【作者】 李阳；

【导师】 田彦涛； 陈万忠；

【作者基本信息】 吉林大学 ， 控制理论与控制工程， 2012， 博士

【摘要】 智能仿生手臂指能够根据人的大脑发出的意念自由地控制假手完成各种动作，在抓取物体时能够感知物体的外部特征，自动调整假手系统的参数，使假手完成灵活、可靠的动作的高性能新一代假肢。目前用于智能仿生手臂的控制信号有很多，其中表面肌电信号易于获取且控制假肢具有直接、自然的特点，已成为应用最为广泛的控制信号源。表面肌电信号的处理质量直接关系着智能仿生手臂控制的安全性与可靠性，本文主要针对表面肌电信号进行分析处理。由于其具有信号微弱、易受噪声干扰、个体差异性大等特点，为提高多运动模式识别率，本文主要进行了以下几个方面的研究：寻找肌肉与不同运动模式间的对应关系，确定最佳肌电电极摆放位置，检测合理活动段。人体肌肉与运动的关系较为复杂，每一块肌肉所对应的功能不尽相同，且相互影响。表面肌电信号采集效果受电极位置影响较大，准确的采集肌电信号使其能清晰的反映不同运动模式下肌肉的运动情况是肌电信号处理的前提。本文借助人体解剖学、生物医学等知识选择手臂指浅屈肌和桡侧腕屈肌作为电极摆放位置，降低了信号采集的随机性，保证了信号的准确性。文中利用MQ-8型表面肌电采集系统，对8位健康测试者采集6种运动模式的表面肌电信号，并用移动平均值法处理表面肌电信号序列的瞬时能量，通过结合阈值进行比较，最终确定信号的有效活动段。经过上述简单必要的预处理过程，找到信号中表示动作起始和终止的位置，即检测到有效活动段，有助于后续表面肌电信号分析。提出非抽样双密度小波变换和双参数软阈值去噪的方法。表面肌电信号在采集过程中噪声无法避免，虽然通过硬件采集系统可去除部分噪声，使有用信号不至于被噪声淹没，但在信号传输过程中仍存在大量仪器噪声。针对此情况，本文采用非抽样双密度小波变换对信号进行分析，此方法具有严格平移不变性，避免了信号重构过程中人为噪声的引入，且与离散小波变换相比，具有两个小波函数，能够在同一尺度内多获得一组信号细节信息，有助于表面肌电信号精细去噪。采用双参数软阈值法对分解的信号高频部分去噪，与传统软阈值方法相比，此方法便于调节，在阈值附近具有较为平滑的过渡带，符合自然信号连续特性，能克服软阈值函数存在的缺陷。特别在小于阈值部分，没有将函数值至零，而是根据表面肌电信号特性，最大程度的保留信号细节信息，符合其高精准度去噪的要求。提出基于傅里叶级数和FFT盲辨识的表面肌电信号建模方法。肌电信号建模是研究肌电信号特性的一种重要途径，目前常用的肌电信号数学模型有线性系统模型、非平稳模型、双极型模型以及集中参数模型4种。以上肌电信号建模的方法从不同方面描述肌电信号与肌肉力的关系，提供了判断肌肉活动状况的理论依据，被广泛应用于临床医学、康复医学与工程及运动医学和运动生物力学领域。上述建模方法着重从生物医学方面描述肌电信号，有助于判断运动疲劳、肌肉异常等状况，无法表现肌电信号与不同运动模式之间的关系，很少在仿生假肢领域使用。针对此情况，文中尝试采用傅里叶级数法和FFT盲辨识法分别建立肌电信号模型，寻找信号与运动间的关系，并采用模型系数作为信号特征，通过实验可见，本文方法建立的模型系数对多运动模式的表面肌电信号具有较好的表征能力，将其作为信号特征有助于后续多运动模式识别。提出半监督boosting多模式识别方法。由于标注样本较难获取，且相似运动模式的表面肌电信号较难区分，为获得较高模式识别率，对分类器性能要求较高。针对此情况，文中采用半监督学习法，其与监督学习和无监督学习相比，可通过少量标注样本和大量无标注样本训练基分类器，使其具有较高的识别能力。且本文方法与现有半监督方法主要针对两分类问题不同，其可直接处理多类问题，由于无需将多类问题转化为多个两类问题，避免了分类结果处于多个尺度无法相加以及将某一样本划分为多类的情况。Boosting算法是半监督学习领域较常用的一种方法，其可被认为是一种自我训练过程，通过某一框架结构联合多个现有分类器，可实现提高分类性能的目的。与现有半监督boosting仅依据分类置信度调整无标注样本的伪类别不同，本文方法将分类置信度与样本间相似度相结合共同作为无标注样本伪类别的调整依据，其最大程度的有效利用了多种假设与聚类假设。实验证明，本文算法对不同基分类器、不同标注样本数均具有较强的鲁棒性。以表面肌电信号模式识别结果作为控制信号，采用自抗扰方法控制Adams虚拟手臂完成指定动作，验证肌电信号手臂控制的准确性和稳定性。利用Adams仿真软件设计开发具有16自由度假手，并依据此建立其动力学模型，采用自抗扰控制器调整参考输入与Adams实际输出间的误差，形成稳定的闭环控制。自抗扰控制器由PID控制器演变而来，其采用了PID误差反馈控制的核心理念。该方法不依赖于被控对象精确的数学模型，算法简单，在未知强非线性和不确定强扰动作用下能保证控制精度，适用于具有较高自由度的仿生手臂控制。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】 According to the ideas issued by brain, intelligent bionic arm can control the prosthetichand to complete a variety of actions freedomly. The new generation prosthetic hand cansense the external characteristics of the objects, adjust the parameters of the hand systemautomatically and complete proposed action flexibly and reliably. At present, there are lots ofintelligent bionic arm control signals. Surface eletromyographic signal (sEMG) obtain easilyand control directly and naturally which has proven to be an effective control input forpowered upper limb prostheses. The sEMG signal processing quality influences the securityand reliability of control, directly. This paper mainly aims at using sEMG processing toextract more information from the input signals to increase the accuracy of the controllers. Itsmain contents are presented as follows:Looking for the corresponding relationship between muscle and different movementmodes，identify the best position for putting myoelectricity electrode, testing for reasonableactivities section. The relationship between movements and human muscles is complicated,the corresponding function of each muscle is different, and they influence each other. Theposition of electrode might have impact during collecting EMG. It is the premise of EMGsignals processing that accurate collecting EMG signals to reflect muscle movements underdifferent movement patterns correctly. In this paper, the musculus flexor digitorum sublimisand the musculus flexor carpi radialis are chosen as the position for putting electrode throughhuman anatomy and biological medicine, which can reduce the randomness of the signalcollection and guarantee the accuracy of the signal. In this paper, the sEMG of six movementmotions for eight healthy testers are collected using MQ-8sEMG collection system. Thenmoving average method is used to process instantaneous energy of sEMG sequence. Finallythe signal effectively activities are determined through the comparison with threshold. Afterthe above pretreatment process, the effective activities section is detected, which can help thesEMG analysis in the subsequence.Undecimated double density wavelet transform (UDDWT) and double-parameter softthreshold are proposed to meet the requirement of sEMG nondestructinve de-noisnig. Although the hardware system can remove a part of noise, which made the desired signal ofsEMG not be drowned. It is sill existente during the acquisition and signal transmissionprocess, which have a lot of instrumental noise. In view of this situation, the undecimateddouble density wavelet analysis of signal is used in this paper. It has strict shift invariance,which can avoid the artifacts in signal reconstruction. UDDWT has two wavelet functionswhich can get more detail in every scale compared with discrete wavelet transform (DWT).Double-parameter soft threshold method, which has a relatively smooth transition zone, isused for denoising the high frequency of the signal in the signal decomposition. It cansuppress the noise efficiently meanwhile keep useful detail information to the greatest extent.Experiment results show that the proposed methods, which get high signal-to-ratio whileretaining the signal characteristics, are very suitable for multi-channel sEMG of similar handmovements de-noising.Fourier series and FFT-based blind identification methods are proposed to establishsEMG model. sEMG signal modeling has become an important approach to study of EMGsignal characteristics. Now there are four methds are commonly used in sEMG signalmathematical model, which are linear system model, nonstationary model, bipolar model andthe lumped parameter model respectively. The methods mentioned above, usually describethe relationship between EMG and muscle force from the different aspects to provide thetheoretical basis for judging of muscle activity status, have been widely applied in clinicalmedicine, rehabilitation medicine, sports medicine, neurophysiology and ergonomics. Thosemodeling methods are described emphatically from the biomedical aspects of EMG signals,which can help to distinguish the motion fatigue or muscle abnormalities. They cannot reflectto the relationship between EMG signals with different movement patterns, therefore they arerarely used in intelligent bionics. In these circumstances, Fourier series method and FFT blindidentification method are proposed to build EMG model in this paper to explore relationshipbetween signals and different motions. The model coefficients are used as the signalcharacteristics, the method proposed in this paper has good characterization capabilitiesthrough experiments. By using these as the signal features contribute to subsequentmovement pattern recognition.Semi-supervised boosting algorithm is proposed to classify different movement motions.Traditionally, machine learning is categorized as two paradigms i.e. supervised versusunsurpervised learning. Supervised learning finds out a rule for the predictive relationshipbetween input and output from a set of finite examples in the format of input-output pairs,while unsupervised learning seeks a structure of interests underlying a data set. In general, supervised learning requires many training examples to establish a learner of the satisfactorygeneralization capability. The acquisition of training examples is nontrivial for supervisedlearning, which needs to annotate input data with appropriate labels. In sEMG, the labeledinput data is often difficult, expensive, and time-consuming, especially when it has to be donemanually by experts. On the other hand, there is often a massive amount of unlabeled dataavailable. In order to exploit unlabeled data, semi-supervised learning has become a novelparadigm by using a large number of unlabeled points together with a small number oflabeled examples to build a better learner. Most semi-supervised learning algorithms havebeen designed for binary classification, and are extended to multi-class classification byapproaches such as one-against-the-rest. The main shortcoming of these approaches is thatthey are unable to exploit the fact that each example is only assigned to one class. Additionalproblems with extending semi-supervised binary classifiers to multi-class problems includeimbalanced classification and different output scales of different binary classifiers. A newsemi-supervised boosting algorithm is proposed in this paper that directly solves thesemi-supervied multi-class learning problem. Compared to the existing semi-supervisedboosting methods, the proposed algorithm is advantageous in that it exploits bothclassification confidence and similarities among examples when deciding the pseudo-labelsfor unlabeled examples. The experiments show that the proposed algorithm performs betterthan the state-of-the-art boosting algorithms for semi-supervised learning.Active disturbance rejection control (ADRC) develops the kenel of the PID control, andcombines the observer of modern control ideas. It is a new control method, which algorithmis simple and adaptability, and can automatically compensate internal and externaldisturbances. It can be widely used in industrial applications. ADRC is introduced into bionicaritficail arm in this paper. Adams is used to design artificial arm with16DOF. Themathematical model is obtained from the established artificial arm by using kinematics anddynamics analysis. ADRC adjusts the error between reference input and the Adams actualoutput, to form a stable closed loop control, according to the mathematical model motionedabove. The control method does not depent on the system model, and has the strongadaptation, robusness and operability.Finally, the main content of this dissertation is summarized, and the further researchesare discussed.更多还原