Abstract: Because the sEMG( Surface Electromyography) is very sensitive to muscle fatigue，different patients
and electrode displacement，it is an arduous task to design a reliable robust and intelligent hand rehabilitation
device． To address these difficulties，a neural decoding method of rehabilitation gestures based on deep learning
is presented by using sEMG on the forearm of patients and CNN ( Convolutional Neural Network) to recognize the
movement intention． A combined feature extraction method is proposed to extract the combined features of each
channel of 8-channel sEMG． The combined feature includes 32 features which are wavelet packet decomposition
energy features，time-domain features and frequency-domain features． The eight channel features are formed into
an 8 × 32 numerical matrix and grayscale processed into a feature map，to train the convolutional neural network．
For five different gestures recognition，the classifier’s accuracy reached 98． 1%． Finally，according to the
classification results，STM32 I /O port outputs the corresponding PWM ( Pulse Width Modulation) signal，which
shows the feasibility of this method and laying a foundation for further control of rehabilitation glove movement．

Key words: surface electromyography ( sEMG) , convolutional neural network ( CNN) , wavelet package transformation, feature extraction, neural decoding