To improve the prediction accuracy of crew fatigue and enable early warning of abnormal states， this study proposes a novel fatigue prediction method integrating the temporal convolutional network-long short-term memory （TCN-LSTM） with a hierarchical attention mechanism. The approach employs a multi-scale temporal feature extraction module to capture time-frequency characteristics and long-term dependencies in electroencephalography （EEG） signals， while the hierarchical attention mechanism effectively fuses EEG and psychological scale data. Differential prediction strategies are designed for operating pilots and monitoring pilots. Experimental results demonstrate that the proposed method achieves a 15.3% higher prediction accuracy compared to single-modality models， with an early warning time window advanced to 12.5 min. Specifically， the TCN-LSTM hybrid network improves temporal feature extraction efficiency by 22.7% over conventional LSTM networks， and the hierarchical attention mechanism enhances multimodal fusion performance by 18.4%. This hybrid architecture outperforms single deep learning models in both timeliness and accuracy， demonstrating applicability for real-time monitoring in aviation human factors engineering.