BDANet: A Binary-Dimensional Aware Network with Multi-Wise Attention for Cognitive Workload Recognition

Yang Shao, Yueying Zhou, Daoqiang Zhang

Abstract: Cognitive workload recognition (CWR) enables operators to achieve optimal performance with limited resources. EEG signals are particularly favored for this purpose due to their affordability and direct correlation with neural activity. However, much existing researches primarily harness single-dimensional information from EEG signals. Techniques such as local characterization through Convolutional Neural Networks (CNNs) and global analysis via Transformers often neglect the integration of both global and local dimensions. And they also overlook the significance of diverse features. To address these limitations, this paper introduces the Binary-dimensional Aware Network with Multi-wise Attention (BDANet), which aims to extract more discriminative features of cognitive workload EEG. BDANet initially employs channel-wise attention to adaptively assess the significance of EEG electrodes. Then binary branches that utilize bidirectional LSTM (BiLSTM) and CNN as core modules are incorporated to extract global and local features, respectively. Distinct attention mechanisms are tailored for each branch to further enhance performance. Experimental results of classifying underload, normal, and high overload of cognitive workloads using the brain-computer interface competition dataset reveal that BDANet achieves an average accuracy of 91.41%, with a peak accuracy of 98.89% at the subject level. These results significantly surpass those of other leading methods. Our code is available at https://github.com/prestyan/BDANet.