基于Hill肌肉模型的人体关节力矩智能预测

Human joint moment prediction based on artificial neural network

目的人体关节力矩是康复评估和人机交互中非常关键的因素之一。它可以通过人工神经网络(artificial neural network,ANN)模型以肌电等信号作为输入进行预测,但是由于人体结构的复杂性导致缺乏有效方法确定ANN模型的输入变量。为此本文提出了一种基于Hill肌肉模型获取关节力矩神经网络预测最优输入变量的方法。方法采用Hill肌肉模型结合人体几何学知识建立关节力矩,智能预测输入输出关系的数学模型,把Hill肌肉模型中肌肉纤维长度和收缩速度,以及以关节自由度为支点的肌肉力臂等不可活体测量的输入变量,转换成关节自由度所关联肌肉的肌电信号(electromyography,EMG)以及这些肌肉所驱动关节自由度角度和角速度等可在线测量的变量。实验中以本文获取的变量作为极限学习机(extreme learning machine,ELM)的输入,对由1位在跑步机上以0.4、0.5、0.6、0.7和0.8 m/s等5个不同速度行走的右下肢偏瘫患者所获的数据进行测试。为了评估本文所提方法智能预测的泛化能力,实验在两个不同的泛化水平下进行,它们分别是只把3个低速数据(0.4、0.5和0.6 m/s)和全部5个速度的数据用于神经网络的训练并预测所有速度下的关节力矩值。实验通过预测值与反向动力学计算值之间的归一化绝对误差和互相关系数评估。结果本文获取的输入变量与其他关节角和角速度作为输入的方法相比,预测关节力矩值更精准。除右踝关节内外翻外,其他关节力矩预测结果的最大归一化绝对误差为12.93%,最小平均互相关系数为0.89。结论该方法比目前通用多体反向动力学的输入变量少,且可实现关节力矩值的在线预测,为运动康复中实时步态分析和外骨骼机器人控制提供技术支持。

Objective Human joint moment is one of the most important factors in rehabilitation assessment and human-robot interaction,which can be predicted using an artificial neural network model.However,challenge remains as lack of effective methodologies to determine the input variables for the artificial neural network(ANN)model.Methods This study develops a novel method to determine the optimal input variables to the ANN based on the Hill muscle model for estimating lower extremity joint moments.In this method,we translate the muscle-tendon moment arm,velocity and length in the Hill muscle model to the online measurable variables,i.e.muscle span joints’angles and angular velocities signals.We use moment-associated muscles’electromyography(EMG)signals together with these variables of the muscle as the inputs to the ANN for joint moment prediction.The method is tested on the experimental data collected from a subject with a highly functional hemiplegic stroke,who is walking on a treadmill with different speeds,i.e.0.4,0.5,0.6,0.7,and 0.8 m/s.We first use the data collected in all speeds as the training data to predict the joint moment.To evaluate the generalization ability of our method,we then repeat the same test using only low speed(0.4,0.5 and 0.6 m/s)data to train the ANN model for the joint moment prediction at all speeds.The accuracy of prediction is evaluated by using the normalized root-mean-square error and cross correlation coefficient between the predicted joint moment and multi-body dynamics moment.Results This method can predict joint moments with a higher accuracy than those obtained by using other joint angles and angular velocities as inputs,except the right ankle inversion-eversion(IE),the maximum normalized root mean square error of each joint moment of the lower limb is 12.93%,and the minimum average cross correlation coefficientρis 0.89.Conclusions The proposed method provides us with a useful tool to predict joint moment using online measurable variables,which uses less variables than the commonly used multi-body dynamics approach with a comparable accuracy.This method may facilitate the research on real-time gait analysis and exoskeleton robot control in motor rehabilitation.