Hybrid KF-LSTM Model for sEMG-Based Handwriting Numeral Traces Reconstruction

For the purpose of reconstructing handwriting traces from neuromuscular activities effectively, a well-integrated Kalman filter modified long-short-term memory network hybrid method (KF-LSTM) is proposed, which can train and decode the sEMG (surface electromyography) signals to the corresponding coordinates of handwriting numeral traces. Five participants were recruited for the between-group and within-group experiments. After synchronously collecting the sEMG signals and coordinates in the handwriting process, the KF-LSTM pre- diction models were constructed. The decision coefficient and the subjective identifiability were calculated as the evaluation indices. The performance of the KF-LSTM models was compared with the LSTM models, the NN (neural network) models and the KF (Kalman filter) models. The experiment results show that the proposed KF-LSTM method perform better than the other 3 methods, improve the reconstruction accuracies and make re- constructed traces much smoother.