Dynamic Topology Networks for Driving Vigilance Recognition to Enhance Human–Machine Interaction

Electroencephalography (EEG) signals directly reflect the cognitive states of the subject and provide a critical pathway for enabling intelligent and adaptive human-machine interaction in consumer electronics. However, existing approaches exhibit limitations in cross-frequency feature extraction and dynamic brain functional connectivity modeling, which restrict their accuracy, robustness, and applicability in consumer-level scenarios. To address these issues, we develop a novel framework, termed FBT-DGSN (Frequency-Band-Topological Dynamic Graph Structure Network), for driving vigilance detection. The proposed framework integrates cross-frequency fusion with a dynamic topological network architecture. Specifically, multiple brain functional connectivity graphs are constructed from frequency-band-decomposed EEG signals. A topology-enhancement module is then employed to capture and fuse spatiotemporal–spectral features across different bands. Subsequently, a dynamic weighting module generates inter-channel association matrices. Unlike traditional adaptive graph convolutions, this module explicitly models inhibitory and excitatory inter-channel relationships to capture complex neural interactions. Multi-layer graph convolutions are used to extract deep feature representations. Finally, by integrating topological brain network features, we construct a human–machine interaction feedback mechanism. Experiments conducted on the SEED-VIG dataset demonstrate that the proposed FBT-DGSN achieves superior performance under both subject-dependent and subject-independent validation protocols. Furthermore, comprehensive analyses of computational efficiency (FLOPs and inference latency) confirm the model’s suitability for real-time deployment in automotive HMI systems. Moreover, quantitative interpretability analyses reveal the critical involvement of prefrontal and parietal regions during vigilance decline.