采用变分模态分解与领域自适应的表面肌电信号手势识别

Gesture Recognition of Surface Electromyography Based on Variational Mode Decomposition and Domain Adaptation

针对传统机器学习在表面肌电信号手势识别领域的适应性和准确性不足,以及新用户因个体生理和行为差异在已有模型上表现不佳的问题,提出一种利用卷积神经网络模型并有效克服肌电数据分布差异的算法,用于提升手势识别的性能。首先对肌电信号进行变分模态分解,构建易于识别的表面肌电图像,并提出了一种卷积神经网络模型进行手势识别,提升用户相关的肌电信号手势识别准确率;同时利用迁移学习中的领域自适应和模型微调技术,提升用户无关的肌电信号手势识别准确率,并将所提算法在NinaPro DB1肌电数据集中进行了3分类、4分类、5分类和12分类共4组评估验证。结果表明:在4组评估验证中,用户相关的肌电信号手势识别平均准确率分别达到了99.28%、99.30%、98.39%和93.40%,用户无关的肌电信号手势识别平均准确率分别达到了94.05%、92.60%、88.38%和70.03%,表明本文提出的算法在表面肌电信号手势识别中具有良好的效果,为实现人机交互中的普适性的肌电设备开发提供了一种可行的方案。

In view of the lack of adaptability and accuracy of traditional machine learning in the field of gesture recognition of surface electromyography(sEMG),as well as the poor performance of new users on existing models due to individual physiological and behavioral differences,an algorithm using convolutional neural network model to effectively overcome the difference in the distribution of EMG data is proposed to improve the performance of gesture recognition.A novel approach combining variational mode decomposition(VMD)of sEMG signal and convolutional neural network is proposed to enhance the performance of gesture recognition in the subject.In addition,domain adaptation and model fine-tuning techniques of transfer learning are used to increase the accuracy of gesture recognition of intra-subject sEMG.The proposed algorithm is evaluated and validated on four evaluation groups of the benchmark dataset DB1 with 3,4,5 and 12 classifications.In the four sets of evaluation,the average accuracies of inter-subject are 99.28%,99.30%,98.39%and 93.40%respectively,and the average accuracies of intra-subject are 94.05%,92.60%,88.38%and 70.03%respectively.Through the experimental results and own data validation,it is proved that the proposed algorithm has a good effect on the gesture recognition of sEMG signals,which is more conducive to the realization of human-computer interaction and provides a feasible solution for the development of universal sEMG equipment.