面向中高频SSVEP脑机接口的编解码算法研究

基于稳态视觉诱发电位（steady-state visual evoked potentials，SSVEPs）的脑-机接口系统（brain-computer interface，BCI）通常使用低频强闪烁刺激诱发强特征脑电信号。尽管相关数据处理技术日臻成熟，但是系统使用舒适度差，训练时间较长。提高刺激频率能够有效缓解受试者的视觉疲劳，提高系统友好度，然而现有中高频SSVEP系统又存在指令集数量少、信息传输率（information transfer rate，ITR）低等缺陷。针对以上问题，本文基于中高频SSVEP脑电特征，提出并使用了包含空码的Code Words编码范式与集成任务相关成分分析（ensemble task-related component analysis， eTRCA）解码算法，并研究了该套编解码方法的适用性与可扩展性。本研究选择中高频段的4个频率（20、24、30、40 Hz）分别构建脑控字符拼写系统，单个频率的闪烁刺激可独立构建多达6个控制指令，联合多个频率理论上可实现指令集数量的成倍扩增。共有10位健康受试者参与了离线脑电实验，利用18~60 Hz带通滤波对脑电数据进行预处理，使用eTRCA算法进行特征识别。18指令集系统的理论平均分类准确率为96.71±1.69 %，理论平均ITR达86.94±6.07 bits/min。以上结果表明，本研究提出的编解码算法能够有效诱发并准确识别中高频SSVEP的时-频-相多维特征，在此基础上通过增加编码单元频率种类、提高有效编码率、改进解码算法等方式有希望进一步提升系统性能。 

Research on Encoding and Decoding Algorithms for Medium/high-frequency SSVEP-based Brain-computer Interface

? ?Brain-computer interface systems （BCIs） based on steady-state visual evoked potentials （SSVEPs） usually use strong flickering stimuli with low frequencies to evoke signals with strong features. Although the relevant data processing technologies are becoming more and more mature， SSVEP systems with low frequency stimuli are still uncomfortable for using and the time for acquiring enough training samples is quite long. Increasing the frequencies of stimuli can effectively relieve the visual fatigue of the subjects. Hense the friendliness of BCI systems would be greatly improved. However， the existing brain-computer interface systems based on SSVEP with medium or high frequencies have defects such as insufficient recognizable commands， low information transfer rate （ITR）， and so on. In response to the problems listed above， this study proposed a set of coding and decoding methods containing the Code Words paradigm with blank codes and ensemble task-related component analysis （eTRCA） algorithm based on the characteristics of medium and high-frequency SSVEP， of which the applicability and scalability were studied. Four frequencies in the medium and high-frequency bands （20， 24， 30 and 40 Hz） were selected to construct brain-controlled spelling systems （BCI Spellers） respectively， and our method could independently construct 6 commands by using flickering stimulation with a single frequency. It was theoretically possible to multiply the number of identifiable targets by combining Code Words with different frequencies. A total of 10 healthy subjects participated in EEG experiments. After a specified trials of electroencephalography （EEG） data were collected， those data were preprocessed by a bandpass filter （bandwidth 18~60 Hz）. Then we used eTRCA algorithm to perform the target identification process. According to our results， the 18-target offline system achieved an averaged theoretical accuracy of 96.71 ± 1.69 % and a theoretical ITR of 86.94 ± 6.07 bits/min. The above results indicate that the codec method proposed in this study could effectively induce and accurately identify the multi-dimensional features in the time， frequency， and phase domain of medium/high-frequency SSVEP. Besides， increasing the frequency types of coding units， increasing the effective coding rate and improving the decoding algorithm are promising to further improve the system performance.