An EEG-based Workload Recognition Method for Civil Aviation Student Pilots

The workload of civil aviation student pilots directly impacts flight safety. To address the limitations of existing electroencephalogram (EEG) based pilot workload recognition methods, such as poor model generalization and insufficient utilization of cross-band and spatial features. This study investigates and develops an EEG-based approach for workload recognition in civil aviation student pilots. An integrated subjective-objective evaluation framework is established. EEG signals and NASA Task Load Index (NASA-TLX) data are collected from student pilots under different task scenarios in a simulated flight environment to concurrently acquire both objective physiological measurements and subjective workload assessments. To overcome the limitation of traditional studies that isolate individual frequency bands and neglect inter-band interactions, an independent samples t-test is applied to identify EEG features with significant differences (P < 0.05). Furthermore, by incorporating whole-brain power spectral density activation maps, the neural response mechanisms, and spatial distribution patterns of the θ, δ, α, and β bands, as well as cross-band power ratios, are analyzed under varying workload levels. Third, the extracted EEG features from the full frequency band and each sub-frequency band are used for model training, and a hybrid model based on convolutional neural network (CNN) and long short-term memory (LSTM) for workload recognition is developed to achieve accurate recognition of workload states. Experimental results showed that the selected features could distinguish the neural regulation modes under different workloads. At high workload, the spectral power of the α, θ, and β bands increased in civil aviation student pilots, while the spectral power of the δ band decreased. Specifically, the θ rhythm facilitated the priority allocation of resources through a frontal-parietal and right temporal circuit, while α rhythms exhibited enhanced interference suppression along the left temporal-parietal-prefrontal pathway. The constructed model successfully captured both the spatial and temporal dynamics of EEG signals. Moreover, the hybrid model achieved a test accuracy of 94.5%, outperforming traditional single models such as CNN, LSTM, and Transformer. Notably, using α-band features alone yielded a test accuracy of 95.5%, confirming the efficacy of the proposed method in identifying pilot workload states.