Researches Of Multi-action Pattern Recognition And Prosthetic Hand Control Based On Mechanomyographic Signal

MMG (Mechanomyographic), which is a sound created by body’s muscle contraction, can be used as prosthetic control signal source. In recent years, a number of domestic and international researches show that MMG signal has an excellent prospect in prosthetic hand control. It is a big step forward if it can be used in practice.In this subject, the acquisition system structure of MMG signal has been built, including the selection of sensors, and the design of hardware and software. Through Matlab software, recognition system program has been written, by which the researches of the multi-action pattern recognition of hands can been done. Collect the MMG signals of four musles (FCU, FCR, ECR, EDC) on the forearm of32testers, who do the4kinds of actions (Hand hold, Stretch hand, Ruler to wrist, Extension of wrist), using4channels of TD-3piezoelectric accelerometers. The feature space constituted by a number of time domain and frequency domain can be extracted. And then by using PCA (Principal Component analysis), make its dimensionality reduction. Finally, make the assessment of results accuracy by the linear classifier and a10-fold cross validation. The researches in6kinds of actions (Hand hold, Stretch hand, Ruler to wrist, Extension of wrist, Cut wrist, Lsterotorsion of wrist) show that the expansion of the feature space can improve the identification rate. However, the feature space is too complexed, and will lead poor real-time. So make the analysis for the original feature space by the feature selection method based on CMI (Conditional Mutual Information). By the hardware circuit designing and embedded programming, the real-time platform based on TMS320F2812DSP has been built up.The above experimental results show that pattern recognition methods on4actions can achieve more than95%identification rate, and can achieve the best identified result when using3channels. What is more, the locations of four muscles selected on the final recognition result are not affected. For further researches about the six pattern recognition, the result shows that the expansion of feature space has an effective advantage on improving the identification rate. And the optimal subspace selected between different testers is different, but the best subspace of the same tester is relatively stable. Compared with PCA, the recognition rate is worse when using feature selection methods, but it provides the benefical reference to realize the real-time system. The prosthetic hand can be driven and make the appropriate action according to the DSP classification results. And prosthetic hand would create the force signals, which give the feedback to the system. The feedback signal can control the prosthetic hand to do some simple actions. It is evident that the practical application of MMG signal can be used as a reliable signal source to control prosthetic hand.