Thirty students from Shanghai University majoring in construction engineering management were selected for the pilot test to confirm the feasibility and validity of the experiment. Forty construction workers from six Shanghai construction engineering enterprises were recruited for the formal experiment. Among all the subjects, 3 are construction workers and the other 34 were workers in supervisory positions. There are 3 subjects were below undergraduate level, 27 were undergraduates and 10 were postgraduates. All subjects had received safety management training of construction work. The studies involving human participants were reviewed and approved by Ethics Committee of Shanghai University. Participants selected for the pilot and formal experiment were based on the following criteria: (1) Familiar with the construction industry or have long-term working experience on construction site, and familiar with the operation codes on the construction site; (2) Physically and mentally healthy without any psychological disorders; (3) All are right-handed; (4) All provided written informed consent. The basic information of the subjects are shown in Table 1.

The experiment was carried out in a closed construction site conference room without interference. Forty rounds of experiments were included in this study. Each round contains three parts that are positive emotions, neutral emotions, and negative emotions. There are 120 samples in total. After informed consent, all the participants were attached electrodes for the physiological measurement. The experiment guidance is displayed by computer, which explains the purpose and procedures of the experiment. Following the instruction, a picture of the targeted emotion will be shown on the screen to stimulate the empathic effect of the participant, and the picture will last for 6 s, 15 pictures will be shown each time. Subsequently, participants completed the Positive and Negative Affect Schedule (PANAS) (Watson et al., 1988), followed by the identification and assessments of safety hazards in the construction pictures. The experimental procedure is shown in Figure 1.

Arousal of individual emotions using picture stimuli is one common forms of emotional stimulation (Gerdes et al., 2014). The International Affective Picture System (IAPS) database (Lang et al., 1988) was used to evoke different emotional states of the construction workers. Thirty images each of positive, neutral and negative pictures were selected from the IAPS as emotional arousal stimulus material. Researchers have found that incidental emotions pervasively carry over from one situation to the next, affecting decisions that unrelated to that emotion, known as the carryover of incidental emotion (Loewenstein and Lerner, 2003; Lerner and Tiedens, 2006; Keltner and Lerner, 2010). In the IAPS database, values of valence indicating the level of enjoyment and the values of arousal indicating the level of excitement. The positive pictures selected in this paper include life scenes, animal activity pictures and baby pictures. Neutral pictures include pictures of static objects, abstract artwork and pictures of natural environment. Negative pictures include catastrophic events, violent and brutal scenes and pictures of disabled individuals. Images of the IAPS database cannot be displayed as a result of a confidentiality agreement. The valence value of the negative neutral and positive mood pictures were 1.78, 4.92, 7.83, and the arousal value was 6.36, 3.37, 5.14, respectively. The mean squared deviation of the pictures was less than 2.4, ensuring the validity of the pictures (Zhang et al., 2018). The subjects were randomly shown one set of emotional pictures, and level of emotional arousal was evaluated using the PANAS Emotional Self-Rating Scale; subjects whose emotions were not aroused were excluded from the results of the experiment, the emotional scale is shown in Table 2.

After the 15 emotion pictures were displayed, the participants will give a self-evaluation of their emotions by filing an emotion scale adapted from PANAS, with a scale ranging from 1 to 9. The higher scores indicate stronger positive emotions, while the lower scores indicate stronger negative emotions. The emotional scale is shown in Table 3.

After cleaning the skin surface, the participants were attached the sensor to the construction worker’s finger. The electrodes were stick in the sensor around the index and middle finger. The galvanic skin response data were recorded and sent to a computer. The picture of the GSR equipment and the electrode attachment location is shown in Figure 2. The galvanic skin response at 5–6 s after the emotional arousal was used for identifying and classifying of the emotions. The experiment was carried out in a laboratory at a room temperature of 22°C, the subjects sitting still in front of a computer for the galvanic skin response measurement, with the temperature and humidity remaining constant throughout the experiment. The signal processing procedures are elaborated in Section 3.

The recognition of safety hazards was measured by identifying the safety hazards from construction site pictures, and images of construction sites containing five types of safety hazards were collected for this study from 12 construction sites in Shanghai, China. These images were obtained from safety incident reports, and the opinions of 10 experts were collected to evaluate the two dimensions of the selected images, (1) whether the images visually represented the safety hazards of that type of construction and (2) whether the images were prevalent in the construction site. The 120 images were retained after deleting the lower scoring images. The images were displayed randomly according to category, and 16 images were presented on screen in a set for the subjects to evaluate the safety hazards. The distribution of the 16 images by category was: 7 fall from height, 2 electric shock and fire, 2 object strikes, 1 collapse hazard, 1 mechanical injury and 3 no hazard images. To avoid a learning effect in the subjects, each picture was presented only once at random, with no repeated presentations in all sets for one participants. Some examples of the images chosen are shown in Figure 3.

The cognitive level of safety hazards was measured by behavioral experiment. The participants were requested to view the construction site images and determine if the pictures contain safety hazards (Tixier et al., 2014). The participants pressed “1” on the keyboard if they think there are safety hazards, while “0” on the keyboard if there are not. The computer automatically records the time taken to identify safety hazards and the accuracy of the safety hazards assessment. The perception level of safety hazards of construction workers was measured by safety hazards perception assessment form, which was completed simultaneously when workers believe there is a safety hazards in the given picture. Hallowell pioneered the use of this form by quantifying safety hazards perception as the product of the expected frequency and severity of injury (Hallowell, 2010). The corresponding scores of perception level of safety hazards are shown in Table 3.

Forty subjects participated in this study, excluding three who failed in emotional arousal with 37 remaining subjects. After 37 sets of experiments, each containing 3 categories of emotional stimuli, 16 pictures of safety hazards for each category, a total of 1776 data were collected. Finally, 1,650 valid data were obtained with a 92.9% validity after removing invalid questionnaires.

The collected data were smoothed and filtered with a median filter and a third-order Butterworth low-pass filter. A cut-off frequency of 0.3 HZ was used to filter out abnormal signals, and the signals were eliminated from the baseline interference to reduce the influence caused by the measurement instrument itself and the current and voltage. The filtered galvanic skin response signals were extracted and normalized from both time domain signal features and descriptive features, and then, principal component analysis is applied to reduce the dimensionality of the acquired features to obtain the signal features for classification.

The reaction time to safety hazards, accuracy of safety hazards identification and perception level of safety hazards of construction workers in different emotional states were analyzed. One-way ANOVA was used to investigate the differences in recognition ability of safety hazards among workers within different age groups under different emotional states. Pearson correlation and regression analyses were used to quantify the effects of emotions on the reaction time to safety hazards, the accuracy of safety hazards identification, and perception level of safety hazards, respectively.

The collected data were smoothed and filtered with a median filter and a third-order Butterworth low-pass filter, and the cut-off frequency was set as 0.3 HZ to filter out the abnormal signals and retain the valid signals. The original waveform, the waveform after median filtering and the waveform after low-pass filtering are shown in Figure 4; the obtained signals were de-baselined to reduce the effects caused by the measurement instrument itself and the current and voltage; the filtered skin electrical signals were extracted from both time domain signal features and descriptive features. The filtered electrical skin signal was extracted from both time domain signal features and descriptive features, and the extracted features are shown in Table 4. Principal component analysis was used to reduce dimension, the results are shown in Table 5, and the coefficient of the feature GSR_diff_Std is low, so the feature is deleted from the subsequent classification training. The final 10 features retained after normalization were obtained for the subsequent Support Vector Machine classification.

The 120 sample points were used as the training data set; the Support Vector Machine (SVM) model was applied for classification training. In a ratio of 2:1, these 120 samples are divided into a training set (90 samples containing 30 positive, 30 neutral, and 30 negative samples) and a test set (30 samples). Both classification accuracy and model validation were improved by classified labelling of these 120 sample points and supervised learning of the model, which was implemented by Matlab R2016b. In the prediction experiment, 30 sample were validated and the results are shown in Figure 5; Table 6. The vertical coordinates 1,2,3 correspond to negative, neutral and positive samples, respectively. When calculating the sensitive, specificity and precision of the examples, for each emotion, the emotion itself is considered a positive example, while the remaining two emotional states are considered negative examples. The classification results indicate that the picture triggering method in this study achieves effective emotional arousal.

The statistical results of recognition ability of safety hazards of construction workers in different emotional states are shown in Table 7. Workers in the positive emotional state had the shortest reaction time (5.61 s), and workers in the negative and neutral emotional states required longer reaction time (8.08 s and 6.91 s, respectively). Construction workers in the neutral emotional state had the highest identification accuracy of safety hazards (92.25%) and perception level of safety hazard (24.52), while in the negative emotional state, the construction workers had the lowest identification accuracy of safety hazards (75.41%) and perception level of safety hazards (0.75) than other emotional states.

The study found that the reaction time for identifying safety hazards was longer in negative emotions than in neutral and positive emotions. When construction workers were in a positive emotion, the feedback time for identifying safety hazards in construction site pictures was 5.61 s, which was less than the 6.91 s in a neutral emotion and 8.08 s in a positive emotion. This conclusion is consistent with Fredrickson’s findings that positive emotions serve to improve individual’s physical, intellectual, and perceptions (Fredrickson, 1998). Although the reaction time to safety hazards became shorter, the accuracy identification of safety hazards decreased when the emotional valence increased from negative to neutral. The identification accuracy of safety hazards was 92.25% under neutral emotional state, which was greater than the negative emotional state (75.41%) and the positive emotional state (80.60%). The findings are similar to the findings on the effect of emotion on driver performance, with drivers performing better in a neutral state (Jeon et al., 2014) hazards.

Given that the construction workers’ emotions may be influenced by the age (Kappes et al., 2017), age was chosen as the independent variable and a one-way ANOVA was used to explore the effect of emotions on construction workers’ recognition ability of safety hazards at different working ages.

According to the results of safety hazards reaction time, identification accuracy of safety hazards and perception level of safety hazards, correlation and regression analysis were used to explore the quantitative relationship between emotional valence and construction workers’ recognition ability of safety hazards.

The correlations between the positive emotion, negative emotion and the perception level of safety hazards were explored, respectively, by Pearson correlation coefficient. The results indicate a low negative correlation between positive emotion and perception level of safety hazards (r = −0.256, p = 0.000), while negative emotion is moderately negative correlated with the perception level of safety hazards (r = −0.520, p = 0.000).

The correlation analysis shows that the relationship between the emotional state and perception level of safety hazards is an inverted U-shape, so curvilinear regression is used to establish a quadratic expression for regression analysis. The regression model passed the F-test (p = 0.000), and the emotional valence explains 35.7% of the workers’ perception level of safety hazards (R² = 0.357, SE = 0.802, F = 365.999). The regression results are shown in Table 13, and quantitative relationship between perception level of safety hazards and emotional valence of construction workers is shown in Equation (3).

Where

The coefficient of the quadratic term( γ 2 = − 0.617) between the emotional state and the perception level of safety hazards is negative, indicating that the perception level of safety hazards increases and then decreases as the emotional valence increases, reaching the highest point when the emotional state of workers is in neutral emotion as shown in Figure 11. Specifically, when construction workers are in positive emotion, the perception level of safety hazards decreases by 1.93 units for every each increase in their emotional state. When construction workers are in negative emotion, the perception level of safety hazards decreases by 1.772 units for each unit decrease in their emotional state. Therefore, the perception level of safety hazards of construction workers in negative and positive emotions is lower than that in neutral emotions. As construction workers’ emotions change from extremely negative to extremely positive, the perception level of safety hazards changes from low to high and then lower.

There was a negative correlation between reaction time to safety hazards and emotional valence, while the accuracy of safety hazards identification and the perception level of safety hazards had an inverted “U” shape relationship with emotional valence. When workers are under positive emotional valence, the findings are consistent with the affective generalization theory (Johnson and Tversky, 1983), where positive emotion drive construction workers to make optimistic judgements about the construction environment, and therefore lower levels of perception of safety hazards. When under the negative emotional valence, the findings are more in line with the Mood Maintenance Hypothesis (Isen and Patrick, 1983), where construction workers tend to take more risky and aggressive decisions in order to escape from their current negative emotional state, thus underestimating the risks of the environment and lowering the level of safety hazards perceived by workers. In addition, the findings of this study are support the Affective states as information hypothesis proposed by Schwarz and Clore (1981). This theory suggests that emotions simplify people’s risk decision-making process, which people judge things based on their feelings rather than their features, and that emotions can lead to overestimation of events of the same valence. This finding is related to information acquisition and cognitive processes, and subsequent research could be further explored from this perspective. Workers should avoid overexcited emotional states, for each unit increase in emotional valence, the reaction time to safety hazards reduced by 0.556 units. Meanwhile the identification accuracy of safety hazards reduced by 1.35 units, and the level of safety hazards perception reduced by 1.93 units when the emotional valence shift by one unit from neutral emotional state. Workers with high emotional valence have a more relaxed and pleasurable state, with increased reaction speed but reduced ability to judge and perceive safety hazards due to inattentiveness. Construction workers also need to avoid negative emotions such as excessive sadness and grief, for each unit decrease in emotional state, the feedback time for safety hazards recognition increases by 0.556 units. Meanwhile the identification accuracy of safety hazards decreases by 1.308 units and the perception level of safety hazards decreases by 1.772 units. When workers are immersed in a state of loss and frustration, the attention allocated to safety hazards identification decreases, prolonging their own judgment time, with a concomitant decrease in the accuracy of safety hazards identification and the level of safety hazards perception. Therefore, neutral emotions are the safest emotions.