摘要
针对运动想象脑电信号特征提取操作繁琐及解码精度低等问题,提出一种基于多视角深度森林的运动想象脑电解码算法。首先,通过子频带滤波及时间窗口划分对原始信号进行细粒度分析,生成空时频能量特征。然后,对上述空时频能量特征分别进行稀疏选择和时序扫描得到重要的浅层能量特征及多示例先验类别特征。继而,将上述两类特征进行融合构建运动想象脑电多视角特征集。最后,利用级联森林的逐层特征变换挖掘深层次的抽象特征进行脑电解码。根据脑机接口竞赛数据和自行采集的数据进行算法测试,并与单视角特征模型、传统共空间模式方法以及深度神经网络算法进行对比。在2个脑机接口竞赛数据集和1个真实数据集上分别取得了91.4%、75.2%和70.7%的最高平均分类准确率,结果表明该文所提多视角深度森林算法具有更优的分类识别准确率。
A decoding algorithm based on multi view deep forest is proposed for solving the problems of complicated feature extraction and low decoding accuracy of motor imagery EEG signal.The energy feature in spatial-temporal-frequency domains which is generated by fine-grained analysis using subband filtering and time window division are processed with sparse selection and temporal scanning for obtaining important shallow energy features and multi instance a priori category features to construct a multi view feature set.Then the hierarchical feature transformation of cascaded forests is used to mine deep level abstract features for EEG coding.The algorithm is tested on two BCI competition datasets and a self-collected dataset,and it is compared with single view feature models,traditional CSP methods and deep neural network algorithms.The proposed method achieves the highest average classification accuracy(91.4%,75.2%and 70.7%,respectively)in 2003BCI competition datasetⅢ,2008 BCI competition dataset 2b and self-collected dataset,which suggests that the decoding algorithm based on multi view deep forest has better classification performance.
作者
郑龙鑫
苗敏敏
徐宝国
胡文军
ZHENG Longxin;MIAO Minmin;XU Baoguo;HU Wenjun(School of Information Engineering,Huzhou University,Huzhou 313000,China;Zhejiang Province Key Laboratory of Smart Management and Application of Modern Agricultural Resources,Huzhou 313000,China;School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China)
出处
《中国医学物理学杂志》
CSCD
2022年第9期1159-1166,共8页
Chinese Journal of Medical Physics
基金
国家自然科学基金(62101189,U20A20228,61772198)
江苏省前沿引领技术基础研究专项(BK20192004)
浙江省基础公益研究项目(LGN18F020002)。
关键词
运动想象脑电信号
解码算法
多视角
特征提取
深度森林
motor imagery EEG signal
decoding algorithm
multi view
feature extraction
deep forest
