Abstract: Mental fatigue is a condition caused by prolonged mental work. Electroencephalogram(EEG) is often used as the preferred tool for detecting mental fatigue. The key problem in the mental fatigue recognition is how to search for suitable EEG features. For the problem of information loss caused by the manual extraction of features in traditional pattern recognition, the convolutional neural network in deep learning was introduced. Due to the temporal and spatial characteristics of EEG, deep convolution neural network and shallow convolutional neural network with time domain convolution kernel and spatial domain convolution kernel were designed in this paper. They combined feature extraction and classification. Both of them classified normal and fatigued EEG data and we visualize the spatial domain convolution kernel of convolutional neural networks. The results show that the average classification accuracy of shallow convolution neural network is 98.868%, and the average classification accuracy of deep convolutional neural network classification is 98.217% . Both of them are higher than the accuracy of the traditional classification method. We can find the degree of participation of different electrodes in the network by means of visualization of the spatial domain convolution kernel. This shows that the model is highly effective in mental fatigue recognition tasks. The idea, at the same time, verifies that convolutional neural network provides a new method for the detection of mental fatigue state.