基于多模态信息融合的肘关节连续运动估计

Continuous Motion Estimation of Elbow Joint Based on Multi-Modal Information Fusion

目的针对目前上肢康复训练设备多为被动式、训练方式单一、患者主动参与度较低等问题,提出一种基于多模态信息融合的上肢连续运动估计算法,实现对肘关节力矩的准确估计。方法首先,在4种角速度下,采集受试者的表面肌电信号和姿态信号,提取信号的时域特征并利用主成分分析方法进行特征融合;其次,通过附加动量法和自适应学习率对反向传播神经网络(back propagation neural network,BPNN)进行改进,使用粒子群算法(particle swarm optimization,PSO)对神经网络进行优化,构建基于PSO-BPNN的连续运动估计模型;最后,以第2类拉格朗日方程计算的关节力矩作为准确值,对模型进行训练,并与传统BPNN模型进行性能对比。结果传统BP神经网络模型均方根误差为558.9 mN·m,R2系数为77.19%,优化模型后的均方根误差和R2系数分别为113.6 N·m、99.12%,力矩估计准确度进一步提高。结论本文提出的肘关节连续运动估计方法能够准确地识别运动意图,为上肢外骨骼康复机器人的主动控制提供切实可行的方案。

Objective Aiming at the problems of lacking initiative in upper limb rehabilitation training equipment,single training mode,and low active participation of patients,an upper limb continuous motion estimation algorithm model based on multi-modal information fusion was proposed,so as to realize accurate estimation of elbow joint torque.Methods Firstly,the surface electromyography(sEMG)signals and posture signals of participants were collected at four angular velocities,and the time domain characteristics of the signals were extracted.The principal component analysis was adopted for multi-feature fusion.The back propagation neural network(BPNN)was optimized through the additional momentum and the adaptive learning rate method.The particle swarm optimization(PSO)algorithm was used to optimize the neural network and a continuous motion estimation model based on PSO-BPNN was constructed.Finally,the joint torque calculated by the second type of Lagrangian equation was used as the accurate value to train the model.The performance of the model was compared with the traditional BPNN model.Results The root mean square error(RMSE)of the traditional BPNN model was 558.9 N·m,and the R2 coefficient was 77.19%,whereas the RMSE and the R2 coefficient of the optimized model were 113.6 mN·m and 99.12%,respectively.Thereby,the accuracy of torque estimation was improved apparently.Conclusions The method for continuous motion estimation of the elbow joint proposed in this study can estimate the motion intention accurately,and provide a practical scheme for the active control of upper exoskeleton rehabilitation robot.