基于肌电信号的手臂动作识别及虚拟仿真

发布时间：2018-07-07 12:52

本文选题：表面肌电信号 + 特征提取　；参考：《山西大学》2017年硕士论文

【摘要】：表面肌电信号作为一种重要的人体生理电信号可以应用于智能系统,实现对虚拟假肢或智能假肢的控制。目前,在智能仿生假肢研究中,利用肌电信号特征进行动作识别和运动控制是其研究的关键。论文主要围绕表面肌电信号采集与预处理、肌电特征提取与识别、虚拟手臂3D模型设计以及虚拟假肢动作仿真等方面进行了相关研究,主要研究内容如下:1)表面肌电信号采集与预处理。设计了前臂外旋、前臂内旋、握拳、展拳、上切、下切、内翻和外翻8种动作模式,分别采集指伸肌,尺侧腕屈肌,掌长肌,屈指浅肌4块前臂肌肉群活动的表面肌电信号,利用小波变换结合自适应滤波的方法,对实验采集的不同动作对应的表面肌电信号进行预处理,获得纯净的表面肌电信号。2)表面肌电信号特征提取。提取了表面肌电信号的绝对积分平均值、均方根值作为时域特征;提取平均功率频率、中值频率作为频域特特征。采用db3小波进行5层小波分解,分别计算第5层近似系数与第3、4、5层细节系数的均方根值与方差作为表面肌电信号的时频特征,并对手臂不同动作模式对应的表面肌电信号特征的差异性进行统计分析。3)基于表面肌电信号特征的动作识别。首先利用BP神经网络算法,分别采用肌电信号的时域特征、频域特征以及时频特征,进行8种手臂动作的分类识别,获得了89%、77%、91%的平均识别正确率。然后,设计栈式自编码深度学习算法,分别利用表面肌电信号时域特征、频域特征和时频特征,进行手臂动作模式分类,平均分类正确率为95%、91%和96%。结果表明:肌电信号时频特征能够较好体现不同动作模式之间的差异,同时,栈式自编码深度学习算法应用于表面肌电信号特征分类与动作识别要优于BP神经网络算法。4)虚拟手臂3D假肢设计与动作仿真。采用改进的D-H方法建立了连杆机械手臂模型,并对其进行了运动学分析和轨迹规划,利用Matlab软件仿真平台,实现连杆机械手臂执行连续喝水动作的模拟。使用SolidWorks软件设计了虚拟3D假肢模型,并在虚拟现实环境中,通过计算机仿真验证了虚拟3D假肢模型的合理性。最后,将表面肌电特征的动作识别结果,通过计算机网络,借助Java跨平台传输给虚拟假肢,控制虚拟假肢连续执行前臂内旋、内翻、外翻、前臂外旋与前臂内旋90o五个动作,模拟仿真结果表明了识别方法的有效性以及动作执行的完整性和精确性。论文的研究成果能够应用于人工智能、人机交互、仿生机器人、智能假肢等领域,具有科学和应用双重价值。
[Abstract]:As an important physiological signal of human body, surface electromyography (EMG) signal can be used in intelligent system to control virtual prosthesis or intelligent prosthesis. At present, in the research of intelligent biomimetic prosthesis, it is the key to use EMG signal to recognize and control the motion. This paper mainly focuses on the surface EMG signal acquisition and pretreatment, EMG feature extraction and recognition, virtual arm 3D model design and virtual artificial limb movement simulation. The main research contents are as follows: 1) Surface EMG signal acquisition and preprocessing. Eight modes of forearm external rotation, forearm internal rotation, fist grip, extended fist, upper cut, bottom cut, varus and valgus were designed. The surface EMG signals of four forearm muscles were collected, including extensor digitorum muscle, flexor Carpi ulnar muscle, palmar longus muscle and flexor superficial muscle. Wavelet transform combined with adaptive filtering is used to preprocess the surface EMG signal corresponding to different actions collected in the experiment, and the pure surface EMG signal (.2) surface EMG signal feature extraction is obtained. The absolute integral mean value of surface EMG signal is extracted, the root mean square value is taken as the time domain feature, and the average power frequency and median frequency are extracted as the special features in frequency domain. The db3 wavelet is used to decompose the five layer wavelet, and the root mean square value and variance of the fifth and the third layer detail coefficients are calculated respectively as the time-frequency characteristics of the surface EMG signal. The differences of EMG signal characteristics corresponding to different arm movements are analyzed statistically. 3) the action recognition based on SEMG features is presented. Firstly, the BP neural network algorithm is used to classify and recognize eight arm movements by using the time domain feature and frequency domain feature of EMG, respectively, and the average recognition accuracy of 89777% is obtained. Then, a stack self-coding depth learning algorithm is designed to classify the arm motion patterns using the time-domain, frequency-domain and time-frequency features of surface EMG signals, respectively. The average classification accuracy is 95% and 96% respectively. The results show that the time-frequency characteristics of EMG signals can well reflect the differences between different action modes, and at the same time, The self-coding depth learning algorithm based on stack is better than BP neural network algorithm in feature classification and motion recognition of surface EMG signal. 4) 3D artificial limb design and motion simulation of virtual arm. An improved D-H method is used to establish the model of the connecting rod manipulator, and the kinematics analysis and trajectory planning are carried out. The simulation of the continuous water movement of the connecting rod manipulator is realized by using Matlab software simulation platform. The virtual 3D prosthesis model is designed with SolidWorks software, and the rationality of the virtual 3D prosthesis model is verified by computer simulation in the virtual reality environment. Finally, the recognition results of the surface electromyoelectric features are transmitted to the virtual prosthesis by means of the Java platform through the computer network. The virtual prosthesis is controlled to perform the five actions of forearm internal rotation, varus, valgus, forearm external rotation and forearm internal rotation 90o continuously. Simulation results show the effectiveness of the method and the integrity and accuracy of the action execution. The research results can be applied to artificial intelligence, human-computer interaction, bionic robot, intelligent prosthesis and so on.
【学位授予单位】：山西大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R318;TN911.7

【参考文献】