A Real-Time Exoskeleton Control Strategy for Multiple Gaits Based on Continuous State Variable Driving and MiniRocket Recognition

Abstract

In recent years, exoskeletons have been increasingly utilized for daily ambulation. Many existing methods such as kinetic modelling strategy have been implemented to achieve control of exoskeletons. However, most of existing system control algorithms are computationally cumbersome, which results in a poor control effectiveness in terms of real-time. To address the above issue, a united gait control strategy based on continuous state variable and MiniRocket for three common gaits is proposed. Gait data from healthy subjects is formulated into a mapping between gait phase and expected position using Fourier function formulae. Upon user mode transition, MiniRocket identifies and obtains new auxiliary contour curves. The hip joint difference δ is employed as a state variable. Gait phase is computed using arctangent function through the mapping relationship between δ and gait phase. To validate the strategy’s effectiveness, the exoskeleton CC-Walk is designed and data from three subjects is collected. The RRMSE of gait phase estimation for walking, ascending stairs, and descending stairs are 3.49%, 2.66% and 5.84% respectively, which overall outperforms the results estimated by the ANN (4.10%, 4.58%, 5.32%). The gait prediction accuracy of improved MiniRocket is 99.46%, with a single sample prediction taking 1.52 ms. The training duration is 5.89 s. These results superior to CNN, LSTM and the previous version of MiniRocket. Furthermore, the state variable algorithm for gait phase estimation involves only simple inverse function operations, ensuring extremely high real-time control and practicality.