Artificial Neural Network for the Identification of Postural Instability in Subject Wearing Lower Limb Exoskeleton

Abstract

Real-time processing of human response to external disturbances assumes a key role in the design of control systems for lower limb exoskeletons. However, the automatic recognition of human intention movement is still an untapped issue in robotics, especially when focusing on stability. In this study, we investigated the feasibility of using Artificial Neural Networks (ANN) to predict human postural response to perturbations in different directions. Fourteen healthy adults underwent standard baseline perturbations via a benchmarking system, B.E.A. T, while wearing the EXO-H2, a lower-extremity exoskeleton. Lower limb kinematics were measured using seven inertial sensors. The B.E.A.T. platform provided four perturbative scenarios with 8° tilt steps and tilt directions, following the four cardinal directions of north, east, south, and west. A set of5 ANNs with different kernels was tested to predict the four perturbative responses. Features extracted from lower limb joint angles were used to train and test the algorithms. The highest accuracy (95.3%) was obtained when applying the Narrow Neural Network. The lowest/highest values of true-positive/false-negative rates were found in the north direction. Our results provide relevant information to implement a similar algorithm in the control system of lower limb exoskeletons for instability management.