As an emerging interactive technology, brain-computer interface (BCI) has been widely used in various fields. The study of electroencephalogram (EEG) can not only improve people’s cognition of the brain, but also establish a new way for the brain to connect with the outside world. On the basis of the high dimensional, non-linear and spatiotemporal features of EGG signals, a novel method called parallel stacking encoded convolutional network is proposed for feature extraction and recognition of motor imagery EEG signals. The proposed method combines the unsupervised learning of stacked denoise autoencoder (SDA) with the supervised learning of convolutional neural network (CNN). First, the EEG data are applied to complete the parallel unsupervised training of multiple SDAs. The pre-training process can individually extract EEG signal features of each channel and avoid the interference between the various channels of EEG signals. Then the supervised convolutional training of the network based on gradient descent algorithm effectively fuses the features of multi-channel EEG signals. The experimental results indicate that the proposed algorithm has better accuracy than mainstream algorithms of EEG recognition.