This paper presents an approach for identifying neurophysiological markers of fatigue based on the analysis of EEG sensorimotor rhythms using deep learning methods. A methodology is proposed for generating spectrograms from multichannel EEG signals, including normalization, preprocessing, and data segmentation. Three adapted convolutional neural network architectures (EEGSpecResNet2D, EEGSpecMobileNetV3, EEGSpecAlexNet) are compared in both multiclass and binary classification tasks related to fatigue phases. EEGSpecResNet2D demonstrated the highest accuracy and robustness against overfitting. Additionally, training dynamics were analyzed, and pattern recognition features during intermediate task phases were visualized. The results confirm the effectiveness of residual architectures and anisotropic signal processing for monitoring cognitive load. The study highlights the potential of neural network models for developing BCI interfaces aimed at assessing and predicting operator functional states.