中国科学院自动化研究所机构知识库

脑卒中是一种由于脑部血管的阻塞或破裂而引起的脑血流循环障碍和脑组织功能结构受损的疾病，大多数脑卒中患者留有不同程度的运动功能障碍，其中上肢运动功能障碍的发生率最高，严重影响其生活自理能力及生活质量。相关临床研究表明，运动康复训练有助于脑卒中患者的脑功能重塑，从而实现对肢体运动行为的有效控制。与传统的人工辅助康复训练相比，机器人辅助康复训练擅长实现高强度、高精度、个性化的功能训练，通过集成自动化的康复评定方法可以构建一个智能康复辅助系统，是提高未来康复诊疗水平的重要手段。本文针对当前机器人辅助康复训练过程中存在的重点难点问题，围绕上肢康复机器人控制策略及运动功能评价方法展开研究，主要内容和创新点如下：

（1）针对机器人辅助康复训练中的动力学模型分析问题，提出了末端牵引式上肢康复机器人系统的动力学建模以及参数辨识方法，实现了优于最小二乘法的九参数动力学模型的精准辨识。该方法首先利用拉格朗日方法对本课题组研发的平面二自由度上肢康复机器人系统CASIA-ARM II进行动力学建模。在得到系统的动力学模型后，将其线性化为关于未知动力学参数的方程组，利用优化的遗传算法对未知参数进行辨识。通过引入自适应遗传操作和精英个体保留策略，在保证全局收敛性的基础上，增强了算法对未知参数的探索能力。另外，为了提高动力学参数辨识的精度，在辨识过程中利用非线性摩擦模型刻画了机器人关节换向前后的摩擦力特性，并采用卡尔曼滤波算法对未知运动学信息进行估计。基于上述动力学分析结果，针对机器人辅助被动康复训练中的轨迹跟踪问题，设计了一个基于关节空间动力学模型的复合控制器。实验结果表明，在该控制器的作用下，CASIA-ARM II可以辅助患者完成精准且稳定的预设轨迹跟踪运动。

（2）针对机器人辅助康复训练中的参考训练轨迹规划问题，提出了基于人体运动规律分析的轨迹规划方法，实现了结合患者主动运动意图与健康人运动模式的运动训练任务的自适应调整。对于末端牵引式上肢康复机器人，首先设计实验探究人体上肢与康复机器人交互过程中的末端运动规律。在此基础上，将够取运动中的急动度优化原则与基于上肢阻抗参数辨识的运动意图估计方法结合，实现了主动训练中任务空间内的末端轨迹规划；对于外骨骼式上肢康复机器人，首先设计实验探究人体上肢复合动作中的冗余控制规律。在此基础上，将统计分析得到的抬升角优化规律与基于最小急动度原则的末端执行器轨迹规划方法结合，实现了主动训练中关节空间内肩部与肘部的轨迹规划。仿真结果表明，以上两种轨迹规划方法可以基于人体正常运动规律生成顺应患者运动意图的参考训练轨迹，有助于提高患者运动再学习过程中的主动参与度。

（3）针对机器人辅助康复训练中的辅助力水平优化问题，提出了一种基于自适应频率振荡器和径向基函数神经网络的自适应控制方法，实现了以患者主动运动意图与运动实现能力为驱动的按需辅助运动训练。该方法首先利用自适应Hopf振荡器从人机交互的实际运动轨迹中提取稳定的运动模式信息，实现无传感器的主动运动意图辨识，并将其作为康复训练过程的驱动输入，对期望运动轨迹进行自适应地调整。与此同时，利用高斯径向基函数拟合患者主动输出能力，通过迭代更新神经网络的权值向量在线学习人机间动力学关系，根据患者的运动功能障碍程度的变化对机器人的辅助力水平进行优化。仿真和实验结果表明，在该自适应按需辅助控制器的作用下，CASIA-ARM II可以准确地获取患者的主动运动意图，在此基础上迭代优化期望训练轨迹和辅助力水平，使患者可以最大程度地利用其自身运动能力完成正常肢体运动模式的再学习。

（4）针对脑卒中康复训练中患者上肢异常运动模式的自动化康复评定问题，提出了一种基于运动协同分析和多模态数据融合的运动功能分析方法，实现了脑卒中偏瘫异常运动模式的准确检测与运动功能障碍程度的量化分析。该方法首先设计了包括15位脑卒中偏瘫患者和15位健康志愿者的临床对照试验，在受试者完成多目标变方向够物任务的过程中，同步采集上肢的运动学和电生理数据，记录关节运动和肌肉活动两个层面的信息。在此基础上，分别对关节运动层面中单个关节的运动轨迹和多个关节之间的协同运动模式进行特征提取，对肌肉活动层面中单块肌肉的收缩模式和多块肌肉之间的协同激活模式进行特征提取，利用监督学习算法构造单模态的初级分类器和集成分类器，获得脑卒中偏瘫患者运动功能障碍程度的分层决策。为了最大程度地利用不同模态信息之间的互补性，基于单模态的最佳预测输出构造多模态决策融合模型，得到一个基于预测概率的上肢运动功能评分，全面地量化了患者的上肢运动功能障碍程度。实验结果表明，该方法得到的上肢运动功能评分与Fugl-Meyer上肢运动功能量表评分之间存在显著的临床相关性，可以为康复治疗中的临床诊断提供重要依据。

Stroke is a chronic disease induced by cerebral haemorrhage or infarction. Approximately two thirds of post-stroke patients have severe deficits in upper-limb movements, which affect their performance in activities of daily living. Results of clinical studies suggest that motor rehabilitation is able to encourage neural plasticity and in turn, yields fast possible functional recovery. Compared with traditional manual rehabilitation, robot-assisted rehabilitation is believed to be more promising, due to the fact that robotic devices can provide high-intensity and subject-adaptive training, as well as evaluate motor function of post-stroke patients. This paper focuses on the control strategies and motor function assessment for robot-assisted upper-limb stroke rehabilitation, aiming to solve the problems in the robot-assisted rehabilitation training. The main contributions and innovations of this research are as follows:

(1) To perform the dynamic analysis of rehabilitation robots for passive training, the dynamics modeling and identification method are proposed for a novel upper-limb rehabilitation robot called CASIA-ARM II. To start with, the Lagrange-based dynamics modeling method is applied to obtain the dynamic equations of the parallel robot. Then the robot dynamic model is rearranged into a linear form with respect to the unknown dynamic parameters, and a genetic algorithm-based method is developed for parameter identification, which adopts the adaptive mutation rate and elite preservation policy to ensure algorithm evolve towards the optimal direction. For accurate identification, smooth joint velocities and accelerations are computed by Kalman filter, and a non-linear friction model is used to quantify the characteristics of frictions. Based on the estimated dynamic model, a model-based PD computed-torque controller is designed for passive training. Experimental results demonstrated that the actual trajectories and desired trajectories almost overlapped each other, which validates the effectiveness of the proposed the dynamics modeling and control method.

(2) To investigate the trajectory planning of rehabilitation robots for active training, two types of trajectory generation algorithms are proposed by integrating the motion intention of the patient and motion restrictions existing in healthy humans. For end-effector based upper-limb rehabilitation robot, interaction experiments of human-dominant mode and robot-dominant mode are designed to investigate the invariant law when humans physically interact with robots. Experimental results demonstrated that the minimum-jerk principle is a human preferred pattern in motor control for goal-directed reaching movements. The human arm impedance parameters are identified to reflect the movement intention, and then the optimal reference trajectory can be generated by integrating the upper-limb stiffness and minimum-jerk model. For exoskeleton-type upper-limb rehabilitation robot, experiments are designed to explore the redundancy resolution adopted by healthy humans. Experimental results demonstrated that the swivel angle can be approximated to the mean value in resolving the arm redundancy problem. By integrating with the minimum-jerk trajectory of end-effector, the optimal reference trajectories of shoulder and elbow joints were generated.

(3) To customize the assistance level of rehabilitation robots for active training, a subject-adaptive control framework is proposed using an adaptive frequency oscillator and radial basis function networks. The adaptive frequency Hopf oscillator acts as a sensorless estimator keeping the state variable in phase with respect to the external input to estimate the desired movement rhythm of post-stroke patients. By integrating the extracted movement phase and the minimum-jerk principle, the reference trajectory can be updated in real-time according to the patient's active effort. Then the Gaussian radial basis networks are included in the control framework so that it can compliantly assist patients by simultaneously learning a model of the patient's motor abilities contributing to task execution. A series of simulation and experimental results showed that the difficulty level of reference trajectories was modulated to meet the requirements of subjects' intended motion, furthermore, the robotic assistance was compliantly optimized in response to the changing performance of subjects' motor abilities, highlighting the potential of adopting the framework into clinical application to promote patient-led motor learning.

(4) To objectively quantify the upper-limb motor impairments in patients with post-stroke hemiparesis, a novel assessment approach is proposed based on motor synergy quantification and multi-modality fusion. Fifteen post-stroke hemiparetic patients and fifteen age-matched healthy persons participated in this study. During different goal-directed tasks, kinematic data and surface electromyography (sEMG) signals were synchronously collected from these participants. Motor features are separately extracted from three-dimensional trajectory and muscle activity to form the input vectors of local single-modality classifiers. In addition, kinematic synergies and muscle synergies are quantified to provide in-depth analysis of the coactivated features responsible for observable movement impairments. Based on the local output and motor synergy analysis, the ensemble classifier can be created for each modality to generate quantitative assessment of upper-limb motor impairments. Furthermore, the complementarity between the evaluation outcomes from different modalities is exploited by a multi-modal fusion scheme to provide more comprehensive assessment of upper extremity functional status. Experimental results demonstrated that the pathological movement patterns in patients with post-stroke hemiparesis can be reliably identified and the assessment result is well correlated with the score of traditional Fugl-Meyer clinical tests.