A green vehicle auditing model and method based on multi-dimensional feature cross analysis

Abstract

The application relates to the technical field of intelligent transportation, and discloses a green vehicle auditing model and method based on multi-dimensional feature cross analysis, which comprises a feature enhancement module, a multi-dimensional feature vector construction module, a comprehensive scoring module, a weight training module and a risk grading module. The method corresponds to the model. According to the application, by constructing a five-dimensional feature vector of time, space, behavior, load and history, relying on core algorithms such as cosine similarity algorithm for quantifying time period matching degree and exponential decay function for mapping path deviation degree, cross analysis of multi-dimensional features is realized, and with the aid of XGBoost machine learning, dynamic optimization of feature weight and risk grading threshold is realized, hidden correlations between unreasonable seasons and path deviations and other features are captured through a self-adaptive coupling module, and the weight of recent abnormal behavior is strengthened by using time decay credit points, so that the misjudgment rate and the missed judgment rate are effectively reduced, and the problem of missed detection of composite risks in traditional single-dimensional detection is solved.

Description

Technical Field

[0001] The present application relates to the field of smart transportation technology, and specifically to a green pass vehicle audit model and method based on multi-dimensional feature cross-analysis. Background Art

[0002] In the field of intelligent auditing of green-pass vehicles on highways, the existing technical system has multi-dimensional structural defects, making it difficult to meet the dual needs of precise supervision and efficient passage. Specific problems are as follows:

[0003] First, the manual inspection mechanism faces efficiency bottlenecks and a lack of standardization. Currently, Green Pass audits still rely heavily on manual on-site inspections, a model with significant limitations. First, single-lane inspections can take as long as 10-15 minutes, easily causing congestion during peak hours and hindering road network efficiency. Second, inspection results rely heavily on the operator's experience and judgment, lacking objective quantitative standards. This leads to inconsistent results across different sites and personnel, and persistently high rates of false positives and missed detections. This subjective bias not only compromises the fairness of policy implementation but also creates opportunities for criminals to circumvent inspections.

[0004] Second, there are systemic vulnerabilities in single-dimensional feature recognition. Existing automated detection solutions often focus on a single feature dimension (such as cargo type identification or load detection) and employ simple threshold matching logic. Such approaches struggle to address complex evasion behaviors: for example, mixing compliant and noncompliant cargo, falsifying origin information, and other methods can easily circumvent single-dimensional detection rules. Due to the lack of a cross-validation mechanism for multi-dimensional features, the system's recognition rate for these types of covert evasion behaviors is less than 30%, creating a significant regulatory blind spot.

[0005] Third, there's a conflict in the adaptability of static models to dynamic transportation characteristics. Agricultural product transportation has significant dynamic characteristics, but existing models employ a fixed threshold system, making it impossible to achieve dynamic adaptation. In terms of time, the distribution of transportation periods for different agricultural product categories varies seasonally (e.g., 70% of fresh produce is shipped in the morning in summer, while it's concentrated in the afternoon in winter). In terms of space, factors such as the migration of major production areas and adjustments to cross-regional distribution routes lead to dynamic changes in transportation flows. In terms of behavior, characteristics such as transportation frequency and load fluctuations are closely linked to market supply and demand. This mismatch between static thresholds and dynamic characteristics results in the system's false alarm rate fluctuating by over 40% across different scenarios, severely impacting audit accuracy.

[0006] Fourth, the depth of historical data mining is insufficient. Existing systems limit their use of historical vehicle behavior data to basic statistics and lack in-depth modeling and analysis. The system lacks a time-series-based behavioral pattern profile, making it difficult to identify abnormal patterns such as periodic, high-frequency, short-distance transport. It also lacks a multi-dimensional correlation analysis mechanism, making it unable to capture compound risk signals such as time-period anomalies and route deviations. Furthermore, the system lacks a dynamic credit evaluation system, and the weighting of historical inspection results does not decay dynamically over time, leading to a disconnect between credit assessments and actual behavior. These shortcomings make it difficult to effectively identify covert toll evasion (e.g., circumventing thresholds by gradually adjusting parameters such as load weight and frequency), resulting in persistent toll loss.

[0007] These technical bottlenecks collectively contribute to the overall inefficiency of the Green Pass audit system, preventing it from accurately combating toll evasion and impacting the transit experience of legitimate vehicles. Therefore, there is an urgent need to build a smarter and more comprehensive audit system through technological innovation. Summary of the Invention

[0008] The purpose of this application is to provide a green pass vehicle audit model and method based on multi-dimensional feature cross-analysis to solve the technical problems raised in the above background technology.

[0009] To achieve the above objectives, this application discloses the following technical solutions:

[0010] In the first aspect, the present application discloses a green pass vehicle audit model based on multi-dimensional feature cross-analysis, including:

[0011] The feature enhancement module is configured to process the original features of the green pass vehicle through the feature enhancement algorithm and calculate the time features respectively. , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics The original features include original data related to time features, original data related to space features, original data related to traffic behavior, original data related to load, and original data related to historical inspection records;

[0012] The multi-dimensional feature vector construction module is configured to: transform the time feature , spatial characteristics , the behavioral characteristics , load characteristics and historical credit characteristics Combination, constructing the multi-dimensional feature vector of green pass vehicles , ;

[0013] Comprehensive scoring module, for Build a comprehensive scoring model and output a comprehensive score , ,in, is the linear weighted term of the standardized eigenvalue, is the adaptive coupling term, is the weight parameter used to balance the normalized eigenvalue linear weighting term and the adaptive coupling term;

[0014] A weight training module is configured to: train and adjust the weight parameters in the feature enhancement algorithm and the comprehensive scoring model through the XGBoost machine learning method;

[0015] The risk grading module is configured to: The risk level of green pass vehicles is determined through risk grading strategy.

[0016] Preferably, the time feature The calculation process is:

[0017]

[0018] in, is the weight parameter within the time feature; is the time period matching degree, which is calculated by the cosine similarity between the current time period vector and the historical time period distribution vector; Seasonal deviations for transported goods.

[0019] Preferably, the spatial characteristics The calculation process is:

[0020]

[0021] in, is the weight parameter within the spatial feature; The rationality of the origin of the goods and the rationality of passing through the entrance toll station and the rationality of the exit toll booths The indicator function combination is calculated; To determine the rationality of the transportation path, the path deviation is mapped by an exponential decay function.

[0022] Preferably, the behavioral characteristics The calculation process is:

[0023]

[0024] in, is the weight parameter within the behavioral feature; is the actual number of times the vehicle passes during the current statistical period, The average number of times the same type of vehicles pass through the same statistical period; Counting short bursts of traffic; is the time dispersion variance, which is calculated based on the mean of the vehicle travel time interval; is the penalty coefficient for short-term intensive traffic; is the influencing factor for adjusting the sensitivity of time dispersion variance.

[0025] Preferably, the load characteristics The calculation process is

[0026]

[0027] in, is the weight parameter inside the load feature; is the individual historical deviation characteristic item; It is the deviation characteristic item of the same vehicle model and cargo group.

[0028] Preferably, the historical credit characteristics The calculation process is:

[0029]

[0030] in, is the dynamic weight; is the result of the i-th inspection; To check the fluctuation coefficient of the result sequence; is the number of inspections; is the volatility penalty index.

[0031] Preferably, the calculation process of the standardized eigenvalue linear weighted term is:

[0032]

[0033] in, 、 、 、 and Time characteristics , spatial characteristics , behavioral characteristics , load characteristics and dynamically adjust weight parameters based on historical credit characteristics.

[0034] Preferably, the calculation process of the adaptive coupling term is:

[0035]

[0036] in, is the dynamic strength coefficient, ; and are the eigenvalues ​​of different features respectively; The service tolerance threshold.

[0037] Preferably, the weight parameters of the feature enhancement algorithm and the weight parameters of the comprehensive scoring model include:

[0038] Weight parameters within the time feature , weight parameters within spatial features , weight parameters within behavioral features , weight parameters within the load feature , dynamic adjustment weight parameters of time characteristics , dynamic adjustment weight parameters of spatial features , Dynamic adjustment weight parameters of behavioral characteristics , Dynamically adjust weight parameters of load characteristics , Dynamic adjustment weight parameters of historical credit characteristics and the weight parameter used to balance the normalized eigenvalue linear weighting term and the adaptive coupling term ;

[0039] The weight parameters of the feature enhancement algorithm and the weight parameters of the comprehensive scoring model are modified by optimizing the loss function The training is obtained, among which, is the total amount of training samples, is the true label of the i-th sample, The model predicts the probability that the i-th sample is abnormal, is the L1 regularization term or the L2 regularization term;

[0040] The risk grading strategy is: When , the risk level is low risk; when When the risk level is medium risk; when When the risk level is high, 、 and is a learnable parameter threshold.

[0041] In a second aspect, the present application discloses a green pass vehicle audit method based on multi-dimensional feature cross analysis, which is applicable to the green pass vehicle audit model based on multi-dimensional feature cross analysis as described above. The method comprises the following steps:

[0042] The original features of green-pass vehicles are processed by feature enhancement algorithm, and the time features are calculated respectively. , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics The original features include original data related to time features, original data related to space features, original data related to traffic behavior, original data related to load, and original data related to historical inspection records;

[0043] The time feature , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics Combination, constructing the multi-dimensional feature vector of green pass vehicles , ;

[0044] Based on the multidimensional feature vector Build a comprehensive scoring model and output a comprehensive score , ,in, is the linear weighted term of the standardized eigenvalue, is the adaptive coupling term, is the weight parameter used to balance the normalized eigenvalue linear weighting term and the adaptive coupling term;

[0045] Training and adjusting the weight parameters in the feature enhancement algorithm and the comprehensive scoring model by using the XGBoost machine learning method;

[0046] Based on the comprehensive score The risk level of green pass vehicles is determined through risk grading strategy.

[0047] Beneficial effects: The green pass vehicle audit model and method based on multi-dimensional feature cross-analysis of this application greatly reduces manual intervention and improves audit efficiency through automated processing of original features, and avoids the deviation of manual subjective judgment based on quantitative analysis of objective features; secondly, by combining time features, spatial features, behavioral features, load features and historical credit features into multi-dimensional feature vectors, combining standardized eigenvalue linear weighting terms and adaptive coupling terms, cross-analysis of multi-dimensional features is achieved, and complex risks that cannot be covered by single-dimensional detection are effectively identified; at the same time, the feature enhancement algorithm and comprehensive scoring model are enhanced through the XGBoost machine learning method The weight parameters in the model are dynamically adjusted to enable the model to adapt to dynamic characteristics such as different cargo types and seasonal changes, thereby improving the detection accuracy in different scenarios; in addition, by extracting historical credit features and incorporating them into multi-dimensional feature vectors, combined with the deep learning ability of XGBoost machine learning on historical data, a dynamic evaluation of the vehicle's historical behavior patterns is achieved, which effectively identifies hidden and periodic fee evasion behaviors and improves the timeliness and accuracy of risk warnings; and, based on the comprehensive score, the risk level is determined through a risk grading strategy, providing clear risk guidance for green pass vehicle inspections, enabling inspection resources to focus on high-risk vehicles and improving the utilization efficiency of inspection resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 This is a structural block diagram of the green pass vehicle audit model based on multi-dimensional feature cross analysis provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] The following is a clear and complete description of the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0051] In this document, the term "comprising" is intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0052] In a first aspect, this embodiment provides a Figure 1 The green pass vehicle audit model based on multi-dimensional feature cross analysis shown in the figure includes a feature enhancement module, a multi-dimensional feature vector construction module, a comprehensive scoring module, a weight training module, and a risk classification module. In the green pass vehicle audit model based on multi-dimensional feature cross analysis, the connection relationship between the modules is as follows:

[0053] The feature enhancement module, multidimensional feature vector construction module, and comprehensive scoring module are connected in sequence. The output of the feature enhancement module serves as the input of the multidimensional feature vector construction module. The latter generates a multidimensional feature vector by combining the features output by the feature enhancement module, realizing the transformation from a single feature to a composite feature vector. The multidimensional feature vector output by the multidimensional feature vector construction module serves as the core input of the comprehensive scoring module. The comprehensive scoring module constructs a comprehensive scoring model based on the multidimensional feature vector and outputs a comprehensive score through weighted calculation of the standardized eigenvalue linear weighting term and the adaptive coupling term.

[0054] The comprehensive scoring module is connected to the risk grading module. The comprehensive score output by the comprehensive scoring module serves as the input of the risk grading module. The risk grading module uses the comprehensive score and the risk grading strategy to determine the risk level of the green pass vehicle.

[0055] The weight training module is connected to the feature enhancement module and the comprehensive scoring module. Using the XGBoost machine learning method, the weight training module trains and optimizes the weight parameters involved in the feature enhancement algorithm and the weight parameters in the comprehensive scoring model. The trained weight parameters are then fed back to the feature enhancement and comprehensive scoring modules, enabling dynamic iteration of model parameters and improving feature processing accuracy and comprehensive scoring accuracy.

[0056] In the green pass vehicle audit model based on multi-dimensional feature cross-analysis of this embodiment, each module realizes the audit function through the forward data flow of original feature processing → feature vector construction → comprehensive score calculation → risk level determination, and forms a closed loop through the reverse parameter optimization of the weight training module to ensure that the model performance continues to adapt to the actual scenario.

[0057] Specifically

[0058] The feature enhancement module is configured to process the original features of the green pass vehicle through the feature enhancement algorithm and calculate the time features respectively. , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics The original features include original data related to time features (such as travel time periods, transportation months, etc.), original data related to spatial features (such as entrance toll stations, exit toll stations, actual travel routes, toll stations corresponding to the origin of goods, etc., which reflect the rationality of the spatial correlation of green pass vehicle transportation), original data related to travel behavior (such as the number of passes within the statistical period, the time interval between adjacent passes, etc.), original data related to load (such as current load, historical load records, similar vehicle model-cargo load records, etc.), and original data related to historical inspection records (such as historical inspection results, inspection time, etc.).

[0059] The feature enhancement algorithm consists of time feature algorithm, space feature algorithm, behavior feature algorithm, load feature algorithm and history feature algorithm. , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics The range is between (0,1].

[0060] The core content of the time feature algorithm is to quantify the time feature value of the current vehicle by integrating time period matching and seasonal rationality. Time period matching uses cosine similarity to calculate the matching degree between the current time period vector and the historical time period distribution vector; seasonal rationality uses the absolute difference between the production season of the goods and the current month to calculate the seasonal deviation. The closer it is to 1, the smaller the risk, and vice versa.

[0061] Based on this, the time characteristics of this embodiment The calculation process is:

[0062]

[0063] in, is the weight parameter within the time feature (used to balance time period matching and seasonal rationality); is the time period matching degree, which is calculated by the cosine similarity between the current time period vector and the historical time period distribution vector (i.e. ),in, is the current time period vector, which is the current transportation time period vector using 24-bit one-hot encoding (24 bits correspond to the 24 hours of a day, the bit corresponding to the current time period is 1, and the rest are 0); is the historical time period distribution vector, which is the time distribution vector of the historical transportation of this type of goods (statisticing the proportion of the passage frequency of this type of goods in each hour period in the historical transportation); The seasonal deviation of the transported goods (i.e. the absolute difference between the production season of the goods and the current month. For example, if the production season of a certain goods is May and the current month is July, then The cosine similarity is used to calculate the matching degree of the current traffic period in the historical period distribution, so that the matching degree ranges from [0, 1]. When it is closer to 1, the matching degree is higher and the risk is lower. When it is closer to 0, the matching degree is lower and the risk is greater.

[0064] The above time characteristics The calculation process of cosine similarity The current period vector encoded by 24-bit one-hot encoding and historical period distribution vector The quantitative matching accurately measures the consistency between the current transportation period and the historical high-frequency mode, avoids the rigid constraint of the static period threshold, and can adapt to the time distribution characteristics of different goods; at the same time, seasonal rationality Considering the absolute difference between the production season of goods and the current month Quantify seasonal deviations, combine the maximum deviation value that may exist in the seasonal range of transported goods in the current month, that is, the denominator 6 to map rationality to the interval [0,1]. When it is closer to 1, the rationality is higher and the risk is lower. When it is closer to 0, the rationality is lower and the risk is greater, thus realizing dynamic evaluation of cross-seasonal transportation. In addition, the weight parameter within the time feature Through the optimization of the XGBoost machine learning method, the importance of time period and seasonal characteristics is balanced, and time dimension anomalies such as transportation in non-historical high-frequency periods and off-season transportation are effectively identified, which makes up for the problem of traditional manual inspection's lack of sensitivity to time patterns.

[0065] The core content of the spatial feature algorithm is: by integrating the rationality of the origin of goods and the deviation of the transportation path, a weighted calculation is used to comprehensively quantify the spatial feature values ​​in the process of cargo transportation. The rationality of the origin of goods is evaluated by matching the relationship between the entrance / exit toll station and the origin of goods; the deviation of the transportation path is calculated based on the difference ratio between the actual path and the shortest billing mileage of the network toll collection. When the spatial feature The closer it is to 1, the smaller the risk, and vice versa.

[0066] Based on this, the spatial characteristics of this embodiment The calculation process is:

[0067]

[0068] in, is the weight parameter within the spatial feature (used to balance the rationality of the origin of goods and the rationality of the transportation route); The rationality of the origin of the goods and the rationality of passing through the entrance toll station and the rationality of the exit toll booths The indicator function combination is calculated (i.e. ,in, 、 , is the indicator function; M is the toll station set corresponding to the main production place of this type of goods. When x belongs to the goods production place dataset M, , otherwise it is 0; when and hour, ;when and ,or and hour, ;when and hour, . 、 and is an adjustment coefficient in the interval [0,1] and satisfies ); To determine the rationality of the transport path, the path deviation is mapped by an exponential decay function (i.e. ,in, is the path deviation, which is calculated by the actual mileage of the vehicle The shortest toll distance from entrance to exit calculated by the network toll collection system The difference ratio is calculated. , is the attenuation coefficient of path deviation, , used to control the impact of deviation on rationality. The larger the value, the more sensitive the deviation is to the penalty of score).

[0069] The above spatial characteristics The calculation process of the goods origin is reasonable Through the matching relationship between the entry / exit toll station and the cargo origin dataset M, combined with The step coefficient is used to quantify the rationality of the association between the starting point, the end point and the place of origin, and accurately identify abnormal flows in and out of non-places of origin; at the same time, the rationality of the transportation path Through the exponential decay function Mapping path deviation , the attenuation coefficient of path deviation Control the deviation penalty intensity so that the path rationality decays nonlinearly with the increase of deviation, and sensitively capture hidden anomalies such as detour transportation; in addition, the weight parameter within the spatial feature Dynamically balance the impact of origin matching and path deviation, achieve cross-validation in spatial dimensions, and effectively identify complex spatial risks such as reasonable origin but abnormal path, normal path but inconsistent origin, etc.

[0070] The core content of the behavior feature algorithm is: extracting the traffic records within the current vehicle time window, and quantifying the degree of abnormality of the vehicle transport vehicle traffic behavior through traffic count statistics, normal traffic count threshold comparison, time concentration variance analysis, and dynamic adjustment of abnormal amplification coefficient. The closer it is to 1, the smaller the risk, and vice versa.

[0071] Based on this, the behavioral characteristics of this embodiment The calculation process is:

[0072]

[0073] in, is the weight parameter within the behavioral feature (used to balance the rationality of the number of trips and the regularity of the trip time); The actual number of times the current vehicle passes during the statistical period (the statistical period can be set to 1 month). The average number of times vehicles of the same type (same model, same cargo type) pass through the same statistical period; Counting short-term intensive traffic ( ,in, is the i-th travel time, is the minimum reasonable time interval threshold (set according to the characteristics of cargo transportation, for example, 6 hours), is an indicator function. When the time interval between two consecutive passes is less than 1 if yes, 0 otherwise); is the time dispersion variance, which is calculated based on the mean of the vehicle travel time interval ( , the larger the variance of time dispersion, the more irregular the travel time. , Indicates the total number of vehicle passage intervals in the current statistical period, that is, ); is the penalty coefficient for short-term intensive traffic ( , the larger the value, the heavier the penalty for dense traffic); is the influencing factor for adjusting the sensitivity of time dispersion variance ( , the larger the value, the stronger the influence of variance on the results).

[0074] The above behavioral characteristics The calculation process of the number of passes is reasonable through the actual number of passes. Average with similar Ratio and penalties for short-term intensive traffic , quantify frequency deviation and dense traffic anomaly, penalty coefficient for short-term dense traffic Strengthen the identification of suspicious behaviors such as high-frequency short-distance round trips; at the same time, the regularity of travel time is analyzed through the variance of time dispersion. With the exponential function , mapping the volatility of the time interval to the interval [0,1), adjusting the influencing factor of the time discreteness variance sensitivity Enhance the sensitivity to irregular high-frequency traffic; in addition, the weight parameters within the behavioral characteristics By balancing the number and time characteristics, it can achieve dynamic characterization of vehicle behavior patterns and effectively identify hidden behavioral anomalies such as frequencies exceeding similar frequencies, intensive round trips in a short period of time, and irregular traffic intervals, making up for the shortcomings of traditional manual inspections in capturing behavioral patterns.

[0075] The core content of the load characteristic algorithm is to quantitatively reflect the rationality of the vehicle's transport load by integrating individual historical deviation characteristics and the distribution characteristics of similar vehicle models and cargo groups. Individual historical deviation characteristics are based on the mean and standard deviation of the vehicle's historical load; group distribution characteristics are based on the statistical distribution of similar vehicle models and cargo loads. When it is close to 1, the risk is smaller, otherwise the risk is greater.

[0076] Based on this, the load characteristics of this embodiment The calculation process is:

[0077]

[0078] in, is the weight parameter within the load feature (used to balance the individual historical deviation feature and the group deviation feature); is the individual historical deviation characteristic item ( , where Q is the original deviation value of the target vehicle load, , is the current load of the target vehicle, is the mean of the vehicle’s historical load data, , is the i-th historical load data of the vehicle, is the sample size of historical load data, is the standard deviation of the vehicle’s historical load data, , is the influencing factor of individual bias sensitivity, , the larger the value, the higher the sensitivity to individual deviations); is the deviation characteristic item of the same vehicle type and cargo group ( ,in, is the group deviation normalized value, , is the average load of the same type of vehicle and cargo (i.e. the average load of vehicles of the same type and cargo type). is the standard deviation of the load of the same type of vehicle and cargo (i.e., the dispersion of the load of vehicles of the same type and cargo type), is the standard normal distribution function (indicating the probability that a variable is less than or equal to Z).

[0079] The above load characteristics The calculation process of individual historical deviation characteristic item Through the Sigmoid function Mapping the original deviation value Q of the target vehicle load, the influencing factor of individual deviation sensitivity Adjust the sensitivity to smooth out the abnormal fluctuations of individual loads and avoid rigid misjudgment of static thresholds; at the same time, the deviation feature items of the same vehicle type and cargo group Through the standard normal distribution function Normalize the group bias Convert to two-sided probability , and then through the rationality of the two-sided probability mapping group, the horizontal comparison across vehicles is achieved; in addition, the weight parameter within the load feature It balances individual and group characteristics, identifies individual historical load mutations, and captures anomalies that significantly deviate from the load of similar groups, solving the problem of poor individual adaptability and neglect of group differences in traditional single load detection.

[0080] The core content of the historical feature algorithm is to achieve dynamic evaluation of vehicle credit score through time decay credit score calculation, weighted evaluation of historical pass rate, and analysis of the latest inspection pass rate. The closer it is to 1, the higher the historical inspection pass rate is and the lower the risk is. The closer it is to 0, the higher the historical inspection pass rate is and the lower the risk is.

[0081] Based on this, the historical credit characteristics of this embodiment The calculation process is:

[0082]

[0083] in, is a dynamic weight (used to reflect the attenuation effect of inspection time, ,in, is the current time, is the time of the i-th inspection (unit: day), is the time decay coefficient, ranging from 0.01 to 0.05. The larger the value, the faster the historical data decays); is the result of the i-th inspection (a result of 1 indicates qualified, and a result of 0 indicates unqualified); is the fluctuation coefficient of the inspection result sequence ( , that is, counting the number of changes in two adjacent inspection results); is the number of inspections (i.e. the total number of historical inspection records); is the volatility penalty index ( , the larger the value, the heavier the penalty for fluctuation).

[0084] The above historical credit characteristics The calculation process of dynamic weight Exponential decay is used to strengthen the impact of recent inspection results, highlight recent inspection results, make credit assessment closer to the current status of the vehicle, automatically weaken the impact of outdated data, and solve the timeliness problem of historical data; at the same time, the fluctuation penalty item By counting the number of changes in adjacent inspection results, credit downgrades are made for unstable behaviors (such as frequent alternation between qualified and unqualified), and the volatility penalty index Amplify the negative impact of fluctuations; in addition, the comprehensive weighted calculation method converts the historical pass rate into a credit value in the [0,1] range, effectively identifying high-risk vehicles with low historical pass rates, frequent recent failures, and large fluctuations in inspection results, making up for the shortcomings of traditional manual inspections in tracing back historical behaviors.

[0085] The multi-dimensional feature vector construction module is configured to: transform the time feature , spatial characteristics , the behavioral characteristics , load characteristics and historical credit characteristics Combination, constructing the multi-dimensional feature vector of green pass vehicles , .

[0086] Comprehensive scoring module, for Build a comprehensive scoring model and output a comprehensive score , , ,in, is the linear weighted term of the standardized eigenvalue, is the adaptive coupling term (i.e., the computational term used to quantify the interaction between features), is the weight parameter used to balance the normalized eigenvalue linear weighting term and the adaptive coupling term.

[0087] In this embodiment, the calculation process of the normalized eigenvalue linear weighted term is:

[0088]

[0089] in, 、 、 、 and Time characteristics , spatial characteristics , behavioral characteristics , load characteristics Dynamic adjustment weight parameters of historical credit characteristics (used to reflect the role of each characteristic in the comprehensive score) The relative importance of ).

[0090] The calculation process of the linear weighted items of the standardized eigenvalues ​​is optimized by dynamically adjusting the weight parameters using the XGBoost machine learning method. This method can adaptively adjust the importance of each feature based on different cargo types and transportation scenarios (e.g., fresh cargo may place more emphasis on time characteristics, while heavy cargo may place more emphasis on load characteristics). In addition, the linear weighted calculation method quantifies the explicit contribution of each feature, preventing any feature from being excessively neglected, ensuring that single-dimensional risks such as temporal anomalies and spatial anomalies are all included in the comprehensive assessment, and improving the model's coverage of diverse abnormal scenarios.

[0091] Furthermore, the calculation process of the adaptive coupling term is:

[0092]

[0093] in, is the dynamic strength coefficient, (used to weaken the coupling effect of high-resolution scenes); and are the eigenvalues ​​of different features ( ,Right now and are the values ​​of any two different features among T, S, B, W, and H); is the business tolerance threshold (set according to the actual business's tolerance for feature differences. The larger the value, the higher the tolerance for feature differences. The value can be, for example, ).

[0094] The calculation process of the above adaptive coupling term, the state intensity coefficient Weaken the coupling effect of high-scoring scenarios (i.e., low-risk types) to avoid excessive punishment of normal vehicles; at the same time, the Gaussian kernel function Quantify the differences between different feature values. When there is an implicit correlation between features (such as unreasonable time and route deviation, abnormal load and low historical credit), the difference is reduced, which increases the coupling item value and strengthens the identification of complex risks. In addition, the business tolerance threshold By controlling the sensitivity of feature differences, we can accurately capture weak single feature anomalies but strong cross-anomalies, making up for the shortcomings of traditional linear models in identifying feature interaction risks. The Gaussian kernel is calculated for the difference, and the coupling term value is obtained by summing. According to the properties of the Gaussian kernel function, when the feature difference When it increases, When the result approaches 0, the coupling term value decreases; conversely, when the difference is small, the coupling term value approaches the number of feature pairs (for example, there are 10 pairs of 5 features, and the maximum value is 10).

[0095] The weight training module is configured to train and adjust the weight parameters in the feature enhancement algorithm and the comprehensive scoring model through the XGBoost machine learning method (eXtremeGradientBoosting, an efficient machine learning algorithm based on the gradient boosting framework, widely used in tasks such as classification, regression, and sorting).

[0096] In this embodiment, the weight parameters of the feature enhancement algorithm and the weight parameters of the comprehensive scoring model include:

[0097] Weight parameters within the time feature , weight parameters within spatial features , weight parameters within behavioral features , weight parameters within the load feature , dynamic adjustment weight parameters of time characteristics , dynamic adjustment weight parameters of spatial features , Dynamic adjustment weight parameters of behavioral characteristics , Dynamically adjust weight parameters of load characteristics , Dynamic adjustment weight parameters of historical credit characteristics and the weight parameter used to balance the normalized eigenvalue linear weighting term and the adaptive coupling term .

[0098] Furthermore, the above weight parameters are optimized to correct the loss function The training is obtained, among which, is the total amount of training samples, is the true label of the i-th sample (1 represents an abnormal vehicle, 0 represents a normal vehicle), The model predicts the probability that the i-th sample is abnormal, is the L1 regularization term or the L2 regularization term (used to prevent model overfitting).

[0099] That is, the process of weight parameter optimization using the XGBoost machine learning method is as follows:

[0100] Input: standardized multidimensional feature vector ;

[0101] Output: binary risk probability y∈{0,1}, indicating the probability that the sample belongs to the positive class;

[0102] Corrected loss function: ;

[0103] Output correction weight: 、 、 、 、 、 、 、 、 as well as .

[0104] The above weight parameters are optimized and the loss function is corrected Iterative optimization through machine learning 、 、 、 、 、 、 、 、 as well as , avoid dependence on artificial experience, make parameters adapt to the actual data distribution, and improve the generalization ability of the model.

[0105] The risk grading module is configured to: The risk level of green pass vehicles is determined through risk grading strategy.

[0106] In this embodiment, the risk classification strategy is: When , the risk level is low risk; when When the risk level is medium risk; when When the risk level is high, 、 and is a learnable parameter threshold (obtained through iterative optimization of training data, e.g. 、 、 ).

[0107] The aforementioned risk grading strategy features learnable parameter thresholds optimized through training data, adapting to the risk distribution characteristics of different regions and time periods (e.g., differentiated risk thresholds during peak and off-season periods). Furthermore, the risk grading strategy maps comprehensive scores to low, medium, and high risks, providing clear guidance for inspections and focusing resources on high-risk vehicles, thus resolving the inefficiency caused by indiscriminate manual inspections.

[0108] To summarize, the green pass vehicle audit model based on multi-dimensional feature cross-analysis in this embodiment constructs five-dimensional feature vectors of time, space, behavior, load, and history, and relies on core algorithms such as the cosine similarity algorithm to quantify time period matching and the exponential decay function to map path deviation to achieve cross-analysis of multi-dimensional features. It also uses XGBoost machine learning to dynamically optimize feature weights and risk grading thresholds. On the technical level, the model captures the implicit correlation between features such as seasonal irrationality and path deviation through an adaptive coupling module, and uses time-decayed credit points to strengthen the weight of recent abnormal behaviors, effectively reducing the false positive rate and missed positive rate, and solving the problem of missed detection of compound risks in traditional single-dimensional detection.

[0109] In a specific application case, a certain unit used the data analysis and early warning service of the Green Pass system based on the Green Pass vehicle audit model based on multi-dimensional feature cross-analysis of this embodiment to achieve feature analysis of the Green Pass data flow, post-audit, and accurate identification of suspected fake Green Pass vehicles. The system background automatically generates a key focus list and high-frequency inspection records for driving licenses. When a suspected vehicle enters a toll station, the system automatically triggers a real-time alarm and accurately pushes it to the on-site inspection front-end APP, providing a decision-making basis for the front-end Green Pass inspection and key audit. According to incomplete statistics, as of December 31, 2024, a total of 2,202 license plates were included in the province's key focus list. By using the Green Pass inspection and early warning function to assist in inspection, a total of approximately 500,000 suspected vehicles were detected, and toll losses of more than 200 million yuan were recovered. The Green Pass unqualified detection rate and inspection efficiency have been significantly improved. The specific performance is as follows:

[0110] (1) Precision implementation of policies

[0111] The average time for single-lane green pass inspections has been shortened from 11 minutes to 7.5 minutes, with efficiency increased by 33%, effectively alleviating traffic congestion pressure, ensuring the accurate implementation of the green channel policy for the transportation of fresh agricultural products, and significantly improving the travel efficiency of legal drivers.

[0112] (2) Optimization of charging order

[0113] As of December 31, 2024, it has assisted in the inspection of approximately 15.96 million vehicles on roads across the province, involving a free green pass inspection amount of approximately 17.88 billion yuan, an unqualified detection rate of 3.2%, and recovered toll losses of more than 200 million yuan, which not only ensured the operational efficiency of the road network, but also purified the traffic environment for compliant vehicles.

[0114] (3) Improved management effectiveness

[0115] The reduction in manual data entry workload and the increase in inspection efficiency, coupled with the analysis of inspected vehicles by the intelligent audit reasoning model, have effectively maintained the toll collection order for green pass vehicles on highways, while also protecting the rights and interests of legitimate transport entities and the efficiency of road network operations.

[0116] In a second aspect, this embodiment provides a green pass vehicle audit method based on multi-dimensional feature cross-analysis, which is applicable to the green pass vehicle audit model based on multi-dimensional feature cross-analysis described above. The method includes the following steps:

[0117] The original features of green-pass vehicles are processed by feature enhancement algorithm, and the time features are calculated respectively. , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics The original features include original data related to time features, original data related to space features, original data related to traffic behavior, original data related to load, and original data related to historical inspection records;

[0118] The time feature , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics Combination, constructing the multi-dimensional feature vector of green pass vehicles , ;

[0119] Based on the multidimensional feature vector Build a comprehensive scoring model and output a comprehensive score , ,in, is the linear weighted term of the standardized eigenvalue, is the adaptive coupling term, is the weight parameter used to balance the normalized eigenvalue linear weighting term and the adaptive coupling term;

[0120] Training and adjusting the weight parameters in the feature enhancement algorithm and the comprehensive scoring model by using the XGBoost machine learning method;

[0121] Based on the comprehensive score The risk level of green pass vehicles is determined through risk grading strategy.

[0122] It should be noted that the green pass vehicle audit method based on multi-dimensional feature cross-analysis of this embodiment corresponds to the aforementioned green pass vehicle audit model based on multi-dimensional feature cross-analysis. Therefore, the parts that are not specifically described in the green pass vehicle audit method based on multi-dimensional feature cross-analysis of this embodiment (including but not limited to specific technical means and technical effects) can be referred to the relevant description in the aforementioned green pass vehicle audit model based on multi-dimensional feature cross-analysis, and this text will not go into details here.

[0123] In the embodiments provided herein, it should be understood that the embodiments described herein can be implemented using hardware, software, firmware, middleware, code, or any appropriate combination thereof. For hardware implementation, the processor may be implemented in one or more of the following: an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a processor, a controller, a microcontroller, a microprocessor, or other electronic units designed to implement the functionality described herein, or any combination thereof. For software implementation, some or all of the processes of the embodiments may be performed by a computer program instructing the relevant hardware. During implementation, the program may be stored in a computer-readable storage medium or transmitted as one or more instructions or codes on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media includes any medium that facilitates the transmission of a computer program from one location to another. The storage medium may be any available medium that can be accessed by a computer. Computer-readable storage media may include, but are not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing the desired program code in the form of instructions or data structures and accessible by a computer.

[0124] Finally, it should be noted that the above is only a preferred embodiment of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent replacements for some of the technical features therein. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A green pass vehicle audit model based on multi-dimensional feature cross analysis, characterized by: include: The feature enhancement module is configured to process the original features of the green pass vehicle through the feature enhancement algorithm and calculate the time features respectively. , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics The original features include original data related to time features, original data related to space features, original data related to traffic behavior, original data related to load, and original data related to historical inspection records; The multi-dimensional feature vector construction module is configured to: transform the time feature , spatial characteristics , the behavioral characteristics , load characteristics and historical credit characteristics Combination, constructing the multi-dimensional feature vector of green pass vehicles , ; Comprehensive scoring module, for Build a comprehensive scoring model and output a comprehensive score , ,in, is the linear weighted term of the standardized eigenvalue, is the adaptive coupling term, is the weight parameter used to balance the normalized eigenvalue linear weighting term and the adaptive coupling term; A weight training module is configured to: train and adjust the weight parameters in the feature enhancement algorithm and the comprehensive scoring model through the XGBoost machine learning method; The risk grading module is configured to: Determine the risk level of green pass vehicles through risk grading strategy; The time characteristics The calculation process is: in, is the weight parameter within the time feature; is the time period matching degree, which is calculated by the cosine similarity between the current time period vector and the historical time period distribution vector; Seasonal deviations for transported goods; The spatial characteristics The calculation process is: in, is the weight parameter within the spatial feature; The rationality of the origin of the goods and the rationality of passing through the entrance toll station and the rationality of the exit toll booths The indicator function combination is calculated; The rationality of the transportation path is obtained by mapping the path deviation through the exponential decay function; The behavioral characteristics The calculation process is: in, is the weight parameter within the behavioral feature; is the actual number of times the vehicle passes during the current statistical period, The average number of times the same type of vehicles pass through the same statistical period; Counting short bursts of traffic; is the time dispersion variance, which is calculated based on the mean of the vehicle travel time interval; is the penalty coefficient for short-term intensive traffic; is the influencing factor for adjusting the sensitivity of time dispersion variance; The load characteristics The calculation process is in, is the weight parameter inside the load feature; is the individual historical deviation characteristic item; It is the deviation characteristic item of the same vehicle type and cargo group; The historical credit characteristics The calculation process is: in, is the dynamic weight; is the result of the i-th inspection; To check the fluctuation coefficient of the result sequence; is the number of inspections; is the volatility penalty index.

2. The green pass vehicle audit model based on multi-dimensional feature cross analysis according to claim 1 is characterized in that: The calculation process of the standardized eigenvalue linear weighted term is: in, 、 、 、 and Time characteristics , spatial characteristics , behavioral characteristics , load characteristics and dynamically adjust weight parameters based on historical credit characteristics.

3. The green pass vehicle audit model based on multi-dimensional feature cross analysis according to claim 2 is characterized in that: The calculation process of the adaptive coupling term is: in, is the dynamic strength coefficient, ; and are the eigenvalues ​​of different features respectively; The service tolerance threshold.

4. The green pass vehicle audit model based on multi-dimensional feature cross analysis according to claim 1 is characterized in that: The weight parameters of the feature enhancement algorithm and the weight parameters of the comprehensive scoring model include: Weight parameters within the time feature , weight parameters within spatial features , weight parameters within behavioral features , weight parameters within the load feature , dynamic adjustment weight parameters of time characteristics , dynamic adjustment weight parameters of spatial features , Dynamic adjustment weight parameters of behavioral characteristics , Dynamically adjust weight parameters of load characteristics , Dynamic adjustment weight parameters of historical credit characteristics and the weight parameter used to balance the normalized eigenvalue linear weighting term and the adaptive coupling term ; The weight parameters of the feature enhancement algorithm and the weight parameters of the comprehensive scoring model are modified by optimizing the loss function The training is obtained, among which, is the total amount of training samples, is the true label of the i-th sample, The model predicts the probability that the i-th sample is abnormal, is the L1 regularization term or the L2 regularization term; The risk grading strategy is: When , the risk level is low risk; when When the risk level is medium risk; when When the risk level is high, 、 and is a learnable parameter threshold.

5. A green pass vehicle audit method based on multi-dimensional feature cross analysis, applicable to the green pass vehicle audit model based on multi-dimensional feature cross analysis as described in any one of claims 1 to 4, characterized in that: The method The following steps are involved: The original features of green-pass vehicles are processed by feature enhancement algorithm, and the time features are calculated respectively. , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics The original features include original data related to time features, original data related to space features, original data related to traffic behavior, original data related to load, and original data related to historical inspection records; The time feature , spatial characteristics , behavioral characteristics , load characteristics and historical credit characteristics Combination, constructing the multi-dimensional feature vector of green pass vehicles , ; Based on the multidimensional feature vector Build a comprehensive scoring model and output a comprehensive score , ,in, is the linear weighted term of the standardized eigenvalue, is the adaptive coupling term, is the weight parameter used to balance the normalized eigenvalue linear weighting term and the adaptive coupling term; Training and adjusting the weight parameters in the feature enhancement algorithm and the comprehensive scoring model by using the XGBoost machine learning method; Based on the comprehensive score The risk level of green pass vehicles is determined through risk grading strategy.

Images