The purpose of this study was to develop a deployable neural network (NN) to predict hip, knee, and ankle Net Joint Moments (NJM) from Ground Reaction Force (GRF) data during the hang-power clean. Thirteen male lacrosse players performed the hang-power clean exercise at 70% of their one-repetition maximum while GRF and 3-D motion data were acquired. An inverse dynamics procedure was used to calculate hip, knee, and ankle NJM. Center-of-mass velocity, position, and power were calculated from the GRF data and used as inputs to a NN that predicted hip, knee, and ankle NJM. Predicted NJM from the trained NN exhibited moderate root mean squared errors, but produced large percentage differences between predicted and calculated peak NJM when tested on new data, which likely resulted from overfitting during open loop training or insufficient closed loop training.
© Copyright 2018 ISBS Proceedings Archive (Michigan). Northern Michigan University. Veröffentlicht von International Society of Biomechanics in Sports. Alle Rechte vorbehalten.