Research on driving fatigue detection based on multi-scale wavelet packet-inspired convolutional network under small...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Driving fatigue poses a significant safety risk.
• Accurate detection of fatigue using electroencephalogram (EEG) signals is crucial for driver safety.
• Existing methods may struggle with noisy EEG data and complex feature extraction.

Purpose of the Study:

• To propose a novel deep learning network, MWPICNet, for enhanced driving fatigue detection using EEG signals.
• To extract robust time-frequency features indicative of fatigue.
• To improve the accuracy and reliability of fatigue detection systems.

Main Methods:

• Development of a Multi-scale Wavelet Packet Inspired Convolutional Network (MWPICNet).
• Alternating multi-scale wavelet packet convolution and soft threshold activation layers for feature extraction.
• Incorporation of frequency band weighting and global power pooling for enhanced feature recognition.
• Validation on SEED-VIG and SADT datasets.

Main Results:

• MWPICNet achieved high detection accuracies of 98.67% (SEED-VIG) and 99.39% (SADT).
• The network maintained over 95% accuracy under significant noise interference.
• MWPICNet demonstrated superior performance compared to traditional CNN and GCN methods, improving accuracy by 5.99% and 4.79%, respectively.

Conclusions:

• MWPICNet offers a powerful and robust solution for EEG-based driving fatigue detection.
• The proposed network effectively extracts salient time-frequency features for fatigue assessment.
• This approach holds significant potential for real-time driver monitoring systems to enhance road safety.