Conventional fatigue recognition methods struggle to capture the nonlinear characteristics of electroencephalogram (EEG) signals, thereby limiting the model’s ability to represent complex neural dynamics. Meanwhile, approaches based on shallow convolutional neural networks (CNNs) are constrained by the receptive field of convolutional kernels, focusing primarily on local temporal features while failing to model long-term temporal dependencies. In addition, existing EEG analysis techniques often overlook the role of frequency information in spatial feature modeling, leaving the potential coupling between spectral features and electrode spatial structures underexplored. To address these challenges, this paper proposes STS-Net, an EEG-driven multi-branch framework for joint spatio-temporal-spectral modeling in fatigue recognition. First, a parallel Temporal Feature Extraction Module (TFEM) is designed, consisting of multi-scale convolutional submodules and a CNN-BiLSTM submodule, which extract short-term local features and long-range temporal dependencies, respectively. Second, the Nonlinear Feature Enhancement Module (NFEM) introduces an entropy-aware enhancement layer to improve the model’s capacity to capture nonlinear dynamic patterns in EEG signals. This module also incorporates a temporal attention mechanism to enable effective fusion of multi-scale temporal features. Finally, a Spectral-Spatial Graph Modeling Module (SSGM) is proposed, which dynamically constructs inter-channel adjacency matrices guided by frequency-domain information. Through graph convolution, this module captures functional connectivity between electrodes, enabling the joint modeling of spectral features and spatial structures. Experimental results demonstrate that STS-Net achieves a mean accuracy (mACC) of 94.27% and a mean F1-score (mF1) of 94.06% on the fatigue dataset.