文章摘要

文章信息/Info

Title:: Research on feature transfer learning algorithm of motor imagery EEG signal based on Riemannian space

Author(s):: GAO Nuo; WANG Yunhui; （Information & Electrical Engineering Department, Shandong Jianzhu University, Jinan 250101, China）

Keywords:: Brain-computer interface; Motor imagery; Transfer learning; Riemannian space; Joint distributed adaptation

摘要:: 由于脑电信号具有低信噪比、非平稳等特点，传统脑机接口需对用户执行长时间的校准训练，才能建立可靠、准确的分类模型。针对当前迁移学习在脑电信号上分类准确率低的问题，本研究提出了基于黎曼空间特征迁移学习（Riemannian space feature transfer learning，RFTL）的运动想象脑电信号分类算法。该算法首先在黎曼空间对源域和目标域数据进行分布对齐后，利用联合分布适配减少不同域间的数据分布差异，构建适用于目标域任务的域不变分类器模型。实验结果表明，RFTL算法可有效解决跨域分布的不一致性，显著提高运动想象脑电信号跨对象的识别准确率，改善脑机接口研究中的通用性问题。

Abstract:: Due to the low signal-to-noise ratio and non-stationary characteristics of electroencephalography(EEG) signals, traditional brain-computer interfaces（BCI） need to perform long-term calibration training for the current user in order to establish a reliable classification model.To solve the low classification accuracy of current transfer learning algorithms, we proposed a motor imagery EEG signal classification algorithm based on Riemannian space feature transfer learning (RFTL). Firstly the distribution of source domain and target domain data were aligned in Riemannian space, then the joint distribution daptation was used to reduce the difference in data distribution between different domains, a domain invariant classifier model suitable for target domain tasks was constructed.The experimental results showed that the RFTL algorithm could effectively solve the inconsistency of cross-domain distribution. It significantly improves the recognition accuracy of motor imagery EEG signals across objects, and improves the generality of BCI research.