Continuous Prediction of Lower-Limb Joint Parameters for Robotic Rehabilitation Based on Horizontal Position of Center of Mass

Continuous prediction of joint parameters is replacing discrete gait phase control to be the mainstream in rehabilitation robot control field. The sensor-based methods which use inertial measurement units (IMU), surface electromyography (sEMG) and so on are widely used in gaining joint parameters for robot control. However, those methods introduce many sensors attached to patients and affect the walking during training. To reduce the number of sensors needed, a method is proposed to use centre of mass (CoM) horizontal position to predict angles, angular velocities and accelerations of ankle, knee, and hip joints. Long short-term memory (LSTM) is a kind of recurrent neural network (RNN) widely used in predicting time series data. To gain the most suitable model to predict joint parameters of each joint, the performances of Autoencoding-LSTM (combining encoder-decoder with LSTM), CNN-LSTM (combining convolutional neural network with LSTM) and stacked LSTM with different input window sizes on predicting joint parameters of each joint are compared, and the models with optimal performances are selected as a model pack to achieve high quality prediction of joint parameters. The number of sensors needed is reduced by 50% with the accuracy equal to those methods using IMU or sEMG sensors. And the results additionally show that different models perform variously on predicting different joint parameters of different joints.