摘要
运动想象脑电(MI-EEG)信号在无创脑机接口(BCI)中扮演重要角色,已被广泛应用于临床辅助康复训练。作为一个主动刺激范式,MI-EEG的样本采集成本高且个体差异大,同时具有复杂时变性和低信噪比等特点,使得构建跨被试MI-EEG解码模型成了研究关键。然而,当前绝大多数的跨被试解码模型采用单阶段的对抗学习策略,并且仅考虑学习边缘分布或条件分布最小化的深度表征,进而严重制约了MI-EEG解码性能。因此,提出一种多阶段分布适应(MSDA)的跨被试MI-EEG解码模型。首先,采用样本协方差对齐被试者之间的样本边缘分布差异;其次,通过预训练特征提取器和域鉴别器获取边缘分布不变的深度表征;最后,通过L2距离构建深度表征的联合分布不变映射,并交替训练联合这些映射和分类器学习联合分布不变的深度表征用于跨被试MI-EEG解码。MSDA模型分别从样本边缘分布、深度表征边缘分布和深度表征联合分布这3个阶段进行被试者之间的分布适应,从而有效地应对单阶段分布适应的挑战。在BCI Competition IV-2a和BCI Competition IV-2b公开数据集上的实验结果表明,MSDA模型在准确率和Kappa系数这2个指标均超越了近期提出的解码模型。可见,MSDA模型提升了跨被试域不变深度特征的学习能力,为构建MI-BCI提供了新的选择。
Motor Imagery ElectroEncephaloGraph(MI-EEG)signals play a significant role in non-invasive Brain-Computer Interface(BCI)and have been utilized in clinical rehabilitation training widely.As one of the subjective paradigms,MI-EEG has high sample collection costs and large individual differences with complex time variability and low signal-to-noise ratio,so that constructing cross-subject MI-EEG decoding models have become a critical research focus.However,most of the recent cross-subject decoding models adopt the single-stage adversarial learning strategy,and only consider to learn deep representations with marginal and conditional distribution minimization,which constrain the MI-EEG decoding performance seriously.Therefore,a Multi-Stage Distribution Adaptation(MSDA)model was proposed for cross-subject MI-EEG decoding.Firstly,sample covariance was employed to align marginal distribution differences between subjects.Secondly,marginal distribution-invariant deep representations were obtained through pre-trained feature extractor and domain discriminator.Finally,a joint distribution-invariant mapping of deep representations was constructed using L2-distance,and such mapping and classifiers were trained alternately to learn joint distribution-invariant deep representations and used for cross-subject MI-EEG decoding.In MSDA model,distribution adaptation between subjects were conducted in three stages,including sample’s marginal distribution,deep representations’marginal distribution and deep representations’joint distribution,thereby addressing the challenge of single-stage distribution adaptation effectively.Experimental results on the BCI competition IV-2a and BCI Competition IV-2b public datasets demonstrate that MSDA model surpasses the latest decoding models in both accuracy and Kappa coefficient.The above indicates that MSDA model enhances the learning ability of cross-subject domain-invariant deep representations,which offers a new option for building MI-BCI.
作者
何敏
罗天健
HE Min;LUO Tianjian(College of Computer and Cyber Security,Fujian Normal University,Fuzhou Fujian 350117,China;Digital Fujian Environmental Monitoring IoT Laboratory(Fujian Normal University),Fuzhou Fujian 350117,China)
出处
《计算机应用》
北大核心
2025年第12期4064-4072,共9页
journal of Computer Applications
基金
国家自然科学基金资助项目(62106049)
福建省自然科学基金资助项目(2022J01655)。
关键词
运动想象
脑电信号
跨被试
域不变深度表征
域适应
Motor Imagery(MI)
ElectroEncephaloGraph(EEG)signal
cross-subject
domain-invariant deep representation
domain adaptation
