Unilateral Upper Limb Fine Motor Imagery Recognition via Symmetric Cross-Modal Attention EEG–EMG Fusion

Precise sensing and decoding of unilateral upper limb fine motor intentions are critical for developing robust Brain-Computer Interfaces. However, non-invasive electroencephalography (EEG) faces inherent limitations, including low spatial resolution and significant signal overlap. These constraints hinder unimodal methods from effectively decoupling complex motor features. To address these challenges, this paper proposes CATA-Net, a multimodal fusion architecture combining EEG and electromyography (EMG). The model operates on an intention-execution synergistic mechanism to exploit the complementarity between central neural and peripheral muscle signals. The framework constructs modality-specific feature extractors. The EEG branch employs depthwise separable convolution with Atrous Spatial Pyramid Pooling to capture multi-scale spatiotemporal dependencies. The EMG branch utilizes a Dilated Residual Temporal Convolutional Network to extract dynamic muscle activation patterns. Crucially, unlike standard cross-attention frameworks that merely perform generic semantic feature exchange, the proposed Symmetric Cross-Modal Attention module explicitly addresses the inherent spatiotemporal misalignment caused by variables. It establishes a bidirectional channel between cortical intention and muscle execution to actively anchor EEG spatial representations to EMG temporal dynamics for mutual feature calibration. Experiments with 25 subjects covered three real movement tasks and two motor imagery tasks. CATA-Net achieved 98.3% accuracy in real movement scenarios. Furthermore, in challenging motor imagery grasping and twisting tasks, it outperformed the unimodal ATCNet baseline by 7.7% and 14.7%. Analysis confirms the effective integration of EEG spatial coding with EMG fine force control. This study provides a high-precision decoding solution offering novel perspectives on multimodal neural synergies.