This study utilized a video-guided experimental paradigm that enhances participant immersion and attentiveness during signal collection. By integrating visual and auditory stimuli, this approach intensifies emotional induction and prolongs the duration of emotional experiences, ultimately leading to the generation of higher-quality EEG signals.

The videos were filmed from a first-person perspective to simulate a realistic driving experience. The videos designed to induce driving stress included scenarios such as high-speed driving, extreme weather conditions, and emergency avoidance maneuvers. In contrast, the videos aimed at eliciting anger portrayed situations such as traffic congestion, traffic violations, and provocations from other drivers. To mitigate potential loss of interest among participants due to excessive video duration, which could negatively impact the emotional induction process, the length was intentionally limited to 3 min, thereby helping to sustain participants’ attention and engagement (Xu et al., 2024).

Prior to the experiment, videos intended to induce stress and anger were meticulously screened. The video materials were sourced from the internet. Participants’ emotional states were assessed using a five-point Likert scale immediately after viewing the inducing videos, with a score of 3 indicating successful emotional induction. Specifically, a score of 1 represented “Calm,” a score of 2 indicated “Almost not,” a score of 3 indicated “Slight,” a score of 4 represented “Moderate,” and a score of 5 indicated “High.” Only videos with an induction success rate exceeding 80% were selected as emotional stimuli. The Likert scale used in this study is presented in Table 1.

Given that this experiment requires participants to drive while experiencing altered emotional states—where real-world driving presents significant safety risks and operational challenges—this study employed a driving simulation experiment as a viable alternative. The primary advantages of this approach include enhanced safety, precise control over experimental conditions (such as temperature, lighting, and audio), and reduced costs. Additionally, driving simulators can effectively replicate real driving scenarios, thereby ensuring the reliability and validity of the experimental results. Numerous studies have shown that the physiological responses observed in driving simulations closely resemble those in actual driving environments (Bobermin et al., 2021). Consequently, this study utilizes driving simulators instead of real-world driving.

This study utilized the Forza Horizon 5 software developed on the EA platform, in conjunction with the Laisida V99 driving simulator, to create a simulated driving environment. The simulator effectively replicates real-world traffic conditions and comprises three main components: the vehicle operation system, the visual display system, and the audio system. The vehicle operation system includes a steering wheel, gear shifter, accelerator pedal, brake pedal, and clutch. The visual display system features an LCD monitor that provides a first-person perspective, while the audio system offers immersive surround sound effects to enhance the simulated driving experience.

The experimental setting is illustrated in Figure 1. The laboratory features adequate ventilation and sufficient lighting.

The emotional induction videos were presented using E-Prime (Richard and Charbonneau, 2009; Verdonschot et al., 2019), which is integrated with the EEG data collection module to form a synchronized data acquisition system. The electrode placements on the EEG cap followed the international 10-10 system, with CPz and the End electrodes serving as the reference and ground, respectively. During the experimental tasks, electrode impedance was consistently maintained below 5 kΩ. Additionally, the regulating music was played through headphones connected to E-Prime, while olfactory stimulation was delivered using an olfactory experience testing device.

The music used for the regulation is titled “Rose Petals.” This gentle piano piece is set in a major key and features a slow, soft rhythm, with a 30-s excerpt used. Soft music is effective in alleviating negative emotions (Liao, 2024; Xiao et al., 2024). The fragrance used is lemon fragrance, which promotes positive emotions and inner calm while soothing the nervous system (Godfrey, 2018; Smith and Nicholson-Lord, 2024). The current experiment also innovatively combines the soothing melodies of soft music with the refreshing lemon fragrance to investigate their synergistic effects on regulating emotional states.

This study employed a rigorous three-phase screening process to recruit 54 drivers, consisting of 27 males and 27 females, aged between 20 and 27 years. Participants with pre-existing central nervous system disorders, rhinitis, or auditory impairments were excluded through subjective assessment tests to minimize potential confounding factors. Additionally, individuals who had participated in similar experimental settings previously were also excluded to ensure the selection of appropriate candidates. Before the study commenced, all participants signed an informed consent form that outlined the study’s objectives and the specific tasks they would be required to undertake. The 54 drivers were divided into two groups: the first group of 27 (Female:15, Male:12.) individuals underwent a stress emotion regulation experiment, while the second group of 27 (Female:12, Male:15.) individuals participated in an anger emotion regulation experiment. Comprehensive information regarding the participants is presented in Table 2.

The stress emotion regulation experiment was conducted first. Prior to the formal initiation of the experiment, participants were given 30 s to calm their mood, during which their EEG signals were recorded in this relaxed state. Following the calming phase, participants viewed a stress-inducing video. After 60 s of video playback, the experimenters assessed participants’ perceived stress levels using a five-point Likert scale. If participants reported a stress level of 3 (slight stress) or higher, their EEG signals were recorded for the subsequent 30 s. After this, a regulation phase lasting 30 s was conducted, during which additional EEG data were collected. Each stage—calm, induction, and regulation—was clearly defined. A total of 27 participants took part in the stress emotion regulation experiment, divided into three groups of nine. Each group experienced a different regulatory modality: fragrance, music, or a combination of both. Throughout the regulation phase, the stress-inducing video continued to play, and participants’ stress levels were reassessed using the Likert scale after the regulation phase concluded. The results of this assessment were subsequently documented.

Following the completion of the stress emotion regulation experiment, the anger emotion regulation experiment was conducted in a similar manner. The grouping information is presented in Table 3.

The experimental process is shown in Figure 2. This figure illustrates the sequence of events in the experiment, including the preparation stage, the calm stage, the emotional induction stage, the emotion regulation experiments carried out by the participants, and the subsequent emotional state assessments. Each stage is clearly depicted, allowing for a better understanding of how the study was conducted and the specific procedures followed by the participants. The right-pointing arrow signifies that the experiment proceeds from left to right.

This study was approved by the Ethics Committee of Chongqing University of Arts and Sciences (Approval No. CQWL202424). All procedures involving human participants were conducted in accordance with the ethical standards of the 2024 Helsinki Declaration and its subsequent amendments, as well as applicable national and regional regulations. Prior to implementation, the research protocol underwent rigorous review by an independent institutional review board. Written informed consent was obtained from all participants, explicitly addressing the following aspects: (1) participation was voluntary, with the right to withdraw at any stage without penalty; (2) personal information was kept strictly confidential through data anonymization techniques; (3) research data were securely stored in encrypted formats accessible only to authorized investigators; and (4) collected data were exclusively used for the stated scientific purposes.

The machine learning workflow comprises three main stages: data preprocessing, feature extraction, and classification (Alazrai et al., 2019). Feature extraction focuses on capturing both time-domain and frequency-domain characteristics of the signals. Ultimately, the selected features are input into classification algorithms to derive the classification results.

The performance of K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Backpropagation Neural Network (BPNN) in recognizing driving stress and anger in terms of accuracy, precision, recall, and F1 score is presented in Table 5.

According to the results presented in Table 3, the Backpropagation Neural Network (BPNN) demonstrates exceptional performance in recognizing driving stress and anger. The BPNN model outperforms all other models across all metrics, exhibiting the highest accuracy, precision, recall, and F1 score. Therefore, the BPNN model is selected as the final model for recognizing stress and anger emotions.

The confusion matrix analysis results for the BPNN stress emotion recognition model developed in this study indicate that the model achieved 100% prediction accuracy in both the high Stress and moderate stress categories, with no misclassifications. In the slight stress category, the model correctly identified 67% of the samples, while 33% were misclassified as almost not stress. For the almost not stress category, the model had a correctness rate of 61%, with 39% of samples incorrectly classified as calm. Additionally, the prediction accuracy for the calm stage category also reached 100%. Overall, the model performed well in the emotion recognition task, demonstrating its effectiveness in handling different emotional states.

And the confusion matrix for the BPNN anger emotion recognition mode. In the high anger category, the model achieved a prediction accuracy of 97%, with only 3% misclassified as moderate anger. For the moderate anxious category, 78% of samples were correctly classified, while 22% were misclassified as Slight Anger. In the Slight anger category, the model correctly identified all samples, achieving 100% accuracy. The Almost not anger category also saw a perfect classification rate of 100%. For the Calm category, 22% of the samples were classified as Almost not anger, while 78% were correctly identified as Calm. Overall, the model demonstrates strong performance in identifying emotions, particularly in the Highly Anxious and Slightly Anxious categories, highlighting its effectiveness in emotional recognition tasks.

In the stress emotion recognition model, the input layer comprises six neurons, each corresponding to one of the six features: mean, variance, skewness, kurtosis, root mean square, and power spectral density. The output layer includes five neurons, representing the five classification labels: calm, nearly no stress, slight stress, moderate stress, and high stress. The hidden layer consists of five neurons and was trained over 24 iterations.

In the anger emotion recognition model, the input layer consists of six neurons, each representing one of the six features: mean, variance, skewness, kurtosis, root mean square, and power spectral density. The output layer contains five neurons, corresponding to the five classification labels: calm, nearly no anger, slight anger, moderate anger, and high anger. The hidden layer comprises five neurons and was trained over 29 iterations.

The stress and anger emotion models, classified using the Back Propagation Neural Network (BPNN) algorithm, have been saved in preparation for further regulatory analyses.

To further validate the effects of music and aroma interventions, we conducted a time-frequency analysis of EEG signals across various emotional and stress states. Specifically, we amalgamated slight, moderate, and high-stress conditions into a single stress category, while consolidating slight, moderate, and high-anger conditions into a unified anger category to facilitate analysis.

Our findings indicate a significant increase in the activity parameter under stress conditions, reaching a value of 9.3961, compared to 6.2258 in calm states. This suggests a strong relationship between stress and heightened brain activity. The mobility parameter exhibited minimal variation between the two states, with values of 1.9424 in calm states and 1.8722 during stress. Additionally, the complexity parameter demonstrated a slight increase from 1.5921 in calm states to 1.6432 under stress, reflecting a negligible change overall. In the anger state, the activity level substantially increased to 13.3234, indicating a higher level of physiological activation. Conversely, the mobility and complexity parameters were recorded at 2.0213 and 1.5367, respectively, showing no significant changes when compared to the calm condition.

Next, the activity, mobility and complexity after receiving music, fragrance, and combined regulation were statistically analyzed, as shown in Figures 5, 6.

Observations indicated that the Hjorth activity parameter of the participant, after receiving music, aroma, and combined interventions, was lower compared to the levels observed during tense and emotional states, approaching the activity levels characteristic of calm states.

The analysis of Hjorth parameters indicates that, overall, the use of music alone is the strategy that shows the greatest impact on alleviating stress, while the use of fragrance alone is particularly effective for mitigating anger. These findings further support the role of both music and fragrance in alleviating abnormal emotions and confirm the reliability of the EEG-based models for identifying emotional states, specifically stress and anger, as well as the emotion regulation assessment system. The multidimensional assessment of EEG data enhances the authenticity and accuracy of the results.

Whether through music, fragrance, or a combination of both, all interventions demonstrate a certain degree of effectiveness in alleviating various emotional states. These findings align with research conducted by Pei et al. (2024) and Cook et al. (2019), which similarly validated the efficacy of music and fragrance in mitigating abnormal emotions.

However, studies by Ning et al. (2022) and Sanyal et al. (2013) indicate that music does not always alleviate abnormal emotions. Certain emotionally charged music, such as sad or intense pieces, may evoke negative feelings or memories, leading to emotional fluctuations and increased anxiety, which can exacerbate stress or anger. Additionally, fast-paced or overly complex music can induce excessive cortical excitation, resulting in physiological responses such as increased heart rate and rapid breathing, potentially hindering emotional relief. In contrast, the gentle music employed in this study effectively reduces heart rate and respiratory frequency, facilitating relaxation and helping to mitigate negative emotions. Similarly, research conducted by Barnes et al. (2018), Wang et al. (2024), and Soars (2009) reveals that not all scents are effective in alleviating drivers’ abnormal emotions; for instance, odors from car engines, body odor, and the scent of new cars can intensify negative feelings. However, the lemon fragrance used in this study enhances driving pleasure, effectively promotes positive emotions, and fosters a sense of inner calm, providing a gentle soothing effect on the nervous system.

The Hjorth parameters are essential tools for signal processing and are widely used in the analysis of electroencephalogram (EEG) signals, particularly in studying emotional states (Fatih et al., 2023). By analyzing Hjorth parameters, researchers can objectively assess and understand the characteristics of EEG activity in drivers experiencing different emotional conditions. This study further employs Hjorth parameter analysis to investigate the effects of music and fragrance on emotional regulation. The results indicate that music alone is the most effective intervention for alleviating stress, while a singular fragrance intervention is most effective in calming anger. Moreover, this study validates the efficacy of both music and fragrance in mitigating abnormal emotions and confirms the reliability of EEG-based models for identifying emotional states (stress and anger) as well as the emotion regulation assessment system.

Hjorth activity serves as an indicator of the intensity of EEG signals and effectively reflects the degree of emotional arousal. In this study, Hjorth activity parameters significantly increased when drivers experienced tension and anger, indicating heightened brain activation in response to abnormal emotions, which aligns with the findings of Mehmood et al. (2022). In contrast, changes in Hjorth mobility and complexity were relatively subtle. This suggests that, while subjects experienced emotional state changes, the effects on EEG signals may be concentrated in specific frequency bands or features, leading to a lack of variability in these parameters.

This study has several limitations. First, the relatively small sample size may restrict the generalizability and reliability of the findings. Second, the investigation focused exclusively on gentle music and lemon fragrance as stimuli, without examining a broader range of musical genres, scents, and their interactions with various emotional states. Importantly, we did not perform statistical hypothesis testing to confirm the stimulus effects due to the exploratory nature and sample size constraints of this study; the observed effects should therefore be interpreted as preliminary trends requiring future validation. Additionally, while EEG provides direct measurements of brain activity, the complexity and multidimensional nature of emotions require the incorporation of additional physiological indicators (such as heart rate variability) and behavioral data to comprehensively assess the impact of music and analyze the predictive capacity of various signal combinations. Future research will aim to develop more refined modeling frameworks and optimization methods to enhance predictive accuracy, including rigorous statistical verification of stimulus effects through controlled experimental designs with adequate power.