Recognising drivers? mental fatigue based on EEG multi-dimensional feature selection and fusion

Detecting the mental state of a driver using electroencephalography (EEG) signals can reduce the probability of traffic accidents. However, EEG signals are unstable and nonlinear and fatigue detection based on one-dimensional features may provide insufficient information, resulting in low recognition efficiency. To resolve these challenges, this paper proposes an EEG-based multi-dimensional feature selection and fusion method to recognise mental fatigue in drivers. First, the EEG signals were decomposed into alpha, beta and theta bands, and then the corresponding time domain, frequency domain and nonlinear features were generated respectively. Furthermore, a three-layer feature-selection method based on Logistic Regression, one-way Analysis of Variance and Recursive Feature Elimination (logistic-ARFE) was proposed to solve the feature redundancy. Logistic-ARFE is designed to automatically select the optimal subset of mental fatigue features. Principal component analysis was used to fuse the optimal feature subset from different dimensions to obtain the fusion feature at a cumulative contribution ratio of 90%, which was used as the final feature to express the recognition accuracy of eight conventional machine learning models. A publicly available EEG dataset for driver fatigue was used to validate the proposed method. The final results show that six of the eight models achieve high recognition accuracy, which indicates that the Logistic-ARFE feature selection algorithm has applicability widely. In particular, compared with other studies using the same dataset, the Gaussian SVM proposed in this study based on time-frequency-nonlinear fusion features achieves the highest recognition accuracy, which is improved by 6.32% and 6.11% respectively.