Functional near-infrared spectroscopy (fNIRS), with its ability to directly measure cerebral hemodynamic changes and advantages such as portability, non-invasiveness, and resistance to interference, holds unique promise for driving fatigue detection. Despite the significant progress of deep learning in fNIRS modeling, two key challenges remain: how to simultaneously capture local temporal details and global dependencies effectively and how to integrate frequency-domain information typically overlooked by conventional time-domain approaches. To address these issues and enhance detection performance, we propose a novel fNIRS signal modeling method that incorporates structured encoding and frequency-domain feature enhancement, specifically designed for driving fatigue detection. The method employs Patch Embedding and Patch Merging techniques to efficiently extract multi-scale temporal features. Additionally, a learnable Fourier-based frequency weighting mechanism is introduced to selectively emphasize task-relevant spectral information. The fusion of temporal and frequency-domain features significantly improves both the accuracy and robustness of fatigue state recognition. Visualization of frequency-domain weights reveals that within-subject models capture personalized and dispersed spectral responses, whereas cross-subject models focus on stable and global frequency features, validating the effectiveness of our frequency modeling module. To further explore the underlying neural mechanisms of driving fatigue, we conducted both brain activation and functional connectivity analyses. Brain activation patterns indicate that during fatigue, oxyhemoglobin (HbO) signals in the prefrontal cortex exhibit negative activation, suggesting neural resource depletion, while positive deoxyhemoglobin (HbR) activation reflects the presence of compensatory mechanisms. Functional connectivity analysis shows enhanced HbO inter-channel connections under fatigue, suggesting compensatory integration within task-positive networks, whereas reduced HbR connectivity indicates neural desynchronization and impaired information transfer. By integrating FFTNet with brain activation and functional network analyses, this study provides a comprehensive understanding of how driving fatigue influences brain function. The dual mechanisms of functional compensation and dysregulation highlighted here offer new insights into the complex neurophysiological regulation underlying fatigue states.