多源域迁移学习的肌电-惯性特征融合及手势识别

Multi-source domain transfer learning for electromyography-inertial feature fusion and gesture recognition

在跨用户手势识别研究中,针对单源域迁移学习存在的负迁移和模型泛化性能差的问题,本研究创新性地提出了一种基于肌电-惯性特征融合的多源域迁移学习策略,关键创新点在于整合多个源域的数据,并在此基础上采用域特有特征对齐与域分类器对齐的技术手段。这一方法旨在强化模型在不同用户间的手势识别性能,进而显著提升跨用户手势识别系统的准确性。首先,引入长短时记忆(long short-term memory, LSTM)网络模型,提取肌电-惯性信息的平均绝对值、方差、峰值等时序特征;其次进行域特有特征对齐与域分类器对齐,利用多个源域数据完成对目标域的特征提取;最后融合分类损失、域特有特征差异损失和域分类器差异损失3个损失函数,协同优化整体损失。实验结果表明,所提方法与单源域、源域组合等多种传统方法相比,识别平均率有所提高,在NinaPro DB5数据集上,目标用户的手势识别平均准确率达到80%以上。

In the realm of cross-user gesture recognition research,addressing the challenges of negative transfer and subpar model generalization observed in single-source domain transfer learning,this study presents a multi-source domain transfer learning strategy centered around the fusion of EMG and inertial features.The pivotal innovation lies in amalgamating data resources originating from diverse source domains,and subsequently employing techniques for domain-specific feature alignment and domain classifier alignment.The primary objective of this approach is to bolster model performance in gesture recognition across different users,thus significantly enhancing the accuracy of cross-user gesture recognition systems.Initially,the long short-term memory(LSTM)network model is introduced to extract time series features,encompassing metrics such as average absolute value,variance,and peak value derived from EMG and inertia data.Subsequently,domain-specific feature alignment and domain classifier alignment procedures are executed,facilitating feature extraction within the target domain utilizing data from multiple source domains.Lastly,the fusion of three loss functions—classification loss,domain-specific feature difference loss,and domain classifier difference loss—is undertaken to collectively optimize the overall loss.The experimental results demonstrate that the proposed method exhibits an improvement in average recognition rate compared to various traditional methods,such as single-source domain and source domain combination approaches.On the NinaPro DB5 dataset,the average gesture recognition accuracy for the target users exceeds 80%.