Individualized Gait Trajectory Prediction Based on Fusion LSTM Networks for Robotic Rehabilitation Training

Abstract

Robot-assisted gait training is promising to help patients recover from stroke. One key problem is how to design an adaptive and coordinated gait trajectory for each subject. In this paper, we utilize long-short term memory (LSTM) neural network with feature-level fusion, to effectively learn the multi-source motion characteristic data of lower limbs and adapt to the individual gait. Experiments are implemented on healthy subjects with motion capture system to get the joint data and electromyography acquisition equipment to collect the muscle signals simultaneously. The extracted features are input into the adopted neural network for fusion, and then train the model through a large amount of data. This learning-based approach can predict knee joint trajectory in conformity with individual gait patterns by combining kinematic data and biological signals. Experimental results indicate that this model can achieve a superior prediction performance compared with other traditional neural networks and the trained LSTM model also presents better adaptability between individuals.