Dynamic driving risk assessment method and system based on DNC and NGBoost-PA-WOE

Abstract

The invention provides a dynamic driving risk assessment method and system based on DNC and NGBoost-PA-WOE, and the system comprises a database which is used for storing an urban road traffic map, historical traffic accident data and historical traffic violation data, and can automatically update the data; the data acquisition module is used for acquiring vehicle dynamic data and environment characteristics; the data preprocessing and cleaning module is used for preprocessing and cleaning the data; the degree centrality factor (DNC) analysis module is used for calculating a DNC value of a target road node; the evidence weight module is used for calculating WOE values corresponding to the accident rate, the violation rate, the vehicle dynamic data and the environment characteristics of the nodes; and the risk assessment module is used for predicting the risk score of the road node according to the DNC value, the WOE value, the accident rate and the violation rate. According to the method, the NGBoost-PA-WOE model is adopted to predict the risk score of the road node, the model combines the NGBoost algorithm, the FPA and the WOE, and the prediction precision of the risk score can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of road traffic safety, and in particular to a dynamic driving risk assessment method and system based on DNC and NGBoost-PA-WOE. Background Art

[0002] With the rapid development and widespread application of intelligent transportation systems, how to effectively evaluate and manage driving risks in traffic networks has become an important topic in the current research field. Traditional static risk assessment methods often cannot timely reflect the impact of dynamic changes in traffic conditions on driving safety, especially in high-density and complex traffic environments. This limitation is more obvious. Therefore, dynamic risk assessment methods based on real-time data have gradually become the focus of research and application.

[0003] Existing dynamic risk assessment methods mainly rely on on-board sensors and traffic monitoring equipment to collect real-time vehicle data, and process and analyze it through machine learning or data analysis technology. These technologies can provide instant traffic status information, helping traffic managers and drivers better understand the current traffic conditions and make safer decisions. However, despite the success of these methods, they still have some shortcomings.

[0004] First, most existing methods only consider vehicle operation data (traffic flow, speed, and driving distance), but do not fully integrate the impact of environmental characteristics (weather, road conditions, and lighting) on ​​driving safety. Second, existing dynamic risk assessment systems usually refer to historical traffic accident data to predict the current traffic risk level in real time, but the prediction accuracy still needs to be improved. Summary of the invention

[0005] In view of the problems of poor accuracy and insufficient reference factors of the existing dynamic driving risk assessment system, the present invention provides a dynamic driving risk assessment method and system based on DNC and NGBoost-PA-WOE. The system adopts the NGBoost-PA-WOE model to predict the risk score of road nodes. The model combines the NGBoost algorithm, penalty attributes (FPA) and weighted evidence (WOE), which can effectively improve the prediction accuracy of risk scores.

[0006] The present invention achieves the above technical objectives through the following technical means.

[0007] A dynamic driving risk assessment system based on DNC and NGBoost-PA-WOE, comprising:

[0008] A database for storing urban road traffic maps, historical traffic accident data and historical traffic violation data at various road nodes in the city, and automatically updating data as traffic accidents and violations occur in real time at various road nodes in the city;

[0009] The data acquisition module includes sensors, on-board equipment and traffic monitoring systems deployed on road nodes and vehicles, which are used to collect vehicle dynamic data and environmental characteristics at each road node in real time;

[0010] The data preprocessing and cleaning module is used to remove noise, fill missing values, and detect and correct abnormal data from the raw data obtained by the data acquisition module;

[0011] Degree Centrality Factor (DNC) analysis module, used to calculate the DNC value of each road node based on the urban road traffic map;

[0012] The evidence weight module is used to calculate the evidence weight WOE value, accident rate, and violation rate corresponding to the vehicle dynamic data and environmental characteristics at each road node based on historical traffic accident data and historical traffic violation data, and normalize the calculation results;

[0013] The risk assessment module is the NGBoost-PA-WOE model, which is used to predict the risk score at each road node based on the DNC value of each road node, the WOE value, accident rate, violation rate corresponding to the vehicle dynamic data and environmental characteristics at the corresponding road node, and the vehicle dynamic data and environmental characteristics at each road node collected by the data acquisition module, and give traffic control recommendations based on the risk score.

[0014] Furthermore, the data acquisition module includes vehicle-mounted sensors, traffic monitoring equipment and a GPS system; the vehicle dynamic data includes vehicle flow, vehicle speed and driving distance; and the environmental characteristics include weather, road conditions and lighting.

[0015] Furthermore, the DNC value of the target road node is calculated as follows: the road nodes in the urban road traffic map in the database are taken as nodes, the road sections between the road nodes are taken as edges, the target road node is recorded as node i, and all first-order neighbor nodes connected to node i are recorded as nodes j. The degree centrality value of node i is calculated as follows:

[0016] A1: Count the number of nodes j, denoted as k i ;

[0017] A2: Calculate the local clustering coefficient of all nodes j, denoted as C j , and its calculation formula is:

[0018]

[0019] Among them, E j is the number of interconnected edges between all first-order neighbor nodes of node j, k j is the number of all first-order neighbor nodes connected to node j;

[0020] A3: The degree centrality value of node i, denoted as DNC(i), is calculated as follows:

[0021] Where a is the correlation coefficient and a>0, is the sum of the local clustering coefficients of all nodes j, F i is the set of first-order neighbor nodes of node i.

[0022] Furthermore, the WOE value is calculated as follows:

[0023]

[0024] Among them, N pix1 is the total number of event pixels belonging to a specific factor category, N pix2 is the total number of pixels that are not inside a specific factor category but outside the event, N pix3 is the total number of pixels not associated with the event, N pix4 N is the total number of pixels that are neither in a specific factor category nor associated with an event. pixX N pix3 or N pix4 , when the proportion of samples without accidents or violations under a certain feature is not lower than the proportion of samples with accidents or violations, when calculating the WOE value of the weight of evidence corresponding to the feature, N pixX Take N pix3 ; When the proportion of samples with accidents or violations under a certain feature is higher than the proportion of samples without accidents or violations, when calculating the WOE value of the weight of evidence corresponding to the feature, N pixX Take N pix4 .

[0025] Furthermore, the method for constructing the NGBoost-PA-WOE model comprises the following steps:

[0026] B1: Preprocess and clean the historical traffic accident data and historical traffic violation data of each road node, and randomly divide them into training sets and test sets;

[0027] B2: Calculate the weight of evidence WOE value, accident rate, violation rate corresponding to the vehicle dynamic data and environmental characteristics at each road node, as well as the DNC value corresponding to each road node, and normalize the calculation results;

[0028] B3: Use the training set data to establish the NGBoost-PA-WOE initialization model based on the NGBoost algorithm; use the test set data to verify the performance of the NGBoost-PA-WOE initialization model, and adjust the model parameters according to the verification results;

[0029] B4: Use K-fold cross-validation technology to verify the model performance again, and adjust the model parameters again according to the verification results, and finally establish the NGBoost-PA-WOE model; the verification indicators include AUC, accuracy, recall rate and F1 score.

[0030] Furthermore, in step B3, the method for establishing the NGBoost-PA-WOE initialization model is as follows:

[0031] B3.1: Use the loss scoring rule S to determine the initial model parameters θ by optimizing the loss function S (0) , the initial parameters of the model θ (0) The parameters of the marginal distribution corresponding to the training set;

[0032]

[0033] Where n is the number of samples; y i is the label of the i-th sample in the training set;

[0034] B3.2: Use penalty attribute (FPA) to analyze the importance of each feature from historical traffic accident data and historical traffic violation data, and assign it an initial λ value; then calculate the weight distribution of the feature:

[0035]

[0036] Among them, WR λ is the feature weight range, λ is the feature level, p is the correlation coefficient and p>0;

[0037] B3.3: For each sample i, calculate the gradient of the loss function S with respect to the model parameters θ Apply Fisher information matrix I s (θ) to adjust the gradient and obtain the natural gradient

[0038]

[0039] Each iteration m ranges from 1 to M, where m represents the current iteration number; M represents the total number of iterations; λi 表 The initial weight adjustment factor λ calculated by the penalty attribute (FPA) i ;

[0040] B3.4: Use Natural Gradients To train the base learner f (m) , using DNC value, WOE value, accident rate, and violation rate as input data:

[0041]

[0042] Among them, x i Represents each sample in the test set;

[0043] B3.5: Determine the scaling factor ρ by minimizing the adjusted loss scoring rule S (m) ;

[0044]

[0045] B3.6: Adjust the scaling factor ρ by combining the weight of the penalty attribute (FPA) (m) , the adjusted formula is:

[0046]

[0047] B3.7: Using learning rate n and scaling factor ρ (m) Update model parameters θ;

[0048]

[0049] B3.8: For each sample xi in the test set, calculate the final model parameters θ(xi);

[0050]

[0051] B3.9: Use the final model parameters θ(xi) to generate the predicted probability distribution for each sample;

[0052] B3.10: Calculate risk score based on predicted probability distribution:

[0053] RiskScore i =P θ(xi) (y=1|xi)

[0054] Among them, RiskScore i represents the risk score; p represents the predicted probability.

[0055] A dynamic driving risk assessment method based on the dynamic driving risk assessment system described in any one of the above items comprises the following steps:

[0056] S1: Real-time collection of vehicle dynamic data and environmental characteristics at each road node, and removal of noise, filling of missing values, and detection and correction of abnormal data;

[0057] S2: The risk assessment module obtains the vehicle dynamic data and environmental characteristics at each road node collected by the data acquisition module, and predicts the risk score at each road node based on the DNC value of each road node, the WOE value, accident rate, and violation rate corresponding to the vehicle dynamic data and environmental characteristics at the corresponding road node, and gives traffic control recommendations based on the risk score.

[0058] Furthermore, during operation, the NGBoost-PA-WOE model updates the DNC value of each road node, the WOE value of the evidence weight corresponding to the vehicle dynamic data and environmental characteristics, the accident rate, and the violation rate based on the changes in the urban road traffic map and the recent historical traffic accident data and historical traffic violation data, and updates and optimizes the NGBoost-PA-WOE model.

[0059] Furthermore, the specific method of giving traffic control recommendations based on the risk score is: comparing the risk score of each road node with the preset safety value. If it exceeds the safety value, based on the degree of influence of the vehicle dynamic data on the risk score, the NGBoost-PA-WOE model uses the preset safety value, the accident rate and violation rate of the risk node, and the environmental characteristics of the current risk node, and uses the vehicle dynamic data that significantly affects the risk score as the adjustment parameter to calculate the target value of the vehicle dynamic data that currently makes the risk score at the risk node reach the preset safety value, and determine the traffic control recommendations based on the target value.

[0060] The beneficial effects of the present invention are as follows:

[0061] 1. The NGBoost-PA-WOE model adopted in the present invention is an innovative combination model that combines the NGBoost algorithm, penalty attributes (FPA) and weighted evidence (WOE), wherein the NGBoost algorithm is responsible for providing probabilistic predictions and handling model uncertainty. It uses natural gradients to optimize the prediction distribution to adapt to different types of data and complex data relationships. The penalty attribute (FPA) is still retained as an integral part of the model to control the importance of different attributes in the model decision-making process by adjusting the λ parameter. This helps to refine the performance of the model when dealing with attributes with different importance. The weighted evidence (WOE) evaluates the risk level of each node by calculating specific positive and negative weights, which is particularly important for risk assessment, especially in scenarios where the influence of risk factors needs to be accurately calculated. The NGBoost-PA-WOE model that combines the NGBoost algorithm, penalty attributes (FPA) and weighted evidence (WOE) has greatly improved the prediction accuracy of road node risk scores.

[0062] 2. The NGBoost-PA-WOE model in the present invention introduces a penalty mechanism. On the one hand, by reducing the reliance on these risk features, the model can better generalize to different traffic conditions without being limited to known high-risk features. On the other hand, it reduces the reliance on independent high-risk features and tends to look for interactions between features. In addition, by reducing the reliance on features with negative WOE values, it helps the model maintain stability and avoids the overall prediction effect being affected by the uncertainty of a few features. Unless "night driving" is combined with other high-risk features to significantly improve the prediction ability, the model will tend to reduce its reliance on this feature to ensure the accuracy and robustness of the overall prediction. In this way, the model can effectively avoid the overall prediction effect being affected by the uncertainty of a few features while maintaining its generalization ability, thereby achieving better prediction performance.

[0063] 3. The present invention inputs the weight of evidence WOE value, accident rate, violation rate corresponding to the vehicle dynamic data and environmental characteristics at the road node, and the DNC value of the road node into the NGBoost-PA-WOE model to obtain the risk score of the corresponding node. Compared with the traditional dynamic driving risk assessment system, the present invention can capture more dynamic factors that affect driving safety, thereby enhancing the accuracy and reliability of risk prediction. In addition, the present invention also analyzes the interaction between different factors and improves the accuracy and stability of risk assessment by optimizing the weight distribution. Ultimately, this module can effectively identify high-risk areas and conditions that are closely related to the occurrence of traffic accidents, and provide forward-looking risk warnings and decision-making support for traffic management.

[0064] 4. The present invention provides real-time safety decision-making suggestions based on risk scores, such as reducing vehicle speed, maintaining a certain distance between vehicles, adjusting traffic signals, etc., to reduce the risk score at the road node. This strategy not only helps to reduce the occurrence of traffic accidents, but also improves the efficiency and fluency of the entire urban transportation system, providing a scientific and systematic solution for urban traffic management, and greatly enhancing the adaptability and management efficiency of the transportation system to complex environments.

[0065] 5. The present invention initializes the NGBoost-PA-WOE model based on historical traffic accident data. During operation, the model can also be updated and iterated using real-time traffic accident data to ensure that the model can automatically adjust its model parameters to better fit the current traffic pattern, so that the model has the ability of self-learning and improvement, avoiding prediction bias caused by outdated data, thereby improving the accuracy of risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a module diagram of the dynamic driving risk assessment system based on DNC and NGBoost-PA-WOE described in the present invention.

[0067] Figure 2 This is a schematic diagram of the DNC value calculation principle of each node in the traffic network calculated in a specific embodiment of the present invention.

[0068] Figure 3 for Figure 2 A display diagram of the weight of evidence WOE value, accident rate, and violation rate corresponding to the vehicle dynamic data and environmental characteristics at the middle node 8.

[0069] Figure 4 Flowchart for constructing the NGBoost-PA-WOE model.

[0070] Figure 5 for Figure 2 A risk score graph of each node in the transportation network of the specific embodiment is shown. DETAILED DESCRIPTION

[0071] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0072] The dynamic driving risk assessment system based on DNC and NGBoost-PA-WOE of the present invention comprises a database, a data acquisition module, a data preprocessing and cleaning module, a degree centrality factor (DNC) analysis module, an evidence weight module and a risk assessment module. Figure 1 This is a module diagram of the dynamic driving risk assessment system.

[0073] The database is used to store the city road traffic map, historical traffic accident data and historical traffic violation data at various road nodes in the city, and automatically updates the data as traffic accidents and violation events occur in real time at various road nodes in the city.

[0074] The data acquisition module includes sensors, on-board equipment and traffic monitoring systems deployed on road nodes and vehicles, which are used to collect vehicle dynamic data and environmental characteristics at each road node in real time. The vehicle dynamic data includes vehicle flow, vehicle speed and driving distance. The environmental characteristics include weather, road conditions and lighting.

[0075] The data preprocessing and cleaning module is used to remove noise, fill missing values, and detect and correct abnormal data from the original data obtained by the data acquisition module.

[0076] The degree centrality factor (DNC) analysis module is used to calculate the DNC value of each road node according to the urban road traffic map. The present invention uses the degree centrality factor (DNC) to measure the importance and connection strength of the node in the traffic network. The higher the degree centrality value, the greater the importance of the node in the network, and the higher the traffic flow and interaction frequency. Based on the calculation results of the degree centrality factor, the key nodes that have a significant impact on the overall network risk are identified, thereby providing an important basis for subsequent risk assessment and decision-making.

[0077] The weight of evidence module is used to calculate the weight of evidence WOE value, accident rate, and violation rate corresponding to the vehicle dynamic data and environmental characteristics at each road node based on historical traffic accident data and historical traffic violation data, and normalize the calculation results. By encoding the weight of evidence (WOE) of vehicle dynamic data and environmental characteristics, the impact of these variables on the risk of accidents is quantified.

[0078] The risk assessment module is the NGBoost-PA-WOE model, which is used to predict the risk score of each road node based on the DNC value of each road node, the WOE value, accident rate, violation rate corresponding to the vehicle dynamic data and environmental characteristics at the corresponding road node, and the vehicle dynamic data and environmental characteristics at each road node collected by the data acquisition module, and give traffic control suggestions based on the risk score

[0079] Furthermore, in the degree centrality factor (DNC) analysis module, the DNC value is calculated as follows: based on the urban traffic network diagram, the road nodes in the traffic network diagram are defined as nodes, and the road sections between the road nodes are defined as edges. The local clustering coefficient of each node is calculated using complex network theory to measure the degree of connectivity between the node and its neighbors.

[0080] The target road node is recorded as node i, and all first-order neighbor nodes connected to node i are recorded as nodes j. The degree centrality value of node i is calculated as follows:

[0081] A1: Count the number of nodes j, denoted as k i .

[0082] A2: Calculate the local clustering coefficient of all nodes j, denoted as C j , and its calculation formula is:

[0083]

[0084] Among them, E j is the number of interconnected edges between all first-order neighbor nodes of node j, k j is the number of all first-order neighbor nodes connected to node j.

[0085] A3: Calculate the degree centrality value of node i, denoted as DNC(i), and its calculation formula is:

[0086] Where a is the correlation coefficient and a>0, is the sum of the local clustering coefficients of all nodes j, F i is the set of first-order neighbor nodes of node i.

[0087] In order to more intuitively demonstrate the calculation process of degree centrality value, a directed graph is established to simulate the traffic network, such as Figure 2 The DNC value calculation principle diagram of each node in the traffic network shown in FIG. 8 is taken as the node to be calculated in this embodiment. First, the degree of node 8, that is, the number of edges directly connected to node 8, is determined to be 5. Then, the first-order neighbor nodes of node 8 are determined, which are node 0, node 2, node 3, node 11 and node 14, and the local clustering coefficient of each first-order neighbor node is calculated. Taking node 0 as an example, the local clustering coefficient of node 0 is recorded as C0, and its calculation formula is:

[0088]

[0089] Among them, E0 is the number of interconnected edges between all first-order neighbor nodes of node 0, and k0 is the number of all first-order neighbor nodes connected to node 0. Figure 3 In the equation, the first-order neighbor nodes of node 0 are node 6, node 8, node 9, and node 12, and the number of neighbors k0 = 4. Since node 6 and node 8 are not connected, node 6 and node 9 are not connected, node 6 and node 12 are not connected, node 8 and node 9 are connected, node 8 and node 12 are not connected, and node 9 and node 12 are not connected, E0 = 1. Therefore Similarly, we can obtain C2=0.4,C3=0.5,C 11 =0.1, C 14 =0.2. The local clustering system is summarized into a table, and the results are shown in Table 1.

[0090] Table 1 Summary of local clustering coefficients of first-order neighbors of node 8 in the traffic network

[0091]

[0092] Finally, the degree centrality value of node 8 is calculated, denoted as DNC(8), and its calculation formula is: Let a = 1, and we get:

[0093] Furthermore, in the weight of evidence module, the WOE value is calculated as follows:

[0094]

[0095] Among them, N pix1 is the total number of event pixels belonging to a specific factor category, N pix2 is the total number of pixels that are not inside a specific factor category but outside the event, N pix3 is the total number of pixels not associated with the event, N pix4 N is the total number of pixels that are neither in a specific factor category nor associated with an event. pixX N pix3 or N pix4 , when the proportion of samples without accidents or violations under a certain feature is not lower than the proportion of samples with accidents or violations, when calculating the WOE value of the weight of evidence corresponding to the feature, N pixX Take N pix3 ; When the proportion of samples with accidents or violations under a certain feature is higher than the proportion of samples without accidents or violations, when calculating the WOE value of the weight of evidence corresponding to the feature, N pixX Take N pix4 .

[0096] by Figure 2 Take the traffic network diagram shown as an example, where the accident rate at node 8 is 4.5% and the violation rate is 29.4%. The features of node 8 collected by the data acquisition module include foggy weather, slippery roads, nighttime, low-speed driving, and sparse traffic. The weight of evidence WOE values, accident rates, and violation rates corresponding to each feature are as follows:

[0097] Foggy Days:

[0098] WOE=1.80: Indicates that the risk of accidents increases significantly under foggy conditions.

[0099] Accident rate = 7.5%: In foggy conditions, the proportion of traffic accidents is higher.

[0100] Violation rate = 45%: The highest rate of violations occurred in foggy conditions.

[0101] Slippery roads:

[0102] WOE = 0.60: indicates that the accident risk increases when the road is slippery, but is lower than in foggy days.

[0103] Accident rate = 5.0%: The accident rate on slippery roads is significantly higher than the average level.

[0104] Violation rate = 28%: The proportion of violations on slippery roads is higher.

[0105] Less traffic:

[0106] WOE = -1.00: Under conditions of sparse traffic, the risk of accidents is low and the overall environment is relatively safe.

[0107] Accident rate = 1.2%: When traffic is light, the accident rate is significantly lower than the average level.

[0108] Violation rate = 10%: Due to relatively loose road conditions and fewer violations, drivers tend to abide by traffic rules.

[0109] at night:

[0110] WOE = -1.50: The risk of accidents while driving at night is relatively low, but you still need to pay attention to safety.

[0111] Accident rate = 6.8%: Despite the negative WOE value, the accident rate is still high at night.

[0112] Violation rate = 35%: The violation rate is higher for nighttime driving, especially in poorly lit areas.

[0113] Low speed driving:

[0114] WOE = 0.20: When driving at low speeds, the risk of an accident is relatively small, but there is a certain risk of an accident due to the slow speed.

[0115] Accident rate = 2.0%: The accident rate is lower when driving at low speeds, and the risk control is better.

[0116] Violation rate = 15%: Violations when driving at low speeds are relatively rare, but you still need to pay attention to obey traffic rules and avoid distractions.

[0117] Figure 3 for Figure 2 A display diagram of the weight of evidence WOE value, accident rate, and violation rate corresponding to the vehicle dynamic data and environmental characteristics at the middle node 8.

[0118] The NGBoost-PA-WOE model is constructed as follows Figure 4 As shown, the following steps are included:

[0119] B1: Preprocess and clean the historical traffic accident data and historical traffic violation data of each road node, and randomly divide them into training set and test set in a ratio of 7:3.

[0120] B2: Calculate the weight of evidence WOE value, accident rate, violation rate corresponding to the vehicle dynamic data and environmental characteristics at each road node, as well as the DNC value corresponding to each road node, and normalize the calculation results.

[0121] B3: Use the training set data to establish the NGBoost-PA-WOE initialization model based on the NGBoost algorithm. Use the test set data to verify the performance of the NGBoost-PA-WOE initialization model, and adjust the model parameters according to the verification results. Specifically, the construction method of the NGBoost-PA-WOE initialization model is as follows:

[0122] B3.1: Use the loss scoring rule S to determine the initial model parameters θ by optimizing the loss function S (0) , the initial parameters of the model θ (0) Parameters of the marginal distribution corresponding to the training set.

[0123]

[0124] Where n is the number of samples; y i is the label of the i-th sample in the training set.

[0125] B3.2: Use penalty attribute (FPA) to analyze the importance of each feature from historical traffic accident data and historical traffic violation data, and assign it an initial λ value; then calculate the weight distribution of the feature:

[0126]

[0127] Among them, WR λ is the feature weight range; λ is the level of the feature; p is the correlation coefficient and p>0. The correlation coefficient p is used to adjust the weight range to avoid overlap between different levels. p ensures that as λ increases, the new weight range does not overlap with the weight range of the previous level, which helps to maintain the independent influence of each attribute level. With this setting, the algorithm can maintain the hierarchy and clarity of the attribute weights in the decision tree. The base e is used to pass the exponential function Forming a nonlinear weight scaling that varies with λ, as λ increases, will decrease, making the weight range more concentrated and small. This exponential form can make the weight adjustment more sensitive and diverse, adapting to various attributes and their importance. At the same time, the correlation coefficient p is introduced to avoid overlap between features of different levels and help maintain the independent influence of each feature level. Through this setting, the algorithm can maintain the hierarchy and clarity of the weights of each feature in the random forest.

[0128] B3.3: For each sample i, calculate the gradient of the loss function S with respect to the model parameters θ Apply Fisher information matrix I s (θ) to adjust the gradient and obtain the natural gradient

[0129]

[0130] In each iteration, m ranges from 1 to M, where m represents the current iteration number; M represents the total number of iterations; λ i Represents the initial weight adjustment factor λ calculated by the penalty attribute (FPA) i .

[0131] B3.4: Use Natural Gradients To train the base learner f (m) , using DNC value, WOE value, accident rate, and violation rate as input data:

[0132]

[0133] Among them, x i represents each sample in the test set.

[0134] B3.5: Determine the scaling factor ρ by minimizing the adjusted loss scoring rule S (m) :

[0135]

[0136] B3.6: Adjust the scaling factor ρ by combining the weight of the penalty attribute (FPA) (m) , the adjusted formula is:

[0137]

[0138] B3.7: Using learning rate n and scaling factor ρ (m) Update model parameters θ:

[0139]

[0140] B3.8: For each sample xi in the test set, calculate the final model parameters θ(xi):

[0141]

[0142] B3.9: Use the final model parameters θ(xi) to generate the predicted probability distribution for each sample.

[0143] B3.10: Calculate risk score based on predicted probability distribution:

[0144] RiskScore i =P θ(xi) (y=1|xi)

[0145] Among them, RiskScore i represents the risk score; p represents the predicted probability.

[0146] B4: Use K-fold cross validation technology to verify the model performance again, and adjust the model parameters again according to the verification results, and finally establish the NGBoost-PA-WOE model. The verification indicators include AUC, accuracy, recall and F1 score.

[0147] Since over-reliance on risk features may lead to overfitting, negative WOE value features have limited predictive effects when used alone, and in order to balance the complexity and stability of the model, the present invention introduces a penalty mechanism. On the one hand, by reducing reliance on these risk features, the model can be better generalized to different traffic conditions, not limited to known high-risk features. On the other hand, it reduces reliance on independent high-risk features and tends to look for interactions between features. In addition, by reducing reliance on negative WOE value features, it helps the model maintain stability and avoids the overall prediction effect being affected by the uncertainty of a few features. Unless "night driving" is combined with other high-risk features (such as slippery roads) to significantly improve the prediction ability, the model will tend to reduce reliance on this feature to ensure the accuracy and robustness of the overall prediction. In this way, the model can effectively avoid the overall prediction effect being affected by the uncertainty of a few features while maintaining generalization ability, thereby achieving better prediction performance.

[0148] This embodiment also relates to a dynamic driving risk assessment method based on the above-mentioned dynamic driving risk assessment system, comprising the following steps:

[0149] S1: Collect vehicle dynamic data and environmental characteristics at each road node in real time, remove noise, fill missing values, and detect and correct abnormal data.

[0150] S2: The risk assessment module obtains the vehicle dynamic data and environmental characteristics at each road node collected by the data acquisition module, and predicts the risk score at each road node based on the DNC value of each road node, the WOE value, accident rate, and violation rate corresponding to the vehicle dynamic data and environmental characteristics at the corresponding road node, and gives traffic control recommendations based on the risk score.

[0151] During operation, in order to enable the NGBoost-PA-WOE model to better fit the current traffic pattern and avoid prediction bias caused by outdated data, the NGBoost-PA-WOE model updates the DNC value of each road node, the WOE value of the evidence weight corresponding to the vehicle dynamic data and environmental characteristics, the accident rate, and the violation rate based on the changes in the urban road traffic map and the recent historical traffic accident data and historical traffic violation data, and updates and optimizes the NGBoost-PA-WOE model.

[0152] Furthermore, during operation, the risk score of each road node is obtained by inputting the DNC value of each road node, the WOE value, accident rate, violation rate corresponding to the vehicle dynamic data and environmental characteristics at the corresponding road node, and the vehicle dynamic data and environmental characteristics at each road node collected by the data acquisition module. Figure 2 Take the traffic network diagram shown as an example. Figure 5 for Figure 2 The risk score diagram of each node in the traffic network is shown, in which the safety values ​​are set to 0.33 and 0.66, and the target road nodes are divided into low-risk areas, medium-risk areas or high-risk areas according to 0-0.33, 0.33-0.66, and 0.66-1. The low-risk area means that the risk level is low and only irregular monitoring is required. The medium-risk area means that the risk is still within the controllable range, but there is a certain degree of uncertainty and more attention needs to be paid. The high-risk area means a higher probability of accidents or potential risks, and vehicles in this area require strict intervention measures.

[0153] After obtaining the risk scores of the above-mentioned nodes, the risk assessment module can also give traffic control suggestions based on the risk scores. The specific method is: compare the risk score of each road node with the preset safety value. If it exceeds the safety value, based on the degree of influence of the vehicle dynamic data on the risk score, the NGBoost-PA-WOE model calculates the target value of the vehicle dynamic data that makes the risk score at the risk node reach the preset safety value based on the preset safety value, the accident rate and violation rate of the risk node, and the environmental characteristics of the current risk node, and uses the vehicle dynamic data that significantly affects the risk score as the adjustment parameter, and determines the traffic control recommendations based on the target value.

[0154] Driving advice for traffic managers includes:

[0155] ① Reduce speed limits at high-risk road nodes. Speed ​​limits can reduce the possibility of accidents because lower speeds provide more reaction time.

[0156] ② Adjusting the duration of traffic lights and extending the green light time at high-risk road nodes can ease traffic congestion, reduce vehicle dwelling time, and reduce accident risks.

[0157] ③ Adjust the traffic flow distribution, guide vehicles to alternative routes, reduce the vehicle density at high-risk road nodes, increase traffic fluidity, and reduce the chance of collision. At the same time, restrict large vehicles or use them in different time periods to reduce the traffic load at high-risk road nodes.

[0158] Driving advice provided to drivers includes:

[0159] ① Remind the driver to reduce the current speed to below the speed limit at high-risk road nodes, for example, to 80% to 90% of the speed limit.

[0160] ② Based on the extension of the green light duration, the driver will be notified of the green light extension information in advance through the vehicle-mounted system, prompting him or her that it is safe to pass and avoid accidents caused by sudden braking.

[0161] ③ When the system detects that the vehicle density near a high-risk road node has reached a critical value, for example, more than 100 vehicles per square kilometer, the on-board system will remind the driver to maintain a safe distance.

[0162] The embodiments are preferred implementations of the present invention, but the present invention is not limited to the above-mentioned implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essential content of the present invention belong to the protection scope of the present invention.

Claims

1. A dynamic driving risk assessment system based on DNC and NGBoost-PA-WOE, characterized in that: include: A database for storing urban road traffic maps, historical traffic accident data and historical traffic violation data at various road nodes in the city, and automatically updating data as traffic accidents and violations occur in real time at various road nodes in the city; The data acquisition module includes sensors, on-board equipment and traffic monitoring systems deployed on road nodes and vehicles, which are used to collect vehicle dynamic data and environmental characteristics at each road node in real time; The data preprocessing and cleaning module is used to remove noise, fill missing values, and detect and correct abnormal data from the raw data obtained by the data acquisition module; Degree Centrality Factor (DNC) analysis module, used to calculate the DNC value of each road node based on the urban road traffic map; The evidence weight module is used to calculate the evidence weight WOE value, accident rate, and violation rate corresponding to the vehicle dynamic data and environmental characteristics at each road node based on historical traffic accident data and historical traffic violation data, and normalize the calculation results; The risk assessment module is the NGBoost-PA-WOE model, which is used to predict the risk score at each road node based on the DNC value of each road node, the WOE value, accident rate, violation rate corresponding to the vehicle dynamic data and environmental characteristics at the corresponding road node, and the vehicle dynamic data and environmental characteristics at each road node collected by the data acquisition module, and give traffic control recommendations based on the risk score.

2. The dynamic driving risk assessment system according to claim 1, characterized in that: The data acquisition module includes vehicle-mounted sensors, traffic monitoring equipment and a GPS system; the vehicle dynamic data includes vehicle flow, vehicle speed and driving distance; and the environmental characteristics include weather, road conditions and lighting.

3. The dynamic driving risk assessment system according to claim 1, characterized in that: The DNC value of the target road node is calculated as follows: the road nodes in the urban road traffic map in the database are taken as nodes, the road sections between the road nodes are taken as edges, the target road node is recorded as node i, and all first-order neighbor nodes connected to node i are recorded as nodes j. The degree centrality value of node i is calculated as follows: A1: Count the number of nodes j, denoted as k i ; A2: Calculate the local clustering coefficient of all nodes j, denoted as C j , and its calculation formula is: Among them, E j is the number of interconnected edges between all first-order neighbor nodes of node j, k j is the number of all first-order neighbor nodes connected to node j; A3: The degree centrality value of node i, denoted as DNC(i), is calculated as follows: Where a is the correlation coefficient and a>0, is the sum of the local clustering coefficients of all nodes j, F i is the set of first-order neighbor nodes of node i.

4. The dynamic driving risk assessment system according to claim 1, characterized in that: The calculation method of the WOE value is: Among them, N pix1 is the total number of event pixels belonging to a specific factor category, N pix2 is the total number of pixels that are not inside a specific factor category but outside the event, N pix3 is the total number of pixels not associated with the event, N pix4 N is the total number of pixels that are neither in a specific factor category nor associated with an event. pixX N pix3 or N pix4 , when the proportion of samples without accidents or violations under a certain feature is not lower than the proportion of samples with accidents or violations, when calculating the WOE value of the weight of evidence corresponding to the feature, N pixX Take N pix3 ; When the proportion of samples with accidents or violations under a certain feature is higher than the proportion of samples without accidents or violations, when calculating the WOE value of the weight of evidence corresponding to the feature, N pixX Take N pix4 .

5. The dynamic driving risk assessment system according to claim 1, characterized in that: The method for constructing the NGBoost-PA-WOE model comprises the following steps: B1: Preprocess and clean the historical traffic accident data and historical traffic violation data of each road node, and randomly divide them into training sets and test sets; B2: Calculate the weight of evidence WOE value, accident rate, violation rate corresponding to the vehicle dynamic data and environmental characteristics at each road node, as well as the DNC value corresponding to each road node, and normalize the calculation results; B3: Use the training set data to establish the NGBoost-PA-WOE initialization model based on the NGBoost algorithm; use the test set data to verify the performance of the NGBoost-PA-WOE initialization model, and adjust the model parameters according to the verification results; B4: Use K-fold cross-validation technology to verify the model performance again, and adjust the model parameters again according to the verification results, and finally establish the NGBoost-PA-WOE model; the verification indicators include AUC, accuracy, recall rate and F1 score.

6. The dynamic driving risk assessment system according to claim 5, characterized in that: In step B3, the method for establishing the NGBoost-PA-WOE initialization model is as follows: B3.1: Use the loss scoring rule S to determine the initial model parameters θ by optimizing the loss function S (0) , the initial parameters of the model θ (0) The parameters of the marginal distribution corresponding to the training set; Where n is the number of samples; y i is the label of the i-th sample in the training set; B3.2: Use penalty attribute (FPA) to analyze the importance of each feature from historical traffic accident data and historical traffic violation data, and assign it an initial λ value; then calculate the weight distribution of the feature: Among them, WR λ is the feature weight range, λ is the feature level, p is the correlation coefficient and p>0; B3.3: For each sample i, calculate the gradient of the loss function S with respect to the model parameters θ Apply Fisher information matrix I s (θ) to adjust the gradient and obtain the natural gradient In each iteration, m ranges from 1 to M, where m represents the current iteration number; M represents the total number of iterations; λ i Represents the initial weight adjustment factor λ calculated by the penalty attribute (FPA) i ; B3.4: Use Natural Gradients To train the base learner f (m) , using DNC value, WOE value, accident rate, and violation rate as input data: Among them, x i Represents each sample in the test set; B3.5: Determine the scaling factor ρ by minimizing the adjusted loss scoring rule S (m) ; B3.6: Adjust the scaling factor ρ by combining the weight of the penalty attribute (FPA) (m) , the adjusted formula is: B3.7: Using learning rate n and scaling factor ρ (m) Update model parameters θ; B3.8: For each sample xi in the test set, calculate the final model parameters θ(xi); B3.9: Use the final model parameters θ(xi) to generate the predicted probability distribution for each sample; B3.10: Calculate risk score based on predicted probability distribution: RiskScore i =Pθ (xi) (y=1|xi) Among them, RiskScore i represents the risk score; p represents the predicted probability.

7. A dynamic driving risk assessment method based on the dynamic driving risk assessment system according to any one of claims 1 to 6, characterized in that: The following steps are involved: S1: Real-time collection of vehicle dynamic data and environmental characteristics at each road node, and removal of noise, filling of missing values, and detection and correction of abnormal data; S2: The risk assessment module obtains the vehicle dynamic data and environmental characteristics at each road node collected by the data acquisition module, and predicts the risk score at each road node based on the DNC value of each road node, the WOE value, accident rate, and violation rate corresponding to the vehicle dynamic data and environmental characteristics at the corresponding road node, and gives traffic control recommendations based on the risk score.

8. The dynamic driving risk assessment method according to claim 7 is characterized in that: During operation, the NGBoost-PA-WOE model updates the DNC value of each road node, the WOE value of the evidence weight corresponding to the vehicle dynamic data and environmental characteristics, the accident rate, and the violation rate based on the changes in the urban road traffic map and the recent historical traffic accident data and historical traffic violation data, and updates and optimizes the NGBoost-PA-WOE model.

9. The dynamic driving risk assessment method according to claim 7, characterized in that: The specific method of giving traffic control recommendations based on risk scores is as follows: the risk score of each road node is compared with the preset safety value. If it exceeds the safety value, the NGBoost-PA-WOE model uses the preset safety value, the accident rate and violation rate of the risk node, and the environmental characteristics of the current risk node as the adjustment parameter to calculate the target value of vehicle dynamic data that makes the risk score at the risk node reach the preset safety value, and determine the traffic control recommendations based on the target value.

Images