P300脑机接口的在线半监督学习算法与系统研究

发布时间：2018-04-12 18:18

本文选题：脑机接口 + 半监督学习　；参考：《华南理工大学》2015年硕士论文

【摘要】：脑机接口(BCI)可以不依赖于人的外部肌肉以及神经组织,直接通过脑与外部环境建立通信机制,因此主要应用于有运动障碍的残疾人以及神经障碍病人的功能辅助与康复。经过数十年的发展,BCI的系统框架和基本方法已经较为完善,但在应用领域仍有大量亟待解决的实际问题。本文针对BCI系统训练时间长的问题以及在线BCI系统的高分类准确率与实时性的特定要求,从两个方面展开了研究工作:1)基于LDA算法提出了一种半监督线性判别分析算法—SUST-ILDA。为了减少系统使用之前的训练时间,采用半监督学习的方式。首先,利用少量有标记样本训练初始LDA分类器模型。之后,采用自训练的方法,利用在线获得的无标记样本逐步对分类器进行更新,以改善分类器的性能;为了减少在线计算复杂度,推导了基于LDA的增量更新模型。理论上,与现有在线半监督学习算法SUST-LSSVM相比,计算复杂度大幅降低且稳定。采用第三次脑机接口竞赛数据进行实验分析,证实了随着在线样本数增加,所提出的算法可以取得与SUST-LSSVM相似的分类准确率,并且稳定收敛。2)设计并实现了一个基于P300的在线半监督字符输入脑机接口系统。与传统有监督P300字符输入脑机接口系统相比,本系统的特点是:1)只需要较短时间的有监督训练之后,即可自动切换到输入模式,大大减少了用户使用系统之前单调沉闷的训练过程。并且随着字符的输入,系统的分类准确率不断提高,达到稳定;2)在系统正常使用阶段,利用在线无标签样本对分类器不断进行更新,在某种程度上对脑电信号的非平稳变化具有一定的自适应性(通过实验观察,尚未经过理论证明)。而一般系统在使用过程中,分类器不再进行更新。其性能由于脑电信号的非平稳变化而降低,进而影响系统的性能。该系统利用了SUST-LSSVM算法在小样本集下更高的分类准确率和SUST-ILDA算法极低的计算复杂度以及足量样本下高的分类准确率。在实现上,采用双分类器更新以及多进程和多线程的处理方式。
[Abstract]:Brain-computer interface (BCI) can directly establish communication mechanism between brain and external environment without relying on human external muscles and neural tissue, so it is mainly applied to the disabled with motor disorders and the functional assistance and rehabilitation of patients with neurological disorders.After decades of development BCI system framework and basic methods have been relatively perfect but there are still a large number of practical problems to be solved in the application field.Aiming at the problem of long training time in BCI system and the special requirement of high classification accuracy and real-time of online BCI system, this paper presents a semi-supervised linear discriminant analysis algorithm (-SUST-ILDAA) based on LDA algorithm.In order to reduce the training time before the use of the system, semi-supervised learning is adopted.First, the initial LDA classifier model is trained with a small number of labeled samples.After that, the self-training method is used to update the classifier step by step by using the unlabeled samples obtained online to improve the performance of the classifier. In order to reduce the computational complexity on line, an incremental update model based on LDA is derived.In theory, compared with the existing online semi-supervised learning algorithm SUST-LSSVM, the computational complexity is greatly reduced and stable.By using the data of the third BCI contest, it is proved that the proposed algorithm can achieve the classification accuracy similar to that of SUST-LSSVM with the increase of the number of online samples.And stable convergence. 2) designed and implemented a P300-based on-line semi-supervised character input brain computer interface system.Compared with the traditional P300 character input brain-computer interface system, the characteristic of this system is that the system only needs a short period of supervised training, then it can switch to input mode automatically.Greatly reduces the user before using the system monotonous training process.And with the input of characters, the classification accuracy of the system is improved continuously, reaching the stability of the system in the normal use phase, the online unlabeled samples are used to update the classifier continuously.To some extent, it has a certain adaptability to the non-stationary change of EEG signal (through experimental observation, it has not been proved by theory.In general, the classifier is no longer updated during the use of the system.Its performance is reduced because of the nonstationary change of EEG signal, which affects the performance of the system.The system makes use of the higher classification accuracy of SUST-LSSVM algorithm in small sample set, the lower computational complexity of SUST-ILDA algorithm and the higher classification accuracy of sufficient sample set.In the implementation, double classifier update and multi-process and multi-thread processing are adopted.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：R49;TP334.7

【参考文献】