[home] [Personal Program] [Help]

---

Session: Poster Session 1 (Even numbers)
Session starts: Thursday 26 January, 16:00
Presentation starts: 16:00

---

Marit van Dijk (Delft University of Technology)
Louise Heringa (Delft University of Technology)
Marco Hoozemans (Vrije Universiteit Amsterdam)
Monique Berger (The Hague University of Applied Sciences)
DirkJan Veeger (Delft University of Technology)

Abstract:
One important performance determinant in wheelchair sports is the mechanical power exchanged between the wheelchair athlete and the environment. Recently, we developed a feasible method to determine power in the field based on rolling resistance estimates. Although this method shows to be accurate, rolling resistance estimates are wrong for the first two or three pushes. Probably this is caused by upper body movements. The aim of current study was to develop a model that accurately estimates the instantaneous rolling resistance based on two inertial sensors during wheelchair propulsion on a treadmill. Twenty-five healthy participants performed five different 120s-trials in an instrumented sports wheelchair on a treadmill. The five trials differed with respect to mass (+0, +5 or +15 kg) or tire pressure (1.75, 3.5, 5.25 bar). Before each trial, body mass inertial parameters were obtained and wheel friction coefficients were determined based on drag tests. During the trials, inertial sensors were attached to the participants’ trunk (thorax) and wheelchair (wheel), upper body kinematics were monitored using an optical motion capture system and a load pin integrated in the front wheels of the wheelchair measured the instantaneous load on the wheels. Based on the instantaneous front wheel load and center of mass kinematics, rolling resistance was determined. Accordingly, a deep learning model was trained to estimate the proportion of front wheel load from inertial sensor data. The training set included data of 20 participants and three trials. The rolling resistance of the other participants and other trials was used to evaluate the model by comparing model estimates with actual instantaneous rolling resistance. Based on the four most predictive features from the inertial sensor data, a machine learning model was trained and evaluated. The root-mean-squared-error (RMSE) turned out to be less than 5% of the body mass for the participants that were excluded from the training set. For the trials that were excluded, the RMSE was slightly higher. Overall, the rolling resistance estimate was improved significantly. To conclude, by combining a machine learning model with existing mechanical models, the changes in mechanical power can be determined accurately based on two inertial sensors.