Abstract: Uses the one-channel SEMG signal to recognize such actions as elbow flextion, elbow extension, forearm supination and forearm pronation, of which the signal sampling is 20 times as large as the highest frequency of SEMG. Although such an oversampling procedure reduces the background noise, it brings heavy job to calculation. However, the decimation and filtering can convert the sampling rate into Nyquist frequency and maintain the signal with lower noise. Using wavelet transform, the five relative energies in five sub-bands and the total energy of SEMG are obtained as primary characteristic vector comprising 6 dimensions. Then by K-L expansion, the primary characteristic vector with 6 dimensions is changed to that with 2 dimensions. Using BP networks as the classifier, the characteristic vector with 2 dimensions can maintain highly accurate recognition rate which isn't lower than the rate of multi-channel systems.