The Study Of Hand Actions Based On SEMG Dentification Technique

When the human action behavior occurs, neuromuscular bioelectric signal can be recorded by electrodes on the surface of skeletal muscle, which is surface Electromyography (sEMG). The signal is a direct reflection of the state of human nerve and muscle function, therefore, the prosthesis controlled by myoelectric signal is biomimetic and convenient. Beacause the sEMG identification tech-nologies is still immature, there is some distance between theoretical research and practical applications.This article in view of myoelectric prosthetic hand, based on sEMG dentifi-cation technique, is intended to identify a variety of hand action pattern. The main contents include: acquisiing sEMG and establishing platform, pretreatment method, and action recognition.Overviews about newly developed prosthetic hands and typical EMG con-trol methods are given in detail. Then, the production and characteristics of EMG is introduced. Departure from the freedom of normal staffing, hand gestures are re-planning in a scientific way. The electrodes position is selected on the bsis of muscle kinematics principle. Furthmore, the platforms of acquisition hardware and software is built and the experiment is implemented.In the preprocessing, the disadvantages of segmentation method based on time window is discussed and the Dynamic Cumulative Sum (DCS) is researched. Then some feature extraction methods are compared, inculding typical time-domain characteristics (mean, rms, over the number of zeros) and the frequency domain values (power spectrum estimation value and mean power frequency).The experimental results show that the calculation method is simple and fast, and the separability is strong in time-domain. Considering the response of speed and recognition rate, time-domain characteristics are more suitable for sEMG. In pattern classification, the artificial neural network is selected as identifi- cation method. The shortcomings in the standard BP neural network Sum such as slow convergence speed and easy to fall into local extremum have been improved. Based the improved BP neural network, eighteen kinds of common hand gestures are recognized respectively used single and three networks. The results show that it is difficult to identify variety actions for a single neural network but easy to three. In the method of simultaneously using three neural networks, the structure is simple, the real-time performance is excellent and classification results are re-liable.