Action Pattern Classification Of Semg

Surface electromyography (SEMG) is bioelectricity of skeletal muscle acquired at the point of surface electrode when people make some movements. It includes a large amount of information of nerve and muscle's performance. EMG not only has been widely used in clinical medical and sport medical, but also conducts a wide research in modern controlling of prosthetic limb. Because any misoperation is very dangerous to amputee, high accuracy is the key of prosthetic limb controlling system, which is also the main purpose of this paper.Firstly, the former preamplifier circuit was revised. The use of "floating power ground" improves the ability to suppress common signal while enlarging differential signal. In addition, a high speed DAQ system consisted of two pieces of AD7862 and CY7C68013 working at GPIF mode was designed. In this paper, SEMG of 8 muscles were acquired by surface electrodes. Switching time between channels is very short (approximately 0.01 ms), which can be negligible, so that those SEMG can be considered as being acquired at the same time. Then those data were transported to PC automatically in bulk transfer mode, which offered signals for pattern identification.In preprocessing, energy threshold value was adopted for action segmentation. To remove a variety of noise interference, Butterworth low-pass filter, adaptive notch filter, based on EMD wavelet denoising filter were designed. From the spectrum of SEMG filtered and unfiltered, it can be seen that valid component has been extracted and noise has been removed.In feature extraction, three characteristics of SEMG were compared. Those were time domain statistics including mean, variance and zero-crossing number, power spectral ratio of AR model and energy of wavelet decomposition band. A conclusion was drawn that the last one is more stable. What's more, using energy of wavelet decomposition band as characteristic of SEMG for classification is promoted initially in this paper.Finally, power spectral ratio of AR model and energy of wavelet decomposition band were used as input of BP neural network for action pattern classification. Both reached high accuracy, and the latter was more reliable realitively. As to BP network's slow convergence, additional momentum and variable learning rate were used in this paper.