基于主成分分析和小波神经网络预测跑步中垂直地面反作用力

Predicting Vertical Ground Reaction Force During Treadmill Running Using Principal Component Analysis and Wavelet Neural Network

目的建立基于主成分分析(principal component analysis,PCA)和小波神经网络模型(wavelet neural network,WNN)预测跑台上人体所受垂直地面反作用力(vertical ground reaction force,vGRF)的方法。方法选取9名后足跑者在跑台上以12、14与16 km/h速度跑步,通过红外运动捕捉系统和测力跑台同步采集运动学数据与vGRF。以Morlet函数作为激活函数并构建3层神经网络,将大腿、小腿与足的环节质心速度与髋、膝与踝关节的关节角度输入到模型。使用重相关系数(coefficient of multiple correlation,CMC)以及误差值评价预测模型的准确性,使用Bland-Altman方法分析vGRF预测峰值与测量峰值间一致性。结果不同速度下vGRF预测曲线与测量曲线间CMC>0.99,预测值与测量值间均方根误差(root mean squared error,RMSE)为0.18~0.28 BW,标准均方根误差(normalized root mean squared error,NRMSE)为6.20%~8.42%。不同速度下冲击力与推进力峰值NRMSE<15%。Bland-Altman结果显示,12 km/h推进力峰值的预测误差以及14 km/h冲击力和推进力峰值的预测误差在95%一致性区间。结论构建的PCA-WNN模型可准确预测出跑台跑步时人体所受vGRF。研究结果为在跑台上获得动力学数据和实时监测提供新途径,对研究跑步损伤及康复治疗有较大意义。

Objective To establish the method of predicting the vertical ground reaction force(vGRF)during treadmill running based on principal component analysis and wavelet neural network(PCA-WNN).Methods Nine rearfoot strikers were selected and participated in running experiment on an instrumented treadmill at the speed of 12,14 and 16 km/h.The kinematics data and vGRF were collected using infrared motion capture system and dynamometer treadmill.A three-layer neural network framework was constructed,in which the activation function of the hidden layers was the Morlet function.Velocities of mass center of the thigh,shank and foot as well as joint angles of the hip,knee and ankle were input into the WNN model.The prediction accuracy of the model was evaluated by the coefficient of multiple correlation(CMC)and error.The consistencies between predicted and measured peak GRF were analyzed by Bland-Altman method.Results The CMC between the predicted and measured GRF at different speeds were all greater than 0.99;the root mean square error(RMSE)between the predicted and measured vGRF was 0.18-0.28 BW;and the normalized root mean square error(NRMSE)was 6.20%-8.42%;the NRMSE between the predicted and measured impact forces and propulsive forces were all smaller than 15%.Bland-Altman results showed that the predicted peak errors of propulsive force at 12 km/h and that of impact force and propulsive force at 14 km/h were within the 95%agreement interval.Conclusions The PCA-WNN model constructed in this study can accurately predict the vGRF during treadmill running.The results provide a new method to obtain kinetic data and perform real-time monitoring on a treadmill,which is of great significance for studying running injuries and rehabilitation treatment.