Lower Limb Trajectory Prediction for Exoskeleton with AutoEncoder-Deep Gaussian Process

Gait trajectory prediction is a crucial component in compensating control system delays, reducing interactive resistance, and achieving real-time assistance in exoskeleton robotics. In this paper, an adaptive gait prediction model based on the autoencoder deep Gaussian process (AE-DGP) is proposed for joint angle estimation using measurement data from inertial measurement units (IMUs). The DGP process involves stochastic gradient Hamiltonian Monte Carlo (SGHMC) calculations and Monte Carlo Expectation-maximization (MCEM) optimization at a lower computational cost. The kernel of the model is the multi-feature fusion input based on autoencoder and the construction of the DGP model with an adaptive sliding window. To suppress spikes in predicted deviations, a weighted averaging strategy is employed to estimate future stable gait trajectories. Benefiting from the design, the proposed model can enhance the accuracy and adaptability of gait prediction at different speeds. Experiments involving four participants at four speeds indicate that the proposed method achieves hip and knee joint prediction with less than 0.16° and 0.35° root mean squared errors respectively. Compared to the Gaussian process (GP) and long short-term memory (LSTM) models, the proposed method exhibits superior performance in adaptive gait trajectory prediction at different speeds.