A Multi-stream convolutional neural network for fatigue detection using EEG signals

Background and Aim : Driver fatigue is becoming a serious problem resulting in an increase in the number of accidents. Therefore, it is vital to detect driver fatigue. The most common and effective method among all objective and subjective ones is to identify the fatigue status and operation of the brain based on EEG signals. Recently, the availability of large EEG datasets and advances in machine learning have both led to the development of deep learning architectures, particularly in analyzing EEG signals and understanding the information they may have for brain function. Automated and robust classification of these signals is an important step in making EEG application more practical.Methods : The signal is recorded through a 32-channels electrode cap with 30 effective channels and two reference channels. The electrode locations are based on the international 10-20 system. This recording is done through a wireless device in a driving simulator from 12 healthy people into two states of fatigue and alert (https://figshare.com/articles/The_original_EEG_data_for_driver_fatigue_detection/5202739). The EEG signals in both conditions are sectioned to 1-second time-slices, and all fatigue detection process is evaluated on these time-slices. The main step of signal pre-processing is done through fully online and automated artifact removal for brain-computer interfacing (FORCe) Toolbox to remove eye-blink noise and also all other noises such as baselines. Then, the power spectral density feature of each channel of the signal is estimated by 4-order Autoregression (AR). Given the brain activity in the 1 to 40 Hz frequency band, this signal range is considered for further investigation. The human brain has five main lobes, including Frontal, Occipital, Central, Parietal, and Temporal. Each of these lobes contains specific information; Hence, the 5-stream middle fusion convolutional network is designed with five inputs that each input is extracted features from one of the lobes. Each input goes through three one-dimensional convolutional blocks. Then, the outputs of the convolutional layers are concatenated. Finally, the concatenated output is vectorized and classified using two fully connected layers. Meanwhile, a single stream deep convolutional network is designed to examine brain areas to compare the effect of each lobe on fatigue detection. The network consists of three convolutional layers and two fully connected layers, where the extracted features of the electrodes of each region are applied to the network, and the accuracy is calculated in that area. 80% of data has been selected randomly for training, and the evaluation is performed on the rest of the data. Results : The average accuracy, sensitivity, and specificity were 93.49%, 95.90%, and 90.92%, respectively. The frontal area with accuracy, sensitivity, and specificity of 81.88%, 80.04%, and 83.61%, respectively, had the most impact on diagnosis.Conclusion : This paper presents a new method for fatigue detection by EEG signals using a deep neural network. For comparison, the extracted features were classified with different proposed classifiers, and the results show that the accuracy and robustness of the proposed new system are higher than the machine learning approaches such as Support Vector Machine and K-Nearest-Neighborhood.

کلیدواژه ها:

Fatigue detection ، EEG signals ، CNN network ، Deep learning