Abstract:Because surface electromyography (sEMG) has nonstationary, aperiodic and chaotic characteristics, the traditional feature extraction method is difficult to be compatible in realtime characteristic and accuracy. In this paper, an improved energy kernel feature extraction method based on sEMG is proposed to process the acquired EMG signals. Firstly, based on the EMG oscillator model, the newly proposed Threshold Matrix Count (TMC) feature extraction method is described in detail. Then, the myoelectric sensors were stuck on the surfaces of 10 different muscles of the leg to detect the EMG signal during different motion processes of the lower limb. After acquiring the required EMG signals, the EMG signal characteristics of the 10 muscles were extracted and ten different feature vectors xk can be obtained. After analysis, four muscles were selected as effective muscles. Finally, the effective muscle feature vectors xk were combined to obtain a feature matrix Xk, which is inputted into the BP neural network for training, and four motion patterns were identified. The experiment results show that the calculation efficiency of the proposed energy kernel feature extraction method is improved by 13 times and 9 times compared with those of the traditional two energy kernel feature extraction methods. At the same time, compared with the commonly used time and frequency domain feature extraction methods, after training the obtained model possesses better stability and the average recognition accuracy reaches 952%.