A Deep LSTM Based sEMG-to-Force Model for a Cable-Driven Rehabilitation Robot

Abstract

The estimation of human motion intention is one major problem for rehabilitation robots. Long Short-term Memory (LSTM) network, which has been long used in sequential data processing, is currently considered to be a promising method to estimate the human voluntary force from the surface electromyography (sEMG) signal. In this research, a deep LSTM network (DLN) is proposed to build the three-dimensional sEMG-to-force model, which is then cascaded with an admittance model to determine the desired trajectory of an upper-limb cable-driven rehabilitation robot (CDRR). The control scheme of the human-robot system is implemented on a real-time platform, termed dSpace, which is characterized by strong processing power and high processing speed. The experiments involving two healthy subjects demonstrate that the proposed method can be applied in the CDRR for tracking a rectangular trajectory. This study will help the CDRR continuously estimate the voluntary force of patients after stroke with short-delay and high-accuracy, thereby promoting the active participation in the rehabilitation training.