Towards a more efficient human-exoskeleton assistance

There is evidence that the energy expended by humans can be reduced by wearing lower limb exoskeletons with user-oriented assistance strategies, such as human-in-the-loop (HITL) controllers. HITL algorithms can be implemented in exoskeletons for the automatic and online optimization of controller parameters, such as the torque profile, depending on the energy expenditure (EE) measured in real-time. This way, it is possible to minimize the EE and tailor the exoskeleton assistance for each specific user. But measuring EE is not trivial. It is more commonly estimated by indirect calorimetry, however, this method requires expensive equipment, takes too long, and is infeasible for everyday use in the real world. Therefore, this study explores machine and deep learning regression models (RMs) as EE estimators in different motor activities based on data acquired by wearable sensors and anthropometric features. Several inputs were tested but the best performance was achieved by the heart rate, the 3-axis acceleration of the chest, wrist, thigh, and ankle, and the body mass index. Results from a public dataset are presented, after the preprocessing of the data. The best-performing RM was an exponential Gaussian process regressor (GPR), that obtained root-mean-squared errors of 0.56 W/kg, 0.45 W/kg, and 0.60 W/kg for the standing, sitting, and walking activities, respectively. The GPR model outperformed a support vector machine, a boosted decision tree, a bagged decision tree, and a convolutional neural network.