QTFD与DenseNet相结合的运动想象分类方法

QTFD and DenseNet Based Motor Imagery Classification Method

运动想象脑机接口(Motor Imagery Brain Computer Interface,MI-BCI)技术近年来在医疗康复、娱乐等许多领域得到了广泛的运用。然而,如何处理非平稳的脑电信号(Electroencephalography,EEG),并从中提取可辨识的特征并分类仍然是主要难点之一。针对这个问题,本研究提出了一种基于二次型时频分布(Quadratic timefrequency distributions,QTFD)和密接型网络(DenseNet)的新型MI-EEG分类模型。具体地,我们首先使用QTFD初步提取MI任务相关的脑电时频特征,并构造得到EEG片段的高分辨率时频表示。常用的线性时频分析方法往往会忽略部分非线性信息,难以准确地描述MI信号的能量分布。与线性时频分析方法相比,QTFD方法以二次型变换的形式将信号从时域投影到时频域,能更好地描述信号的能量分布,其对时间和频率的变化具有不变性,能提供较为稳定准确的时频特征。随后,本研究采用了轻量级网络模型DenseNet对时频表示的浅层和深层特征进行逐级提取并整合。DenseNet可训练参数量较少,适用于数据量较少的MI-BCI应用,它在每层网络之间都建立了直接的连接,每一层网络都可以访问之前所有网络的特征图,从而得到更具有区分性的特征表示。最后,本研究在BCI竞赛IV数据集上进行了实验验证,将提出的分类模型与各先进对比算法进行了比较。结果表明,我们所提出的方法在使用脑电通道数更少的情况下,获得了更好的分类性能。

Motor imagery(MI)-based brain-computer interfaces(BCIs)have gained immense popularity in a wide range of fields,including medical rehabilitation and entertainment.However,effectively processing non-stationary Electroencephalography(EEG)signals and extracting recognizable features from them remain major obstacles.In this study,our primary objective was to address this challenge by employing Quadratic Time-Frequency Distribution(QTFD)as an initial step to extract relevant EEG time-frequency features associated with MI tasks.By utilizing QTFD,we constructed high-resolution time-frequency representations of EEG fragments,enabling a more comprehensive analysis.Conventional linear timefrequency analysis methods often overlook nonlinear information and struggle to accurately depict the energy distribution of MI signals.In contrast,QTFD projects the signal from the time domain to the frequency domain using quadratic transformations,resulting in a more precise description of the signal’s energy distribution.Moreover,QTFD exhibits invariance to changes in time and frequency,offering relatively stable and accurate time-frequency characteristics.To further enhance feature extraction,we adopted the lightweight DenseNet network model.This model progressively extracts and integrates shallow and deep features from the time-frequency representations.DenseNet,with its small number of trainable parameters,is particularly well-suited for MI-BCI applications with limited available data.A notable advantage of DenseNet is its direct interconnection between each layer of the network,allowing feature graphs from all preceding layers to be accessed.Consequently,this facilitates the generation of a more distinguishable feature representation.Finally,we conducted experimental validation on the BCI Competition IV dataset,comparing the performance of our proposed classification model with various state-of-the-art algorithms.The results conclusively demonstrate that our method achieves superior classification performance,even when utilizing a reduced number of EEG channels.These findings underscore the efficacy of our approach in effectively processing non-stationary EEG signals and extracting informative features for MI-BCI applications.