Fatigue Driving State Detection Based on Spatial Characteristics of EEG Signals

Abstract

Monitoring the driver’s physical and mental state based on wearable EEG acquisition equipment, especially the detection and early warning of fatigue, is a key issue in the research of the brain–computer interface in human–machine intelligent fusion driving. Comparing and analyzing the waking (alert) state and fatigue state by simulating EEG data during simulated driving, this paper proposes a brain functional network construction method based on a phase locking value (PLV) and phase lag index (PLI), studies the relationship between brain regions, and quantitatively analyzes the network structure. The characteristic parameters of the brain functional network that have significant differences in fatigue status are screened out and constitute feature vectors, which are then combined with machine learning algorithms to complete classification and identification. The experimental results show that this method can effectively distinguish between alertness and fatigue states. The recognition accuracy rates of 52 subjects are all above 70%, with the highest recognition accuracy reaching 89.5%. Brain network topology analysis showed that the connectivity between brain regions was weakened under a fatigue state, especially under the PLV method, and the phase synchronization relationship between delta and theta frequency bands was significantly weakened. The research results provide a reference for understanding the interdependence of brain regions under fatigue conditions and the development of fatigue driving detection systems.

1. Introduction

According to statistics from the Ministry of Public Security of China, as of the end of September 2023, China’s motor vehicle ownership reached 430 million, with an increase of more than 10 million over the last year, and the trend in growth has been maintained every year. The continuous growth in the number of motor vehicles has brought tremendous pressure to traffic safety. Among the many factors that lead to traffic accidents, fatigue driving is one of the most common causes. Fatigue is defined as a state of decreased physical or mental activity, which usually occurs when people engage in physical or mental activities for a long time, and affects their daily life [1]. Fatigue driving refers to the decline in physical function caused by the driver’s continuous driving for a long time or insufficient rest [2]. The state may make it difficult for the driver to stay alert, resulting in the driver’s lack of concentration, increased reaction time, and weakened judgment and decision-making ability, thereby increasing the risk of traffic accidents. According to statistics, about 20–30% of traffic accidents are caused by fatigue driving, 5% to 15% of accidents result in death, and about 50% of people have experienced fatigue driving [3].

Therefore, it is crucial to monitor the driver’s physical condition through wearable sensing devices and provide accurate warnings and interventions on possible fatigue states in order to improve the reliability of the driving system and ensure driving safety. At present, the detection methods for fatigue driving are roughly divided into subjective detection methods and objective detection methods [4]. Subjective detection methods include filling out questionnaires and subjective evaluation scales [5,6]. In objective detection, the first category is based on analysis of the vehicle’s motion parameters [7], such as lane position and the standard deviation of steering wheel motion [8]. The second category is based on analysis of the driver’s facial expressions and eye tracking data [9,10]. The third category is based on the driver’s physiological signals, including Electrocardiogram (ECG) [11], Electrooculogram (EOG) [12], Electromyography (EMG) [13], and Electroencephalogram (EEG) [14]. Subjective detection methods usually serve as auxiliary means in fatigue driving detection. Among objective detection methods, the methods based on vehicle motion parameters and human external features are easily affected by road conditions and lighting levels. In contrast, detection based on physiological signals can accurately assess the driver’s fatigue and mental state. Among them, EEG signals can directly and objectively reflect the current mental state of the brain due to their high temporal resolution and the wearable and portable nature of the acquisition equipment. It is considered to be the most effective method for fatigue driving identification [14].

In recent years, scholars have conducted a series of tests on the process of fatigue driving based on EEG signal collection and analysis. Zeng et al. [14] used two feature extraction methods, power spectral density (PSD) and empirical mode decomposition (EMD). They decomposed the data into multiple intrinsic mode function (IMF) components through EMD, extracted the features of each IMF component, combined these components at the feature points to form the feature vector of the EEG, and used Support Vector Machine (SVM) and K-Nearest Neighbors (KNN) for classification, with an average accuracy of over 80%. Chinara et al. [15] proposed a sleepiness detection model based on single-channel EEG signals, using wavelet packet transform to extract the time domain features of EEG signals, and achieved a relatively good accuracy. Rashid et al. [16] proposed a new channel-selection algorithm based on a correlation coefficient and an integrated classifier based on the random subspace k-nearest neighbor to improve the classification performance of EEG data in driver fatigue detection. The robustness of the method was verified by extracting features using PSD. Experimental results show that the combination of a correlation coefficient, PSD, and KNN can improve the accuracy of a driver fatigue detection system based on EEG signals. Gao et al. [17] proposed a convolutional recurrent neural network based on a log-Mel spectrum (LogMel-CRNN) for fatigue driving detection. First, EEG signals are subjected to one-dimensional convolution to achieve short-time Fourier transformation, and the log-Mel spectrum is obtained through the Mel filter group. The obtained log-Mel spectrum is then input into the fatigue detection model to complete the fatigue detection task of EEG signals, with an accuracy rate of over 80%. Zhang et al. [18] proposed a multidimensional feature selection and fusion method based on EEG signals to identify the mental fatigue of drivers. The corresponding time domain, frequency domain, and nonlinear features were generated in the corresponding bands. On this basis, a three-layer feature selection method based on logistic regression, one-way analysis of variance, and recursive feature elimination was proposed to solve the problem of feature redundancy, and a good accuracy rate was achieved on SVM. Wang et al. [19] proposed a driving fatigue detection method based on a brain functional network. In the construction of a brain functional network, partial directional coherence was used to calculate the correlation between EEG signal channels. Then, a support vector machine was used as a classifier to extract graph-related features for fatigue detection. The results showed that low-frequency rhythms are easier to identify than high-frequency rhythms. Wang et al. [20] studied the effect of fatigue driving on the reorganization of dynamic functional connections through temporal brain network analysis, and quantitatively compared the alert and fatigue states in a simulated driving experiment. The study found that the spatiotemporal topological structure of dynamic functional connections was significantly disintegrated, which was manifested as reduced global temporal efficiency and increased local temporal efficiency in the fatigue state. In particular, the frontal lobe and parietal lobe were shown to be the brain areas most susceptible to fatigue driving. Qin et al. [21] proposed a method based on directed brain networks of EEG signals to reveal the impact of driving fatigue on the brain’s information processing ability. Based on the EEG signal current source density obtained by source analysis, a directed brain network of fatigue driving was constructed using a directed transfer function. The study found that the average clustering coefficient and average path length gradually increased with the increase in driving time. Especially in the theta band, there were obvious differences in the networks under awake and fatigue driving states. Zhou et al. [22] proposed a driving fatigue detection method based on a functional brain network and studied the optimal combination of features and algorithms. First, the correlation between each EEG signal channel was calculated using the Pearson correlation coefficient to construct a functional brain network. Then, the functional brain network features were extracted and combined, and different machine learning algorithms were used as classifiers for fatigue detection. Finally, the optimal combination of features and algorithms was obtained.

At present, the feature extraction of fatigue driving occurs mostly in the time domain, frequency domain, and time–frequency domain. There are few descriptions of the spatial domain activity characteristics of the driver’s brain activity in the fatigue state during driving. More in-depth discussions are needed on the spatial characteristics of brain activities corresponding to the driving process. The mining and learning of spatial domain features can provide new ideas and methods for representing and identifying fatigue driving characteristics. In view of this, this paper proposes a fatigue driving detection method based on the spatial characteristics of EEG signals. Firstly, based on the PLV and PLI methods, the functional brain networks of the five frequency bands of EEG signals are constructed and the characteristics are analyzed, respectively, to reveal the spatial network structure characteristics corresponding to the driving process. Then, through the statistical analysis of the network feature parameters, the feature parameters sensitive to fatigue and alertness are selected to form feature vectors. Finally, the machine learning algorithm is used to classify and identify the above results. The results of this study demonstrate that the PLV method is an efficient and successful EEG signal feature extraction technique. It also shows that EEG signals can be a promising tool for fatigue driving detection in actual driving tasks and have great potential in future research. The roadmap of this study is shown in Figure 1.

2. Materials and Methods

2.1. Description of Dataset

This paper uses the public dataset of simulated driving tasks released by National Chiao Tung University [23], which includes 52 simulated driving experiments and EEG data collection. The subjects were aged between 22 and 28 years old and were required to participate in a continuous driving task at a speed of 100 km/h. The experiment simulated a driving situation on a four-lane highway at night, requiring the subjects to keep the vehicle in the center of the lane. Lane deviation events were randomly introduced during driving, which induced the vehicle to deviate from the original driving lane to both sides. When the deviation occurs, the subject needs to immediately correct the vehicle trajectory by adjusting the steering wheel, which is the start of the reaction, and drive the vehicle back to the original driving lane, which is the end of the reaction, thus completing a trial experiment. The interval between two experiments is 5–10 s. While the subjects were completing the above driving tasks, their EEG signals were collected synchronously. During driving, EEG signals were collected using 32-channel international standard Ag/AgCl EEG electrodes, including 30 EEG signal electrodes and 2 reference electrodes. The contact impedance between all electrodes and the skin was kept below 5 kΩ, and the online sampling rate of the EEG signals was 500 Hz. The experimental scene and paradigm are shown in Figure 2.

2.2. Data Preprocessing

The EEGLAB toolbox was used to preprocess the EEG data. Based on the conclusions in the literature [24,25], the first 10 min after the start of the experiment were defined as the alert driving state, and the last 10 min of the experiment were defined as the fatigue driving state. There are four types of events in the experiment, deviation onset (left), deviation onset (right), response start, and response end. In this study, only events with the response start type are considered. Referring to the conclusions of the existing literature [26,27], the 2 s before the response start and the 4 s after the response start are taken as one epoch (6 s). The 52 experiments were divided into two groups according to the experimental duration. The first group had an experimental time of 60 min, and the second group had an experimental time of 90 min. This was to facilitate a comparative analysis of the experimental results. According to the results in the literature [28], the driving process of the subjects in both groups of experiments reached the level of fatigue. Since the number of epochs in each experiment was different, the minimum number of epochs in all experiments was 33. In order to facilitate unified processing, 33 was selected as the number of epochs corresponding to alert driving and fatigue driving for each subject. In the end, a total of 3432 epochs were obtained in the two groups of experiments, which were used to complete the subsequent spatial network structure feature analysis. Finally, the EEG signal was filtered into five frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–50 Hz).

2.3. Functional Brain Network Construction

The brain is the human body’s advanced nerve center and the biological basis of human intelligence and complex thinking. Since driving behavior is a comprehensive and complex cognitive process, which includes decision-making ability, execution ability, spatial perception, and visual information processing, it requires the coordination of each brain region to complete. Therefore, measuring the brain activity state and fatigue level during driving from the perspective of the whole brain can provide new specific indicators for fatigue driving detection and identification. This paper will use the phase locking value and phase lag index algorithms to construct a functional brain network.

Phase synchronization analysis is often used to evaluate the interdependence between different brain regions. Compared with analysis methods based on amplitude correlation, it is less susceptible to volume conduction effects and provides a richer representation of interdependence. Lachaux et al. [29] proposed a method that can directly quantify the transient phase locking between two neural electrical signals. The phase locking value (PLV) is calculated using the relative phase difference to characterize the phase synchronization degree and information transmission process between the two signals. It is very suitable for measuring the synchronization between EEG signals of each channel. Phase-based methods have achieved good results in other areas of EEG signals. Wang et al. [30] proposed a multi-channel EEG emotion recognition method based on a PLV and graph convolutional neural network. The multi-channel EEG features were modeled as graph signals using a brain network built based on a PLV, and EEG emotion classification was performed on this basis. Experiments were conducted on the emotional EEG datasets SEED and DEAP datasets, and the accuracy of both datasets exceeded 70%. Cui et al. [31] calculated the PLV of EEG signals under different emotional states and combined the CNN model to classify the emotional states in five frequency bands. The results showed that the model can recognize EEG emotions with an accuracy rate of 85%, proving the effectiveness of the phase-based method. Xi et al. [32] designed and conducted a series of memory experiments with different memory loads or target forms, while collecting EEG signals. After slicing the EEG signals, the PLV of the Gamma frequency band was calculated, and the brain functional network was constructed after binarization. The extracted network characteristics were input into the machine learning model for classification, and good performance was achieved.

The following are used the steps to calculate PLV:

First, the instantaneous phase of the signal is calculated using the Hilbert transform; ${\theta }_a\left(t\right)$ represents the instantaneous phase locking value at time point a:

$${\theta }_a\left(t\right)=\arctan {\displaystyle \frac{\tilde{x}(t)}{x(t)}}$$

(1)

$\tilde{x}(t)$ is the Hilbert transform of $x(t)$, and the transformation formula is as follows:

$$H[x(t)]=x(t)\ast {\displaystyle \frac{1}{\pi t}}={\displaystyle \frac{1}{\pi }}{\displaystyle {\int }_{-\infty }^{+\infty }\frac{x(\tau )}{t-\tau }}d\tau =\tilde{x}(t)$$

(2)

Then, calculate the PLV:

$$\mathrm{PLV}={\displaystyle \frac{1}{N}}\left|{\displaystyle \sum _{t=1}^N{e}^{j\left[{\theta }_a\left(t\right)-{\theta }_b\left(t\right)\right]}}\right|$$

(3)

where ${\theta }_a\left(t\right)-{\theta }_b\left(t\right)$ represents the phase difference between two time points; n represents the total number of sampling points of the EEG signals. In this study, since the epoch length is 6 s and the sampling rate is 500 Hz, n = 500 × 6 = 3000. The PLV ranges from 0 to 1. The larger the value, the stronger the correlation between the two channels.

The Phase Lag Index (PLI) is used to measure the asymmetry of the phase difference distribution between two signals. The value of the PLI is also between 0 and 1. A value of 0 means no coupling or coupling with a phase difference centered on 0 mod $\pi$. The closer to 1, the stronger the phase synchronization. The following are the steps to calculate the PLI:

$$\mathrm{PLI}=\left|〈sign\left[\Delta \phi \left(t\right)\right]〉\right|={\displaystyle \frac{1}{N}}\left|{\displaystyle \sum _{t=1}^{N}sign\left[\Delta \phi \left(t\right)\right]}\right|$$

(4)

$\Delta \phi \left(t\right)$ represents the phase difference between two time points; n is also 3000.

For the 30-channel EEG data, using all epoch data segments corresponding to the fatigue state and alert state, construct the corresponding functional brain network based on the PLV and PLI for the five sub-bands of EEG signals, and expresses it in the form of an adjacency matrix (30 × 30), and complete the calculation of the PLV matrix and PLI matrix for all subjects in turn. Figure 3 shows the electrode position diagram of 30-channel EEG signals.

The electrodes corresponding to rows 1–30 of the adjacency matrix are FP1, FP2, F7, F3, FZ, F4, F8, FT7, FC3, FCZ, FC4, FT8, T3, C3, CZ, C4, T4, TP7, CP3, CPZ, CP4, TP8, T5, P3, PZ, P4, T6, O1, OZ, and O2. Among them, FP1, FP2, F7, F3, FZ, F4, and F8 are located in the frontal lobe; FC3, FCZ, FC4, C3, CZ, C4, CP3, CPZ, CP4, P3, PZ, and P4 are located in the parietal lobe; FT7, FT8, T3, T4, TP7, TP8, T5, and T6 are located in the temporal lobe; and O1, OZ, and O2 are located in the occipital lobe.

2.4. Network Characteristic Measures

Functional brain topology networks can intuitively characterize the interdependence between different brain regions, but the quantitative analysis of network structure needs to be carried out through the statistical description of the characteristic parameters of the network and their relationships, which helps to describe the changing process of the interdependence between brain regions. There are four common characteristic parameters in brain networks:

(1)

Degree (D): indicates the importance of the node in the network.

(2)

Betweenness Centrality (BC): Reflects the importance of a node. The calculation formula is as follows:

$$BC\left(i\right)={\displaystyle \sum _{i\ne s\ne t}\frac{p_{st}^i}{p_{st}}}$$

(5)

where $p_{st}$ represents the number of all shortest paths from node s to t, and $p_{st}^i$ represents the number of shortest paths passing through node i among all the shortest paths from node s to t.

(3)

Local Efficiency (LE): Represents the efficiency of information transmission between each node and its neighbors in the network, reflecting the connection strength and efficiency of the node in its local neighborhood. The calculation formula is as follows:

$$LE={\displaystyle \frac{1}{N_{G_i}(N_{G_i}-1)}}{\displaystyle \sum _{i\ne j\in G_i}\frac{1}{d_{ij}}}$$

(6)

where $G_i$ represents the graph consisting of node i and its neighboring nodes; $d_{ij}$ represents the distance between node i and node j.

(4)

Clustering Coefficient (CC): Measures the clustering of nodes in the network, which measures the degree of interconnection between a node and its neighboring nodes. The clustering coefficient calculation formula of nodes in the network is as follows:

$$C_i={\displaystyle \frac{E_i}{\left[k_i(k_i-1)\right]/2}}={\displaystyle \frac{2E_i}{k_i(k_i-1)}}$$

(7)

where $k_i$ represents the number of adjacent nodes of node i; $E_i$ represents the actual number of edges between node i and its $k_i$ neighbor nodes.

The clustering coefficient of the network can be expressed as

$$C_N={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N{C}_i}$$

(8)

where n represents the number of nodes in the network.

3. Experiments and Results

3.1. PLV Average Matrix and PLI Average Matrix

The PLV matrix and PLI matrix constructed in each frequency band under the alert state and fatigue state in the two groups of experiments were averaged. Figure 4a,b and Figure 5a,b show the averaged PLV matrix and PLI matrix of the first and second groups of experiments, respectively. The PLV intensity and PLI intensity of the EEG channel are represented by color bars ranging from blue (close to 0) to yellow (close to 1).

It can be seen intuitively from the figure that in each group of experiments, there are certain differences in each sub-band under the alert driving state and the fatigue driving state. It is more obvious in the PLV method, especially in the connection strength of the delta and theta bands. The value of PLV in the alert state is significantly higher than that in the fatigue state. This shows that in the alert state, the synchronization between various brain regions is stronger and information transmission is more effective. In the fatigue state, this synchronization is weakened, which may lead to a decline in brain function and a decrease in reaction speed. The difference in connection strength in the gamma band is not as obvious as in other bands, which indicates that the brain may work relatively stably in the high-frequency band in alert and fatigue states, and is not as susceptible to state changes as the low-frequency band. Because it is the result after the overall average, this intuitive difference shows obvious specificity in the network structure.

3.2. Threshold Selection

In the process of network construction, as shown in Figure 3, the electrode positions of the 30 channels were defined as nodes, and the PLV and PLI values between different nodes were defined as the connecting edges of the network, so as to obtain the corresponding functional brain network topology. In the process of using PLV and PLI to measure the phase dependence between two electrodes, due to the presence of noise in the signal, there are often certain weak connection edges, which cannot represent the true connection relationship between brain regions. Therefore, it is necessary to select a threshold for screening and then eliminate the weak connection edges.

Shown in Figure 6a,b are the probability distribution diagrams of the PLV and PLI values of all epochs of the five sub-bands corresponding to the two groups of experiments. This paper selects the upper quartile in the statistical distribution diagram as the threshold of the corresponding network. As can be seen from the figure, the thresholds of each band in the two groups of PLV experiments are 0.47, 0.48, 0.45, 0.36, and 0.32 and 0.46, 0.47, 0.46, 0.36, and 0.32, respectively, and the thresholds of each band in the two groups of PLI experiments are 0.16, 0.07, 0.04, 0.05, and 0.03 and 0.09, 0.07, 0.04, 0.05, and 0.03, respectively.

3.3. Network Binarization

Based on the threshold values determined above for each state, the weights in the PLV average matrix and the PLI average matrix that are greater than the threshold are set to 1 (yellow), and the weights that are less than the threshold are set to 0 (red), thereby obtaining the binarization matrix corresponding to each state. Shown in Figure 7a,b are the PLV binarization matrices of the two groups of experiments, and shown in Figure 8a,b are the PLI binarization matrices of the two groups of experiments.

As can be seen from the figure, there are obvious differences between each sub-band in the alert state and the fatigue state in the two groups of experiments, especially in the delta and theta bands. Because the delta and theta bands often correspond to sleepiness and fatigue, the alpha band corresponds to the resting state, the beta band is related to the state of concentration, and the gamma band often corresponds to a high cognitive state. After a long period of driving, drivers will experience some fatigue, which is why the delta and theta bands are more different in the two states.

3.4. Network Topology Analysis

The functional brain network was constructed for the binarized matrix on each sub-band of the two groups of experiments. The functional brain network topological connection diagrams based on the PLV and PLI of the two groups of experiments are shown in Figure 9a,b and Figure 10a,b respectively.

The frontal lobe is the primary area for executive function and decision-making, and its connections with other brain regions are used to perform cognitive tasks and regulate behavior. The parietal lobe is related to functions such as spatial perception and body sensation. The temporal lobe is associated with functions such as hearing, language comprehension, and memory. The occipital lobe is primarily responsible for visual processing, and therefore has connections with other areas for the reception and interpretation of visual information.

As shown in the figures, in the alert state, there is uniform connection between major brain regions such as the frontal lobe, parietal lobe, temporal lobe, and occipital lobe, which reflects the brain’s need for functional coordination and information transmission in the alert state. The even distribution of these connections suggests that when in the alert state, the brain is in a highly active and efficient working state, with fast and effective communication and coordination between different functional areas. However, in the fatigue state, the connections between brain regions are weakened and the brain structure is relatively sparse, which is particularly evident in the brain network constructed by the PLV method. This reflects the reduction in brain functional activity in a state of fatigue. Fatigue may lead to a decline in cognitive and executive functions, affecting aspects such as attention, memory, and decision-making.

3.5. Results of Statistical Analysis

This paper calculated and discussed the above four brain network characteristic parameters for a total of 3432 epoch data segments of EEG signals in each band under the two experimental states obtained by preprocessing, and carried out paired sample t-tests on the above characteristic parameters in the alert state and fatigue state. The results are shown in

Table 1. (a): Differences in the characteristic parameters of the first group of experiments. (b): Differences in the characteristic parameters of the second group of experiments.

(a)
	PLV				PLI
	BC	CC	D	LE	BC	CC	D	LE
delta	/	*	*	/	/	*	*	/
theta	*	*	*	*	*	*	*	*
alpha	*	*	*	*	*	*	*	*
beta	*	*	*	*	*	*	*	*
gamma	/	/	*	/	/	/	*	/
(b)
	PLV				PLI
	BC	CC	D	LE	BC	CC	D	LE
delta	/	*	*	/	/	*	*	/
theta	*	*	*	*	*	*	*	*
alpha	*	*	*	*	*	*	*	*
beta	*	*	*	*	*	*	*	*
gamma	/	/	*	/	/	/	*	/

Notice: / indicates no significant difference; * indicates p < 0.05.

.

In the two groups of experiments, in the delta band, there is no significant difference between BC and LE, but there is a significant difference between CC and D. In the gamma band, except for D, there is no significant difference in the other three network parameters. There are significant differences in the network characteristic parameters in the theta, alpha, and beta bands. Finally, the characteristic parameters with significant differences were extracted to form the final feature matrix for subsequent machine learning to complete classification and recognition.

3.6. Analysis of the Results for Classification

For the above two groups of experiments and their corresponding five sub-band data, the brain network characteristic parameter indicators with significant differences were screened out, and the feature matrix was composed of the above four network characteristic parameters and the mixed features (MIX) composed of these four characteristic parameters. The dimension of each network characteristic matrix is 1716 × 30, where 1716 represents the network feature parameters calculated from 33 epochs of 26 subjects in each group under alert and fatigue states, and 30 represents the number of channels. The matrix was sent to the SVM, Naive Bayes (NB), and linear discriminant analysis (LDA) for classification and identification, so as to complete the identification of alert states and fatigue states. In Figure 11a,b, the average accuracy of the network indicators with significant differences under the two methods in the three classification algorithms after ten-fold cross validation is shown.

As can be seen from Figure 11a, among the network characteristic parameters of the brain network constructed based on the PLV method, when SVM is used to classify the mixed features of the beta band of the second group of experiments, the recognition accuracy can reach up to 89.5%. When LDA is used to classify the betweenness centrality features of the alpha band of the first group of experiments, the accuracy is the lowest, but it can also reach 75.4%. From Figure 11b, among the network characteristic parameters of the brain network constructed based on the PLI method, when SVM is used to classify the mixed features of the beta band of the second group of experiments, the recognition accuracy can reach up to 82.5%. When the LDA is used to classify the betweenness centrality features of the alpha band of the first group of experiments, the accuracy was the lowest, but it could still reach 70.5%. The accuracy rates of different network indicators and different classification algorithms are all above 70%, which shows that good results have been achieved. This also shows that the method proposed in this paper is effective. In addition, we adopt two other evaluation criteria, precision and recall, to evaluate the classification performance, as shown in

Table 2. (a): Classification performance evaluation based on PLV. (b): Classification performance evaluation based on PLI.

(a)
		The First Group of Experiments					The Second Group of Experiments
		BC	CC	D	LE	MIX	BC	CC	D	LE	MIX
Precision (%)	theta	77.91	80.45	80.68	82.11	84.25	84.25	83.49	85.82	85.03	86.89
	alpha	75.63	78.52	78.30	80.89	83.08	82.67	82.96	84.41	84.28	84.66
	beta	78.26	81.78	81.96	82.15	85.02	86.24	84.52	87.29	84.93	88.12
Recall (%)	theta	80.85	84.35	84.36	85.14	88.06	87.81	87.42	89.33	88.15	90.75
	alpha	77.62	81.28	81.43	84.87	86.16	86.22	86.28	88.02	87.83	87.92
	beta	81.15	85.22	85.61	85.66	88.72	89.32	87.89	90.25	88.46	91.78
(b)
		The First Group of Experiments					The Second Group of Experiments
		BC	CC	D	LE	MIX	BC	CC	D	LE	MIX
Precision (%)	theta	72.85	73.55	76.28	74.20	78.49	78.31	77.11	78.58	74.72	80.08
	alpha	72.16	72.29	75.26	71.79	75.82	75.43	75.42	77.64	74.25	78.86
	beta	73.89	73.78	77.58	75.36	78.96	76.85	77.36	80.06	75.92	80.73
Recall (%)	theta	76.68	76.21	80.45	77.96	82.42	82.06	80.26	82.51	78.68	84.06
	alpha	75.14	75.29	78.96	75.63	79.94	78.67	79.71	81.63	78.22	82.46
	beta	77.43	76.62	81.54	79.04	82.85	80.24	81.33	83.74	79.83	84.89

.

4. Discussion

Vehicle driving requires high coordination and control of the brain, and cooperation among various brain regions is required to complete it. As shown in Figure 9a,b and Figure 10a,b, the functional brain network topological relationships corresponding to alertness and fatigue states are constructed in this paper based on the PLV method and PLI method, respectively. The intuitive analysis shows that, compared with the initial alert state, the interdependence between brain regions in the fatigue state of the driver after 60 min of driving is gradually weakening from the perspective of phase synchronization, in terms of both the low-frequency components of the EEG signal or the corresponding high-frequency components. Compared with the fatigue state after 90 min of driving, the weakening of the delta and theta bands appears to be faster, while the changes in alpha, beta, and gamma bands are not so obvious. A comparison of alert and fatigue states revealed that the changes in low-frequency delta, theta, and alpha bands between fatigue and alert were more obvious, which is consistent with the view that delta and theta waves often appear in deep sleep and drowsiness. As fatigue occurs, the driver’s state tends to be drowsy, the delta and theta band activities increase, but the activity of neurons in each brain region tends to be chaotic, and the phase synchronization relationship between signals gradually weakens. By comparing the PLV method and the PLI method, it is found that the above changes are more obvious in the PLV method.

Comparative analysis found that the recognition effect of mixed-feature parameters is better than that of single-network-feature parameters. This shows that the characteristic parameters of spatial functional brain networks can be used to quantitatively analyze the state of fatigue driving, but the selection of network parameters requires the integration of multi-angle information. The specific combination of parameters is expected to further improve the recognition accuracy, which needs further research.

By comparing the results of the three types of machine learning classifiers, it was found that the recognition effect of the SVM was relatively good. In addition, the recognition effect on the beta band was the best among the five sub-bands. The main reason may be that the beta band mainly corresponds to the state of concentration and is a key band in fatigue detection, which is consistent with the previous research results [33]. Similarly, the recognition effect of the second group of experiments is better than that of the first group of experiments, mainly because the driving time of the subjects in the second group of experiments is 30 min longer than that in the first group of experiments, and the fatigue is relatively stronger. The difference in the brain network characteristics of the EEG signals corresponding to the fatigue state is more obvious, which makes it easier to be detected by the machine learning model.

In order to further illustrate the effectiveness of the method proposed in this paper, a comparative analysis was conducted on the research results of this paper and other methods using the same dataset, as shown in

Table 3. Comparison between studies.

Researchers	Research Methods	Average Accuracy (%)
Paulo et al. [28]	Recurrence Diagram/Gramian angular field	75.87/74.53
Li et al. [34]	Spectral spatial model	70.60
Liu et al. [35]	Maximum independence Domain adaptation	73.01
This paper	Spatial characteristics (PLV)	86.47 (SVM)

. Through comparative analysis, it was found that although the method and research ideas proposed in this paper still have room for improvement, on this dataset, the method adopted in this paper achieved relatively ideal results, and the recognition efficiency was better than the existing methods.

Based on the EEG data of simulated driving experiments, this paper completed the characteristic analysis of EEG signals corresponding to alert driving and fatigue driving by constructing brain networks and calculating network characteristic parameters. First, the functional brain network was constructed based on the phase locking value and phase lag index method on the five bands of EEG signals, and the network was quantitatively analyzed based on the characteristic parameters. The characteristic parameters with significant differences were selected to construct feature vectors, and machine learning models such as SVM, NB, and LDA were used for classification and recognition. The recognition accuracy was higher than 70%, and the highest could reach 89.5%. This result proves the effectiveness of this method in fatigue driving recognition.

Through the analysis of the topological structure of brain networks, it was found that there were obvious differences in the functional brain networks between alert and fatigue states. In the alert state, the interaction between brain regions was relatively uniform. This is because the driving process involves decision-making, executive function, auditory function, and visual processing, which require synergy. The connectivity of brain networks in the fatigue state was weakened, and the topological structure was relatively sparse, especially in the delta and theta bands. This is because long-term driving produces fatigue, leading to the depletion of neural resources, which in turn affects the decline of cognitive and executive functions as well as attention, memory, and decision-making. With the advent of the era of intelligent transportation, combined with the development of portable EEG equipment, this study will provide new ideas for the classification and identification of fatigue driving based on EEG signals, and will have a certain reference value for the development of fatigue driving detection systems based on brain–computer interfaces.