Predict Hip Joint Moment Using CNN for Hip Exoskeleton Control

Abstract

The assistance of the hip exoskeleton is closely related to hip joint moment, and the corresponding assistance can be generated by hip joint moment directly. However, it is difficult to obtain the hip joint moment directly from the sensor during walking. An effective method is to use deep learning models to predict hip joint moment using hip joint angle information. But there are still many challenges in applying this method to practical control systems. This study designs specific parameters to train a deep neural network for predicting hip joint moments and discusses the effects of the number of convolutional layers (N) and the length of sampling time (SL) on the prediction accuracy and the computation time in a real control system. Experiments show that for N $\le$ 5, using angle information with SL $\le$ 1.0 s achieves the highest prediction accuracy. However, for larger values of N (N > 5), the optimal SL becomes to be longer, where it increases to 2.0 s and the corresponding RMSE is 0.06260 Nm/kg and 0.05977 Nm/kg, respectively. These results indicate the importance of SL for prediction accuracy. Furthermore, a comparative analysis of the computation time with models using different N and SL is performed, and the relationship between computation time and prediction accuracy is discussed. This study provides specific model training parameters and detailed comparative analysis, which facilitates further application of deep learning models on hip exoskeletons.