抄録
A real-time performance and strictness is important to apply myoelectric signal for control of a well-fare robot such as a prosthetic hand, rehabilitation and robot therapy. Past researches also has tried to discriminate a human motion from myoelectric signal in a moment for real-time performance, so that the signal can be apply to robot control. There is a paper to describe a method to discriminate six motions from myoelectric signals within 100ms. It's probability, however, of some motions are less than about 80 %. In order to apply myoelectric signal to robot control, it is serious to discriminate a human motion correctly. Moreover it should be 100 % as closer as possible. Additionally there is a paper to report that the accuracy of motion discrimination falls when myoelectric signal comes through couple neighboring muscles is analyzed by neural network method. Therefore, it might be said that to apply a neural network method for such a problem like past trials would have difficult to reach near 100 % accuracy of discrimination of human motions from myoelectric signal. The purpose of this paper is to develop a motion discrimination method from myoelectric signal within 80 ms data from start to measure, with almost 100 % probability for a real-time performance and strictness. In this paper, the myoelectric signal generated in muscles of a forearm of right hand is analyzed using combination of morlet wavelet analysis and three-layered neural network and the motions which are "Grasp", "Flexion of wrist" and "Pronation" of a right hand are discriminated. A bipolar method is used to measure three kinds of myoelectric signal generated in three muscles related with each motion. Through the tests of 87 times for a subject and 300 times for another subject, the validity of the proposed method is confirmed. As a result, for a subject, all of three motions are identified completely without mistake, for another subject, the result showed very high identified completely without mistake, for another subject, the result showed very high identification ability of the proposed method such as 94 % for "Grasp", 100 % for "Flexion" except "Pronation".
