Abstract:Objective To explore the accuracy and stability of principal component analysis and wavelet neural network model to predict the vertical ground reaction force on a treadmill. Methods Nine rearfoot strikers were selected and participated in a running experiment on an instrumented treadmill (Germany, h/p/cosmos mercury) at 12, 14 and 16 km/h. The treadmill has a running area of 150 50 cm and two force plates of Kistler. And the kinematics data and vGRF values, such as the center of mass velocity, joint angle and angular velocity of the lower limbs, were collected. The 273 samples of training set and 42 samples of test set were input into the PCA-WNN model. A three-layer neural network framework is constructed by python 3.8, in which the activation function of the hidden layers is the Morlet function, and the loss function E is optimized by the gradient descent method. The cumulative contribution rate of principal component analysis is set to 95%, and the kinematic data is filtered out and input into the WNN model for training and predicting. The performance of the model is evaluated by the root mean square error, standard root mean square error, mean absolute error, maximum error and correlation coefficients. Results 17 principal components such as velocities of the center of mass of segments and joint angles were input into the wavelet neural network model. On the whole, the vertical ground reaction force curves predicted by the model has a good fit with the measured curve, the error is quite small (RMSE=0.18-0.28 BW,NRMSE=6.2-8.42%) and there is a significant correlation between the two (r> 0.97, P<0.01) ; the errors between the predicted and measured impact force peak and active force peak are all quite small and relatively stable (RMSE=0.08-0.37 BW,NRMSE=2.91-11.14%). In particular, the prediction result of 12 km/h is more similar to the measurement and there is no significant difference(P>0.05). Conclusion The principal component analysis and wavelet neural network model constructed in this study can accurately and stably predict the vertical ground reaction force during running on a treadmill, which provides a new method to obtain kinetic data and to perform real-time monitoring on a treadmill.