Variations in information processing patterns induced by operational directives under varying fatigue conditions within the cerebral cortex can be identified and analyzed through electroencephalogram (EEG) signals. The inherent complexity of EEG signals poses significant challenges in the effective detection of driver fatigue across diverse task scenarios. Recent advancements in deep learning, particularly the Transformer architecture, have shown substantial benefits in the retrieval and integration of multi-dimensional information. Nevertheless, the majority of current research primarily focuses on the application of Transformers for temporal information extraction, often overlooking other dimensions of EEG data. In response to this gap, the present study introduces a Multidimensional Adaptive Transformer Recognition Network specifically tailored for the identification of driving fatigue states. This network features a multidimensional Transformer architecture for feature extraction that adaptively assigns weights to various information dimensions, thereby facilitating feature compression and the effective extraction of structural information. This methodology ultimately enhances the model's accuracy and generalization capabilities. The experimental results indicate that the proposed methodology outperforms existing research methods when utilized with the SEED-VIG and SFDE datasets. Additionally, the analysis of multidimensional and frequency band features highlights the ability of the proposed network framework to elucidate differences in various multidimensional features during the identification of fatigue states.