Electroencephalogram (EEG) contains a wealth of temporal, spatial and frequency information, which is the most accurate physiological signal to reflect the emotional state, and plays an important role in the field of emotion recognition. Because the single feature EEG emotion recognition method has the problem of missing information, the EEG information processing method of three-dimensional fusion features was proposed. The EEG signal differential entropy frequency domain features and eight kinds of time domain features were fused, and according to the electrode position information for spatial arrangement, the EEG signal three-dimensional fusion feature was constructed. The attention mechanism was introduced into the multi-task convolutional neural network (MTA-CNN), and the constructed three-dimensional features were used as the input for test and analysis. The results show that the proposed model and method can significantly improve the classification accuracy in the valence dimension and arousal dimension of DEAP data set.