Long-term flight would result in cognitive fatigue of pilots, which is a critical influence factor to aviation incidents. Thus, accurately and timely detecting their fatigue status during flight has become an urgent and important issue. Brain activity exhibits high sensitivity to cognitive state variations, rendering them an effective biomarker for fatigue detection. However, substantial inter-individual variability in neural responses poses challenges in identifying specific brain regions that accurately reflect individualized cognitive fatigue patterns. In the current study, eighteen pilots were recruited to participate in a long-term simulated flight task combining the Psychomotor Vigilance Task and airfield traffic pattern task to induce cognitive fatigue, while synchronously collecting brain activity of their frontal-parietal brain regions using functional Near-Infrared Spectroscopy (fNIRS). Then, a fatigue-fNIRS Net (FF-Net) based on self-attention mechanism and Long Short-Term Memory network was proposed to extract individual features related to cognitive fatigue by integrating time- and channel-dependent information, and achieved an average accuracy of 94.08%, which is 7.32%, 3.54%, 14.02%, and 11.07% higher than that of LSTM, Transformer, fNIRSNet and fNIRS-T, respectively. The brain regions, including the middle frontal gyrus, supplementary motor area and precentral gyrus, were identified and played an important role in representing cognitive fatigue in most pilots, while the parietal gyrus and precuneus contributed more in few pilots. The findings indicated that the proposed FF-Net not only effectively detects the cognitive fatigue status in pilots, but also identifies brain regions associated with individual fatigue, providing insights for real-time cognitive state monitoring during flight.