基于小波降噪和时序数据图像化的表面肌电信号识别  

Surface Electromyography Recognition Based on Wavelet Denoising and Time-series Imaging

作　　者：菅小艳[1] 韩素青[1] 杨红菊[2,3] JIAN Xiaoyan;HAN Suqing;YANG Hongju(School of Computer Science and Technology,Taiyuan Normal University,Jinzhong 030619,China;School of Computer and Information,Shanxi University,Taiyuan 030006,China;Computational Intelligence and Chinese Information Processing of Ministry of Education,Shanxi University,Taiyuan 030006,China)

机构地区：[1]太原师范学院计算机科学与技术学院,山西晋中030619 [2]山西大学计算机与信息技术学院,山西太原030006 [3]山西大学计算智能与中文信息处理教育部重点实验室,山西太原030006

出　　处：《山西大学学报（自然科学版）》2024年第1期103-111,共9页Journal of Shanxi University(Natural Science Edition)

基　　金：国家自然科学基金(61976128);山西省教育厅(2022-008);山西省教育科学“十四五”规划课题(GH-220176)。

摘　　要：表面肌电信号(sEMG)是人体肌肉收缩时发出的信号,能很好地反映人体肌肉功能,因此被广泛应用在临床、假肢控制、康复评估等领域。由于受采集器、佩戴位置和环境等因素的影响,电脑接收的信号包含随机噪声,严重影响信号的分析和研究。因此,本文提出了一种基于新的小波阙值降噪和时序数据图像化的表面肌电信号识别方法。首先,本文采集五种基本上肢运动的sEMG,并采用小波分解对其降噪。提出了一种新的阈值函数以弥补传统小波分解任务中软阈值函数的失真现象和硬阈值函数会产生振荡的缺陷,并在理论上证明了该函数在阈值处的连续性和与原小波系数的无偏差性。然后,受计算机视觉中卷积神经网络成功应用的启发,本文利用短时傅里叶变换将时序数据转换成图像数据。随后,在原始数据集和不同阈值函数降噪后的数据集上的实验表明,本文方法降噪后的数据集上分类性能更优;在降噪后的数据集上,使用二维卷积神经网络(Two Dimensional Convolutional Neural Networks,2DCNN)模型在四个动作的数据上准确率最高、一个次高。说明本文方法可以有效提高sEMG的识别率,具有较好泛化能力。Surface Electromyography(sEMG)is the signal sent by human muscle contraction,which can well reflect human muscle function,so it is widely used in clinical,prosthesis control,and rehabilitation evaluation etc.However,due to the influence of collector,wearing position,environment and other factors,the signal received by the computer contains random noise,which seriously affects the analysis and research of the signal.In this article,we proposed a sEMG recognition method based on a new wavelet threshold denosing and time sequence data visualization.Firstly,sEMG of five basic upper limb movements were collected and denoised by improved wavelet decomposition.A new threshold function was proposed to make up for the distortion of soft threshold function and the vibration of hard threshold function in traditional wavelet decomposition,and it was proved that the function was continuous at the threshold and non-deviation from original wavelet coefficient.Then,inspired by the successful application of convolutional neural networks in computer vision,we transformed time-series data into image data using Short-time Fourier Transform.Finally.the experimental results on both the original datasets and the datasets denoised by different methods show that the model obtains superior classification results on the datasets denoised by the proposed method.The Two-dimensional Convolutional Neural Networks(2DCNN)model has highest accuracy on four action datasets and second highest accuracy on one action datasets.Therefore,the proposed method can effectively improve the recognition rate of sEMG and has good generalization.

关 键 词：表面肌电信号 小波降噪 时序数据 图像数据 短时傅里叶变换 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]