中文摘要：

下肢运动预测对于步行康复机器人患者主动训练控制系统的设计具有重要意义。提出一种基于广义回归神经网络（GRNN）的利用肌电信号预测踝关节角轨迹算法：分别用肌电图仪和三维运动捕捉仪同步采集踝关节做屈伸运动时周围五块肌肉的肌电信号和踝关节角度，并对肌电信号进行特征提取。基于主分量分析的数值算法对肌电数据进行降维，得到肌电主分量信号。基于肌电主分量信号利用GRNN算法预测踝关节角轨迹，用黄金分割搜索算法确定GRNN中的最佳平滑参数or。采用小波消噪算法对踝关节角预测轨迹进行滤波以提高预测精度。用上述算法对9名志愿者进行实验的结果表明：该方法预测精度较高，与BP神经网络预测算法相比运算时间短且预测误差较小，因而更适用于下肢关节角轨迹的在线预测。

英文摘要：

The prediction of lower limb movement is important to the design of patient-positive training control system of walking rehabilitation robot. An ankle joint angle trajectory prediction algorithm using surface electromyography （SEMG） based on the general regression neural network （GRNN） is proposed. Electromyography and 3D motion cap- ture instrument are adopted to synchronously acquire the SEMGs of five muscles around ankle joint and the joint angle trajectory when the ankle joint does flexion-extension movement. The root mean squares of SEMGs are obtained as their features. The principle component analysis （PCA） based numerical algorithm is used to reduce the dimension of SEMGs, and the principle component signals of SEMGs are obtained. GRNN algorithm is used to predict the trajectory of ankle joint angle based on the principle component signals of SEMGs and golden section searching algorithm is used to determine the best smooth parameter tr of GRNN. The wavelet noise elimination algorithm is used to filter the predic- ted trajectory of ankle angle and improve the prediction accuracy. The above algorithm was used to carry out experi- ments on 9 volunteers, and experiment results show that this method has high prediction accuracy;compared with the BP neural network prediction algorithm, the proposed method has shorter computation time and smaller prediction error. Consequently ,the proposed method is more suitable for the online-prediction of lower limb joint angle trajectory.