Classification Methods of sEMG Through Weighted Representation-Based K-Nearest Neighbor

As a valuable bio-electrical signal, surface electromyography (sEMG) can be adopted to predict the user’s motion gestures in human-machine interaction, but its validity severely depends on the accuracy of patterns recognition. In order to improve the recognition accuracy, the paper introduced a weighted representation-based k-nearest neighbor (WRKNN) to classify different hand motion patterns based on the forearm sEMG signals. All the signals were collected from 8 able-bodied volunteers through a sEMG acquisition system with 16 channels, and the root mean square (RMS) feature with the window size of 300 ms and the window shift of 100 ms were used to acquire the feature data. Based on the average classification accuracy and the standard deviation, the proposed algorithm and its improved version (weighted local mean representation-based k-nearest neighbor, WLMRKNN) were compared with k-nearest neighbor (KNN) and BP neural network. The experimental results show that WRKNN and WLMRKNN are superior to KNN and BP network with the best classification accuracy, and can be widely applied in the pattern recognition of sEMG in future.