Abnormal toll collection monitoring and recognition system for expressway
By constructing a noise-resistant graph model and graph neural network, and combining it with a game-theoretic decision-making module to optimize the interception strategy, the problems of accuracy in identifying ghost data and toll evasion in highway toll collection systems and system response have been solved, achieving efficient and self-iterative anomaly monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GANSU XINLUGANG TECH CO LTD
- Filing Date
- 2025-12-23
- Publication Date
- 2026-05-12
AI Technical Summary
Existing highway toll monitoring systems struggle to accurately identify anomalies and find the optimal solution between interception and information acquisition when faced with ghost data injection and toll evasion behavior feature drift, leading to decreased judgment accuracy and passive system response.
A noise-resistant graph model is constructed, ghost nodes are eliminated through topological inconsistency energy, adversarial feature drift rate is calculated using graph neural network, and the interception benefits and social congestion costs are dynamically weighed by the game decision module to achieve optimal interception execution. The release sample is updated through the policy execution module.
It significantly improves the recognition accuracy and robustness of the highway toll collection system, dynamically optimizes the interception strategy, avoids congestion caused by over-control, and realizes the system's self-iteration capability.
Smart Images
Figure CN122020277A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent traffic management technology for highways, specifically to a system for monitoring and identifying toll anomalies on highways. Background Technology
[0002] Currently, highway toll monitoring mainly relies on roadside units and gantry systems to obtain vehicle trajectory flow and road network topology data. It identifies routine anomalies by calculating the matching degree between physical distance and time, and executes interception or release based on established rules. This method can achieve basic operation and audit functions in a single, static tolling scenario.
[0003] However, with the emergence of highly adversarial toll evasion methods such as ghost data injection, monitoring systems based on static physical rules have exposed significant weaknesses in related technologies. On the one hand, maliciously generated ghost nodes exist logically but are physically infeasible, easily interfering with the accuracy of traditional models. On the other hand, the feature drift rate of toll evasion behavior is accelerating, making it difficult for fixed models to capture new variants. Furthermore, existing strategies lack game theory thinking and cannot find the optimal solution between interception benefits, social congestion costs, and the value of acquiring unknown attack information, resulting in a passive system response during congestion periods and an inability to transform unknown threats into training data for self-iteration. Therefore, a solution is urgently needed to address the problems existing in current technologies.
[0004] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention discloses a toll collection anomaly monitoring and identification system for highways. Specifically, the technical solution of this invention includes:
[0006] The data acquisition module is used to acquire vehicle trajectory flow data and basic road network topology data;
[0007] The first processing module is used to construct a noise-resistant graph model based on vehicle trajectory flow data and road network topology data, calculate topological inconsistency energy, and generate a purified subgraph based on the topological inconsistency energy.
[0008] The second processing module is used to input the purified subgraph into the graph neural network model, calculate the adversarial feature drift rate, and solve the posterior probability that the target vehicle is a real toll evader.
[0009] The game decision-making module is used to determine the total expected utility based on the drift rate of adversarial features and posterior probability, combined with the preset interception benefits and social congestion cost, and to solve for the optimal interception execution probability that maximizes the total expected utility.
[0010] The strategy execution module is used to respond to the optimal interception execution probability: when the optimal interception execution probability indicates that the interception benefit is greater than the information acquisition benefit, physical interception is performed; when the optimal interception execution probability indicates that the information acquisition benefit is greater than or equal to the interception benefit, a release operation is performed and the current sample is used as an adversarial sample to update the graph neural network model.
[0011] Optionally, the first processing module calculates the topology inconsistency energy, including:
[0012] Retrieve the node observation timestamps and node physical locations from the vehicle trajectory flow data;
[0013] Obtain historical average flow velocity and equipment clock jitter statistics for the road segment;
[0014] The time difference between actual observations between computational nodes and neighboring nodes and the time deviation between theoretical physical road network travel time;
[0015] Based on the time deviation, equipment clock jitter statistics, and the output of the vehicle model characteristic mutual exclusion function, the topology inconsistency energy is determined.
[0016] Optionally, the first processing module generates a purified subgraph based on topological inconsistency energy, including:
[0017] Call the preset energy constraint threshold;
[0018] Compare the topologically inconsistent energy with the energy constraint threshold;
[0019] If the energy of a topologically inconsistent node exceeds the energy constraint threshold, the node is identified as a ghost node and removed.
[0020] If the energy is less than or equal to the energy constraint threshold due to topological inconsistency, the nodes are retained to construct a purified subgraph.
[0021] Optionally, the second processing module calculates the adversarial feature drift rate, including:
[0022] Collect the feature centroids of high-risk sample clusters in the current time window and the previous time window;
[0023] Calculate the centroid distance between two feature centroids in Euclidean space;
[0024] Calculate the gradient magnitude of the model loss function relative to the input features;
[0025] Invoke the preset gradient-distance mapping coefficients;
[0026] The adversarial feature drift rate is determined based on centroid distance, gradient magnitude, and gradient-distance mapping coefficient.
[0027] Optionally, the gradient-distance mapping coefficients are used to map the magnitude of the gradient to the feature distance space, wherein the gradient-distance mapping coefficients are determined based on normalization experiments of the adversarial generative network under simulated attacks.
[0028] Optionally, the game decision-making module determines the total expected utility, including:
[0029] The optimal interception execution probability, posterior probability, adversarial feature drift rate, interception benefits, and social congestion cost are all considered.
[0030] The difference between the expected positive benefit from execution interception and the expected negative risk caused by false interception is defined as the direct intervention utility.
[0031] The information value monetization coefficient is invoked, and the information gain utility obtained by abandoning the interception is calculated based on the information value monetization coefficient and the adversarial feature drift rate.
[0032] The total expected utility is determined by a weighted sum of the utility of direct intervention and the utility of information gain.
[0033] Optionally, the information value monetization coefficient is used to convert the observed feature drift into an equivalent intelligence economic value, wherein the intelligence economic value characterizes the short-term toll revenue that the system is willing to sacrifice to obtain a unit of feature drift observation data.
[0034] Optionally, the policy execution module performs physical interception or allows passage, including:
[0035] The product of the adversarial feature drift rate and the information value monetization coefficient is defined as the information acquisition benefit;
[0036] The interception benefit is defined as the product of the interception benefit and the posterior probability, minus the product of the social congestion discount cost and the non-fare evasion probability.
[0037] If the benefit of interception is greater than the benefit of information acquisition, the optimal interception execution probability is determined to be close to 1, and physical interception is executed.
[0038] If the benefit of information acquisition is greater than or equal to the benefit of interception, the optimal interception execution probability is determined to be close to 0, and the release operation is executed.
[0039] Compared with the prior art, the present invention has the following beneficial effects:
[0040] 1. This invention effectively solves the common noise and error problems in highway vehicle trajectory flow data by constructing a noise-resistant graph model and introducing a topological inconsistency energy calculation mechanism. The system comprehensively utilizes node observation timestamps, physical road network distances, and equipment clock jitter statistics to accurately calculate the time deviation between actual observations and theoretical physical laws, and automatically removes ghost nodes to generate a purified subgraph through energy constraint thresholds. This design significantly reduces the interference of environmental noise and equipment failures on the recognition system, greatly improves the purity and reliability of input data, and lays a solid foundation for subsequent accurate recognition.
[0041] 2. This invention utilizes graph neural networks combined with the calculation of adversarial feature drift rates to overcome the limitation of traditional static rules in recognizing novel toll evasion methods. By calculating the feature centroid distance and model gradient magnitude of high-risk sample clusters, the system can keenly capture the subtle drifts and evolutions of toll evasion behavior in the feature space. This gradient-based deep feature analysis enables the system not only to identify routine anomalies but also to effectively combat adversarial attacks that attempt to evade detection through fine-tuning behavior, significantly enhancing the robustness and recognition accuracy of the model in complex and ever-changing scenarios.
[0042] 3. This invention innovatively introduces a game-theoretic decision-making module that incorporates the social congestion cost, changing the previous extensive model that only focused on the benefits of interception. When calculating the total expected utility, the system dynamically weighs the direct economic returns from interception against the social congestion costs caused by false or frequent interceptions, and solves for the optimal interception execution probability. This decision-making mechanism effectively avoids toll station congestion caused by excessive control, ensuring precise auditing while maintaining high traffic efficiency on highways, and maximizing both economic benefits and social public interests.
[0043] 4. This invention establishes an active learning and model iteration mechanism through the cooperation of the strategy execution module and the information value monetization coefficient. When the system evaluation finds that the information gain utility obtained by abandoning the current interception is higher than the benefit of direct interception, it will automatically execute the release and collect samples to update the graph neural network. This strategy of obtaining high-value intelligence by releasing transforms the loss of some benefits in the short term into the improvement of the system's long-term detection capability, enabling the system to continuously evolve itself as the means of evasion change. Attached Figure Description
[0044] The present invention will be further explained below with reference to the accompanying drawings and embodiments:
[0045] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.
[0047] Example 1
[0048] Please see Figure 1 A system for monitoring and identifying toll irregularities on highways, comprising:
[0049] The data acquisition module is used to acquire vehicle trajectory flow data and basic road network topology data;
[0050] The first processing module is used to construct a noise-resistant graph model based on vehicle trajectory flow data and road network topology data, calculate topological inconsistency energy, and generate a purified subgraph based on the topological inconsistency energy.
[0051] The second processing module is used to input the purified subgraph into the graph neural network model, calculate the adversarial feature drift rate, and solve the posterior probability that the target vehicle is a real toll evader.
[0052] The game decision-making module is used to determine the total expected utility based on the drift rate of adversarial features and posterior probability, combined with the preset interception benefits and social congestion cost, and to solve for the optimal interception execution probability that maximizes the total expected utility.
[0053] The strategy execution module is used to respond to the optimal interception execution probability: when the optimal interception execution probability indicates that the interception benefit is greater than the information acquisition benefit, physical interception is performed; when the optimal interception execution probability indicates that the information acquisition benefit is greater than or equal to the interception benefit, a release operation is performed and the current sample is used as an adversarial sample to update the graph neural network model.
[0054] This embodiment provides a toll collection anomaly monitoring and identification system for highways. The system consists of a data acquisition module, a first processing module, a second processing module, a game decision-making module, and a strategy execution module in its logical architecture.
[0055] The data acquisition module is configured to access the raw data streams of the highway roadside unit and toll gantry system in real time to obtain vehicle trajectory flow data and basic road network topology data. The vehicle trajectory flow data includes timestamps, gantry IDs and vehicle type identification codes, and the basic road network topology data includes the physical distances and connections between stations.
[0056] The first processing module is configured to construct a noise-resistant graph model based on vehicle trajectory flow data and road network topology data. In this model, vehicle trajectory points are mapped as graph nodes. The module calculates the topological inconsistency energy of each node based on the principles of physical kinematics, and prunes the graph structure based on the energy value to generate a purified subgraph that has eliminated ghost nodes.
[0057] The second processing module is configured to input the purified subgraph into a pre-trained graph neural network model; this module calculates the adversarial feature drift rate to quantify the evolution speed and adversarial strength of fare evasion behavior; at the same time, this module calculates the posterior probability that the target vehicle is a real fare evader.
[0058] The game decision-making module is configured to determine the total expected utility based on the adversarial feature drift rate and posterior probability, combined with the preset interception benefits and social congestion cost. This module outputs the optimal interception execution probability by solving an optimization problem that maximizes the total expected utility.
[0059] The strategy execution module is configured to respond to the optimal interception execution probability; when the optimal interception execution probability indicates that the interception benefit is greater than the information acquisition benefit, the system controls the field equipment to perform physical interception; when the optimal interception execution probability indicates that the information acquisition benefit is greater than or equal to the interception benefit, the system performs a release operation and updates the graph neural network model with the current sample as an adversarial sample.
[0060] This embodiment utilizes the organic combination of the above modules to resist ghost data attacks by using the physical denoising of the first processing module. Through the cooperation of the game decision-making module and the strategy execution module, dynamic game theory against highly adversarial toll evasion behavior is achieved while ensuring passage efficiency. The strategic passage mechanism transforms unknown toll evasion behavior into high-value training data, enabling the system to have self-iterative capabilities.
[0061] Example 2
[0062] The first processing module calculates the topological inconsistency energy, including:
[0063] Retrieve the node observation timestamps and node physical locations from the vehicle trajectory flow data;
[0064] Obtain historical average flow velocity and equipment clock jitter statistics for the road segment;
[0065] The time difference between actual observations between computational nodes and neighboring nodes and the time deviation between theoretical physical road network travel time;
[0066] Based on the time deviation, equipment clock jitter statistics, and the output of the vehicle model characteristic mutual exclusion function, the topological inconsistency energy is determined.
[0067] The first processing module calls the node observation timestamps and node physical locations in the vehicle trajectory flow data to establish spatiotemporal coordinates; the module obtains the historical average flow velocity of the road segment and the statistical value of equipment clock jitter from the system database.
[0068] The actual observation time difference between the module's computational node and its neighboring nodes is compared with the time deviation between the theoretical physical road network and the travel time. To ensure the consistency of physical dimensions and eliminate equipment errors, this embodiment uses the following formula to calculate the topology inconsistency energy. :
[0069]
[0070] in, This is the dimensionless topological inconsistency energy;
[0071] For nodes A set of neighboring nodes that have spatiotemporal relationships;
[0072] The weights are determined based on road network connectivity, if the nodes With nodes If there are direct physical connections in the road network topology, then ,otherwise ;
[0073] The actual observation time difference derived from the timestamps of the streaming data;
[0074] Physical road network distances derived from high-precision map databases;
[0075] Average flow velocity of road segments derived from historical normal traffic data statistical distribution;
[0076] The standard deviation of the time deviation is derived from the statistical calculation of the device clock accuracy by the system and is used to normalize the deviation in the time dimension into a dimensionless value.
[0077] Let the vehicle model feature be a mutual exclusion function, if the node and Model code The function returns 1 if a logical mutation occurs, otherwise it returns 0.
[0078] The dimensionless feature conflict penalty weights are determined through logistic regression analysis of historically labeled black samples;
[0079] The module determines the topology inconsistency energy based on time deviation, device clock jitter statistics, and the output of the vehicle model characteristic mutual exclusion function; it also introduces an energy function that incorporates physical distance, historical flow rate, and vehicle model mutual exclusion logic. This embodiment can accurately identify ghost nodes that logically exist but are impossible in the physical world; utilizing Normalization effectively eliminates calculation errors caused by asynchronous device clocks.
[0080] Example 3
[0081] The first processing module generates a purified subgraph based on topological inconsistency energy, including:
[0082] Call the preset energy constraint threshold;
[0083] Compare the topologically inconsistent energy with the energy constraint threshold;
[0084] If the energy of a topologically inconsistent node exceeds the energy constraint threshold, the node is identified as a ghost node and removed.
[0085] If the energy of the topologically inconsistent node is less than or equal to the energy constraint threshold, the node is retained to construct a purified subgraph.
[0086] After calculating the energy value of each node, the first processing module calls the preset energy constraint threshold; the energy constraint threshold is set to the 99th percentile value of the topologically inconsistent energy distribution in the historical compliant vehicle trajectory dataset.
[0087] The module compares the topology inconsistency energy with the energy constraint threshold; if the topology inconsistency energy is greater than the energy constraint threshold, the system determines the node as a ghost node and removes it; if the topology inconsistency energy is less than or equal to the energy constraint threshold, the system retains the node to construct a purified subgraph.
[0088] This step eliminates false nodes through physical energy constraints, and the resulting purified subgraph contains only physically real vehicle trajectories, reducing the computational load of subsequent models and eliminating interference from maliciously injected data.
[0089] Example 4
[0090] The second processing module calculates the adversarial feature drift rate, including:
[0091] Collect the feature centroids of high-risk sample clusters in the current time window and the previous time window;
[0092] Calculate the centroid distance between two feature centroids in Euclidean space;
[0093] Calculate the gradient magnitude of the model loss function relative to the input features;
[0094] Invoke the preset gradient-distance mapping coefficients;
[0095] The adversarial feature drift rate is determined based on the centroid distance, gradient magnitude, and gradient-distance mapping coefficient.
[0096] The second processing module samples within the feature space, collecting the feature centroids of high-risk sample clusters in the current and previous time windows. A high-risk sample cluster refers to the posterior probability of toll evasion output after forward propagation calculation by the graph neural network model within the corresponding time window. Greater than the preset warning threshold ,For example The set of all vehicle samples;
[0097] Graph neural network models employ a two-layer graph convolutional network or graph attention network architecture, with input layer feature vectors... The average time deviation between this node and its neighboring nodes Average physical distance from neighboring areas It is pieced together, that is Characteristic centroid This is the arithmetic mean of the node embedding vectors output by the last layer of the graph neural network before the fully connected classification layer, within the sample cluster.
[0098] The module calculates the centroid distance between two feature centroids in Euclidean space. Simultaneously, the module calculates the model loss function, using the class with the highest predicted probability as the gradient magnitude of the pseudo-label relative to the input features. ;
[0099] The module calls the preset gradient-distance mapping coefficients. The adversarial feature drift rate is determined using the following formula. :
[0100]
[0101] in, The drift rate is the adversarial feature, and its dimension is the feature space distance unit.
[0102] and Each is the current window With the previous window High-risk sample cluster feature centroid; This is the loss function for the current model;
[0103] Simultaneously, it captures the macro-evolution of toll evasion methods and micro-adversarial attacks targeting model weaknesses, enabling the system to detect new variants of toll evasion.
[0104] Example 5
[0105] The gradient-distance mapping coefficients are used to map the magnitude of the gradient to the feature distance space. The gradient-distance mapping coefficients are determined based on normalization experiments of adversarial generative networks under simulated attacks.
[0106] Gradient-distance mapping coefficients are used to map the magnitude of the gradient to the feature distance space;
[0107] The gradient-distance mapping coefficient This was determined through normalization experiments of adversarial generative networks under simulated attacks; specifically, a calibration dataset containing simulated attack samples was constructed. ,for For each sample, calculate its characteristic centroid displacement. With gradient magnitude By fitting with linear regression, The optimal coefficient for the condition to hold;
[0108] Determined through experiments The coefficient ensures The comparability of the two indicators in the formula and the reasonableness of their weights.
[0109] Example 6
[0110] The game-theoretic decision-making module determines the total expected utility, including:
[0111] The optimal interception execution probability, posterior probability, adversarial feature drift rate, interception benefits, and social congestion cost are all considered.
[0112] The difference between the expected positive benefit from execution interception and the expected negative risk caused by false interception is defined as the direct intervention utility.
[0113] The information value monetization coefficient is invoked, and the information gain utility obtained by abandoning the interception is calculated based on the information value monetization coefficient and the adversarial feature drift rate.
[0114] The total expected utility is determined by a weighted sum of the utility of direct intervention and the utility of information gain.
[0115] The game theory decision-making module calls the optimal interception execution probability. Posterior probability , Adversarial feature drift rate , intercept revenue and the cost of social congestion ;in, Based on real-time road network congestion index Through a pre-set nonlinear growth function The calculated monetization cost value; the nonlinear growth function adopts an exponential model:
[0116]
[0117] in, The unit price for basic road network toll costs;
[0118] This is a congestion sensitivity coefficient, with a value ranging from 1.5 to 2.0.
[0119] The congestion trigger threshold;
[0120] Total expected utility of module construction Computational model:
[0121]
[0122] The difference between the expected positive benefit from module-based interception and the expected negative risk from false interception is defined as the direct intervention utility, corresponding to the formula in... Partial; wherein, only when interception is performed, i.e. Only then will the system need to bear the cost of congestion risks caused by false alarms. ;
[0123] Module call information value monetization coefficient Based on the information value monetization coefficient and the adversarial feature drift rate, the information gain utility obtained by abandoning the interception is calculated, corresponding to the formula in... part;
[0124] The module determines the total expected utility based on a weighted sum of direct intervention utility and information gain utility; the system determines the total expected utility by... Solve for the optimal solution and find the one that makes The largest ;
[0125] This utility function is introduced by... Reduce the tendency to intercept during severe congestion by introducing When faced with new and highly adversarial attacks, the system proactively chooses to allow passage in order to trap data, thus achieving a balance between the long-term value of model iteration and the short-term revenue from toll fees.
[0126] Example 7
[0127] The information value monetization coefficient is used to convert the observed feature drift into equivalent intelligence economic value, where the intelligence economic value represents the short-term toll revenue that the system is willing to sacrifice to obtain a unit of feature drift observation data.
[0128] Information value monetization coefficient Used to convert observed feature drift into equivalent intelligence economic value; its dimension is monetary per unit feature distance; intelligence economic value represents the short-term toll revenue that the system is willing to sacrifice to obtain a unit of feature drift observation data;
[0129] The parameters quantify the long-tail value of the data, ensuring that the system continuously accumulates high-value samples in highly competitive environments; the information value monetization coefficient. The methods for determining this include: building an offline simulation environment based on historically labeled datasets, and setting... The search range is ,in The maximum interception benefit per instance; with a step size of 0.1. Perform a grid search traversal and calculate each The cumulative total expected utility corresponding to the value Choose the one that maximizes the cumulative total expected utility. The value is used as the system's preset parameter.
[0130] Example 8
[0131] The policy enforcement module performs physical interception or allows passage, including:
[0132] The product of the adversarial feature drift rate and the information value monetization coefficient is defined as the information acquisition benefit;
[0133] The interception benefit is defined as the product of the interception benefit and the posterior probability, minus the product of the social congestion discount cost and the non-fare evasion probability.
[0134] If the benefit of interception is greater than the benefit of information acquisition, the optimal interception execution probability is determined to be close to 1, and physical interception is executed.
[0135] If the benefit of information acquisition is greater than or equal to the benefit of interception, the optimal interception execution probability is determined to be close to 0, and the release operation is executed.
[0136] Module calculates adversarial feature drift rate Information value monetization coefficient The product of these is defined as the information acquisition benefit. ;
[0137] Module calculates interception revenue With posterior probability The product of, and minus the social congestion discount costs. Probability of non-evasion of fees The product of these is defined as the interception revenue. ;
[0138] If the benefits of interception are greater than the benefits of information acquisition, that is... The system determines the optimal interception execution probability. Approaching 1, and performing physical interception;
[0139] If the benefit of information acquisition is greater than or equal to the benefit of interception, that is... The system determines the optimal interception execution probability. The value approaches 0, and the release operation is performed; at the same time, the current sample is marked as a high-value potential adversarial sample and stored in the database. After the real label is obtained through backend audit, it will be used to update the graph neural network model.
[0140] This execution logic enables automated and intelligent hierarchical responses, solving the problems of traditional systems being unable to intercept during congested periods and being slow to react to new types of attacks.
[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.
Claims
1. A toll collection anomaly monitoring and identification system for highways, characterized in that, include: The data acquisition module is used to acquire vehicle trajectory flow data and basic road network topology data; The first processing module is used to construct a noise-resistant graph model based on vehicle trajectory flow data and road network topology data, calculate topological inconsistency energy, and generate a purified subgraph based on the topological inconsistency energy. The second processing module is used to input the purified subgraph into the graph neural network model, calculate the adversarial feature drift rate, and solve the posterior probability that the target vehicle is a real toll evader. The game decision-making module is used to determine the total expected utility based on the drift rate of adversarial features and posterior probability, combined with the preset interception benefits and social congestion cost, and to solve for the optimal interception execution probability that maximizes the total expected utility. The strategy execution module is used to respond to the optimal interception execution probability: when the optimal interception execution probability indicates that the interception benefit is greater than the information acquisition benefit, physical interception is performed; when the optimal interception execution probability indicates that the information acquisition benefit is greater than or equal to the interception benefit, a release operation is performed and the current sample is used as an adversarial sample to update the graph neural network model.
2. The toll collection anomaly monitoring and identification system for highways according to claim 1, characterized in that, The first processing module calculates the topology inconsistency energy, including: Retrieve the node observation timestamps and node physical locations from the vehicle trajectory flow data; Obtain historical average flow velocity and equipment clock jitter statistics for the road segment; The time difference between actual observations between computational nodes and neighboring nodes and the time deviation between theoretical physical road network travel time; Based on the time deviation, equipment clock jitter statistics, and the output of the vehicle model characteristic mutual exclusion function, the topology inconsistency energy is determined.
3. The toll collection anomaly monitoring and identification system for highways according to claim 1, characterized in that, The first processing module generates a purified subgraph based on topological inconsistency energy, including: Call the preset energy constraint threshold; Compare the topologically inconsistent energy with the energy constraint threshold; If the energy of a topologically inconsistent node exceeds the energy constraint threshold, the node is identified as a ghost node and removed. If the energy is less than or equal to the energy constraint threshold due to topological inconsistency, the nodes are retained to construct a purified subgraph.
4. The toll collection anomaly monitoring and identification system for highways according to claim 1, characterized in that, The second processing module calculates the adversarial feature drift rate, including: Collect the feature centroids of high-risk sample clusters in the current time window and the previous time window; Calculate the centroid distance between two feature centroids in Euclidean space; Calculate the gradient magnitude of the model loss function relative to the input features; Invoke the preset gradient-distance mapping coefficients; The adversarial feature drift rate is determined based on centroid distance, gradient magnitude, and gradient-distance mapping coefficient.
5. A toll collection anomaly monitoring and identification system for highways according to claim 4, characterized in that, The gradient-distance mapping coefficients are used to map the magnitude of the gradient to the feature distance space. The gradient-distance mapping coefficients are determined based on normalization experiments of the adversarial generative network under simulated attacks.
6. A toll collection anomaly monitoring and identification system for highways according to claim 1, characterized in that, The game decision-making module determines the total expected utility, including: The optimal interception execution probability, posterior probability, adversarial feature drift rate, interception benefits, and social congestion cost are all considered. The difference between the expected positive benefit from execution interception and the expected negative risk caused by false interception is defined as the direct intervention utility. The information value monetization coefficient is invoked, and the information gain utility obtained by abandoning the interception is calculated based on the information value monetization coefficient and the adversarial feature drift rate. The total expected utility is determined by a weighted sum of the utility of direct intervention and the utility of information gain.
7. A toll collection anomaly monitoring and identification system for highways according to claim 6, characterized in that, The information value monetization coefficient is used to convert the observed feature drift into equivalent intelligence economic value, wherein the intelligence economic value characterization system represents the short-term toll revenue that it is willing to sacrifice to obtain a unit of feature drift observation data.
8. A toll collection anomaly monitoring and identification system for highways according to claim 1, characterized in that, The policy execution module performs physical interception or allows passage, including: The product of the adversarial feature drift rate and the information value monetization coefficient is defined as the information acquisition benefit; The interception benefit is defined as the product of the interception benefit and the posterior probability, minus the product of the social congestion discount cost and the non-fare evasion probability. If the benefit of interception is greater than the benefit of information acquisition, the optimal interception execution probability is determined to be close to 1, and physical interception is executed. If the benefit of information acquisition is greater than or equal to the benefit of interception, the optimal interception execution probability is determined to be close to 0, and the release operation is executed.