多任务脑—机接口导联选择与分类算法研究

发布时间：2018-08-11 19:10

【摘要】：近年来,脑-机接口系统的发展得到了众多国内外研究人员的关注,并成为神经工程领域一个新的研究热点。随着脑-机接口的不断发展,研究人员正试图将BCI技术应用到实际生活中。但是,对基于运动想象的脑-机接口系统,当导联数目过多时会导致共空域模式算法的性能降低,因此脑-机接口系统一直难以在现实生活中应用。另外,随着导联数目的增加,多通道脑电信号的记录需要复杂的准备工作以及复杂的计算过程,既费时又可能降低分类正确率。为解决此问题,本文提出四种导联选择算法：第一种是二进制多目标粒子群优化算法(bMOPSO);第二种是基于培养的多目标微粒群优化算法(cMOPSO);第三种是L1-norm算法；第四种是基于滤波器组的cMOPSO算法(fbMOPSO)。bMOPSO算法和两种cMOPSO算法都是基于微粒群(PSO)的,微粒群优化是一种启发式研究技术,它是对鸟群寻找食物时的活动规律性进行建模与仿真。起初,PSO算法主要用来解决单目标优化问题,MOPSO算法是PSO算法的扩展,它在解决多目标优化问题中应用也越来越广泛。大部分MOPSO算法在整个进化过程中对所有粒子使用固定的权重和加速度,与bMOPSO算法不同,cMOPSO算法通过引入一个培养架构去适应变异粒子的个体飞行参数。第三种算法是典型的L1-norm算法,这个方法对两分类数据能够选择较少的导联数目。另外,本文用三种分类算法对脑电数据进行分类：支持向量机(SVM)、k-近邻法(k-NN)和BP神经网络算法。最后,分别对三种分类算法和四种导联选择算法进行实验仿真。仿真结果表明,三种分类算法中,支持向量机和k-近邻法分类效果比BP神经网络好。四种导联选择算法在没有降低分类正确率的情况下,基于fbMOPSO算法能够选择更少的导联数目。
[Abstract]:In recent years, the development of brain-computer interface system has been paid attention by many researchers at home and abroad, and has become a new research hotspot in the field of neural engineering. With the development of brain-computer interface, researchers are trying to apply BCI technology to real life. However, for the brain-computer interface system based on motion imagination, when the number of leads is too large, the performance of the common spatial mode algorithm will be reduced, so the brain-computer interface system has been difficult to be applied in real life. In addition, with the increase of the number of leads, the recording of multichannel EEG signals requires complicated preparation and complicated calculation process, which may be time-consuming and reduce the classification accuracy. In order to solve this problem, four lead selection algorithms are proposed in this paper: first, binary multi-objective particle swarm optimization algorithm (bMOPSO); second, culture-based multi-objective particle swarm optimization algorithm (cMOPSO); third, L1-norm algorithm; The fourth one is cMOPSO algorithm (fbMOPSO). BMOPSO based on filter banks, and both cMOPSO algorithms are based on PSO (PSO). PSO is a heuristic research technique, which is to model and simulate the regularity of birds' activities when looking for food. At first, the PSO algorithm is mainly used to solve the single objective optimization problem. It is an extension of the PSO algorithm, and it is used more and more widely in solving the multi-objective optimization problem. Most MOPSO algorithms use fixed weights and accelerations for all particles throughout the evolution process. Unlike the bMOPSO algorithm, the CMOPSO algorithm adapts to the individual flight parameters of the mutated particles by introducing a culture framework. The third algorithm is the typical L1-norm algorithm, which can select fewer leads for the two classification data. In addition, this paper uses three classification algorithms to classify EEG data: support vector machine (SVM) k- nearest neighbor method (k-NN) and BP neural network algorithm. Finally, three classification algorithms and four lead selection algorithms are simulated. The simulation results show that support vector machine and k- nearest neighbor method are better than BP neural network in the classification of the three classification algorithms. The four lead selection algorithms can select fewer leads based on fbMOPSO without reducing the classification accuracy.
【学位授予单位】：南昌大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP334.7

【相似文献】