Electroencephalography (EEG) has become a prevalent tool in clinical practice, non-invasively recording brain activity to diagnose brain conditions and motor impairments. This study aimed to address the challenges posed by the low signal-to-noise ratio of EEG signals, which makes them difficult to label and impairs the practicality of using such systems. A pre-training framework, DMAE-EEG, was developed. It consisted of a denoising masked autoencoder to extract spatiotemporal representation from unlabeled EEGs. A brain region topological heterogeneity division method segmented nonuniform signals, improving spatiotemporal feature extraction, while a denoised pseudo-label generator suppressed noise, enhancing the robustness of representation learning. The model was validated on signal quality enhancement, resulting in a normalized mean squared error reduction of 27.78-50.00% under 25%, 50%, and 75% corruption levels. The DMAE-EEG was also validated on motion intention recognition, where it had a relative improvement of 2.71-6.14% in intrasession classification balanced accuracy. In both evaluations, the model outperformed current models. These findings demonstrate that the pre-training framework can facilitate the transferability of learned representations across a range of conditions, improving accuracy and robustness and advancing EEG-aided diagnosis.