In order to meet the needs of complex application scenarios such as active rehabilitation training and human-computer interaction, a hand motion pattern recognition method based on multi-channel surface EMG signal (shorted as sEMG) wavelet packet decomposition is proposed. Through the comparison and analysis via experiments, the optimal sampling layout scheme is determined. The automatic identification algorithm of the EMF signal segment based on the digital filter is designed by collecting the sEMG signal of the forearm surface, and the production of the label can be completed quickly and accurately. The wavelet packet decomposition coefficient of the original sEMG signal is used as the feature vector training classifier. By comparing the influence of different hidden layer nodes on the accuracy of classifier pattern recognition, all the structural parameters of BP neural network model classifier are finally determined. The BP neural network for hand motion pattern classification was designed and trained. The average recognition rate of 9 kinds of hand movements was 93.6% and the calculation time was less than 150ms.