脑卒中和脊髓损伤是导致神经损伤患者肢体运动功能障碍的两大主要病因，近年来的患病率和致残率与日俱增。运动疗法作为康复治疗的重要组成部分，对神经损伤患者的康复起着不可替代的作用。随着科技的进步，运动疗法的实现正逐渐从传统治疗师帮助患者进行运动训练转变为采用康复机器人辅助患者进行康复。临床研究已经证明了主动训练对于患者的运动功能恢复和神经系统康复具有更加积极的效果，如何激发患者在康复训练过程中的自主参与意识、利用多传感器信息融合来实时获取患者的运动意图，是实现人机协调控制的关键。本文以能够直接反映肌肉活动状态的表面肌电信号为主要手段，探讨如何在康复机器人平台上建立有效的人机交互接口和运动控制策略，将人体真正融入由人机构成的一体化系统中。此外，还针对康复机器人在临床应用研究中的重点和难点问题，对机器人的康复训练效果、康复评价等方面展开了研究。本文的主要工作及创新点如下：
       1. 为充分激发患者的康复训练积极性，基于BP神经网络方法建立了肌电信号和关节处主动力矩的量化关系，实现了机器人辅助的下肢主动康复训练。由于动力学信号与真实的运动意图相比往往存在滞后，本文以肌电信号的归一化值、关节角度、关节角速度作为BP神经网络输入来估计髋、膝关节处的主动力矩，然后采用导纳控制方法将主动力矩转化为角度偏差来修正当前的运动轨迹，从而实现以患者的运动意图驱动机器人动作的主动康复训练，扩展了神经损伤患者的临床康复训练形式，提高患者在训练中的主动参与程度。
        2. 针对肌肉骨骼模型建模过程相对复杂、参数优化费时的问题，提出了一种简化的基于EMG信号驱动的肌肉骨骼模型来模拟和预测膝关节运动，并在康复机器人平台上实现了两种用于膝关节康复的自由运动训练。该模型将膝关节近似为沿转轴旋转的单铰链关节，采用肌肉骨骼系统的简化形式对肌肉力和力矩进行动态分析，模型中未知参数的标定采用双种群遗传算法以减小参数优化时陷入局部最优的风险，最后基于导纳控制方法实现了由患者的运动意图驱动的自由运动训练。
       3. 为提高肌电模式分类的准确率，提出了一种基于词包模型的特征处理方法，以形成更加有区分度的特征向量来区分容易混淆的动作模式。原始肌电信号中提取的特征向量在送入分类器之前，使用词包模型从特征向量中提取出相同类别的结构相似性信息，以使得不同类之间的区分特性更加明显，之后将结构信息组合成新的特征向量送入分类器进行分类。通过对5类肩、肘关节的自然动态运动进行模式分类，验证了特征处理方法的有效性。
        4. 通过开展上肢机器人辅助脑卒中患者康复训练的临床实验，验证了机器人辅助疗法的临床应用有效性，此外基于患者运动参数和表面肌电信号对患者的肢体运动功能进行了深入分析和康复评价。临床实验结果表明，机器人辅助疗法对脑卒中患者的上肢运动功能和日常生活活动能力的改善具有积极效果。同时，本文从运动行为层面对患者肢体功能损伤与恢复进行了分析和评价，通过提取患者训练过程中的运动参数建立线性回归模型，较准确地估计出患者的Fugl-Meyer量表评分结果。采用表面肌电信号分析了患者各块肌肉的收缩时序、参与程度，指出表面肌电信号可以对脑卒中患者肢体功能状况提供客观定量的评价。

Stroke and spinal cord injury are two common disease, leading to increasing morbidity and disability in recent years. Exercise therapy plays an irreplaceable role for the rehabilitation of patients with neurological injury. The implementation of exercise therapy is shifting from traditional therapists assist patients with exercise training to rehabilitation robots help patients perform exercise training. The clinical studies have demonstrated that active training have a more positive effect on the motor function recovery and neurological rehabilitation of patients. How to motivate the patients' voluntary participation in the rehabilitation training and obtain the patients' motion intention in real-time is the key to achieving human-machine coordinated control. This paper uses surface EMG as the primary means to explore how to establish an effective human-machine interface and motion control strategy on the rehabilitation robot platform. In addition, pilot studies on rehabilitation training effect and rehabilitation evaluation of robot applied to clinical  practice are conducted. The main contributions of this dissertation are as follows:
    1. The quantitative relationship between surface EMG and joint active torques are established based on neural network, and the robot-assisted active training for lower limb rehabilitation is achieved in order to fully stimulate the enthusiasm of patients with rehabilitation training. Since kinetic signals always lag behind the real motion intention, the method proposed in this study uses the normalized value of EMG, joint angle, joint angular velocity as inputs to the BP neural networks to estimate the active torques of hip and knee joints. Then the active torques are converted into angular deviations based on admittance control to correct the current trajectory. This kind of exercise training could improve patients' active participation in the rehabilitation training.
     2. A simplified EMG-driven musculoskeletal model is proposed to simulate and predict the movement of knee joint, and two robot-assisted exercise training methods for knee rehabilitation are developed. The EMG-driven model approximates the knee joint as a single hinge joint with a center of rotation, which uses the simplified representation of musculoskeletal system to dynamically analyze the muscle forces and torque. The dual population genetic algorithm is applied to calibrate the model parameters to reduce risk of fallen into local minimum. Then the ``patient-driven'' voluntary exercise training methods are achieved based on admittance control.
     3. An effective feature processing method based on the bag-of-words (BoW) model is proposed for improving the accuracy of EMG pattern recognition. The feature extracted from raw EMG is reformulated by the BoW model prior to classification, which extracts structural similarity information within the same class and make the characteristics between different classes more discriminative. Then the structural information is combined into a new feature vector and input to the classifier. The experimental results show that the feature processing method significantly improves the classification accuracies for 5-class natural dynamic motions of shoulder and elbow joints.
     4. The effectiveness of robot-assisted therapy is verified by conducting the clinical trials about upper limb robot-assisted rehabilitation training of stroke patients. The robot-based rehabilitation evaluation is also conducted based on patient's motion parameters and surface EMG. The clinical results show that the robot-assisted therapy has a positive effect on improving the upper limb motor function and activities of daily living in stroke patients. Meanwhile, the sensorimotor recovery of upper limb is analyzed at the behavioral level. The Fugl-Meyer scale score could be approximately estimated by extracting motion parameters to establish linear regression models. The contraction timing and participation of each muscle are analyzed using surface EMG, and points out that surface EMG could provide objective and quantitative evaluation on limb functions in stroke patients.