A comprehensive approach to recognition of mental commands is proposed, which is based on the analysis of frequency modulation of sensorimotor rhythms of electroencephalographic (EEG) signals using modern deep learning architectures. The focus is placed on adaptation and comparative analysis of the models, including the modified AlexNet and simple configuration MobileNetV2, as well as hybrid LSTM–Transformer networks with the potential for capturing temporal and spectral patterns. A specialized dataset has been created based on EEG recordings from 30 volunteers, followed by preprocessing, filtering, and normalization procedures. A key feature of the approach is the use of spectrograms with high temporal and frequency resolution, allowing for the effective extraction of frequency-modulated patterns associated with imagined motor actions. The proposed neural network modifications are designed to increase noise resilience and adapt to individual EEG signal characteristics. The best results have been obtained using the hybrid LSTM– Transformer model, which has demonstrated high classification accuracy on the test dataset and stable validation metrics. The reported findings highlight the potential of deep neural networks for the implementation of brain–computer interface (BCI) systems, especially in applications requiring sensitivity to subtle changes in frequency responses of biosignals. The presented work can underlie the development of adaptive BCIs in rehabilitation and assistive technologies.