Abstract:
In automated human-machine collaboration systems, operators play an increasingly important role in making decisions and strategies. Recently, measures based on physiological signals, such as electroencephalogram (EEG) and electrocardiogram (ECG), have received increasing attention due to their ability to objectively deduce the inherent mental states and sensitively monitor mental workload level within continuous time interval. However, obvious individual differences affect the transferability of trained model to a new subject. In order to solve the problem of poor generalization performance of recognition model in cross subject scenarios, a mental workload recognition model based on multi-source deep domain adaptation is constructed. In view of the excessive number of source domains containing redundant information and the cost of model training, the source domains with similar data distribution to the target domain are firstly selected by the optimization algorithm based on Maximum Mean Discrepancy. Then, a dynamic adversarial domain adaptive network is introduced to simultaneously adapt the marginal and conditional distribution of source domain and target domain data in the form of countermeasure training. Finally, an ensemble learning strategy is used to vote and integrate the classification results of models trained in different source domains, improving the accuracy and stability of mental workload recognition. The experimental results show that the model has good recognition accuracy and robustness in cross subject mental workload recognition task in the WAUC dataset and can be further developed as the operator's mental workload identification system to reasonably allocate work tasks and improve work performance.