In the brain-computer interface technology, MI-EEG is difficult to extract features by hand and the classification accuracy is low due to its high degree of non-stationarity. This paper conducts corresponding research and exploration to solve this problem, and designs a shallow convolutional neural network (SCNN) based on short time Fourier transform (Short time Fourier transform, STFT). Through tuning experiments on the hyperparameters of the network model, the model was tested on the public data set 2b of the BCI competition in 2008 and got good classification results. The classification accuracy rate reached 80.23%, which was much higher than the accuracy of the CNN model without STFT processing of 61.04%. Under the same test indicators, it is superior to the traditional machine learning classification method (CSP+SVM) with a classification accuracy of 73.52%, and it also has more advantages than the same type of deep learning method (CNN-SAE, 77.60%).