作者:徐林森，张恒玮，陈根，汪志焕，眭翔

Authors:XULinsen,ZHANGHengwei,CHENGen,WANGZhihuan,SUIXiang摘要:使用上肢表面肌电信号对上肢动作进行识别是实现康复机器人持续被动运动和主动辅助运动模式的重要方法 。为了提高基于表面肌电信号(surface electromyography, sEMG) 的上肢动作识别精度 ,分别采用了分段时域信号和拼接频谱图的两种肌电动作识别方法 。分段时域信号方法采用融合卷积神经网络(convolutional neural network, CNN)、长短时记忆网络(Long Short-term Memory, LSTM)和注意力机制的自建网络对上肢动作进行识别 ; 拼接频谱图方法将预处理后的时域信号通过短时傅里叶变换(Short-Time Fourier Transform, STFT) 转换为对应频谱图 ,利用两种微调的预训练模型 VGG16 和 Resnet50 对所有通道竖直拼接的频谱图提取特征并将特征拼接 ,结合支持向量机对上肢动作进行识别 。实验结果表明 ,所提出的两种方法在采集的受试者肌电信号数据集上均表现出 90% 以上的识别精度 ,可有效区分不同的上肢动作。
Abstract:The recognition of upper limb action by sEMG is an important method to realize the continuous passive action and active assisted action mode of rehabilitation robot . In order to improve the accuracy of upper limb action recognition based on sEMG signals, two methods of EMG motion recognition, namely segmented time-domain signals and concatenated spectrograms, are adopted. The segmented time-domain signal method uses a self-built network that integrates CNNs, LSTM, and attention mechanisms to recognize upper limb actions. The concatenated spectrum method converts the preprocessed time-domain signal into the corresponding spectrum through STFT. Two fine-tuning pre-trained models VGG16 and Resnet50 are used to extract features from the vertically concatenated spectrum of all channels and concatenate them. Support vector machine is used to recognize upper limb actions. The experimental results show that the two proposed methods both show a recognition accuracy of more than 90% on the collected EMG data sets, and can effectively distinguish different upper limb actions.

PDF全文下载地址:

可免费Download/下载PDF全文