<p>Magnetoencephalography (MEG) and electroencephalography (EEG) are prominent non-invasive brain imaging techniques that have attracted considerable interest in neuroscience research. Unsupervised learning of M/EEG signals has become a critical area of focus. However, existing data transformation strategies often introduce representation space biases, whether through random and mechanical methods or adaptive approaches that rely on prior knowledge, which limits the effectiveness of unsupervised M/EEG signal analysis. To overcome these challenges, we propose a novel unsupervised learning method BrainDEC that enhances the extraction of robust and generalizable M/EEG features. By employing a disentanglement framework, we disentangle invariance and equivariance features in a manner that enables more precise and transparent control over the equivariance training process. This adjustment leads to improved transparency and greater precision in the learned representations. Extensive experiments across multiple M/EEG datasets demonstrate the superior performance of our approach in both linear and semi-supervised evaluation settings. Our method holds great promise for advancing brain signal analysis in downstream tasks, all within an unsupervised learning framework.</p> Graphical Abstract <p>By employing a disentangled equivariance constraint, our method explicitly separates invariant and equivariant features to ensure precise and transparent control over the representation learning process. Extensive experiments across multiple datasets demonstrate that BrainDEC significantly enhances feature robustness and generalization performance in downstream tasks.</p>