ASFT-Transformer: A Fast and Accurate Framework for EEG-Based Pilot Fatigue Recognition

Area of Science:

• Neuroscience and Aviation Safety
• Biomedical Signal Processing
• Machine Learning in Aviation

Background:

• Objective evaluation of pilot fatigue is critical for aviation safety.
• Electroencephalography (EEG) shows promise for fatigue detection but faces challenges with raw signal analysis (data volume, training time, overfitting).
• Existing feature-based methods lack efficient feature and channel selection, leading to redundancy and reduced accuracy.

Purpose of the Study:

• To propose ASFT-Transformer, a framework for fast and accurate pilot fatigue detection.
• To address limitations of direct deep learning on raw EEG and inefficient feature-based methods.
• To provide an objective tool for managing pilot fatigue and improving flight safety.

Main Methods:

• Extraction of time-domain and frequency-domain features from four EEG bands.
• Implementation of a feature and channel selection strategy using ANOVA-SVM to identify relevant features and EEG channels.
• Classification using the FT-Transformer model on selected features, reframing fatigue detection as a tabular data classification problem.

Main Results:

• ASFT-Transformer achieved high average accuracies: 97.24% (cross-clip) and 87.72% (cross-subject).
• The feature and channel selection strategy improved accuracy by 2.45% and 8.07% respectively.
• Training time was drastically reduced from over 1 hour to under 10 minutes.

Conclusions:

• ASFT-Transformer offers a significant improvement over existing machine learning and deep learning models for pilot fatigue detection.
• The proposed method enhances recognition efficiency and accuracy while substantially reducing computational time.
• This framework provides a valuable tool for aviation authorities and operators to objectively manage pilot fatigue and bolster flight safety.