Abstract
Abstract In Industrial 5.0 human-machine collaboration (HMC) scenarios, real-time and accurate assessment of operators’ mental workload is critical to ensuring system safety and operational efficiency. However, due to the pronounced non-stationarity of electroencephalography (EEG) signals and strong noise interference in industrial environments, existing methods often suffer from severe performance degradation in cross-subject transfer tasks where labeled data are scarce. To address this issue, this paper proposes a transferable assessment framework that integrates dynamic graph topology learning with cross-modal attention. Specifically, a cross-modal attention-based complementary module leverages eye-tracking data to calibrate EEG noise; dynamic graph topology learning captures transient reconfigurations of brain networks; and a masked self-supervised pretraining strategy, combined with distribution alignment constraints, is employed to extract cross-subject invariant features. Experiments on the STEW and MAHNOB-HCI datasets demonstrate that the proposed model achieves accuracies of 90.5% and 89.2%, respectively, outperforming state-of-the-art (SOTA) baselines by 1.2–3.8% in cross-subject and multimodal settings. Moreover, under the − 5 dB condition, the model still attains an accuracy of 82.3%, with a smaller degradation than unimodal baselines by approximately 6–7% points. This study effectively overcomes the limitations of cognitive decoding in uncontrolled environments and provides methodological foundations and technical support for building Industrial 5.0 HMC systems with “perception-adaptation” capabilities.