Continuous estimation of upper limb joint angle from sEMG based on multiple decomposition feature and BiLSTM network

In human-robot interaction systems oriented to rehabilitation training, surface electromyogram (sEMG)-based human motion intention recognition has essential application value. Compared with discrete motion classification, continuous motion estimation is more natural, fast, and accurate. However, due to the non-stability, non-linearity, and strong randomness of sEMG, the effective motion information of sEMG is often lost when extracting the time-domain features of sEMG, and there are also cases where sEMG and joint angle data are not completely synchronized in practical applications, all of which affect the performance of continuous motion estimation. To solve the above problems, this paper firstly proposed a multiple decomposition feature (MDF) representation method based on variational mode decomposition (VMD) and wavelet packet transform (WPT), which can extract more hidden motion information of sEMG from multiple frequency scales; then introduced a bi-directional long short-term memory (BiLSTM) network to establish the regression model between sEMG and joint angle to deal with the incomplete synchronization problem between the input and output data. The experimental results showed that the multiple decomposition feature and the BiLSTM network regression model used in this paper could significantly improve the estimation performance in continuous motion estimation.