A Hybrid Model Utilizing Principal Component Analysis and Artificial Neural Networks for Driving Drowsiness Detection

Abstract

The detection of drowsiness while driving plays a vital role in ensuring road safety. Existing detection methods need to reduce external interference and sensor intrusiveness, and their algorithms must be modified to improve accuracy, stability, and timeliness. In order to realize fast and accurate driving drowsiness detection using physiological data that can be collected non-intrusively, a hybrid model with principal component analysis and artificial neural networks was proposed in this study. Principal component analysis was used to remove the noise and redundant information from the original data, and artificial neural networks were used to classify the processed data. Three other models were designed for comparison, including a hybrid model with principal component analysis and classic machine learning algorithms, a single model with artificial neural networks, and a single model with classic machine learning algorithms. The results indicated that the average accuracy of the proposed model exceeded 97%, the average training time was lower than 0.3 s, and the average standard deviation of the proposed model’s accuracy was 0.7%, indicating that the model could detect driving drowsiness more accurately and quickly than the comparison models while ensuring stability. Thus, principal component analysis can help to improve the accuracy of driving drowsiness detection. This method can be applied to active warning systems (AWS) in intelligent vehicles in the future.

1. Introduction

Numerous drivers have driven when drowsy. In the USA, 41% of American drivers reported having “fallen asleep or nodded off” while driving at least once in their lifetime [1]. In Europe, the statistics vary between countries. For example, in Croatia, 6.1% of drivers admitted that they had driven when tired; in Spain, 70% of drivers reported falling asleep while driving [2].

Driving drowsiness is one of the main threats to road traffic safety [3,4,5]. Traffic accidents caused by drowsy driving account for 15–30% of global traffic accidents. Moreover, drowsy driving accidents have high fatality rates. In France, driver drowsiness caused 85% of fatal road accidents in 2011; in Germany, drowsy drivers caused 25% of fatal road accidents [1].

Drowsy driving is harmful, and it is necessary to accurately and quickly detect drivers’ drowsiness in actual driving to minimize these accidents. A significant concern in drowsiness detection today is the exploration of accurate driving drowsiness detecting methods with low interference to drivers. There are three kinds of data that can detect drowsiness: facial images, driving performance, and physiological signals [5,6]. When using facial images to detect drivers’ drowsy state, the results are affected by lighting conditions. In addition, drivers cannot wear glasses, masks, and other objects that block the face, which is unrealistic for actual driving [7,8]. The accuracy of most driving drowsiness detection methods based on driving behavior is around 75%. With these methods, the accuracy is affected by the external environment, and differences in driving habits significantly reduce the robustness of the model [9]. Physiological signals can directly reflect the driver’s state with a minimal time delay and they are not easily affected by the external environment. For this reason, many studies have applied physiological signals for the detection of drowsiness.

In previous studies, electroencephalogram (EEG) and electrocardiogram (ECG) signals have been the most commonly used physiological signals for drowsiness detection. EEG, which is called the “golden standard” of drowsiness detection, can intuitively and effectively reflect the electrical activity information of the brain; thus, it has a wide range of applications in assessing the alertness of the brain [10,11,12]. ECG can be used to calculate heart rate variability (HRV), which refers to the tiny variations in successive heartbeat intervals. HRV is sensitive to drowsiness and is not easily affected by external environments [13]. However, both EEG and ECG are too intrusive to allow for the collection of data during actual driving [14]. Despite their efficiency, it is challenging to apply EEG and ECG for drowsiness detection in real-world driving scenarios.

Several studies have indicated that electrodermal activity (EDA), respiration (RESP), and photoplethysmography (PPG) signals can also reflect drivers’ drowsiness with small and light devices, which present little interference to drivers [15,16,17,18,19].

Electrodermal activity (EDA) refers to the small changes in the skin’s electrical activity [20,21]. Malathi, et al. [16] found that EDA signals became unstable when the participants became drowsy. Respiration (RESP) signal values show changes in the chest when a human breathes. Zhu et al. [15] found that drowsy drivers had a longer respiration period and a larger respiration amplitude than those who were awake. In addition, awake drivers hardly yawned, while drowsy drivers yawned frequently. However, neither Malathi, et al. [16] nor Zhu, et al. [15] built a model to classify the drowsy state; they only performed a comparative analysis of the signals in those two states. Hence, the effectiveness of EDA and RESP in detecting driving drowsiness remains unknown.

Photoplethysmography (PPG) is one of the most common clinical signals used to measure pulse [22]. L Hyeonjeong, et al. [18] collected PPG signals through a wearable device to detect driving drowsiness. The accuracy was 70%, which easily caused misjudgment. Thus, it was not satisfactory in actual driving scenarios.

Table 1. The scientific gap in existing driving drowsiness detection features.

Data	Features	Defects
Driving behavior	Vehicle data (speed, acceleration, steering wheel angle, lane center offset, etc.)	Low accuracy (75%)
Facial images	Facial features Eye movements	Cannot wear glasses or facial masks Does not match natural driving scenes
Physiological signals	Electroencephalogram (EEG) Electrocardiogram (ECG)	Too intrusive to collect during the actual driving process
	Electrodermal activity (EDA) Respiration (RESP)	Accuracy for driving drowsiness detection remains unknown
	Photoplethysmography(PPG)	Low accuracy when only using single signals to identify driving fatigue.

shows the scientific gap in existing driving drowsiness detection features.

Therefore, researchers have attempted to combine these physiological signals to detect drowsiness and guarantee higher accuracy. For example, Xie [23] collected EDA, RESP, and PPG signals from eight rail drivers. He adopted the k-nearest neighbors method and support vector machines to detect the drivers’ drowsiness. In his study, the highest accuracy reached 85%.

As explained above, the combination of EDA, RESP, and PPG is sufficient to improve accuracy but cannot meet the needs of application in real driving scenarios. This may be due to redundant information in the data or defects in the classification algorithms. Too many variables in the data may contain redundant information, which would increase the complexity of computation. Principal component analysis assists in removing the redundant information that may exist in the data. The present study adopted principal component analysis (PCA), one of the most commonly used dimensionality reduction methods [24]. Detecting driver drowsiness requires adequate models. Some classification methods, such as support vector machine and k-nearest neighbors, have defects that reduce the model’s accuracy or take a long time to classify high-dimensional data, leading to low accuracy and a high time delay. Artificial neural networks can quickly and thoroughly approximate arbitrarily complex linear or nonlinear relationships and have strong robustness and fault tolerance [25]. Considering this, we adopted artificial neural networks to classify the noise-removal data.

This study proposed a hybrid model utilizing principal component analysis and artificial neural networks to detect driving drowsiness quickly and accurately. Physiological data (EDA, RESP, and PPG) were non-intrusively collected by wearable devices and used as the model’s input. Then, we designed three comparison models: a single model with artificial neural networks, a hybrid model utilizing principal component analysis and classic machine learning algorithms, and a single model with classic machine learning algorithms. This research analyzed the impact of principal component analysis on driving drowsiness detection and compared the performance of artificial neural networks and classic machine learning algorithms. Two hypotheses were proposed:

Hypothesis 1.

Removing noise that may exist in the data can improve accuracy and shorten training time.

Hypothesis 2.

The accuracy of artificial neural networks is higher than the comparison models used in this research, which were support vector machine and k-nearest neighbors.

2. Methodology

In this research, the proposed hybrid model was based on principal component analysis and classification algorithms. There were two kinds of classification algorithms: artificial neural networks and classic machine learning algorithms. The artificial neural networks included the backpropagation neural network (BPNN) and the cascade forward neural network (CFNN), while the classic machine learning algorithms included the support vector machine (SVM) and k-nearest neighbors (KNN).

Based on noninvasive physiological data, this paper proposed a hybrid model utilizing principal component analysis and artificial neural networks. Then, we compared it with hybrid models based on principal component analysis and classic machine learning algorithms and other single models. Figure 1 shows the constructions and names of the proposed models and the comparison models. The dark purple thick arrows represent the proposed model. The light purple thin arrows represent comparison model 1, which was the single artificial neural network model. The dark purple dotted arrows represent comparison model 2, which was the hybrid model utilizing principal component analysis and classic machine learning algorithms. The light purple dotted arrows represent comparison model 3, which was the single classic machine learning algorithms model.

2.1. Principal Component Analysis

Principal component analysis (PCA) is a linear feature extraction method that is widely used for linear dimensionality reduction. It uses the variance of each feature to find new features to maximize the separability of categories for dimensionality reduction [26]. Its principle is to delete closely related variables and create as few new variables as possible to make the transformed variables uncorrelated. Meanwhile, the transformed variables should reflect the original information as much as possible [27].

Given a training set $X=x_1,x_2,\dots ,x_N(x_i\in {ℝ}^D,i=1,2,\dots ,N)$ and a lower dimension d, we can calculate the average of the training set and the covariance matrix. Then, we can obtain the spectral decomposition of the covariance matrix to obtain the eigenvalues (${\lambda }_1\ge {\lambda }_2\ge \dots \ge {\lambda }_D$) and corresponding eigenvectors (${\xi }_1,{\xi }_2,\dots ,{\xi }_D$). For any $x\in {ℝ}^D$, its new low-dimensional representation can be represented as Equation (1)

$$y=({\xi }_1{}^T(x-\overline{x}),{\xi }_2{}^T(x-\overline{x}),\dots ,{\xi }_d{}^T(x-\overline{x}))\in {ℝ}^d$$

(1)

2.2. Artificial Neural Networks

The artificial neural networks (ANNs) used in this research were the backpropagation neural network (BPNN) and the cascade forward neural network (CFNN).

2.2.1. Backpropagation Neural Network

Backpropagation neural networks (BPNNs) can realize arbitrarily complex nonlinear mapping and are particularly suitable for solving complex internal mechanisms. They can perform complex pattern recognition and function fitting on experimental data. Figure 2 shows a basic diagram of the backpropagation neural network (BPNN). It is a multi-layer neural network trained according to the error backpropagation algorithm [28].

We designed a two-layer backpropagation neural network with a learning rate of 0.01. There were 20 neuron nodes in the first hidden layer, and 10 neuron nodes in the second hidden layer.

2.2.2. Cascade Forward Neural Network

As shown in Figure 3, the cascade forward neural network (CFNN) is similar to the BPNN; however, it includes a connection from the input and every previous layer to the following layers [29]. As with the BPNN, a two-or-more layer cascade network can learn any finite input–output relationship fairly well when given enough hidden neurons.

Here, we designed a two-layer cascade forward neural network with a learning rate of 0.01. There were 20 neuron nodes in the first hidden layer, and 10 neuron nodes in the second hidden layer.

2.3. Classic Machine Learning Algorithms

This research adopted two classic machine learning algorithms: support vector machine and k-nearest neighbors.

2.3.1. Support Vector Machine

Support vector machine (SVM) is a generalized linear classifier that classifies binary data in a supervised learning manner [30]. Its decision boundary is the maximum margin hyperplane for solving learning samples. The basic idea of SVM is to find a hyperplane for category division based on the training set D to separate samples into different categories. The hyperplane is shown in Equation (2).

$${\omega }^{T}x+b=0$$

(2)

The maximum interval for classification is shown in Equation (3).

$$\begin{array}{l}\underset{\omega ,b}{\min }\frac{1}{2}{\omega }^T\omega \\ s.t.\hspace{1em}{y}_i({\omega }^T{x}_i+b)\ge 1,\hspace{1em}i=1,2,\dots ,m\end{array}$$

(3)

To better solve this convex quadratic programming problem, the Lagrange multiplier method was introduced to obtain its dual problem, as shown in Equation (4).

$$\begin{array}{l}\underset{a}{\max }{\displaystyle \sum _{i=1}^m{a}_i}-\frac{1}{2}{\displaystyle \sum _{i=1}^m{\displaystyle \sum _{j=1}^m{\alpha }_i}}{\alpha }_j{y}_i{y}_j{x}_i^T{x}_j\\ s.t.\hspace{1em}{\alpha }_i\ge 0,\hspace{1em}i=1,2,\dots ,m\\ {\displaystyle \sum _{i=1}^m{\alpha }_i}{y}_i=0\end{array}$$

(4)

Most problems in the real world are not linearly separable, which means that there is no reasonable hyperplane in the original sample space to ensure the precise division of samples. Introducing the concept of the kernel function to replace the dual problem and the dot product operation after nonlinear mapping can solve this problem. Standard kernel functions include linear kernel, polynomial kernel, and RBF kernel. In addition, a penalty factor and a slack variable can facilitate the elevation of the accuracy of SVM.

We set the cost of SVM as 1 and the value of γ was 0.01.

2.3.2. K-Nearest Neighbors

K-nearest neighbors (KNN) is a classification algorithm [31]. The main idea of the algorithm is if a sample is most similar to the k samples in a data set, and if most of the k samples belong to the same category, the sample also belongs to this category. Euclidean distance can measure the similarity between two samples. The Euclidean distance between two points $x=(x_1,x_2,\dots ,x_n)$ and $y=(y_1,y_2,\dots ,y_n)$ in n-dimensional space can be represented as shown in Equation (5).

$$d_{xy}=\sqrt{{\displaystyle \sum _{i=1}^n(x_i-y_i)}}$$

(5)

We set the k value as 3, which meant that each point was measured within the 3 nearest points.

3. Experimental Design

As shown in Figure 4, we designed a driving simulation experiment to collect the physiological signals from participants and their drowsiness levels based on the Karolinska sleepiness scale (KSS) and trained observer rating (TOR). After feature extraction and drowsiness measurement, we adopted the hybrid model utilizing principal component analysis and artificial neural networks and the three other comparison models mentioned in Section 2 for the numerical experiment.

3.1. Driving Simulation

3.1.1. Participants

The selection criteria for the participants in this experiment were as follows:

○

Have held a valid driver’s license for at least six months;

○

Good physical condition;

○

No history of taking drugs in the past month; no alcohol, coffee, or functional beverages in the day before the experiment;

○

Have good sleeping habits; sleep no less than 6 h per day.

Nine qualified drivers aged 22–32 years (mean = 24.4, standard deviation = 3.13 years) were recruited strictly according to the selection criteria, including seven males and two females. After completing the simulating experiment, each participant received RMB 100 (about $16) as a subsidy.

3.1.2. Apparatuses

Figure 5 shows the apparatuses used in this research, including the driving simulator and physiological acquisition equipment.

The driving simulator for this experiment was the G29 developed by Logitech. We put the G29 on a computer with three displays with a viewing angle of 270°, which ensured the participants felt like they were driving in reality.

A Logitech 720p HD camera was set facing the drivers to capture facial videos. Moreover, the driving simulation laboratory was equipped with four surveillance cameras to observe the drivers’ behavior. The facial images and the surveillance images were both transmitted to the console.

This study captured the physiological signals using three pieces of physiological acquisition equipment developed by Kingfar Technology, including an electrodermal activity (EDA) recorder, respiration (RESP) recorder, and a photoplethysmography (PPG) recorder. Their sampling rates were 64 Hz. An Ergolab experiment platform was used to combine the signals for real-time data transmission. In addition, the lightweight devices ensured freedom of movement and minimized interference to drivers.

3.1.3. Scenario

Figure 5 also shows the driving scenario. World Editor, developed by 51-world, was used to design driving sections. Studies have shown that drivers are more likely to become drowsy in a monotonous environment, and drowsy driving traffic accidents are more common on highways [32,33].

Therefore, the driving scenario was designed as a free-flow one-way three-lane ring-shaped highway with fences and streetlights on both sides, no brightly colored irritants, and no other traffic interference in the lane. The weather was cloudless. The simulated driving time was consistent with the actual time.

3.1.4. Procedure

It has been shown that drivers are most likely to become drowsy from 5 PM to 6 AM the next day [1]. Therefore, the experiment was carried out after 5 PM. The experimental procedure consisted of a preparation stage and the driving task.

In the preparation stage, experimenters introduced the procedure to participants. Once the participants decided to take part in the experiment, they provided informed consent and wore physiological acquisition equipment under the guidance of the experimenters. Before the driving task, participants had five to ten minutes to adapt to the simulated scene.

During the driving task, participants were required to drive the vehicle from awake to drowsy at a speed of 100–120 km/h. Moreover, they were required to keep driving in the middle lane. In order to ensure the credibility of the participants’ drowsy state, this study combined the Karolinska sleepiness scale (KSS) developed by Azmeh, et al. [34] and the trained observer rating (TOR) method developed by Wierwille, et al. [35] to measure drowsiness.

Table 2. The KSS, TOR, and the corresponding relationship between them according to Zhang, et al. [36].

KSS Level	TOR Level	TOR Indicators
1 Extremely alert	0 Not drowsy	Normal fast eye blinks, often reasonably regular; Apparent focus on driving with occasional fast sideways glances; Normal facial tone; Occasional head, arm, and body movements.
2 Very alert
3 Alert
4 Rather alert
5 Neither alert nor sleepy
6 Some signs of sleepiness	1 Slightly drowsy	Increase in duration of eye blinks; Possible increase in the rate of eye blinks; Increase in duration and frequency of sideways glances; The appearance of a “glazed eye” look; The appearance of abrupt irregular movements—rubbing face/eyes, moving restlessly on the chair; Abnormally large body movements following drowsiness episodes; Occasional yawning.
7 Sleepy, but no effort to keep alert	2 Moderately drowsy	Occasional disruption of eye focus; Significant increase in eye blink duration; Disappearance of eye blink patterns observed during alert state; Reduction in the degree of eye opening; Occasional disappearance of facial tone; Episodes without any body movements.
8 Sleepy, some effort to keep alert	3 Very drowsy	Discernable episodes of almost complete eye closure, eyes never fully open; Significant disruptions in eye focus; Periods without body movements (more prolonged than in level 2) and facial tone followed by abrupt large body movements.
9 Very sleepy, great effort to keep alert, fighting sleep	4 Extremely drowsy	Significant increase in the duration of eye closure; Longer duration of episodes of no body movement followed by significant isolated “correction” movements.

shows the KSS and TOR levels and the corresponding relationship between them.

Participants evaluated their drowsy state during the experiment according to the KSS. It is worth mentioning that errors can appear when asking the participants’ KSS level too frequently or sparsely. Zhang, et al. [36] indicated that a 10-min interval best met the needs of their experiment. For this reason, the participants in the present study reported their self-perceived KSS level every 10 min during the driving task.

To check the reliability of participants’ self-reporting, an experimenter at the console observed each driver’s facial state and body movements in the video from the cameras. Figure 6 shows the display on the console. The experimenter determined the drowsy state of participants according to TOR at the same time as asking the KSS level of participants. When the difference between the corresponding driver’s self-reported score and the experimenter’s score exceeded 2, the experimenter would check the video to confirm the participants’ assessment.

3.2. Feature Extraction

We sliced the data every 2 s and extracted all the features shown in

Table 3. Extracted features.

EDA	RESP	PPG
Fuzzy entropy	Breath rate (mean, standard deviation)	Sympathetic vagal ratio
Wavelet entropy	Amplitude (mean, standard deviation)	Sympathetic ratio
Mean		Vagal ratio
Standard deviation		Heart rate (mean, standard deviation)
		Amplitude (mean, standard deviation)

.

For EDA, we calculated the mean and standard deviation of the signal intensity for each slice [6]. EDA is a series of complex and unstable nonlinear signals, while the entropy index is a parameter used to measure the complexity of the signal [37]. Therefore, we calculated the classical fuzzy entropy and wavelet energy entropy according to Min and Cai [11].

For RESP, we extracted the mean and standard deviation of the breath rate and signal amplitude [15].

For PPG, we calculated the three parameters of HRV shown in Equation (6) and the mean and standard deviation of both heart rate and signal amplitude [6].

$$\begin{array}{l}\mathrm{sympathetic}\_\mathrm{vagal}\mathrm{ratio}=s_{lf}/s_{hf}\\ \mathrm{sympathetic}\mathrm{ratio}=s_{lf}/(s_{vlf}+s_{lf}+s_{hf})\\ \mathrm{vagal}\mathrm{ratio}=s_{hf}/(s_{vlf}+s_{lf}+s_{hf})\end{array}$$

(6)

In Equation (6), $s_{vlf}$ refers to the very low-frequency (0.0–0.04 Hz) power of PPG signals; $s_{lf}$ refers to the low-frequency (0.04–0.15 Hz) power of PPG signals; and $s_{hf}$ refers to the high-frequency (0.15–0.4 Hz) power of PPG signals.

3.3. Measurement of Drowsiness

Although the nine detailed classification levels of the KSS are often beneficial, there is no need to divide the drowsiness level into nine categories in the classification problem [38]. Some studies have classified driver drowsiness into just two categories, with 0 representing no drowsiness and 1 indicating drowsiness [39,40]. As shown in

Table 4. Correspondence between the KSS, TOR, and drowsiness state defined in the present research.

Drowsiness State	KSS Level	TOR Level
0 No drowsiness	1 Extremely alert	0 Not drowsy
	2 Very alert
	3 Alert
	4 Rather alert
	5 Neither alert nor sleepy
1 Drowsiness	6 Some signs of sleepiness	1 Slightly drowsy
	7 Sleepy, but no effort to keep alert	2 Moderately drowsy
	8 Sleepy, some effort to keep alert	3 Very drowsy
	9 Very sleepy, great effort to keep alert, fighting sleep	4 Extremely drowsy

, none of the KSS levels 1–5 indicates sleepiness; thus, they were classified as state 0, or no drowsiness, in this study. KSS levels 6–9 all indicate sleepiness; thus, they were classified as state 1, or drowsiness, here.

3.4. Numerical Experiment

The models were trained on a computer equipped with Intel i9-10900k, 64G RAM, and 1T SSD. In order to eliminate the errors caused by sample differences, we modeled each sample individually to detect driving drowsiness. The original data for each model were the extracted physiological features described in Section 3.2. The training targets were the drowsiness states mentioned in Section 3.3 (0—no drowsiness; 1—drowsiness).

There were two stages in the numerical experiment: drowsiness detection and verification.

• Drowsiness detection

In this stage, the original data were input into the proposed model and the three comparison models. As shown in

Table 5. Data division.

Classification Algorithms		Training Set	Validation Set	Testing Set
Artificial neural networks	BPNN	60%	20%	20%
	CFNN	60%	20%	20%
Classic machine learning algorithms	SVM	60%	0%	40%
	KNN	60%	0%	40%

, data division varied with different classification algorithms. The evaluation indexes of the model are shown

Table 6. Evaluation indexes of numerical experiment.

Index	Description
Accuracy (%)	Percentage of physiological state accurately detected.
Drowsiness recall (%)	Percentage of drowsiness accurately detected.
Drowsiness precision (%)	Percentage of the precise drowsiness output
AUC	Area under the ROC curve; reflects the model’s ability to classify positive and negative examples.
Training time (seconds)	Time required for model training.

.

• Verification

In this stage, drowsiness detection was repeated 100 times, then each evaluation index’s average and standard deviation were calculated to verify the drowsiness detection results and evaluate each model’s stability.

4. Results

4.1. Drowsiness Detection

Table 7. The average accuracies, drowsiness recalls, drowsiness precisions, AUCs, and training times of the models in the drowsiness detection stage.

Models	Hybrid Models				Single Models
	Hybrid Model of Principal Component Analysis and ANN		Hybrid Model of Principal Component Analysis and CMLA		Single Model of ANN		Single Model of CMLA
	PCA-BPNN	PCA-CFNN	PCA-SVM	PCA-KNN	BPNN	CFNN	SVM	KNN
Accuracy (%)	97.2	97.9	94.8	94.6	96.1	96.1	93.3	93.0
Drowsiness recall (%)	98.2	98.5	98.4	97.1	97.2	97.3	97.3	95.8
Drowsiness precision (%)	97.5	98.1	94.3	94.8	96.6	96.4	92.8	93.6
AUC	0.973	0.973	0.914	0.920	0.970	0.970	0.892	0.900
Training time (seconds)	0.182	0.231	50.719	0.107	0.441	0.982	64.963	0.115

* ANN—artificial neural networks; CMLA—classic machine learning algorithms.

shows the average accuracies, drowsiness recalls, drowsiness precisions, AUCs, and training times of the models; the best performance of each evaluation index for the hybrid and single models is bolded. The average accuracies, drowsiness recalls, and drowsiness precisions of the training results from all models were above 90%, and the average AUCs of most of the models were above 0.9. Different models showed different average training times. The average training time of PCA-SVM and SVM was about one minute, while the training times of other models were mainly within one second. Additional information can be found in Appendix A, which shows the results of drowsiness detection.

The hybrid model utilizing principal component analysis and artificial neural networks obtained the highest accuracy, drowsiness recall, drowsiness precision, and AUC, while guaranteeing timeliness. The average accuracy of the hybrid model (PCA-CFNN) was 97.9%, which was up to 4.9% higher than the comparison models (KNN). The training time of PCA-CFNN was 0.231 s, up to 99% shorter than the comparison models (SVM). Although the training time was slightly longer than that for PCA-KNN (0.124 s) and KNN (0.116 s), it still satisfied the needs for application in real driving scenarios.

The artificial neural networks performed better than the classic machine learning algorithms. For hybrid models, the average accuracies, drowsiness precisions, and AUCs of the artificial neural networks (PCA-BPNN and PCA-CFNN) were higher than classic machine learning algorithms (PCA-SVM and PCA-KNN). The drowsiness recalls of the artificial neural networks (PCA-BPNN and PCA-CFNN) were higher than that of PCA-KNN but lower than that of PCA-SVM. The single models had similar performance. The average accuracies, drowsiness precisions, and AUCs of the artificial neural networks (BPNN and CFNN) were higher than those of the classic machine learning algorithms (SVM and KNN). The drowsiness recalls of the artificial neural networks (BPNN and CFNN) were higher than that of KNN but lower than that of SVM. However, PCA-SVM had not only the lowest average drowsiness precision and AUC, but also the longest training time of the hybrid models. SVM showed the same performance in single models. This indicated that both PCA-SVM and SVM had high misjudgment rates, low classification reliabilities, and a long time delay and were not suitable for application in real driving scenarios.

Moreover, the hybrid models performed better than the single models. Figure 7a shows the improvement of the hybrid models compared to the single models in each evaluation index. The accuracy, drowsiness recalls, drowsiness precisions, and AUCs of the hybrid models were mostly higher than the single models. In addition, the training times of the hybrid models were shorter than those of the single models.

Figure 7b takes sample 9 to illustrate the above results in detail. The principal component analysis and artificial neural network hybrid models showed the highest accuracy, drowsiness recall, drowsiness precision, and AUC, with a training time of lower than one second. For the hybrid models, the accuracies, drowsiness precisions, and AUCs of the artificial neural networks (PCA-BPNN and PCA-CFNN) were higher than those of the classic machine learning algorithms (PCA-SVM and PCA-KNN). For single models, the accuracies, drowsiness precisions, and AUCs of the artificial neural networks (BPNN and CFNN) were also higher than those of the classic machine learning algorithms (SVM and KNN). In addition, the hybrid models showed improvements compared to single models in accuracy, drowsiness recall, drowsiness precision, and AUC, and had reductions in training time. This confirmed that the performance of artificial neural networks is better than classic machine learning algorithms and hybrid models have better performance than single models.

In general, the results of the drowsiness detection stage reflect that (1) the hybrid model with principal component analysis and artificial neural networks performed the best; (2) artificial neural networks obtained a higher accuracy than classic machine learning algorithms; (3) the hybrid with principal component analysis and classification algorithms (artificial neural networks and classic machine learning algorithms) enhanced the accuracy and timeliness of drowsiness detection.

4.2. Verification

Table 8. The average accuracies, drowsiness recalls, drowsiness precisions, AUCs, and training times of each model in the verification stage.

Models	Hybrid Models				Single Models
	Hybrid Model of Principal Component Analysis and ANN		Hybrid Model of Principal Component Analysis and CMLA		Single Model of ANN		Single Model of CMLA
	PCA-BPNN	PCA-CFNN	PCA-SVM	PCA-KNN	BPNN	CFNN	SVM	KNN
Accuracy (%)	97.3	97.4	94.8	94.7	96.3	96.5	93.7	93.4
Drowsiness recall (%)	98.4	98.5	98.2	97.0	97.6	97.8	97.8	96.1
Drowsiness precision (%)	97.3	97.4	94.1	94.9	96.5	96.6	93.0	93.8
AUC	0.977	0.978	0.916	0.923	0.973	0.973	0.896	0.903
Training time (seconds)	0.215	0.259	50.849	0.158	0.607	1.073	65.392	0.169

* ANN—artificial neural networks; CMLA—classic machine learning algorithms.

shows the average accuracies, drowsiness recalls, drowsiness precisions, AUCs, and training times of each model in the verification stage. Figure 8a shows the improvements in the hybrid models compared to the single models. Consistent with the results of the drowsiness detection stage, the hybrid model with principal component analysis and artificial neural networks showed the best performance while guaranteeing timeliness; the artificial neural networks performed better than the classic machine learning algorithms; and the hybrid models performed better than the single models in the verification stage. For more information, Appendix B shows the results of the verification.

Table 9. The standard deviation of the accuracies, drowsiness recalls, drowsiness precisions, AUCs, and training times of models in the verification stage.

Models	Hybrid Models				Single Models
	Hybrid Model of Principal Component Analysis and ANN		Hybrid Model of Principal Component Analysis and CMLA		Single Model of ANN		Single Model of CMLA
	PCA-BPNN	PCA-CFNN	PCA-SVM	PCA-KNN	BPNN	CFNN	SVM	KNN
Accuracy (%)	0.8	0.7	0.6	0.6	1.2	0.8	0.6	0.7
Drowsiness recall (%)	0.8	0.7	0.4	0.7	1.0	0.9	0.5	0.8
Drowsiness precision (%)	0.9	0.9	0.7	0.8	1.3	0.9	0.8	0.8
AUC	0.003	0.003	0.011	0.012	0.005	0.004	0.013	0.013
Training time (seconds)	0.042	0.053	1.149	0.011	0.204	0.439	1.203	0.012

* ANN—artificial neural networks; CMLA—classic machine learning algorithms.

shows the standard deviations of the accuracies, drowsiness recalls, drowsiness precisions, AUCs, and training times of the hybrid and single models; the best evaluation indexes in each model are bolded. The average standard deviations of the evaluation indexes of all the hybrid models were lower than 1%, which was lower than the single models. This showed that the hybrid models demonstrated a more stable performance than the single models. In addition, as shown in Figure 8b, most of the standard deviations in the accuracies, drowsiness recalls, drowsiness precisions, AUCs, and training times obtained by the hybrid models were lower than those of the single models. This showed that the performance of the hybrid models was more stable and less volatile than that of the single models.

Figure 9 takes sample 9 to illustrate these findings in detail. The hybrid model with principal component analysis and artificial neural networks had the best performance; the accuracies, drowsiness precisions, and AUCs of the artificial neural networks were higher than the classic machine learning algorithms; and the hybrid models showed improvements compared to the single models in each evaluation index, which was consistent with the drowsiness detection stage. Moreover, the distribution ranges of the evaluation indicators of the single models were more comprehensive than those of the hybrid models. We speculated that the performance of the hybrid models was more stable than that of the single models.

In general, the results of the verification stage were consistent with the results of the drowsiness detection stage. In addition, a hybrid utilizing principal component analysis can improve the stability of the model.

5. Discussion

Drowsy driving is harmful to human society, and it is essential to detect driving drowsiness accurately and quickly. This study proposed a hybrid model of principal component analysis and artificial neural networks to detect driving drowsiness quickly and accurately using non-intrusive physiological data. We collected EDA, RESP, and PPG signals and calculated their features. Then, we used the proposed model and its comparison models to detect drowsiness. We also compared their performance to verify our hypotheses.

Hypothesis 1 was confirmed. The results showed that the performance of the hybrid models was better than that of the single models. We speculate that this is due to the advantages of PCA, which include reducing resource requirements, enhancing data interpretability, and—most importantly—removing noise. In other words, PCA can give the model higher accuracy [41]. In addition, we found that the training times of the hybrid models were shorter than those of the single models. We speculate that the hybrid models reduced the dimensions of the extracted physiological features, leading to shorter computation times for the driving drowsiness detection models. Moreover, the results of the verification showed that the standard deviations of the accuracies, drowsiness recalls, drowsiness precisions, AUCs, and training times of the hybrid models were lower than those of the single models. This showed that PCA could assist the models in adapting to unknown data and enhance their stability. Thus, PCA improved the accuracy, timeliness, and guaranteed stability of the model.

Hypothesis 2 was also confirmed. The results showed that the accuracies, drowsiness precisions, and AUCs of the artificial neural networks were higher than those of the classic machine learning algorithms. In other words, artificial neural networks can detect driving drowsiness more accurately and credibly than classic machine learning algorithms. Thus, artificial neural networks are more appropriate for application to natural driving scenes than classic machine learning algorithms. The internal mechanism behind the detection of driving drowsiness is very comprehensive. Basheer and Hajmeer [42] suggested that since artificial neural networks can realize any complex nonlinear transformation, they are particularly appropriate for problems with complex internal mechanisms.

In summary, we believe that the proposed hybrid model utilizing principal component analysis and artificial neural networks meets the need for fast and accurate drowsiness detection. In the future, the proposed model may be applied in smart cars to detect the drowsiness states of drivers and promptly remind them to pay attention to their physiological states. This method can guarantee road traffic safety in the future.

6. Conclusions

The present study proposed a hybrid model utilizing principal component analysis and artificial neural networks to accurately detect driving drowsiness. A driving simulation experiment was designed to collect the participants’ physiological signals and their drowsiness levels. We extracted the corresponding physiological features as the original data and the drowsiness levels as the training targets of the models. Then, we used the proposed model and three other comparison models for drowsiness detection.

The artificial neural networks obtained higher accuracy and credibility than the classic machine learning algorithms. Hybrid models utilizing principal component analysis can provide higher accuracy, timeliness, and stability in drowsiness detection; our hybrid model with principal component analysis and artificial neural networks reached the goal of this research. The driving drowsiness detection method proposed in this study can be applied to the active warning system (AWS) of smart cars in the future and contribute to road safety. In the future, it will be necessary to conduct research on how AWS systems can alert drivers to stay awake without scaring them. In addition, the hyperparameters of PCA or artificial neural networks can be modified to further improve the accuracy of driving drowsiness detection.