摘要
摘要
随着下肢外骨骼机器人技术的蓬勃发展，其在日常辅助，军事以及医疗康复领域的作用日益凸显。由于表面肌电信号（electromyography, EMG）与人体动作的高度相关性，基于EMG信号控制的外骨骼人机交互具有自主理解人类运动意图的优势。本文针对下肢外骨骼机器人的上层控制，提出了基于EMG信号的人体下肢运动意图感知与预测系统，包括运动意图模式识别模型、下肢关节角度预测模型以及在线验证实验。通过基于本文数据集的离线实验以及现场的在线实验验证，达到了外骨骼上层控制系统需满足的控制信号精度以及实时性的要求。
本文根据目前基于EMG信号的动作模式识别和关节连续角度预估的研究进行了分析，就目前存在的部分突出问题进行总结，确立了本文的研究方向，同时为本文研究提供了大量参考。通过分析EMG信号的生理学特征与下肢步态的相关研究，利用集成惯性测量单元（Inertial Measurement Unit, IMU）的无线肌电电极，脚底压力传感器等设备，建立了包含5种日常生活中常见的下肢运动模式的离线下肢多通道信号数据集。其中包含了9组下肢肌肉，原始EMG与IMU信号，多种EMG信号特征以及IMU信号特征。该数据集用于离线实验的模型训练与测试，以及在线实验的基础模型参数训练。参考相关研究，选取了三种机器学习算法作为运动模式识别模型的参考算法。
本文提出的下肢运动意图感知与预测系统包括运动模式识别模型以及对应不同运动模式的下肢关节角度预测模型。根据本文数据集的数据特点，对运动模式识别模型进行了适用于多分类问题的参数预设，对EMG信号特征，IMU信号特征以及两者的融合特征的识别准确度进行了研究。同时，基于三种识别模型，对保持模型较高识别准确度所需的肌肉通道数目，以及该数目下对应的肌肉组合方式进行了研究，同时简化了所需的EMG信号特征数目与组合，最终建立了运动模式识别模型以及识别所需的最简最佳特征子集。根据肌电-动作延迟（electromechanical delay, EMD）的生理基础确定了下肢关节角度预测方案以及四个预测时间长度的选择。基于长短期记忆神经网络建立EMG信号特征提取器与预测器，二者组合成为下肢关节角度预测模型，该模型可以在不同预测时间下对关节角度进行准确预测。通过与传统的EMG信号时域特征和机器学习回归算法在本文设定的不同预测时间内进行实验比较，证实了本文所提出的下肢关节角度预测模型的优越性。
为证明本文所提出运动意图感知与预测系统的在线实验可靠性，在离线数据集建立的模型基础上，利用现场采集的数据以及迁移学习对离线训练模型进行参数
摘要
优化，将参数优化后的模型导入至便携式人工智能计算卡，利用无线肌电电极及其SDK模块实现EMG与IMU在线采集与使用。通过运动模式识别实验，下肢关节角度预测实验以及运动意图感知与预测系
统的实时性实验对本文提出的模型进行准确度与实时性检验，最终证明了本文面向外骨骼上层控制提出的下肢运动意图感知与预测系统的有效性与可靠性。
关键词：表面肌电信号；运动意图感知；人机交互；下肢关节角度预测
Abstract
Abstract
With the flourish of the lower limb exoskeleton robot technology, its role in the daily assistance, military and medical rehabilitation fields has become increasingly prominent. Due to the high correlation between electromyography (EMG) and human motion, human-robot interaction of exoskeleton bad on EMG signals has the advantage of understanding human motion intentions autonomously. In this paper, aimed at the upper layer control of lower limb exoskeleton robot, a human lower limb motion intention detection and prediction system bad on EMG signals was propod, this system includes the motion intention pattern recognition model, the lower limb joint angle prediction model and the online verification experiment. Through the verification of offline experiments bad on data t propod in this paper and on-site online experiments, the accuracy and real-time requirements of the control signals required by the exoskeleton upper control system a
re achieved.trampoline
Bad on the previous rearch of motion pattern recognition bad on EMG signals and continuous joint angle estimation, this paper summarized some outstanding problems and established the rearch direction of this paper. At the same time, previous rearch provides a lot of reference for this paper. By means of analyzing physiological characteristics of EMG signals and the gait of the lower limbs, an offline lower limb multi-channel signals data t containing five lower limb motion patterns was established using wireless myoelectric electrodes which integrated with inertial measurement unit (IMU) and foot pressure nsors. It contains 9 lower limb muscles, corresponding original EMG and IMU signals, multiple EMG signal features and IMU signal features. This data t was ud for model training and testing in offline experiments, as well as basic model parameter training in online experiments. Referring to the related rearch, three machine learning algorithms are lected as the reference algorithms of the motion pattern recognition model.
The lower limb motion intention detection and prediction system propod in this paper includes a motion pattern recognition model and a lower limb joint angle prediction model corresponding to different motion patterns. According to the data characteristics of the datat in this paper, the motion pattern recognition model is applied to the parameter pret of multi-classification problem, a
好看的ppt背景图片nd the recognition accuracy of EMG signal features, IMU signal features and the fusion features of the two were studied. At the same time, bad on the three recognition models, the number of muscle channels required to maintain the high recognition accuracy of the model, the corresponding muscle combination method under this number was studied, and the required number and combination of EMG signal characteristics were simplified, and finally established. The
Abstractjust beat it歌词
catnap
省力motion pattern recognition model motion pattern and the subt of the simplest and best features required for recognition. At the same time, bad on the three recognition models, the number of muscle channels required to maintain the high recognition accuracy of the model, and the corresponding muscle combination mode under the number are studied, the required number and combination of EMG signal features are simplified. At the same time, bad on the three recognition models, the number of muscle channels required to maintain the high recognition accuracy of the model, and the corresponding muscle combination mode under the number are studied, and the required number and combination of EMG signal features are simplified. A motion pattern recognition model and a subt of the simplest and best features required for recognition were established. According to the physiological basis of electromechanical delay (EMD), the lower limb joint angle pre
diction scheme and the lection of four prediction time lengths were determined. The EMG signal feature extractor and predictor are established bad on the long short-term memory neural network. The two are combined to form the lower limb joint angle prediction model, which can accurately predict the joint angle under different prediction times. Comparing with the traditional EMG signal time domain features and machine learning regression algorithm in different prediction time, the superiority of the lower limb joint angle prediction model propod in this paper is confirmed.
奥巴马与英拉In order to prove the reliability of the online experiment of the motion intention detection and prediction system, bad on the model established by offline data t, the data collected on-site and the transfer learning are ud to optimize the offline trained model parameters, then the parameter optimized model was imported to the portable artificial intelligence computing card, the EMG and IMU are collected and ud online by using the wireless electromyography electrode and its SDK module. The accuracy and real-time test of the propod model were tested through the motion intention pattern recognition experiment, the lower limb joint angle prediction experiment and the real-time experiment of the motion intention detection and prediction system. Finally, the effectiveness and reliability of the lower limb motion intention detection and prediction system propod for the upper-layer exoskeleton control is proved.
北京新东方夏令营
Keywords: electromyography (EMG), motion intention detection, human-robot interaction (HRI), lower limb joint angle prediction
目录
目录
摘要 ............................................................................................................................... I ABSTRACT ................................................................................................................... III 第1章绪论 (1)
1.1 课题来源 (1)
1.2 研究背景及意义 (1)
1.3 国内外研究现状 (2)prospective
1.3.1 基于EMG的运动意图模式识别研究现状 (2)
1.3.2 EMG用于运动意图连续性参数预估的研究现状 (5)issi
1.3.3 国内外研究现状简析 (7)
1.4 本文主要研究内容 (8)
第2章基于EMG的下肢运动意图的信息采集与处理方法 (10)
2.1 EMG信号概述 (10)
duelist
2.1.1 EMG信号的生理学基础 (10)
2.1.2 EMG信号的特点 (11)
2.2 传统EMG信号处理方法 (11)
2.2.1 EMG信号的预处理 (11)
2.2.2 特征提取方法 (12)
2.3 同步多通道信息采集系统 (14)
2.3.1 EMG信号采集系统 (14)
2.3.2 下肢运动学信号采集系统 (15)
2.3.3 脚底压力信号采集系统 (15)
2.3.4 关节角度解算验证系统 (15)
2.3.5 同步触发装置 (15)
2.4 基于EMG的下肢运动意图数据集设计 (16)
2.4.1 下肢动作设计 (16)
2.4.2 步态相位的划分 (17)
2.4.3 动作相关肌肉的选取 (17)
2.5 下肢运动意图信息采集实验 (18)
