Abstract: Aiming at the non-linear and non-stationary characteristics of motor imagery ElectroEncephaloGram (EEG) signals, a novel EEG signal classification method combining wavelet packet transform (WPT) and serial-parallel convolutional neural network(SPCNN) is proposed. In the process of wavelet packet transform, the EEG signal is decomposed in time and frequency, and the frequency band closely related to motor imagination is selected for reconstruction. The reconstructed EEG signal retains effective time-frequency information. Then, considering the features between and within the different channels of the EEG signal, the SPCNN network model is constructed to automatically extract the effective features and classify them. Use the public competition data set BCI competition Ⅳ 2b to verify, the results show that the method can adaptively extract effective features, and the average classification accuracy reaches 84.77%, which is 6.49% higher than the convolutional neural network. It provides a classification method for the research of brain-computer interface systems.