( SKLDA ) . The SWLDA and SKLDA algorithms mainly solve the problem of small sample in LDA algorithm , although it can improve the small sample problem of LDA algorithm to some extent , but its algorithm recognition rate can ' t be maintained effectively . In recent years , with the development of depth learning algorithm , its application in image processing has shown great potential .
A brain - computer interface system based on OMAP3530 is designed and implemented based on SSVEP signal . The research results are as follows : 1 ) Based on LDA , SWLDA and SKLDA algorithm , the classification accuracy of 85 target characters under three algorithms is tested .
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：R749;TN911.23

【相似文献】