Download PDFOpen PDF in browserIEconformer: a Robust Convolutional Transformer for EEG-based Fatigue Driving DetectionEasyChair Preprint 136234 pages•Date: June 11, 2024AbstractFatigue driving detection technology plays a pivotal role in ensuring road safety, and electroencephalography (EEG) signals can be employed as an objective measure of driver fatigue in intelligent vehicles. However, current EEG-based fatigue driving detection methods encounter certain limitations. Firstly, the restricted receptive field of convolutional neural networks struggles to effectively handle the non-stationary nature of fatigue EEG signals for feature extraction. Secondly, real-world training data often suffers from noisy labels, leading to model overfitting on mislabeled data and consequent degradation in the fatigue detection performance. In this paper, we propose the IEconformer ensemble, a robust EEG-based fatigue driving detection model. The IEconformer architecture integrates multi-scale convolutional layers for local feature extraction and the multi-head attention mechanism to capture global feature correlations. To tackle the challenge of noisy data during training, we introduce the co-teaching plus mechanism into our training scheme. This mechanism facilitates cross-updating each IEconformer using disagreement data that yields minimal loss on the respective IEconformer. Experimental results demonstrate the superiority of our proposed IEconformer ensemble over baseline models in fatigue detection. Particularly, the IEconformer ensemble demonstrates high performance even in the presence of noisy data during the training stage, underscoring the practicality of our approach in fatigue driving detection applications for intelligent vehicles. Keyphrases: Brain Computer Interface, Convolutional Neural Network, EEG, Noisy label, fatigue driving detection, robust deep learning, self-attention
|