Multi-pattern recognition of sEMG based on improved BP neural network algorithm

For realizing seven hand gestures classification correctly, wavelet transform is used firstly to eliminate the noise in sEMG, because of its multi-resolution analysis characteristic. Then combine time domain features (such as EMG integral, variance, the third-order AR model coefficients) with frequency domain features (power-spectrum) as the inputs of neural network classifier to discriminate seven motion patterns. According to the shortcoming of traditional BP neural network algorithm which is easily trapped into local minimum, an improved one based on existing BP algorithm and simulated annealing algorithm is proposed in this paper. The experimental results indicate that the correct rate is above 90% by using the above algorithm. Comparing with traditional BP algorithm, the novel one has better recognition capability.