Understanding and decoding movement-related brain activity from EEG is essential for advancing brain-computer interfaces and other neurotechnology applications. We propose a deep learning framework that leverages band-limited mu and beta power of single independent components, enabling the network to autonomously learn discriminative features without relying on handcrafted extraction. By avoiding traditional feature engineering, this approach minimizes bias and preserves richer spatiotemporal information embedded in the brain signal. The sequential structure of EEG data poses computational challenges for neural network models. To address this, we integrate a self-attention mechanism into a bidirectional LSTM architecture. This augmentation enhances the model’s ability to attend to temporally salient regions, improves the representation of long-range dependencies, and yields interpretable attention. Our results demonstrated that the attention-augmented BiLSTM achieves high classification performance metrics and revealed that pre-movement neural activity carries the most discriminative information for decoding motor intention.