节点文献
基于肌肉协同的腕关节力预测研究与实现
Research and Realization on Wrist Force Estimation Based on Muscle Synergy
【作者】 杨柳;
【导师】 娄平;
【作者基本信息】 武汉理工大学 , 信息与通信工程, 2016, 硕士
【摘要】 表面肌电信号中含有丰富的与神经系统相关的信息,所以常用来作为康复设备的控制信号。传统的基于模式识别的控制方式存在一些缺陷:第一,模式识别只能实现离散的控制,而人的运动却是连续;第二,当涉及的肌肉较多的时候,时域特征提取样本集的维度往往较高,且时域特征提取在电极片位置发生偏移时鲁棒性差。近年来有学者认为,人的中央神经系统并不是直接控制每一块肌肉的收缩,而是控制一个更小维数的参数集合,即肌肉协同收缩模型,这种特征使得在利用肌电信号进行模式识别或者做信息预测时,特征的选取有了实际的意义并且能弥补传统时域特征提取的不足。人体上肢从事着复杂且精细的活动,因此本文针对以上问题,以腕关节为研究对象,研究了基于肌肉协同分析的特征提取方法,并提出一种新的权值和阈值初始化方法来改进BP神经网络,同时建立了基于动作判别的力预测模型,最后设计了基于动作判别的力预测软件系统,验证了动作判别和力预测的准确率以及实用性。具体研究工作包括:(1)采集小臂肌电信号和腕关节力信号,并对采集到的信号做预处理(滤波、同步、归一化等)。针对传统时域特征提取和动作识别所存在的缺陷,提出并采用了基于肌肉协同的特征提取方法,利用非负矩阵分解确定了腕关节4个动作(屈、伸、桡侧偏移、尺侧偏移)肌肉协同的同时,通过实验证明了肌肉协同对于电极片位置偏移的鲁棒性要优于传统的时域特征,同时验证了人体运动中肌肉间的协同性。在分解得到肌肉协同的同时,也得到了肌肉协同随时间变化的信息,将该信息作为特征量用于之后的力预测模型。(2)针对多自由度预测问题,设计了两种不同结构的神经网络,把非负矩阵分解得到的系数矩阵作为特征量输入到神经网络,把相应的力作为输出进行训练。针对神经网络权值和阈值随机初始化所造成的预测结果不准确、收敛速度慢和易陷入局部极小点等问题,采用了Levenberg-Marquard算法,并提出了一种新的方法实现权值和阈值的初始化,降低了预测误差,同时加快了网络的收敛速度。最后设计了基于动作判别的力预测模型,在对腕关节动作正确识别的同时获取了每个动作过程中的力信息。(3)研究了力预测在机器人康复领域的应用,设计并实现了基于动作判别和力预测的软件系统,结合该软件,同时针对传统模式识别的控制方法所存在问题,采用基于力预测的比例控制方法,通过力水平反应实验者的运动意图,来控制设备的运动速度,实现对康复设备的连续控制。
【Abstract】 Surface EMG signal is rich in information related to the nervous system,so it was often uesd as the control signal of the rehabilitation equipment.There exist some drawbacks of traditional control methods based on pattern recognition: Firstly,pattern recognition can only achieve discrete control,but human movement is continuous;the secondly,when the muscles are more involved,the time domain feature extraction often have higher dimensions,and time-domain feature extraction have poor robustness in face of the electrode position shifting.In recent years,some scholars believe that the human central nervous system is not directly control every muscle contraction,but rather a set of control parameters which have smaller number of dimensions,namely muscle synergy model.This makes that feature selecting have practical significance and can also make up for the traditional time-domain feature extraction deficiencies when making use of EMG signals to do pattern recognition or prediction information.Human upper limb engaged in complex and elaborate activity,So in this paper,aiming at above problems,we take the wrist joints as the research object,study the feature extraction method based on muscle synergy analysis.and propose a new weights and thresholds initialization method to improve BP neural network.In the meantime establishing a force prediction model based on motion discriminating,and finally design a software system based on the force forecasting to verify the accuracy and utility of the motion recognition and force prediction.Specific studies including:(1)Acquisition EMG signal of forearm and the force signal of wrist,then signal preprocessing was done(filtering,synchronization,normalization,etc.).To overcome the drawbacks of traditional time-domain feature extraction and motion recognition,we propose the use of feature extraction method based on muscle synergy,using non-negative matrix factorization to determine wrist four movements(flexion,extension,radial deviation,ulnar deviation)synergistic muscles and meanwhile through experiments prove that this feature is more robust than the traditional time-domain feature when electrode position shifting taking place,and reveals the human muscle movement synergism.Through matrix decomposition,we also get the information about the changes of the muscle synergy over time,which can be used as the input of the force estimation model.(2)We design two different structures of the neural network,the NMF coefficient matrix obtained as tht input to the neural network,the output was the corresponding force.We use the Levenberg-Marquard algorithm and adaptive learning rate algorithm and then propose a new methods to achieve initialization weights and thresholds to overcome the inaccurate predictions,slow convergence and easy to fall into local minima problems caused by the weights and thresholds random initialization.Not only the prediction error was reduced,the speed of network convergence was accelerated either.Finally we design a force prediction model based on motion discrimination.On the basis of correct idenfication of the wrist motion,the corresponding force was predicted at the same time.(3)Here we study the force application in the field of robotics rehabilitation.We designe and implement a software system based on the motion determination and force estimation.Combined with the software,we adopt the proportional control stratergy based on force prediction to overcome the deficiency of the traditional control methods baesd on the pattern recognition.The stratergy adjust the device speed according to the force,which realize the continurous control of the rehabilitation device.
【Key words】 EMG signals; feature extraction; muscle synergies; force estimation; proportional control;
- 【网络出版投稿人】 武汉理工大学 【网络出版年期】2019年 05期
- 【分类号】TP183;TN911.7;TH77
- 【被引频次】3
- 【下载频次】109