Effectiveness of Adaptive Attention-Based Network for Situation Awareness Recognition

Situation awareness (SA) is directly related to the operating level of dynamic system operators, and electroencephalography (EEG) is frequently employed as the gold standard for SA recognition. Several deep learning models performed well in SA recognition based on EEG features. However, it has limitations such as a limited size of datasets, restricted model interpretability, and low capability of extracting beneficial features. In this work, an adaptive spatial-channel attention mechanism (ASCAM) was introduced in the architectures of a convolutional neural network (CNN). Specifically, ASCAM allows the layers of CNN architectures to fuse various sizes of received information and selectively focus on effective interpretable features. Regarding the problem of the limited size of datasets, combining frequency noise with multivariate variational mode decomposition (MVMD) enhances the generalization capability of models. Experiment results showed that EEGNet embedded in the framework exhibited a relative improvement of 6.02% over the baseline method. The ASCAM contributes to feature extraction and significantly enhances considerable performance. Ablation studies were further implemented to confirm the efficacy of the proposed ASCAM and the MVMD-based data augmentation. Interpretation results indicated that neural network models with embedded attention mechanisms have discovered neurobiological mechanisms related to SA loss. Meanwhile, the proposed lightweight framework is plug-and-play, which can be embedded into any CNN architecture and utilized for various EEG decoding tasks.