Tunnel traffic risk discrimination method combining Pareto law and machine learning

Abstract

The invention discloses a Pareto law and machine learning combined tunnel traffic risk discrimination method, and the method comprises the steps: building a tunnel traffic risk discrimination model based on the fusion of a risk threshold value and machine learning through employing a random forest learning algorithm based on tunnel traffic operation data; according to the method, collision time TTC is selected as an index for measuring traffic risks by analyzing rear-end collision conflicts, the traffic risks are divided into a high level, a middle level and a low level, a risk boundary value is determined by using a cumulative frequency curve method, a decision basis is effectively provided for tunnel traffic managers, and the tunnel traffic managers are helped to timely and accurately take countermeasures; through model training and testing, results show that the random forest model has the best performance, and the testing accuracy reaches 99.45%.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic risk identification, and in particular to a tunnel traffic risk identification method combining the Pareto principle with machine learning. Background Art

[0002] Highway tunnels are typical special structures in the road traffic system, and traffic flow is severely restricted. Compared with ordinary sections of highways, the closed spatial structure characteristics of highway tunnels create a relatively special traffic operation environment, which makes them easy to become traffic "bottleneck" sections on highways and key nodes of the road traffic system. Due to factors such as poor tunnel traffic environment, limited vision, narrow roads, and large traffic volume, the traffic risk in the tunnel is high, and the degree of accident hazard is also greater. The probability of traffic congestion and traffic accidents in the tunnel is significantly higher than that on the outer roads.

[0003] Therefore, in order to solve the above problems, this paper proposes a tunnel traffic risk identification method that combines the Pareto principle with machine learning. Summary of the invention

[0004] The purpose of the present invention is to design a tunnel traffic risk identification method that combines the Pareto principle with machine learning and has certain practicality and universality, which can effectively provide decision-making basis for tunnel traffic managers and help them make timely and accurate response measures.

[0005] In order to achieve the above technical effects, the present invention is implemented by the following technical solutions: a tunnel traffic risk identification method combining Pareto principle with machine learning, characterized in that building the model includes the following steps:

[0006] S1. Analyze highway tunnel surveillance videos based on video analysis algorithms, automatically mark vehicles, conduct real-time and continuous tracking of multiple target vehicles, realize lane-level traffic parameter detection, and obtain sufficient traffic operation data;

[0007] S2. Preprocess the traffic operation data obtained by removing incomplete data, missing values ​​and outliers from the collected trajectories, and clean the data. Label the data trajectories and unify the format of the corresponding feature data to facilitate the subsequent machine learning algorithm modeling;

[0008] S3, based on the analysis of the three basic parameters of traffic flow, namely, traffic volume, density and speed, combined with the processed traffic operation data, the collision time TTC value of each trajectory is calculated;

[0009] S4, using the quantile of the cumulative frequency curve to determine the risk threshold for the calculated collision time TTC value, and then using machine learning technology to process and divide the data set to classify the traffic risk operation status in the tunnel;

[0010] S5. Construct a tunnel traffic risk discrimination model based on the classified data; the tunnel traffic risk discrimination model is a random forest model.

[0011] S6. Model training: Bring the prepared training set into the model for training, and use the out-of-bag error rate OOB data set for evaluation. Use the oobError function to calculate the OOB error rate and use it as another indicator for model evaluation; set Prune to on, and the model will be pruned during the training process; this can reduce the complexity of the model and improve the generalization ability of the model;

[0012] S7. Model validation: K-fold cross-validation method is used to combine different training and validation set data, and then the combined data set with a data volume of N is divided into K subsets. The data volume of each subset is (N / K). Then, K-1 subsets are used to train the RF model, and the remaining subset is used for model validation. This is repeated K times, and the average K results are used as the final generalization error of the model, and the model with the smallest error value is selected as the optimal model.

[0013] Furthermore, in S1, the traffic operation data includes the time when the vehicle passes through the detection section, the vehicle number, the vehicle type, the lane (L, M, R) where the vehicle is located, the x / y coordinates of the vehicle, the running speed of the vehicle when it passes through the detection section, and the distance between the current vehicle and the previous vehicle that passed the detection section.

[0014] Furthermore, in S4, the traffic risk operation status in the tunnel is classified into three categories: high risk, medium risk and low risk; the specific risk threshold indicators for each level of risk are as follows:

[0015] High risk indicator: less than 2.4s;

[0016] Medium risk index: 2.4-12.1s;

[0017] Low risk indicator: above 12.1s.

[0018] Furthermore, in S4, the data sets are processed and divided using machine learning techniques as follows: predictive variables are obtained by calculating and counting these data sets, TTC collision time is used as the response variable, risk classification results are obtained by determining the risk threshold, and division is performed according to the risk classification results.

[0019] Furthermore, in S6, the prediction variables are lane, vehicle speed, occupancy, headway, headway and speed difference.

[0020] Furthermore, in S6, the model training is specifically as follows: initialize the model, and set OOBPrediction to on, numTrees to 100, and Prune to on. Use the TreeBagger function in Matlab to initialize the model; for each data point, the model will use the RF algorithm to classify from numTrees, i.e., 100 decision trees; finally, compare the model's prediction results with the actual results in the test set, and record the prediction accuracy.

[0021] Furthermore, in S7, the model validation is as follows: in the random forest model, each decision tree compares each data point with the high, medium, and low risk thresholds determined according to the Pareto principle, and then obtains its own classification results; finally, the random forest model combines the classification results of each decision tree, and decides the final classification prediction result based on the voting principle of minority obeys majority, and then compares the prediction result label of the random forest model with the real label of the tunnel traffic data set, and records the prediction accuracy of the model as one of the model evaluation indicators.

[0022] The beneficial effects of the present invention are:

[0023] The present invention collects data based on long tunnels, obtains the microscopic operation information of vehicles inside the tunnel, determines the risk threshold by analyzing the characteristics of tunnel traffic flow, and uses the random forest learning algorithm to construct a tunnel traffic risk discrimination model that integrates the risk threshold and machine learning; through the analysis of rear-end collisions, the collision time TTC is selected as an indicator to measure traffic risk, and the traffic risk is divided into three levels: high, medium and low, and the risk boundary value is determined using the cumulative frequency curve method; it effectively provides decision-making basis for tunnel traffic managers and helps them make timely and accurate response measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0025] Figure 1 It is a schematic diagram of the data extraction process in the data collection process of the present invention;

[0026] Figure 2 is a schematic diagram of the distribution of vehicle types in lanes in a tunnel in Example 1 of the present invention;

[0027] Figure 3 is a speed frequency distribution histogram in a tunnel in Embodiment 1 of the present invention;

[0028] Figure 4 is a diagram showing the relationship between speed and occupancy in a tunnel in Example 1 of the present invention;

[0029] Figure 5 is a diagram showing the relationship between speed and traffic volume in a tunnel in Embodiment 1 of the present invention;

[0030] Figure 6 is a graph showing the relationship between occupancy rate and traffic volume in a tunnel in Embodiment 1 of the present invention;

[0031] Figure 7 is a TTC frequency distribution histogram in the tunnel in Example 1 of the present invention;

[0032] Figure 8 The high, medium and low risk lane distribution map in Example 1 of the present invention;

[0033] Fig. 9 It is a schematic diagram of the GBDT model training results in Example 1 of the present invention;

[0034] Fig.10 Schematic diagram of the RF model training results in Example 1 of the present invention;

[0035] Fig.11 It is a schematic diagram of the KNN model training results in Example 1 of the present invention;

[0036] Fig.12 It is a schematic diagram of the structure of testing the trained model in Example 1 of the present invention using new data. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Example 1

[0039] A tunnel traffic risk identification method combining Pareto principle and machine learning is constructed. The steps are as follows:

[0040] Step 1: Use the video parsing algorithm to parse the highway tunnel surveillance video, automatically mark the vehicle, perform real-time and continuous tracking of multiple target vehicles, realize lane-level traffic parameter detection, and obtain sufficient traffic operation data such as speed, vehicle spacing, and flow. The data sample is shown in Table 2. The acquired traffic operation information is: the time when the vehicle passes the detection section, the vehicle number, the vehicle type (car, truck), the lane where the vehicle is located (L, M, R), the x / y coordinates of the vehicle, the running speed of the vehicle when it passes the detection section, and the distance between the current vehicle and the previous vehicle that passed the detection section.

[0041]

[0042] Table 2 Data sample

[0043] like Figure 1 As shown in the figure, the vehicle operation and trajectory research data in the tunnel were obtained through the Yolo v7 target detector, which has a faster running speed and extremely high recognition accuracy.

[0044] After the preliminary processing of the video data in the tunnel, the trajectory data is preprocessed, and the corresponding feature data is formatted in a unified format to facilitate the subsequent machine learning algorithm modeling. The main operations are data cleaning, filtering out missing values ​​and outliers, and labeling the data trajectory. The processed data are shown in Table 3;

[0045]

[0046]

[0047] Table 3 Example of video data format after processing

[0048] Step 2: Analysis of tunnel traffic flow characteristics

[0049] 2.1 Analysis of vehicle distribution characteristics

[0050] According to the one-hour traffic flow data in the tunnel, Figure 2 As shown in the figure, different lanes on the highway are usually used for vehicles of different speeds; it is observed that the outer lanes of large vehicles account for the highest proportion, and the proportion decreases towards the inner lanes. The proportion of different types of vehicles has a significant impact on the characteristics of traffic flow. The traffic flow in the outer lanes is discontinuous and the speed is slower, which increases the possibility of traffic risks; in contrast, the inner two lanes run faster and have a larger traffic volume, so the inner lanes are more prone to traffic safety risk accidents than the outer lanes;

[0051] 2.2 Speed ​​characteristics analysis

[0052] According to the design speed of the road and the driving capacity of the vehicle, the speed distribution is usually limited, so it presents a certain peak value; in the analysis of tunnel traffic operation data, it is calculated that the 85% vehicle speed of the tunnel is 84.5km / h, and the 15% vehicle speed is 61.6km / h. The speed distribution is similar to the normal distribution, with a low frequency distribution of high speeds, and there are also cases of low frequency distribution of low speeds, such as Figure 3 As shown;

[0053] However, since the speed limit in the tunnel ranges from 60km / h to 80km / h, the actual speed distribution in the tunnel shows a large degree of discreteness; the higher the degree of speed discretization, the greater the speed difference, which indicates a higher traffic risk.

[0054] 2.3 Analysis of three-parameter characteristics of traffic flow

[0055] Since the research on traffic operation risks in tunnels is mainly conducted from a microscopic perspective, this paper analyzes the three basic parameters of traffic flow: traffic volume, density, and speed. This is used to monitor traffic congestion in real time, evaluate traffic flow, and formulate traffic management and planning strategies.

[0056] like Figure 4 As shown, Figure 4 This is the case where the left side is the inner lane and the right side is the middle lane. From the distribution of speed-occupancy sample points, it can be seen that there are fewer samples with speed points exceeding 100km / h, and more vehicles exceed the tunnel speed limit of 80km / h. Most samples obey the model distribution within the reference line, and have a trend of decreasing speed with increasing occupancy, that is, congestion. Due to the randomness of drivers' driving habits, when the occupancy rate is low, that is, the tunnel is unobstructed, the speed is low, and there are also more sample points in the speed limit range of 60km / h to 80km / h.

[0057] like Figure 5 As shown, Figure 5 The left side is the inner lane and the right side is the middle lane. From the distribution of speed-traffic volume sample points, it can be seen that the traffic volume in the middle lane is larger and the maximum value of the speed distribution is lower than that of the inner lane. The distribution of the data is not well fitted and the discrete situation is obvious, but the basic nonlinear relationship can be seen.

[0058] like Figure 6 As shown, Figure 6This is the case where the left side is the inner lane and the right side is the middle lane. From the distribution of sample points in the occupancy-traffic volume relationship diagram, it can be seen that there are more large trucks and heavy vehicles in the outer lane. The larger the vehicle type, the slower the speed, forming a discontinuous traffic flow, so the traffic volume and vehicle speed are significantly reduced compared to the inner lane; there are more personalized data, and the overall curve model is similar to the external section. In this case, traffic risks often occur during congestion. When the traffic conditions in the tunnel are congested, the traffic risk of collision and rear-end collision is higher.

[0059] Step 3: Analyze the three basic parameters of traffic volume, density and speed, and calculate the collision time TTC value of each trajectory in combination with the processed traffic operation data;

[0060] According to the Pareto principle, only about 20% of the factors affect 80% of the results; that is to say, among all the variables, only 20% are the most important. Although the remaining 80% account for the majority, the magnitude of their influence is far less than that of the "critical few". This law is well reflected in traffic speed management. On the speed cumulative frequency distribution curve, the speed of vehicles at the 15%-85% position changes relatively smoothly. Management departments often use the 85% speed as the speed limit for certain sections of the road, that is, the maximum speed limit. The significance of this definition is that by setting a relatively high speed value that is not too high to cause safety hazards, it can not only meet the driving needs of most vehicles, but also ensure traffic safety. Based on this, after data preprocessing, the present invention defines medium and low risks based on the 85% quantile of the TTC cumulative frequency curve inside the tunnel, and defines medium and high risks based on the 15% quantile. Since there are large differences in the severity of the potential consequences of traffic conflicts, the construction of a cumulative frequency curve is used as a tool to graphically represent the frequency and severity of events.

[0061] Step 4: Use the quantile of the cumulative frequency curve to determine the risk threshold for the calculated collision time TTC value, and then use machine learning technology to process and divide the data set to classify the traffic risk operation status in the tunnel; specifically:

[0062] according to Figure 7 The TTC frequency distribution histogram inside the tunnel is shown. The boundary value between medium risk and low risk is 12.1 seconds, and the boundary value between medium risk and high risk is 2.4 seconds. The risk threshold definition method based on traffic operation characteristics can effectively divide traffic risks and provide a basis for the evaluation of traffic safety events.

[0063] According to the relationship between the divided risk level and its corresponding lane, a risk distribution map of each lane is drawn; Figure 8It can be seen intuitively that the high and medium risk ratios of the inner and middle lanes are high, and the medium risk ratio of the outer lanes is slightly higher; due to the high speed of the inner and middle lanes and the inability to overtake inside the tunnel, the collision time indicator describing the rear-end collision is obvious; specifically, the high risk ratio of the inner and middle lanes is 8%, the medium risk ratio is 24%, and the low risk ratio of the outer lanes is 71%. Due to the low speed of the outer lanes and the presence of more heavy vehicles, the risk is not obvious.

[0064] Then, the risk identification model based on machine learning can quickly process large amounts of multidimensional data, extract the features and correlations hidden in the data, identify the patterns and laws of traffic risks, and accurately predict the conflict risks on the road; tunnel traffic risk identification is a classification task, and the classification algorithm in machine learning is required to distinguish high, medium and low risks in tunnels.

[0065] Step 5: Construct a tunnel traffic risk discrimination model based on the classified data; construct a gradient boosting tree model (GBDT in this article), a random forest model (RF in this article), and a K-nearest neighbor model (KNN in this article) for the tunnel traffic risk discrimination model; study the judgment accuracy of the gradient boosting tree model, the random forest model, and the K-nearest neighbor model, and select the optimal model in turn; the details are as follows:

[0066] 5.1 Risk Discrimination Model Based on Gradient Boosting Tree (GBDT Model)

[0067] Gradient Boosting Decision Tree (GBDT) is suitable for processing nonlinear relationships of data and can improve the accuracy of prediction in iterations. It has certain advantages in identifying complex features of tunnel traffic risks, so it is selected as the representative Boosting algorithm to complete the classification task; the model training is shown in Table 4.

[0068]

[0069]

[0070] Table 4 GBDT model training parameters

[0071] like Fig. 9 As shown, it can be seen that the training accuracy of the GBDT model reached 100%.

[0072] 5.2 Risk discrimination model based on random forest (RF model)

[0073] Since the random forest model has strong anti-overfitting ability, it can be classified through the voting results of multiple decision trees, has a certain inhibitory effect on overfitting, has good generalization ability, and can process large-scale data. Due to the parallel processing of random forest, it can effectively process large-scale data sets; therefore, it is selected as the representative of the Bagging algorithm to complete the classification task; the model training is shown in Table 5.

[0074]

[0075] Table 5 RF model training parameters

[0076] During the training process, the out-of-bag error rate (OOB) data set is used for evaluation. These data points are not in the training set of any decision tree and are used to evaluate the generalization ability of the model. The OOB error rate is calculated using the oobError function and used as another metric for model evaluation. Setting Prune to on will cause the model to be pruned during training. This can reduce the complexity of the model and improve the generalization ability of the model.

[0077] like Fig.10 As shown, it can be seen that the training accuracy of the RF model has reached 100%;

[0078] 5.3 Risk Identification Model Based on K Nearest Neighbor (KNN Model)

[0079] Since the K nearest neighbor model is simple and easy to implement, it does not require assumptions about the data or complex training processes, and is easy to implement and adjust. It is also adaptable to nonlinear relationships and has a certain adaptability to the complex characteristics and pattern recognition of tunnel traffic risks. Therefore, the K nearest neighbor model is selected as a single classifier to complete the classification task of this study. The model training is shown in Table 6.

[0080]

[0081] Table 6 KNN model training parameters

[0082] Then the model is trained using the training set and validated using the validation set. For each data point, the model calculates its distance from all data points in the training set and classifies it according to the categories of the five nearest neighbors. The final classification result is obtained. Finally, the predicted result label of the KNN model is compared with the true label of the tunnel traffic data set, and the prediction accuracy of the model is recorded as one of the model evaluation indicators; the training results are shown in Figure 2. Fig.11 As shown, it can be seen that the training results of the KNN model have a high accuracy, reaching 99.43%.

[0083] The following classification result confusion matrix is ​​obtained from the above training process. It can be seen that the training process of the GBDT model and the RF model can obtain 100% classification accuracy; the classification accuracy of the KNN model also reaches 99.43%.

[0084] Step 6: Model validation: Using the K-fold cross validation method, by using different combinations of training and validation sets, the performance of the model on different data subsets can be better evaluated, and the bias problem caused by the distribution of a specific data set can be reduced, and overfitting can be avoided. When randomly dividing the data set, if special samples (such as abnormal data) are assigned to the test set, the model results will be abnormally biased or poor. In order to solve these problems, it is necessary to use the cross validation method.

[0085] Step 7: Using the classification model, considering the indicators such as precision, accuracy and recall, the evaluation index results of GBDT model, RF model and KNN model in traffic risk state identification are explored; as shown in Table 7;

[0086]

[0087] Table 7 Model prediction classification results

[0088] In the table, TP is the true positive, FP is the false positive, FN is the false negative, and TF is the true negative. When the true label of the sample in the dataset is high risk and the model predicts that the sample is also classified as high risk, this situation is called TP; when the true label of the sample in the dataset is not high risk (it can be medium risk or low risk) and the model prediction result is not high risk, it is called TF; when the true label of the sample in the dataset is high risk and the model prediction result is not high risk, it is called FN; when the sample in the dataset is not high risk and is predicted to be high risk, it is called FP.

[0089] Precision: Indicates how many of the samples predicted to be positive are true samples, as shown in Formula 1.

[0090]

[0091] Recall: It indicates the ratio of the number of correctly predicted positive samples to the actual positive samples.

[0092] As shown in formula 2:

[0093]

[0094] The F1 value represents the harmonic mean of Precision and recall, as shown in Formula 3:

[0095]

[0096] Precision, recall, and F1 values ​​are used as indicators to verify the classification performance of the model; Table 8 shows the evaluation index results of the RF model, KNN model, and GBDT model in traffic risk state recognition for comparative analysis. In addition, this paper also uses the naive Bayes classifier and decision tree classifier for data training and compares them with the three models. The results show that the naive Bayes classifier has a poor effect, while the decision tree performs well in risk degree classification, so it is used as the basic classifier for RF and GBDT.

[0097]

[0098] Table 8 Model classification performance

[0099] Judging from the model results, the model with the best performance is the random forest model, with a Precision of 0.9658, a Recall of 0.9828, and an F1 value of 0.9702. Among the several trained models, the GBDT model has the worst performance, with a Precision of 9172, a Recall of 0.8378, and an F1 value of 0.8757.

[0100] In order to obtain a more stable and reliable evaluation of the model performance, multiple training and validation are performed; the K-fold cross-validation method is adopted. By using different combinations of training and validation sets, the performance of the model on different data subsets can be better evaluated, and the bias problems caused by the distribution of specific data sets can be alleviated, avoiding overfitting and other situations.

[0101] Table 9 shows the K-fold loss values ​​during model training. It is found that the GBDT model has the largest loss value, which may be due to overfitting during model training. The KNN model has the smallest loss value, and its training process is basically consistent with the test results, which indicates that the model has good performance.

[0102]

[0103] Table 9 Model K loss

[0104] Then, the trained model is tested on the test data set to test the generalization performance of the model and to judge the classification ability of the new data. By comparing the predicted categories with the actual categories, the following confusion matrix can be obtained: Fig.12), we can see that the KNN model training accuracy did not reach 100%, but the classification accuracy of new data can reach 98.76%. The random forest model has the highest classification accuracy, and its training model accuracy reached 100%, and the accuracy of the test set also reached 99.45%. The generalization ability of the GBDT gradient boosting tree model is poor. As a result, the accuracy rate in the discrimination of medium and low risks can reach about 95%, but in the discrimination of high-risk status, 16% of the data will be misclassified as medium risk. It is speculated here that due to the small amount of high-risk trajectory data, the training process does not accurately extract the features of high-risk data, resulting in low classification accuracy;

[0105] In summary, the random forest model performs best, with higher precision, recall rate and F1 value. Therefore, the tunnel traffic risk identification model based on the random forest model finally established can complete the risk identification task.

Claims

1. A tunnel traffic risk identification method combining Pareto principle and machine learning, characterized in that: Building the model involves the following steps: S1. Analyze highway tunnel surveillance videos based on video analysis algorithms, automatically mark vehicles, conduct real-time and continuous tracking of multiple target vehicles, realize lane-level traffic parameter detection, and obtain sufficient traffic operation data; S2. Preprocess the traffic operation data obtained by removing incomplete data, missing values ​​and outliers from the collected trajectories, and clean the data. Label the data trajectories and unify the format of the corresponding feature data to facilitate the subsequent machine learning algorithm modeling; S3, based on the analysis of the three basic parameters of traffic flow, namely, traffic volume, density and speed, combined with the processed traffic operation data, the collision time TTC value of each trajectory is calculated; S4, using the quantile of the cumulative frequency curve to determine the risk threshold for the calculated collision time TTC value, and then using machine learning technology to process and divide the data set to classify the traffic risk operation status in the tunnel; S5. Constructing a tunnel traffic risk discrimination model based on the classified data; the tunnel traffic risk discrimination model is a random forest model; S6. Model training: Bring the prepared training set into the model for training, and use the out-of-bag error rate OOB data set for evaluation. Use the oobError function to calculate the OOB error rate and use it as another indicator for model evaluation; set Prune to on, and the model will be pruned during the training process; this can reduce the complexity of the model and improve the generalization ability of the model; S7. Model validation: K-fold cross-validation method is used to combine different training and validation set data, and then the combined data set with a data volume of N is divided into K subsets. The data volume of each subset is (N / K). Then, K-1 subsets are used to train the RF model, and the remaining subset is used for model validation. This is repeated K times, and the average K results are used as the final generalization error of the model, and the model with the smallest error value is selected as the optimal model.

2. The tunnel traffic risk identification method combining Pareto principle and machine learning according to claim 1 is characterized in that: In S1, the traffic operation data includes the time when the vehicle passes the detection section, the vehicle number, the vehicle type, the lane (L, M, R) where the vehicle is located, the vehicle's x / y coordinates, the vehicle's running speed when passing the detection section, and the distance between the current vehicle and the previous vehicle that passed the detection section.

3. The tunnel traffic risk identification method combining Pareto principle and machine learning according to claim 1 is characterized in that: In S4, the traffic risk operation status in the tunnel is classified into three categories: high risk, medium risk and low risk; the specific risk threshold indicators for each level of risk are as follows: High risk indicator: less than 2.4s; Medium risk index: 2.4-12.1s; Low risk indicator: above 12.1s.

4. The tunnel traffic risk identification method combining Pareto principle and machine learning according to claim 1 is characterized in that: In S4, the machine learning technology is used to process and divide the data set as follows: predictive variables are obtained by calculating and counting these data sets, TTC collision time is used as the response variable, and risk classification results are obtained by determining the risk threshold, and division is performed according to the risk classification results.

5. The tunnel traffic risk identification method combining Pareto principle and machine learning according to claim 1 is characterized in that: In S6, the predictor variables are lane, speed, occupancy, headway, headway, and speed difference.

6. The tunnel traffic risk identification method combining Pareto principle and machine learning according to claim 1 is characterized in that: In S6, the model training is specifically as follows: initialize the model and set OOBPrediction to on, numTrees to 100, and Prune to on. Use the TreeBagger function in Matlab to initialize the model; for each data point, the model will use the RF algorithm to classify from numTrees, i.e., 100 decision trees; finally, compare the model's prediction results with the actual results in the test set, and record the prediction accuracy.

7. The tunnel traffic risk identification method combining Pareto principle and machine learning according to claim 1 is characterized in that: In S7, the model validation is as follows: in the random forest model, each decision tree compares each data point with the high, medium, and low risk thresholds determined according to the Pareto principle, and then obtains its own classification results; finally, the random forest model combines the classification results of each decision tree, and decides the final classification prediction result based on the majority voting principle, and then compares the prediction result label of the random forest model with the real label of the tunnel traffic data set, and records the prediction accuracy of the model as one of the model evaluation indicators.

Images