【作者】 徐卓君；

【导师】 田彦涛；

【作者基本信息】 吉林大学 ， 控制理论与控制工程， 2015， 博士

【摘要】 在人体前臂肌肉的表面肌电信号模式识别研究中，首要任务是给出能够识别日常生活常用动作及情感表达手势的模式识别方法。除此之外，为智能仿生手臂的控制模块提供更多的动作细节信息，如手指关节角、握力等，能够帮助其更好的完成日常生活动作，提高设备的拟人化程度。针对以上设想，本文提出将加速信号及握力信号等其他与描述动作有关的多源信号，与表面肌电信号融合，在模式识别的训练阶段帮助生成更为准确的分类策略，希望对智能仿生手臂的模式识别模块的开发能够起到推动作用。为了实现基于多源信息的智能仿生手臂模式识别研究，从表面肌电信号中提取更多能够用于智能仿生手臂仿生化控制的运动细节信息，本文主要进行了下面几个方面的研究：提出了一种融合双路信号特征的模特征提取算法。在表面肌电信号的手势动作模式识别中，一些识别率的损失是由于信号采集过程中信道顺序颠倒造成的。为了解决这一问题，本文基于复数的概念提出了综合双路信号特征的模特征提取算法，该算法将独立的双路特征作为复数的实部和虚部，计算双路特征的模值作为信号的模特征。该模值能够综合反映双路信号特征，同时对信道顺序不敏感，鲁棒性强。针对表面肌电信号样本标注代价大的问题，本文提出在仅有少量已标注样本的实验条件下采用安全半监督支持向量机的方法进行表面肌电信号模式识别的方法，该方法尝试发掘多个大阈值低密度分类器空间，从而保证半监督学习质量，与有监督学习相比不会出现表现退化的现象，同时保留半监督学习的优势，发掘更多的未标注样本参与到分类中来，最终提高分类的精确度。通过实验结果表明，本文提出的基于模特征的半监督模式识别方法适用于仅有少量已标注样本的实验条件下的鲁棒表面肌电信号模式识别。提出了融合关节角信息的类活动段检测方法。在表面肌电信号-关节角模式识别实验中，采用与手势动作表面肌电信号模式识别的活动段检测方法来处理对应关节角标签的信号分段问题已经不能达到实验要求。针对这一问题，本文提出了一种类活动段检测方法，方法通过融合加速信号帮助准确找到对应关节角标签的表面肌电信号活动段，为神经网络提供更为准确的训练数据，从而提高神经网络的识别效果。实验结果表明，本文提出的类活动段检测方法能够达到理想的信号分段效果，经该活动段分段数据训练的神经网络能够实现较高的识别率。提出了新的表面肌电信号仿生特征提取方法。基于人体上臂肌肉力变化过程分析，本文提出基于窗样本熵和窗峰度值两类特征提取方法，通过建立连续不重叠时间窗的方法，跟踪提取每一时间段的信号特征，从仿生角度让信号特征追随人体肌肉力变化过程的时变特性，衡量表面肌电信号的复杂度及概率密度分布情况，模拟接受大脑指令后不同肌肉力下运动神经元的募集情况，更加直观的体现人脑对产生肌肉力大小的指令意图。针对智能仿生手臂佩戴者的残疾上肢无法提供神经网络训练中完整的输入输出数据的问题，提出采用对侧信号训练的神经网络预测输出肌肉力的实验方案，实验中将左手表面肌电信号输入通过右手实验数据训练好的神经网络中预测左手的输出肌肉力。实验结果表明，本文提出的考虑肌肉力运动特性的表面肌电信号模式识别方法能够实现表面肌电信号-肌肉力的模式识别。建立了具有16个自由度的智能仿生手虚拟样机模型。本文在ADAMS软件中建立了智能仿生手虚拟样机模型，通过运动学仿真实验验证了模型设计的合理性。为了更好的实现智能仿生手虚拟样机模型对模式识别算法和控制算法的调试，本文提出通过ADAMS和MATLAB联合仿真来进行智能仿生手臂虚拟在线仿真的实验方案，实验将智能仿生手臂中应具备的信号采集功能、信号处理功能、模式识别功能及动作控制功能相结合，通过虚拟在线的形式进了完整的在线控制智能仿生手虚拟样机模型的实验。模型按照联合仿真中的模式识别模块对表面肌电信号的识别结果设计控制目标量，通过在SIMULINK中搭建的PID控制系统对模型进行控制完成简单的抓取动作。实验结果证明本文提出的智能仿生手虚拟样机模型在控制算法设计合理、模式识别结果准确的前提下，能够较好的完成在线识别、实现拟人化手部动作。最后，总结了全文所做的工作，提出了今后进一步需要研究的问题。更多还原

【Abstract】 In the study on the pattern recognition of the surface electromyography (sEMG) fromthe human forearm muscles, we make it a priority to find the pattern recognition methodfor identifying the hand common movements in daily life and emotional expressiongestures, in addition, the study can provide much more detailed information of movementfor the control module of the intelligent bionic arm, such as grip strength and joint angle,so as to assist in completing the activities of daily life better and improving how well theequipment mimic the humankind. In view of the above ideas, the fusion concept with thesEMG signal, the acceleration signal and the grip strength signal is proposed in this paperto help generate a more accurate classification strategy in the training phase of patternrecognition, and expect to promote the development of the pattern recognition module ofthe intelligent bionic arm. In order to undertake the research into the pattern recognitionfor the intelligent bionic arm based on multiple sensors information, more details of themotion will be extracted from the sEMG for the bionic control of the intelligent bionicarm. This paper mainly falls into the following aspects：Firstly, a fusion feature extraction algorithm is proposed in this paper. In the patternrecognition of the sEMG for hand movements, a partial loss of accuracy is caused by thereverse sequence of the channels in the process of the signal acquisition. In order to solvethis problem, a modulus feature extraction algorithm based on the concept of plural isproposed in this paper. The algorithm takes the independent characteristics from dualchannels as the real part and the imaginary part of the plural, and this modulus value ofthe dual characteristics is calculated as the signal feature in this algorithm.The modulusvalue can comprehensively reflect dual channels features,and be insensitive to the channelorder, but the robustness is quite powerful. Due to the high costs of the labeled samplesof sEMG, in the condition of a small number labeled samples, a safe semi-supervised support machine method (S4VM) is applied to carry out the pattern recognition. S4VMtry to exploit the candidate low-density separators simultaneously to reduce the risk ofidentifying a poor separator with unlabeled data. overall performance of S4VM are highlycompetitive to semi-supervised support machine (S3VM), while in contrast to S3VM,S4VM are significantly inferior to inductive supervised support machine. The results ofcomprehensive experiments show that the semi-supervised method with the modulusfeature extraction algorithm proposed in this paper is suitable for the pattern recognitionof sEMG with the limited labeled samples.Secondly, An analogous active segment detection method is proposed in this paper.In the experiment of the sEMG-joint angle pattern recognition, the classical activedetection method which is widely used in the sEMG-movement type pattern recognitionhas already been unsatisfactory, so an analogous active segment detection method isproposed in this paper. It can be applied to accurately locate and find the active segmentaccording the joint angle label by fusing the acceleration signal so as to provide the moreaccurate training data for the neural network to improve its recognition performance. Theexperimental result show that, the analogous active segment detection method can exert aquite good effect on signal segment, and the neural network, which is trained by the datafrom analogous active segment detection method, can be at a higher accurate recognitionrate.Thirdly, two new bionic sEMG feature extraction methods are proposed in this paper.This paper drew inspiration from physiological process of muscle force, aiming foraccurate estimation of muscle force through the features extracted from sEMG. A sEMG-muscle force pattern recognition method of upper arm based on bionic concept ofkinesiology is proposed in this paper. The proposed sEMG feature extraction method ofWSE and WK in this paper is different from that in the past which only cut a small part ofa long sEMG signal. This method tracked the sEMG signal during the time throughbuilding a continuous and non-overlapping window, and made the feature follow the time-variation characteristic of sEMG signal to express its change better. The amputates’residual limb couldn’t provide full training data for pattern recognition, therefore, asolution was investigated in this paper that used the neural network trained by right hand’sdata to predict the relation between the features of sEMG and muscle force of left hand.The contralateral performance is much closer to the ipsilateral performance. This schemeprovided unilateral transradial amputees a possibility to train the intelligent bionic limb byusing their own sEMG. The prediction results of both ipsilateral and contralateral experiment show that the new feature extraction and pattern recognition method forsEMG-Muscle force pattern recognition basing on window sample entropy and windowkurtosis is feasible. The method may offer helpful clues to enhance the performance inpattern recognition module for intuitive control development of intelligent bionic hand.Fourthly, a model for intelligent prosthetic hand is designed by ADAMS in this paper,which has a high fidelity simulation of humanoid shape. The proposed model has16degrees of freedom. Kinematic simulation of prosthetic hand model using ADAMS hasbeen presented in this paper. Several movements have been carried out to verify thereasonability of design of the modeling. Simulation results show that virtual prosthetichand model created in ADAMS can represent the actual prosthetic hand. In order to realizethe debugging function of the virtual prototype model on pattern recognition algorithmand control algorithm, a virtual online co-simulation experiment is conducted on thevirtual prototype model with ADAMS and MATLAB/SIMULINK. Such functionspossessed by Intelligent prosthetic hand as signal acquisition, signal processing, patternrecognition and action control are combined together during this experiment. The controlsignal for the PID algorithm in the SIMULINK is provided by pattern recognition of sEMG.Experimental results show that, the intelligent bionic arm virtual prototype modelproposed in this paper can realize anthropomorphic hand movements under the reasonablecontrol algorithm and the accurate pattern recognition results.Finally, the main content of this dissertation is summarized, and the further researchwill be continued.更多还原