BACNet: A multi-attention network for cross-subject and cross-task EEG-based pilot operational intent recognition

Recognizing pilot operational intent is crucial for enhancing flight safety and improving the efficiency of human-machine interaction. Electroencephalography (EEG), known for its high temporal resolution and non-invasive acquisition, has become a prominent modality for this task. However, current approaches often suffer from high model complexity and limited accuracy in EEG feature extraction. This study aims to address these limitations by proposing efficient and accurate neural network architecture for pilot intent recognition based on EEG signals.

We introduce a novel framework, the Balanced Attention Convolutional Network (BACNet), designed to enhance EEG-based intent recognition through collaborative optimization in both channel and spatial dimensions. BACNet features: (1) a three-branch parallel convolutional structure that extracts multi-scale time-frequency features; and (2) dynamic feature modulation mechanisms to adaptively highlight salient channels and spatial locations. EEG data were collected from 15 participants across various simulated flight phases, forming a labeled dataset for model training and evaluation. Five-fold cross-validation was conducted to ensure the robustness of the performance assessment.

BACNet achieved an average classification accuracy of 96.07 % in a three-class EEG-based intent recognition task, outperforming five state-of-the-art baseline methods. The model also demonstrated a significant reduction in computational complexity. Ablation experiments validated the individual and combined contributions of the multi-scale attention modules, highlighting the effectiveness of the collaborative attention design.

With its lightweight architecture and high accuracy, BACNet not only provides a novel solution for pilot operational intent recognition but also demonstrates broad applicability in brain-computer interface (BCI) systems.