采用迁移学习的表面肌电信号手势识别方法

为解决采用表面肌电信号(surface electromyography, sEMG)进行手势识别时电极贴片位移、受试者动作变化等复杂情况下分类识别准确率下降这一问题,提出了一种基于表面肌电信号与迁移学习的手势分类模型。首先对4通道表面肌电信号进行活动段提取与降噪处理,然后提取活动段信号的四种时域特征与两种频域特征。采用流形嵌入分布对齐(manifold embedded distribution alignment,MEDA)方法将源领域和目标领域的特征矩阵嵌入到格拉斯曼流形中进行流形特征学习,减小两域之间的数据差异,消除特征退化；同时根据自适应因子执行动态分布对齐,动态调整数据不同分布差异下边缘分布和条件分布的相对重要性。对多名受试者开展实验以验证所提方法的合理性,实验结果表明:所提方法与决策树(decision tree, DT)、支持向量机(support vector machine,SVM)、k临近(k-nearest neighbor,KNN)三种传统机器学习方法相比,识别准确率分别提高了13%、21%、9%。与未执行流形学习与动态分布对齐的联合分布适配(joint distribution adaptation,JDA)迁移学习方法相比,识别准确率提高了52%,达到93%。

sEMG Gesture Gecognition Method using Transfer Learning

In order to address the problem of decreased classification accuracy due to electrode displacement and variations in subject movements during gesture classification, a gesture classification model based on surface electromyography (sEMG) signals and transfer learning is proposed. Firstly, the four-channel sEMG signals are processed to extract and denoise the active segments. Then, four temporal features and two frequency domain features are extracted from the active segments. The Manifold Embedded Distribution Alignment (MEDA) method is utilized to embed the feature matrices of the source and target domains into the Grassmann manifold for manifold feature learning, aiming to reduce the data differences between the two domains and eliminate feature degradation. Additionally, adaptive factors are employed to perform dynamic distribution alignment, dynamically adjusting the relative importance of marginal and conditional distributions under different distribution disparities. Experimental validation is conducted on multiple subjects to demonstrate the rationality of the proposed method. The results indicate that compared to three traditional machine learning methods, namely Decision Tree (DT), Support Vector Machine (SVM), and k-Nearest Neighbor (KNN), the proposed method achieves an improvement in recognition accuracy of 13%, 21%, and 9% respectively. Furthermore, compared to the Joint Distribution Adaptation (JDA) transfer learning method without manifold learning and dynamic distribution alignment, the recognition accuracy is enhanced by 52% to reach 93%.