Research On Hand Grassping Movmwent Of SEMG Signals For Artifical Limb

Hand is an important tool for work and human communication, and the loss of hand would cause inconvenience in both work and life to the disabled ones.The 27 DOFs in human’s hand and wrist make it easy to achieve flexible gestures and grabbing functions. Its convenience would not cause too much attention from the mind and the outside world. With the development of technology, higher demands to the quality of life are required by the disabled persons. However,prosthetic hand available in the market cannot meet their requirements. To create a prosthetic hand that can realize various actions freely will improve the quality of life of the disabled essentially. Nowadays, the structure designed for the mechanical body has basically meet the requirement of the DOF of human hand,which means that prosthetic hand realizes the action functions of human hand mostly at the perspective of mechanical structure. But the functions achieved via control technique are fewer than those of human hand. It has many disadvantages,for example, the action patterns that can be identified are few; the recognition rate is low, the sensory feedback function is scarce, the sense of the phantom limb is lacking. Taking the EMG signals of forearm of human as the control source, the actions and gestures of human hand is pattern recognized. The main research contents are as follows, physiological basis and characters that EMG signals produces, pretreatment during the process of signal acquisition;decoupling of multi channel EMG; the pattern recognition and the classification of the grabbing action of human hand.Firstly, the research elaborates the control source of prosthetic hand, the mechanical structure of humanoid prosthetic hand, as well as the detection method of EMG, and the physiological basis that EMG signals produces the pattern recognition is elaborated. Aiming at the characteristic of strong noise during signal acquisition of EMG, the possible noise sources are analyzed, the EMG signals collected is spectrally analyzed, the major noise sources(the frequency electric and its harmonic component) are obtained, and the full periodaverage filtering, envelope filtering and homomorphic adaptive filter are adopted.Aiming at the signal loss of EMG during the full period average filtering process,which may affect the pattern recognition, the moving full period average filtering is proposed. Analyzing and comparing three filtering methods mentioned above,the pros and cons and applications of different pattern of filtering are obtained to provide method for the pretreatment of EMG signals extraction.Secondly, to reduce the volume of the subsequent data processing of EMG signals, the pattern recognition was carried out by using the steady state data.And it requires identifying the action conversion point from the collected EMG signals, and then segmenting the motion. In this paper, the short-time energy,DCS and the dynamic finite difference method are used to identify the motion segmentation point, and the recognition rate is analyzed and compared. The short-term energy method can determine the segmentation point of the hand movements, muscle condition and strength. DCS method can accurately determine the action segmentation point and muscle state, but this method needs large amount of calculation and impacts control rapidity. The dynamic finite difference method can accurately determine the action segmentation point without setting threshold and window width. This method reduces the human factors to the action segmentation point of the judgment and improves the accuracy rate of the recognizing the motion segmentation point.Thirdly, since actions of human hand is conducted by the action of dragging the specific bones by specific muscles, to recognize more action patterns of hand needs multi-channel data collection. There are serious problems of signal crosstalk during collecting multi-channel EMG signals. Therefore, the ICA analysis method is adopted to decouple multi-channel EMG signals, and related analysis is used to obtain the matching relationship between the EMG signals decoupled and the original one.Finally, eight kinds of basic grab actions are confirmed to recognize thecatching pattern. The whole cycle moving filtering method is adopted to obtain effective EMG signals and the static EMG signals data of grabbing action is obtained by accurately determining the action segmentation point. By the extraction of characteristic quantities of time domain, frequency domain, wavelet transform, parameter model of the EMG signals, the experimental research ofseparability and recognition rate characteristics, the fusion feature of the wavelet coefficient maximum modulus value and time domain varianceare adopted as the characteristic quantity of the acquisition of pre-grabbing action, and the fourchannel EMG signals BP neural network and dynamic fuzzy neural network are adopted to recognize the eight basic actions of hand. The experiment shows that two kinds of neural networks has no significant difference in the recognition rate but characteristic quantity fusion can effectively improve the separability and the recognition rate of human hand actions.