Hand gestures recognition from surface electromyogram signal based on self-organizing mapping and radial basis function network

Predicting the intention of human hand movements is a practical problem in prosthetic control. In recent years, surface electromyography (sEMG) has been widely used as a signal source in the field of wearable exoskeleton motion recognition and human-computer interaction. However, how to extract the information from sEMG signals and evaluate the intention of human hand movement effectively is still difficult. In order to achieve this goal, this work proposed a processing algorithm based on self-organizing mapping network (SOM) and radial basis neural network (RBF) for feature selection and classification recognition, then the principal component analysis (PCA) to reduce the size of feature vectors was used, finally used for pattern classification from sEMG signals to hand motion. In this research, the classification method mainly used the SOM method to find the hidden nodes centers of the RBF network, the Euclidean distance between the data centers was used to calculate the variance of the node and find the optimal center and radius of the radial basis function, so as to improve the learning performance of RBF network. In the experiment, the MYO armband sensor was used to sample the real sEMG signal data of 6 volunteers under 8 gestures. The experiment result show that the proposed algorithm as a classifier achieves a maximum recognition rate of 100 %, an average recognition accuracy of 96.875 +/- 2.7296 %, and a response time of 0.437 s. Meanwhile, the effects of the proposed method on hand motion recognition with different classifiers (RBF with k-means, K-Nearest Neighbor, Multi-Layer Perceptron with Scaled Conjugate Gradient) were compared. The corresponding average accuracy rates were 95.833 +/- 3.3244 % (RBF with kmeans), 94.583 +/- 2.243 % (KNN) and 88.89 +/- 1.1324 % (MLP with SCG). Compared with existed methods, the advantages of the method proposed in this research are as follows: 1) This research selects the PCA method and threshold value method based on the short-term average energy (STAE) used to detect the active segment of sEMG signal, so as to select the appropriate feature vector; 2) The proposed algorithm of SOM combined with RBF has higher identification accuracy and efficiency than that of RBF with k-means, which is more conducive to distinguish different actions; 3) While ensuring real-time performance, it can accurately classify gestures that are easy to be confused, indicating that this classification method has a good application prospect in prosthetic control and other fields.