Recognition of fatigue status of pilots has important research significance. Aiming at the complexity and accuracy of recognition of fatigue status of pilots, a new deep learning model based on electroencephalogram signals is proposed to recognize fatigue status of pilots. The delta wave (0.5sim4Hz), theta wave (5sim8Hz), alpha wave (7sim14Hz) and beta wave (14sim30Hz) are extracted by multi-scale decomposition of electroencephalogram signals using filters, and the reconstruction signals of them are taken as the input vectors of the model. A deep contractive sparse auto-encoding network-Softmax model is proposed for identifying pilots' fatigue status, and its recognition results are also compared with these of the deep auto-encoding network-Softmax and traditional PCA-Softmax model. The results show that the proposed deep learning model not only has a nice classification, the accuracy of which is up to 91.67%, but also the learned features are stable, and the proposed model is stable and reusable verified.