Abstract
In the human-machine collaboration paradigms of Industry 5.0, the intrinsic non-stationarity and domain drift of electroencephalogram signals pose significant challenges to fatigue monitoring. Consequently, achieving accurate and robust cross-subject fatigue detection is paramount for workplace safety. This paper proposes a Spatio-Temporal Metric-Guided Multi-Stream Expert Domain Adaptation Network. The framework initially employs Singular Value Decomposition-based signal reconstruction to attenuate non-stationary noise, followed by Mixup data augmentation to enhance and smooth the feature space distribution. Furthermore, a hybrid architecture integrating Domain-Specific Batch Normalization with a ResNet-Transformer backbone is constructed to precisely decouple statistical discrepancies between domains. Finally, Multi-Kernel Maximum Mean Discrepancy is introduced as a dynamic routing criterion to facilitate fine-grained, adaptive decision-making for target samples. The baseline experiments on the SEED-VIG dataset yield a subject-wise mean accuracy of 90.54% across all 23 subjects, including S15. Visual analysis of brain topography confirms that the model successfully aligns distributional shifts caused by individual heterogeneity—specifically correcting for specific frontotemporal lobe activation observed in subject S7—and establishes distinct decision boundaries. These findings indicate that synergizing signal reconstruction with Spatio-Temporal metric-guided multi-stream experts effectively mitigates negative transfer, ensuring reliable fatigue detection in complex cross-subject tasks.
Funding Information
  • Xi (23LLRH0044)
  • National Natural Science Foundation of China (62174134)
  • Xi’

This publication has 54 references indexed in Scilit: