Recognition method of surface electromyographic signal based on two-branch network
An enhanced two-dimensional feature based two-branch network (ETDTBN) was proposed aiming at the problems of insufficient detailed information extraction and difficulty in distinguishing similar gestures in surface electromyogram (sEMG) gesture recognition. Discrete features were converted into two-dimensional feature maps by the proposed enhanced two-dimensional method. Then a multi-layer convolutional neural network (ML-CNN) was used to extract the spatial features,while a bidirectional gated recurrent unit (Bi-GRU) was used to extract the temporal features from the original signal. A self-adaptive feature fusion mechanism was introduced to fuse different branches,strengthen useful features and weaken useless features in order to improve the accuracy of sEMG recognition by considering that different features had different degrees of influence on the network. Experiments were used to train and test the ETDTBN in two scenarios of electrode displacement and different subjects comparing with mainstream sEMG gesture recognition models. Results showed that the overall recognition accuracy of ETDTBN were 86.95％ and 84.15％,respectively. Both accuracies are optimal,proving the effectiveness of the model.
万方数据
