运动想象(MI)脑电信号本身是由一组较长且连续的特征值组成的信号序列,传统Transformer模型无法捕捉较长序列之间的依赖,设置固定长度的序列又会产生碎片化问题,因此有待进一步调整和优化。针对上述问题,在传统Transformer模型中加入了片段重用的循环机制和重用之前片段信息的相对位置编码机制,使Transformer模型能够学习更长特征序列的特征信息,同时解决重用片段之间的位置编码信息错乱和重用等问题。然后,通过并行多分支CNN进一步捕捉脑电局部特征。最后,利用竞赛数据集2008 BCI-Competition 2A对改进的Transformer模型性能进行评估。结果表明,在不做任何人工特征提取的前提下,对于四分类数据集,改进Transformer模型的平均准确率和kappa值分别为94.27%和87.34%。
Abstract
The motor imagery (MI) EEG signal itself is a signal sequence consisting of a set of long and continuous feature values. Traditional Transformers cannot capture dependencies between longer sequences, and setting fixed length sequences can lead to fragmentation issues, so further adjustments and optimization are needed. To address the above issues, a fragment reuse loop mechanism and a relative position encoding mechanism for fragment information have been added to traditional Transformers. This can enable the Transformer to learn feature information from longer feature sequences. At the same time, it can also solve problems such as confusion and reuse of positional encoding information between reused fragments. Subsequently, parallel multi branch CNN was used to further capture local EEG features. Finally, the performance of the improved Transformer model was evaluated using the competition dataset 2008 BCI-Competition 2A. The results showed that, without any manual feature extraction, the average accuracy and kappa value of the improved Transformer model for the four classification dataset were 94.27% and 87.34%, respectively.
关键词
计量学 /
脑电信号 /
运动想象 /
脑机接口 /
卷积神经网络 /
Transformer模型 /
片段循环机制 /
相对位置编码
Key words
metrology /
EEG signal /
motor lmagery /
BCI /
convolutional neural networks /
transformer model /
loop mechanism for segment /
relative position encoding
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
［1］Du Y, Xu Y, Wang X, et al. EEG temporal-spatial transformer for person identification ［J］. Scientific Reports, 2022, 12(1): 14378.
［2］Yan J, Li J, Xu H, et al. Seizure prediction based on transformer using scalp electroencephalogram ［J］. Applied Sciences, 2022, 12(9): 4158.
［3］Sun J, Wang X, Zhao K, et al. Multi-Channel EEG Emotion Recognition Based on Parallel Transformer and 3D-Convolutional Neural Network ［J］. Mathematics, 2022, 10(17): 3131.
［4］Wang Z, Wang Y, Hu C, et al. Transformers for EEG-based emotion recognition: A hierarchical spatial information learning model ［J］. IEEE Sensors Journal, 2022, 22(5): 4359-4368.
［5］李妍, 高闽光, 童晶晶, 等. 基于巴特沃斯滤波器的傅里叶变换红外光谱处理方法研究 ［J］. 量子电子学报, 2021, 38(6): 780-787.
Li Y, Gao M G, Tong J J, et al. Research on Fourier transform infrared spectroscopy processing method based on Butterworth filter ［J］. Chinese Journal of Quantum Electronics, 2021, 38(6): 780-787.
［6］Barry R J, De B. EEG frequency PCA in EEG-ERP dynamics ［J］. Psychophysiology, 2018, 55(5): e13042.
［7］骆金晨, 邹任玲. 基于多尺度排列熵的睡眠阶段阈值研究 ［J］. 软件导刊, 2021, 20(3): 119-123.
Luo J C, Zhou R L. A study of sleep stage thresholds based on multi-scale alignment entropy ［J］. Software Guide, 2021, 20(3): 119-123.
［8］Bai H, Shi P, Lin J, et al. Segatron: Segment-aware transformer for language modeling and understanding ［C］// National Conference on Artificial Intelligence, Beijing, China, 2021.
［9］Bengtsson C. Status and trends in transformer monitoring ［J］. IEEE Transactions on Power delivery, 1996, 11(3): 1379-1384.
［10］Bhatt D, Patel C, Talsania H, et al. CNN variants for computer vision: history, architecture, application, challenges and future scope ［J］. Electronics, 2021, 10(20): 2470.
［11］张立峰, 王智. 基于多元经验模态分解与卷积神经网络的气液两相流流型识别 ［J］. 计量学报, 2023, 44(1): 73-79.
Zhang L F, Wang Z. Flow Pattern Recognition Method of Gas-Liquid Two-Phase Flow Based on Multiple Empirical Mode Decomposition and Convolution Neural Network［J］. Acta Metrologica Sinica, 2023, 44(1): 73-79.
［12］Bang J S, Lee M H, Fazli S, et al. Spatio-spectral feature representation for motor imagery classification using convolutional neural networks ［J］. IEEE Transactions on Neural Networks and Learning Systems, 2021, 33(7): 3038-3049.
［13］Li M, Han J, Yang J. Automatic feature extraction and fusion recognition of motor imagery EEG using multilevel multiscale CNN ［J］. Medical & Biological Engineering & Computing, 2021, 59(10): 2037-2050.
［14］周强, 田鹏飞. 基于迁移学习多层级融合的运动想象EEG辨识算法 ［J］. 电子测量与仪器学报, 2021, 35(12): 174-181.
Zhou Q, Tian P F. Motion imagery EEG recognition algorithm based on migration learning multilevel fusion ［J］. Journal of Electronic Measurement and Instrumentation, 2021, 35(12): 174-181.
［15］Agarwal P, Kumar S. Electroencephalography based imagined alphabets classification using spatial and time-domain features ［J］. International Journal of Imaging Systems and Technology, 2022, 32(1): 111-122.
［16］金海龙, 邬霞, 樊凤杰, 等. 基于GST-ECNN的运动想象脑电信号识别方法 ［J］. 计量学报, 2022, 43(10): 1341-1347.
Jin H L, Wu X, Pan F J, et al. A GST-ECNN-based method for recognition of motor imagery EEG signals ［J］. Acta Metrologica Sinica, 2022, 43(10): 1341-1347.
［17］Majidov I, Whangbo T. Efficient classification of motor imagery electroencephalography signals using deep learning methods ［J］. Sensors, 2019, 19(7): 1736.
［18］付荣荣,隋佳新,刘冲,等. 基于共空间模式和均匀流形投影的运动想象脑电信号识别方法［J］. 计量学报, 2022, 43(8): 1103-1108.
Fu R R, Sui G X, Liu C, et al. Recognition of Motor Imagery Electroencephalogram Signal Based on Common Spatial Pattern and UMAP［J］. Acta Metrologica Sinica, 2022, 43(8): 1103-1108.
基金
内蒙古自然科学基金(2022MS06008)
PDF(47348 KB)
