Highway vehicle pass auditing method and system based on "one vehicle one file"

Abstract

The application relates to a highway vehicle passing auditing method and system based on 'one vehicle one file', wherein the method comprises the following steps: S1, constructing a vehicle file, collecting basic information, passing records, characteristic information and auditing information of the vehicle, and updating and verifying the vehicle file according to the collected information; S2, performing file fusion and preprocessing of multi-source data corresponding to each vehicle file; S3, extracting and updating multi-level structured characteristics; S4, performing risk research and judgment: calling an atomized expert rule model to match the structured characteristics, and calling a pre-trained AI model to reason the structured characteristics; obtaining a composite auditing strategy preset in a visual arrangement mode; and S5, synthesizing an electronic evidence chain and associating the electronic evidence chain to an auditing work order. The highway vehicle passing auditing method based on 'one vehicle one file' can realize an automatic, intelligent and accurate highway passing auditing scheme.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic management and data processing technology, specifically to a method and system for highway vehicle passage auditing based on "one vehicle, one file". Background Technology

[0002] With the continuous improvement of my country's expressway network and the widespread adoption of ETC (Electronic Toll Collection), expressway traffic efficiency has been greatly enhanced. However, toll evasion and underpayment have become increasingly diversified and covert, seriously damaging the legitimate rights and interests of expressway operators. Existing auditing methods suffer from the following shortcomings: Data silos: Data collected by various gantries and toll stations are independent, lacking effective correlation and integration, making it difficult to trace complete vehicle passage behavior. Lack of vehicle "profiles": Traditional methods only judge single passage behavior, lacking analysis of long-term characteristics such as vehicle history and credit status, making it impossible to accurately identify and warn of high-risk vehicles. Auditing lag: Most audits rely on post-event manual screening, which is inefficient in the face of massive amounts of traffic data and easily overlooks carefully planned toll evasion. Furthermore, after discovering suspected toll evasion, evidence needs to be manually retrieved from different systems, a cumbersome process that makes it difficult to quickly form a complete and clear chain of evidence for toll collection. Summary of the Invention

[0003] This invention provides a highway vehicle passage audit method based on "one vehicle, one file", which can realize an automated, intelligent and precise highway passage audit solution.

[0004] The present invention provides a highway vehicle passage audit method based on "one vehicle, one file," comprising:

[0005] S1, Construct vehicle files, including: integrating external databases; establishing a unique identifier for each vehicle and creating a vehicle file for each vehicle; collecting the vehicle's basic information, passage records, characteristic information, and audit information, and updating and verifying the vehicle file with the collected information;

[0006] S2, corresponding to the file fusion and preprocessing of multi-source data for each vehicle, including: accessing multi-source heterogeneous data; standardizing, cleaning and normalizing the data; using the unique vehicle identifier combined with timestamps and spatial locations, aligning and associating the data of the entrance, gantry and exit in time and space, and constructing a "trip-data block" time and space association model to describe a single complete passage;

[0007] S3, extract and update multi-level structured features, which include: basic fact features directly extracted from raw data, rule features generated by comparing basic fact features with business rules, and behavioral patterns and risk features mined from historical archives through statistical analysis or machine learning; perform scenario-based feature analysis on the extracted features, and generate a risk profile of the vehicle based on the features;

[0008] S4, Perform risk assessment: Invoke an atomized expert rule model to match the structured features, and invoke a pre-trained AI model to infer the structured features; obtain a pre-defined composite audit strategy through a visual orchestration method, wherein the composite audit strategy defines the combination logic between the matching results of the expert rule model and the inference results of the AI ​​model; perform logical combination operations on the matching results and / or inference results according to the composite audit strategy to generate a suspicion assessment result; dynamically distribute the audit calculation task to the real-time computing layer, near-line computing layer, or offline computing layer for execution according to the complexity, timeliness requirements, and data volume of the composite audit strategy, and dynamically allocate computing resources according to the negative computing resource allocation.

[0009] S5, synthesize the electronic evidence chain and link it to the audit work order.

[0010] As a preferred approach, in S2, when accessing multi-source heterogeneous data, the data sources across the entire domain are defined and categorized for access. Multi-source heterogeneous data includes real-time streaming data sources, batch data sources, and external data sources. Real-time streaming data sources include real-time data from each toll station accessed through a message queue; batch data sources include data uploaded from each toll station and gantry that is periodically extracted; and external data sources include cross-provincial traffic data, vehicle registration information, and vehicle release information from external databases.

[0011] As a preferred option, in S2, a data asset catalog and metadata management are established for the accessed multi-source heterogeneous data.

[0012] As a preferred option, in S2, the standardization and normalization of data includes: performing structured data cleaning on transaction data; and performing unstructured data preprocessing on images / videos.

[0013] As a preferred approach, when the same license plate is identified in different locations within a very short period of time, the spatiotemporal overlap verification technology is used to perform spatiotemporal consistency cross-verification on multi-source data and flow data, with the vehicle's unique identifier, timestamp, and spatial location as the core. The vehicle's trajectory temporal and spatial continuity and rationality are verified based on the time dimension and road network spatial dimension.

[0014] As a preferred method, when traffic volume is high, lane coil triggering signals and video target tracking segments are introduced for auxiliary correction to help correct matching errors between adjacent vehicle images.

[0015] As a preferred option, S2 also includes building a standardized thematic data layer and quality monitoring, including: generating a "vehicle-trip" star-shaped data layer from the associated and fused data, including a trip fact table, a trip detail table, a vehicle file table, and an image feature index table; and establishing full-link data quality monitoring and closed-loop feedback.

[0016] Preferably, in S3, the basic factual features include spatiotemporal facts, transaction facts, and vehicle facts; and / or the rule features include consistency features between entry and exit, compliance features related to toll policies, and labeling features based on simple rules; and / or the behavioral patterns and risk features include statistical features derived from long-term historical archives, aggregation features, and model-derived features output by AI models.

[0017] Preferably, in S3, the automated feature extraction and update engine includes a stream processing framework and a batch processing framework. The stream processing framework is used to calculate the dynamic features of vehicles in transit in real time and update the dynamic feature part in the vehicle file in real time. The batch processing framework is used to scan and calculate the full vehicle file at fixed intervals and update the periodic feature part in the vehicle file.

[0018] Preferably, in S4, each atomic rule is accompanied by metadata, which includes the rule creation and update time, historical trigger count, hit accuracy, associated toll evasion type, and applicable vehicle model range.

[0019] Preferably, S4 further includes: the expert rule model for invoking atomization also includes a rule dynamic management step: real-time monitoring of the hit rate, accuracy and resource consumption of each expert rule, and dynamic adjustment of the priority, activation status or trigger threshold of the rule according to the monitoring results; obtaining newly added evasion cases confirmed by manual review, and iteratively optimizing the parameters of the corresponding expert rule according to the newly added evasion cases.

[0020] Preferably, in S4, calling the pre-trained AI model to infer the structured features includes: extracting multimodal features and assembling them into feature vectors; inputting the feature vectors into the pre-trained AI model to obtain the inference result of suspected evasion; the AI ​​model can learn and provide feedback online.

[0021] Preferably, the multimodal features include spatiotemporal features, transaction behavior features, vehicle features, and related features. Historical audit work orders that have been finally confirmed by human intervention are used as labeled samples and associated with the feature vectors to form a supervised learning dataset.

[0022] As a preferred option, in S4, the pre-trained AI models include: a traffic trajectory anomaly prediction model, which predicts a reasonable path range based on vehicle files and real-time traffic trajectories, and marks anomalies when the actual path deviates significantly; a driving behavior clustering analysis model, which clusters all vehicle traffic behavior features and identifies abnormal groups whose behavior patterns deviate from the mainstream cluster; and an image sequence analysis model, which analyzes subtle changes in vehicles in continuous gantry-captured images to assist in identifying card-switching and license plate-changing behaviors.

[0023] As a preferred approach, online learning and feedback for AI models include: obtaining the final conclusions of auditors after manually reviewing the AI ​​model's prediction results as new labeled samples, automatically feeding them back to the AI ​​model training platform, and periodically using incremental data to fine-tune the AI ​​model.

[0024] Preferably, in S4, obtaining the pre-set composite audit strategy through visual orchestration includes: selecting atomized expert rules and / or AI model output results as strategy components from the model component library through a visual orchestration interface, and connecting the selected strategy components through logic gate components to generate an audit strategy flowchart; performing sandbox simulation on the generated audit strategy flowchart to output a list of simulated hit vehicles and a matching report with historical confirmed cases; and deploying the audit strategy flowchart that has passed the simulation review to the production environment as a new composite audit strategy for subsequent audit scanning tasks.

[0025] Preferably, in step S4, the dynamic distribution of audit calculation tasks includes: allocating high-frequency, high-accuracy models to the real-time streaming computing layer to provide immediate risk warnings for vehicles in transit; and allocating low-frequency, complex models to the offline batch computing layer to periodically perform in-depth scans of the full or sampled vehicle files.

[0026] Preferably, in S5, the synthesis of the electronic evidence chain and its association with the audit work order includes: automatically generating a structured audit trigger command when suspected toll evasion is determined; starting the intelligent collection and scheduling engine, matching the corresponding evidence template according to the suspected toll evasion type, and retrieving multi-source heterogeneous evidence from the vehicle file according to the evidence template; standardizing and preprocessing the retrieved multi-source heterogeneous evidence, adding a timestamp, spatial location tag, and metadata to each piece of evidence to generate a standardized evidence package; sorting and aligning the standardized evidence package according to millisecond-level time precision and spatial order using the vehicle's actual travel path or abnormal travel section as the time axis to generate a visual evidence trajectory map; dynamically synthesizing the sorted and aligned evidence package into an electronic evidence chain, and automatically generating an audit reason summary based on the content of the electronic evidence chain; and binding the electronic evidence chain and the audit reason summary to the audit work order.

[0027] As a preferred approach, before automatically binding the evidence chain and summary to the audit work order, a self-check of the evidence chain is performed, and the missing items and their potential impact are marked if key evidence is missing.

[0028] This invention also provides a highway vehicle passage audit system based on "one vehicle, one file", comprising:

[0029] The data source layer includes toll stations, gantries, and external databases. It collects transaction records, captured images, and videos through toll stations and gantries, and collects vehicle issuance information, license plate records, and vehicle information from the transportation department through external databases.

[0030] The data processing and core engine layer includes a vehicle file center, a data fusion and preprocessing module, an automated feature extraction and update engine, an intelligent audit model library, and an automated evidence chain generation module. The vehicle file center integrates external databases to establish a unique identifier for each vehicle, creates a vehicle file for each vehicle, and collects basic information, travel records, feature information, and audit information for that vehicle. It then updates and verifies the vehicle file using the collected information. The data fusion and preprocessing module accesses multi-source heterogeneous data, performs standardized cleaning and normalization processing on the data, and constructs a "trip-data block" spatiotemporal correlation model. The automated feature extraction and update engine... The new engine extracts and updates multi-level structured features, performs scenario-based feature analysis, and generates risk profiles. The intelligent audit model library includes an expert rule library based on business knowledge graphs, where the rules are atomic expert rules. The intelligent audit model library also includes a deep mining AI model based on multimodal feature learning and a composite audit strategy factory based on visual orchestration. The evidence chain automatic generation module uses the actual travel path of the vehicle as the time axis, captures the required multi-source heterogeneous evidence, performs millisecond-level spatiotemporal alignment and sorting of all evidence, automatically synthesizes an electronic evidence chain containing a visual trajectory map, and electronically binds the evidence chain to the audit work order.

[0031] Preferably, the data processing and core engine layer also includes a closed-loop feedback module, which is used to automatically feed back the final conclusions of the auditors after manually reviewing the prediction results of the AI ​​model as newly labeled samples to the model training platform, and periodically use incremental data to fine-tune the training of the AI ​​model.

[0032] Preferably, the system further includes an application layer, which includes a real-time early warning dashboard, audit work order management, and multi-dimensional analysis reports.

[0033] Compared with existing technologies, this invention has the following advantages: The auditing method of this invention integrates multi-source data to construct a dynamic holographic file for each vehicle. Multi-level features are extracted from the file, and a smart model integrating rules and AI algorithms is used for risk assessment. When suspicion is detected, the system automatically collects and binds multimodal evidence, generating a structured electronic evidence chain. This achieves greater precision in auditing. By creating a "one vehicle, one file" approach, the audit focus shifts from isolated events to the vehicle's entire lifecycle behavior, enabling more accurate identification of disguised and complex toll evasion behaviors. Furthermore, it achieves not only intelligence and automation, but also automatic analysis of massive amounts of traffic data using various analytical models and machine learning technologies. This significantly reduces manual intervention, improves auditing efficiency, and ensures a complete evidence chain. All evidence originates from the vehicle's own file, automatically generated and linked, resulting in a complete and highly credible evidence chain that greatly facilitates subsequent toll collection efforts. Finally, a closed-loop feedback mechanism enables fully automated auditing from data to decision-making, with continuous optimization capabilities. Through continuous learning of behavioral patterns, early warnings can be issued for vehicles or abnormal behaviors exhibiting high-risk characteristics, enabling a shift from post-event auditing to in-process intervention and pre-event warning. Furthermore, the structured data foundation of "one vehicle, one file" facilitates the subsequent introduction of more advanced AI algorithms and the expansion into new audit scenarios. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the multi-source data fusion and preprocessing steps of a highway vehicle passage audit method based on "one vehicle, one file" according to an embodiment of the present invention.

[0035] Figure 2 This is a flowchart illustrating the steps of multi-source data fusion and preprocessing for each vehicle file in a highway vehicle passage audit method based on "one vehicle, one file" according to an embodiment of the present invention.

[0036] Figure 3 This is a flowchart illustrating the steps of extracting and updating multi-level structured features in a highway vehicle passage auditing method based on "one vehicle, one file" according to an embodiment of the present invention.

[0037] Figure 4 This is a flowchart illustrating the risk assessment steps of a highway vehicle passage audit method based on "one vehicle, one file" according to an embodiment of the present invention.

[0038] Figure 5 This is a schematic diagram of the overall architecture of a highway vehicle passage audit system based on "one vehicle, one file" according to an embodiment of the present invention. Detailed Implementation

[0039] This invention provides a highway vehicle passage auditing method based on "one vehicle, one file", including:

[0040] S1. Construct vehicle profiles, including: integrating external databases; establishing a unique identifier for each vehicle. In this embodiment, a vehicle profile is created for each vehicle using "license plate + license plate color". Establishing a unique identifier by integrating external databases effectively solves the problem of identity confusion such as cloned license plates, card swapping, and forged license plates. Basic information, traffic records, characteristic information, and audit information of the vehicle are collected, and the vehicle profile is updated and verified using the collected information. In this way, a reliable behavioral trajectory throughout the vehicle's entire lifecycle can be formed, facilitating long-term tracking and retrospection.

[0041] S2, fusion and preprocessing of multi-source data, includes: accessing multi-source heterogeneous data and registering metadata; and standardizing, cleaning, and normalizing the data. Through these processes, the scattered, heterogeneous, and multimodal raw data in the highway network toll collection system is transformed into unified, clean, high-value, vehicle-granular standardized data assets. This forms an automated, traceable, and strongly error-correcting data governance pipeline. Using the vehicle's unique identifier combined with timestamps and spatial locations, the data at entrances, gantries, and exits are spatiotemporally aligned and correlated. This allows for the precise piecing together of all data fragments from the same vehicle and the same journey into a complete chain, forming a "journey-data block" spatiotemporal correlation model to describe a vehicle's single, complete passage.

[0042] S3 involves using an automated feature extraction and update engine to extract and analyze multi-dimensional features from vehicle records. In this embodiment, these multi-dimensional features form a three-tiered structured feature system. The first tier consists of basic factual features; the second tier comprises rule features generated by combining basic factual features with business rules, serving as the basis for judging the compliance of traffic behavior; and the third tier consists of behavioral patterns and risk features mined from historical vehicle traffic records through statistical analysis and machine learning methods. This step constructs a systematic, automated, and scenario-oriented feature factory, providing accurate input for the intelligent audit model. S3 also includes performing scenario-based feature analysis on the extracted features and constructing a profile, i.e., generating a risk profile for the vehicle.

[0043] S4. Perform risk assessment by calling an atomized expert rule model to match the structured features and a pre-trained AI model to infer the structured features. Obtain a pre-defined composite audit strategy through visual orchestration, which defines the combination logic between the matching results of the expert rule model and the inference results of the AI ​​model. Perform logical combination operations on the matching results and / or inference results according to the composite audit strategy to generate a suspicion assessment result. Through an intelligent scheduler, dynamically distribute audit computing tasks to the real-time computing layer, near-line computing layer, or offline computing layer according to the complexity, timeliness requirements, and data volume of the composite audit strategy. This avoids overwhelming all audit tasks to the real-time computing layer (high cost, difficult to scale), instead selecting tasks based on their characteristics—simple tasks are processed in real time, while complex / large-batch tasks are placed in the near-line or offline layer to balance the overall system load. Real-time computing resources are expensive (high memory and CPU usage). Scheduling non-urgent large-scale audits (such as daily full-volume pass record comparison) to the offline layer can significantly reduce the overall computing cost. In scenarios requiring second-level / minute-level response, such as "instant interception of vehicles evading large tolls," the scheduler allocates these tasks to the real-time computing layer for rapid results. The intelligent scheduler also dynamically allocates computing resources based on available capacity. When the real-time layer is busy, non-urgent tasks are automatically downgraded to the near-line layer; when offline layer resources are idle, it automatically absorbs large batches of backtracking tasks to prevent single-point overload or system crashes. In summary, the intelligent scheduler achieves "on-demand layered computing + dynamic load balancing" for audit tasks, ensuring rapid response for high-efficiency audits while reducing computational costs and improving overall system throughput and stability. Step S4 ultimately forms a three-layer intelligent audit model library construction method based on "rule-driven, AI-mining, and human-machine collaboration." This intelligent audit model not only enables precise targeting of known toll evasion behaviors but also proactively discovers hidden and new toll evasion patterns, deeply integrating expert experience with data intelligence to form a dynamically optimized audit defense.

[0044] S5 generates an electronic evidence chain and links it to the audit work order: Using the vehicle's actual travel path as the timeline, it captures the necessary multi-source heterogeneous evidence, performs millisecond-level spatiotemporal alignment and sorting of all evidence, and automatically synthesizes an electronic evidence chain including a visualized trajectory map, forming a deeply integrated evidence chain. This evidence chain is then electronically bound to the audit work order. Finally, the suspicion warning output by the audit model is transformed into an audit work order with a complete and rigorous evidence chain.

[0045] The auditing method of this invention integrates multi-source data to construct a dynamic holographic profile for each vehicle. Multi-level features are extracted from the profile, and a smart model incorporating rules and AI algorithms is used for risk assessment. When suspicion is detected, the system automatically collects and binds multimodal evidence, generating a structured electronic evidence chain. Finally, a closed-loop feedback mechanism enables fully automated auditing from data to decision-making, and allows for continuous optimization.

[0046] In this embodiment, step S1 is the first step in constructing a vehicle file and establishing a "one vehicle, one file" data system, specifically including:

[0047] 1. Basic Information Collection: By connecting with information from external databases, such as vehicle information from traffic police (vehicle management offices) and vehicle issuance information from issuing units, a preliminary basic vehicle information database is formed through data cleaning and analysis.

[0048] 2. Basic Information Verification and Update: The vehicle basic information database is distributed to the toll station lanes. When a vehicle makes an exit transaction, the system automatically compares the vehicle's basic information. If there is a discrepancy, a special alert is generated to prompt the toll collector to manually verify the vehicle's basic information. Once the vehicle's basic information is confirmed, the system automatically updates the vehicle basic information database.

[0049] 3. Vehicle Traffic Information Collection: At each highway entrance toll station, license plate recognition equipment is used to acquire basic vehicle information such as license plate number and vehicle type, while simultaneously reading vehicle identification information from the toll medium. During the vehicle's journey, real-time vehicle location and passage time are collected through the gantry system along the route. At the exit toll station, the vehicle's relevant information is confirmed again, and the toll amount is recorded. Furthermore, video surveillance systems are used to acquire vehicle exterior feature image data for extracting characteristic information such as vehicle color and brand.

[0050] 4. Vehicle audit information collection: When highway operators generate vehicle audit information through audit analysis, the relevant vehicle audit information is automatically collected.

[0051] 5. Vehicle File Creation and Storage: Vehicle files are created in the database using the vehicle's unique identifier (such as license plate number and color) as an index. Collected basic vehicle information, traffic records, characteristic information, and audit information are stored in the vehicle files according to a standardized format. Distributed database technology is employed to ensure the security and scalability of the file data storage, meeting the storage needs of massive amounts of vehicle data.

[0052] 6. File Updates and Maintenance: Whenever a vehicle has new passage records or audit information, promptly update the relevant information in the vehicle file. Regularly check and clean the file data, deleting expired or invalid data records to ensure the accuracy and timeliness of the file data.

[0053] Step S2, multi-source data fusion and preprocessing, is the cornerstone of building the "one vehicle, one file" intelligent audit model. Its process is as follows: Figure 1 As shown. The multi-source heterogeneous data access and metadata registration first require the establishment of a unified "data access bus" to standardize the access and management of various data sources. Specifically, the access of all data sources is first defined and categorized, as follows:

[0054] Real-time streaming data sources: These are data collected from each toll station via message queues (such as Kafka), including real-time transaction logs, gantry transaction logs, gantry license plate records, and toll station lane events (such as signals triggered by captured images). This type of data requires millisecond-level latency for real-time risk warning.

[0055] Batch data source: Through ETL tools or data exchange platforms, extract high-definition images, video clips, OBU issuance information, vehicle information databases, toll rate parameter tables, etc. uploaded by each toll station or gantry at regular intervals (e.g., every 15 minutes).

[0056] External data sources: Through a secure interface, cross-provincial traffic data from the Ministry of Transport's Road Network Center and vehicle registration information from traffic police (vehicle management offices) are imported.

[0057] Secondly, data asset cataloging and metadata management are implemented, including: establishing a detailed "data profile" for each type of data access, recording information such as its source system, data format, update frequency, field meaning, data quality (e.g., historical missing rate), and business owner. This catalog supports global retrieval and lineage analysis, serving as the basis for all subsequent data fusion, cleaning, and application.

[0058] Data cleaning and standardization / normalization are performed on the raw data to ensure data quality. Specifically, this includes:

[0059] 1. Perform structured data cleaning on transaction data, including: Invalid value filtering: automatically filtering out invalid data such as abnormal timestamps (e.g., future times), empty license plate numbers, or "no license plate" entries; Format standardization: unifying license plate numbers from different sources into a standard format of provincial code + alphanumeric characters, correcting common OCR misidentifications (e.g., "0" and "D", "1" and "I"), and labeling the original values ​​for traceability; Logical consistency verification: performing real-time verification based on simple rules, such as: entry time must be earlier than exit time; in the same transaction, vehicle type code must be within the legal range defined in the rate table.

[0060] 2. Perform unstructured data preprocessing on images / videos, including: Automated quality screening: Using a lightweight AI model, automatically determine whether an image is overexposed, underexposed, blurry, or has an obscured license plate, and mark low-quality images as "awaiting manual reshoot" or "for reference"; Structured extraction of key information: Call vehicle vision AI algorithms to extract structured information such as license plate number (secondary recognition), license plate color, vehicle model, vehicle brand, vehicle color, and vehicle face feature vector from the image, and compare it with the corresponding fields in the data stream to form a cross-validation record.

[0061] Data association and fusion based on spatiotemporal alignment and unique vehicle identifiers specifically include:

[0062] 1. Construct a "trip-data block" spatiotemporal correlation model, including: taking a complete passage (from entrance to exit) as the basic unit, called a "trip". The system uses license plate number and color, and passage time accurate to milliseconds, to spatiotemporally sort and concatenate the flow records scattered at the entrance station, multiple gantries, and exit station, initially forming a trip trajectory skeleton. All images, videos, and event logs captured within this time window and spatial path are precisely associated with the corresponding nodes of the trip skeleton through time synchronization and location tags (toll station / gantry codes), like attaching accessories.

[0063] 2. Identity resolution and error correction in complex scenarios, including: resolving "one vehicle, multiple recognitions": When the same license plate is recognized in different locations within a very short period (physically impossible), the system initiates "spatiotemporal overlap verification," combining vehicle visual features to determine whether it is an identification error or suspected license plate cloning, and creates associated files for possible genuine vehicles; handling "multiple vehicles mixed together": When traffic is heavy, images of adjacent vehicles may be mismatched. The system introduces lane coil trigger signals and video target tracking segments for auxiliary correction to ensure accurate binding of images and traffic flow.

[0064] Step S2 also includes building a standardized subject data layer and quality monitoring, specifically including:

[0065] 1. Construct a star schema data layer for "vehicle-trip" and a standardized thematic data layer. To reorganize the merged data according to business themes and form an easily analyzable data model, this specifically includes:

[0066] Trip Fact Sheet: Core facts for each trip, such as trip ID, license plate number, license plate color, entrance / exit / time, total mileage, total amount, payment method, and key feature tags.

[0067] Trip details: A detailed record of each gantry passed through during the trip, including time, location, transaction amount, and captured image ID.

[0068] Vehicle Profile Table: Vehicle-level information generated based on the trip fact table, such as monthly travel frequency, frequently used routes, and historical anomaly records (this table is the core storage for the subsequent "one vehicle, one profile" system).

[0069] Image Feature Index Table: Stores the structured features and feature vectors of all images, supporting advanced queries such as image search and vehicle model comparison.

[0070] 2. End-to-end data quality monitoring and closed-loop feedback:

[0071] Quality checkpoints are set up at each key stage of the data pipeline (access, cleaning, association, and fusion) to monitor indicators such as data volume fluctuations, latency, error rate, and association success rate. This forms a closed loop of "quality anomaly - root cause analysis - rule optimization." For example, if the correlation rate between license plate data and transaction flow for a certain road segment is found to be consistently low, the system can automatically issue an alert, prompting technicians to check the time synchronization equipment or network for that road segment, and temporarily lower the confidence weight of the data for that road segment according to the rules to prevent dirty data from polluting the global model.

[0072] The specific process of step S3, multidimensional feature extraction and analysis, is as follows: Figure 2As shown. The three layers of features in the "three-level" structured feature system are: Basic factual features: Objective facts directly extracted from or simply derived from the original data, serving as the atomic units for constructing upper-level features. These include: Spatiotemporal facts, such as entrance / exit stations, passage date (whether it's a holiday), time period (whether there are differentiated discount periods), travel duration, theoretical shortest path, etc.; Transaction facts, such as payment method (ETC / CPC), transaction amount, OBU number, vehicle type recognition result, etc.; Vehicle facts, such as license plate number, license plate color, vehicle type, brand, and model, etc. Business rule features: Status judgments generated by comparing basic factual features with toll collection business rules, serving as the direct basis for triggering the rule model. This includes: Consistency features, such as consistency between entrance and exit vehicle types, consistency between the toll collection medium (OBU / CPC) and license plate binding, and consistency between license plate recognition and payment vehicle type; Compliance features, such as whether the vehicle enjoys preferential policies such as green channels / military / police (combined with cargo identification and document information), completeness of the tolling path (whether the gantry sequence is continuous), and whether the minimum toll collection condition is triggered; Tagging features, i.e., labels based on simple rules, such as "long-distance trucks," "high-frequency tolling vehicles," and "inter-provincial vehicles." Behavioral patterns and risk features: Deep patterns and risk indicators mined through statistical analysis and machine learning models based on the long-term historical records of individual vehicles. This includes: Statistical features, such as the mean and variance of tolling frequency in the past 30 days, habitual driving routes, frequently used entrances and exits, and average driving speed preferences; Aggregation features, such as the number of historical suspected toll evasions (different categories), audit work order status, and toll payment records; Model-derived features, which are intermediate results or scores output by the AI ​​model, such as "trajectory anomaly index," "behavioral cluster number," and "probability of license plate / card replacement risk."

[0073] The extracted features form a feature library or feature centers, which are used for subsequent calculations.

[0074] In step S3, to ensure the timeliness and accuracy of the features, an automated feature pipeline is used for calculation, specifically including:

[0075] Feature stream computation: For features with low latency requirements, such as the time interval between the current passage and the previous passage, and real-time anomaly detection of the current path, streaming computing engines such as Flink are used to perform real-time computation when data is received, and the "dynamic features" part of the vehicle file is updated in real time for real-time risk warning.

[0076] Feature Batch Calculation: For complex historical statistical features and model-derived features, such as monthly traffic patterns and behavioral clustering, batch processing engines such as Spark are used to scan and calculate the full archives at fixed intervals, such as every day and night, to update the "periodic features" part of the vehicle archives for in-depth analysis and model training.

[0077] Feature quality monitoring and lineage tracking: Establish a feature quality dashboard to monitor the missing rate, outlier ratio, and stability of key features. Create a data lineage graph for all features, clearly recording the source, calculation logic, and upstream / downstream dependencies of each feature. When raw data is abnormal or business rules change, the impact can be quickly assessed and the recalculation of relevant features can be triggered.

[0078] The extracted features ultimately serve specific audit scenarios and require targeted analysis and synthesis. Step S3, which involves scenario-based feature analysis and risk profile generation, specifically includes:

[0079] Scenario-based feature vector assembly: For different audit models, relevant features are selected from the overall feature library and assembled into customized feature vectors. For example, the feature vector for the "large vehicle, small label" audit model focuses on: the vehicle model difference in this instance, vehicle model information in the vehicle archive, vehicle model recognition from videos / images, and vehicle model information issued by the OBU, etc. The feature vector for the "long-distance toll purchase" audit model focuses on: the relationship between the entrance and exit toll stations, the minimum toll amount identifier, the integrity of the gantry sequence, historical route preferences, and the matching degree between transaction time and travel time, etc.

[0080] Vehicle dynamic risk profile generation: The system will periodically, such as every 24 hours, generate a dynamic risk profile summary for each vehicle. This profile is a comprehensive and visual presentation of multi-dimensional features.

[0081] In step S4, the risk assessment process is as follows: Figure 3 As shown.

[0082] The system utilizes an expert rule base model based on a business knowledge graph (including business specifications and expert knowledge) to deeply deconstruct the clauses in the highway network toll collection operation management specifications and audit work management system, transforming them into formalized logical rule statements. To maintain the flexibility and composability of the rules, they are designed at the "atomic level," meaning each rule is an indivisible minimum rule. All atomic rules are stored in a centrally managed expert rule base. Each rule, in addition to its logical body, includes rich metadata, including: rule creation / update time, historical trigger count, accuracy rate after manual confirmation, associated toll evasion type, applicable road sections or vehicle type range, etc. The system automatically generates rule performance analysis reports periodically (e.g., monthly), highlighting rules that are consistently inefficient (e.g., high trigger, low confirmation) or conflict with the latest policies, prompting management experts to review, optimize, or discontinue them, ensuring the timeliness and simplicity of the rule base.

[0083] The pre-trained AI model is invoked to infer the structured features. This pre-trained AI model is a deep mining model based on multimodal feature learning. As the "intelligent radar" of the audit system, the AI ​​model proactively mines potential and complex anomaly patterns from massive amounts of historical data using machine learning, filling the blind spots that rule-based models cannot cover. The AI ​​model performs the following tasks:

[0084] (1) Extracting multimodal features and assembling them into a feature vector: The cornerstone of the AI ​​model's effectiveness lies in its features. Based on "one vehicle, one file," the system constructs a high-dimensional vector containing multimodal features for each passage behavior, specifically including:

[0085] Spatiotemporal characteristics: travel time (whether it is a holiday or a period with differentiated discounts), travel duration, and speed within a section (whether it is abnormal), etc.

[0086] Transaction behavior characteristics: OBU / CPC card switching frequency, transaction response time, and completeness of inbound / outbound information (Part 2).

[0087] Vehicle characteristics: historical frequency of passage, travel route, vehicle type, etc.

[0088] Association characteristics: Association with other vehicles in the same fleet that frequently enter and exit the same station.

[0089] Historical audit work orders that have been finalized by manual review are used as labeled samples and associated with the aforementioned feature vectors to form a high-quality supervised learning dataset, i.e., a labeled sample library. The feature vectors are then input into a pre-trained AI model to obtain inference results regarding suspected toll evasion.

[0090] Pre-trained AI models refer to multiple specialized AI models trained using an ensemble learning framework for different complex toll evasion scenarios. These specialized AI models include:

[0091] Traffic trajectory anomaly prediction model: Based on a vehicle-to-vehicle profile and real-time traffic trajectory, it analyzes the spatiotemporal characteristics of vehicles and predicts their reasonable path range. When the actual traffic path deviates significantly from the predicted path and cannot be explained by rules, a "suspected path anomaly" message is displayed.

[0092] Driving behavior clustering analysis model: Clustering the traffic behavior characteristics of all vehicles. Vehicles whose behavior patterns significantly deviate from the mainstream cluster (e.g., consistently traveling between adjacent gantries for very short periods) are identified as "abnormal behavior groups" to discover new types of gang-related and tentative toll evasion.

[0093] Image sequence analysis model: Analyze subtle changes in vehicles in continuous gantry-captured images (such as the degree of license plate damage and changes in front decorations), and combine them with transaction signals to help identify "card swapping and license plate exchange" behavior.

[0094] (2) The online learning and feedback loop of the AI ​​model includes: after all the prediction results of the AI ​​model are manually reviewed and a final conclusion is reached by the auditor, the conclusion will be automatically fed back to the model training platform as a new labeled sample. The system regularly (e.g., weekly) uses incremental data to fine-tune the model training, so as to achieve continuous self-evolution of the model. This feedback loop ensures that the AI ​​model can keep up with the changes in toll evasion methods and continuously optimize the rule base.

[0095] In S4, to address the ever-emerging new methods of fee evasion, this invention provides a powerful "strategy factory," allowing audit experts to flexibly combine underlying rules and AI capabilities to quickly create customized composite audit models without writing code. Specifically, it provides a composite audit strategy factory based on visual orchestration, including:

[0096] Visual Strategy Orchestrator: Provides auditors with a graphical interface that includes all "model component libraries" and logic gate components. All atomic rules and AI model outputs are displayed in draggable component form and connected by logic gate components to draw complex audit strategy flowcharts.

[0097] Strategy Simulation and One-Click Deployment: Once orchestrated, strategies can be run in a sandbox simulation within the system over selected historical time periods. The system will simulate the execution of the strategy and output a list of simulated hit vehicles, an estimated number of work orders, and a report on the match between the strategy and historically confirmed cases, allowing the orchestrator to evaluate the strategy's effectiveness and feasibility. Once approved and released, the strategy can be deployed to the production environment with a single click, serving as a new audit model for real-time or batch scanning tasks. The performance data of the new strategy (trigger count, accuracy) will also be included in unified monitoring.

[0098] Tiered Scheduling and Resource Optimization: The system has a built-in intelligent scheduler that manages all active audit models (including rules, AI, and composite strategies) in a tiered manner. High-frequency, high-accuracy models are used for real-time stream computing to provide immediate risk warnings for passing vehicles; low-frequency, complex models are used for offline batch computing to periodically perform in-depth scans of full or sampled vehicle files. The scheduler dynamically adjusts the calculation order and resource allocation of the models based on the load of computing resources to ensure the stability and efficiency of the entire audit system.

[0099] To ensure the objectivity, credibility, and efficient traceability of audit conclusions, this invention transforms the suspected alerts output by the audit model into audit work orders with complete and rigorous evidence chains. The intelligent collection and generation process of the audit evidence chain is as follows: Figure 4 As shown, it includes:

[0100] Precise Trigger Response: When the embedded audit model (such as a rule-based model or an AI prediction model) identifies a vehicle's passage characteristics as highly matching a preset toll evasion type (such as "large vehicle with small license plate," "entry without exit," "long-distance trip with short-distance payment," "fake green channel," "card swapping," etc.) during real-time or batch analysis, the system will automatically generate a structured audit trigger instruction. This instruction not only includes the vehicle's unique identifier (license plate number, license plate color, etc.), trigger time, and suspected toll evasion type, but also carries the initial suspicion level and associated feature index output by the model, providing precise guidance for subsequent targeted evidence collection.

[0101] The intelligent data collection and scheduling engine retrieves evidence: Upon receiving a trigger command, the intelligent data collection and scheduling engine is immediately activated. At its core is a built-in knowledge base mapping "toll evasion type - evidence template." For example, for "card swapping" toll evasion, the template pre-sets the required evidence as: photos of the same license plate being used at different entrance stations within an unreasonable time interval, vehicle feature comparison results, and toll medium number comparison records; for "fake green channel vehicles," the template requires retrieving multi-angle cargo videos of the lane and the vehicle's historical green channel passage records, etc. Based on the template, the engine initiates parallel data retrieval requests to the "one vehicle, one file" database (i.e., the file center) and related image and video storage systems.

[0102] Standardized encapsulation of multi-source heterogeneous evidence: The evidence captured by the system from the archives is original and heterogeneous, including: structured data: toll transaction records (including entrance, exit, and gantry information), vehicle registration information; unstructured data: high-definition captured images of key gantries and toll stations, lane video clips, etc.

[0103] The data acquisition engine performs standardized preprocessing on this evidence, including: assigning uniform timestamps and spatial location (road segment, gantry / station number) tags to all evidence; performing quality checks on image evidence (such as clarity and license plate integrity); and extracting and transcoding key segments of video evidence. Finally, each piece of evidence is packaged into a standardized evidence package containing metadata (source, time, type, hash value).

[0104] Spatiotemporal trajectory retrospection and evidence alignment: The system uses the vehicle's actual route during this passage (reconstructed from gantry transaction records, gantry identification records, etc.) or abnormal passage sections (such as missing gantries) as the timeline, automatically sorting and aligning all captured transaction records, images, and video evidence with millisecond-level time precision and spatial order. For example, for "running a long distance and buying a short distance" toll evasion, the system automatically generates a time-ordered visual evidence trajectory map, accurately marking the image and time when the vehicle was identified at each key gantry, and comparing it with the exit transaction record to intuitively display the vehicle's toll evasion trajectory map.

[0105] Synthetic Dynamic Evidence Chain and Intelligent Summary: After evidence alignment, the system calls the corresponding logical chain model based on the type of toll evasion, dynamically synthesizing discrete evidence packages into an electronic evidence chain with inherent proof logic. This evidence chain is not a simple list, but an interactive and traceable electronic case file. Simultaneously, the system's built-in Natural Language Processing (NLP) module analyzes all evidence and automatically generates a summary of the audit findings, such as: "The system detected license plate XXX at [time]. Its entry identification was a Type 6 truck, but the exit ETC transaction was for a Type 3 passenger vehicle, with a toll difference of XXX yuan. The evidence chain includes: a high-resolution image of the entry (displayed as a red heavy truck), entry transaction records (Type 6), exit ETC transaction records (Type 3), and the vehicle's historical toll registration information for all Type 6 vehicles, indicating suspected 'large vehicle, small license plate' toll evasion."

[0106] Automatic binding and integrity self-check with audit work orders: The synthesized complete evidence chain (including structured data, multimedia files, visual trajectory maps, and intelligent summaries) will be automatically bound to the newly created audit work order as an indivisible data package, forming a strong "one case, one chain" connection. Before binding, the system will perform an evidence chain integrity self-check to verify whether the actually collected evidence meets the minimum requirements of the corresponding "evidence template". If key evidence is missing (such as the inability to obtain a necessary gantry image), the system will clearly mark the missing item and its potential impact in the work order, prompting the auditor to initiate a manual supplementary investigation to ensure that the foundation of each audit conclusion is solid and reliable.

[0107] This invention also provides a highway vehicle passage audit system based on "one vehicle, one file," the overall architecture of which is as follows: Figure 5 As shown, it includes a data source layer, a data processing and core engine layer, and a closed-loop feedback module.

[0108] The data source layer includes toll stations, gantries, and external databases. Transaction records, captured images, and videos are collected through toll stations and gantries, while vehicle issuance information, license plate records, and vehicle information from the transportation department are collected through external databases.

[0109] The data processing and core engine layer includes a vehicle file center, a data fusion and preprocessing module, an automated feature extraction and analysis engine, an intelligent audit model library, and an automated evidence chain generation module. The vehicle file center integrates external databases to establish a unique identifier for each vehicle, creates a vehicle file for each vehicle, and collects basic information, travel records, feature information, and audit information for that vehicle. It then updates and verifies the vehicle file using the collected information. The data fusion and preprocessing module accesses multi-source heterogeneous data, performs standardized cleaning and normalization processing on the data, and constructs a "trip-data block" spatiotemporal correlation model. The multi-dimensional feature calculation engine constructs a multi-level structured feature body. The system is equipped with an automated feature extraction and update engine, which performs scenario-based feature analysis and profile construction. The intelligent audit model library is built with an expert rule base based on a business knowledge graph. The rules in the expert rule base are atomic expert rules. The intelligent audit model library is built with a deep mining AI model with multimodal feature learning. The intelligent audit model library is built with a composite audit strategy factory based on visualization orchestration. The evidence chain automatic generation module uses the actual travel path of the vehicle as the time axis, captures the required multi-source heterogeneous evidence, performs millisecond-level spatiotemporal alignment and sorting of all evidence, automatically synthesizes an electronic evidence chain containing a visualized trajectory map, and electronically binds the evidence chain to the audit work order.

[0110] The closed-loop feedback module takes the final conclusions of the auditors after manually reviewing the AI ​​model's prediction results as newly labeled samples and automatically feeds them back to the model training platform, using incremental data to fine-tune the model on a regular basis.

[0111] The system also includes an application layer, which comprises a real-time alert dashboard, audit work order management, and multi-dimensional analysis reports. The real-time alert dashboard monitors the missing rate, outlier ratio, and stability of key features, and establishes a data lineage graph for all features, clearly recording the source, calculation logic, and upstream and downstream dependencies of each feature. When raw data is abnormal or business rules change, the impact can be quickly assessed and the recalculation of relevant features can be triggered.

[0112] The audit work order management system enables full lifecycle management of work orders, covering automatic generation, intelligent assignment, handling verification, audit and recovery, and archiving and assessment. Based on anomaly characteristic results pushed by real-time alert dashboards, the system automatically generates audit work orders. Work order content includes abnormal vehicle information, missing characteristics or suspected toll evasion, data lineage tracing path, and judgment criteria. After work order generation, the system's built-in rule engine automatically assigns it to the corresponding audit personnel based on the anomaly level and the highway operating unit to which the work order belongs, and supports multi-level review mechanisms (self-review, mutual review, and superior random review) to ensure standardized handling. During the handling process, the system provides one-click access and download of the full evidence chain, including vehicle trajectory, gantry records, and captured images, ensuring a complete record of the handling process. Approved work orders are automatically pushed to the toll collection system, achieving a closed loop for audit and recovery. The system also establishes a work order timeliness early warning and quality assessment system, automatically supervising overdue work orders and incorporating indicators such as accuracy, rollback rate, and timely feedback rate into the assessment.

[0113] The multidimensional analysis reporting module provides statistical and visual displays of key indicators such as the number of work orders, processing time, amount recovered, and refund rate from multiple dimensions, including time (day / week / month / quarter / year), region (toll station, road segment center, operating unit, etc.), type of toll evasion (large vehicle with small label, long-distance toll purchase short-distance, OBU mismatch, etc.), and processing efficiency. It supports report formats such as line charts, bar charts, and pie charts, providing data-driven decision support for audit strategy optimization, human resource allocation, and performance evaluation.

[0114] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Various modifications or equivalent substitutions made by those skilled in the art to the present invention within its spirit and scope of protection also fall within the scope of protection of the present invention.

Claims

1. A highway vehicle passage auditing method based on "one vehicle, one file", characterized in that, include: S1, Construct vehicle files, including: integrating external databases; establishing a unique identifier for each vehicle and creating a vehicle file for each vehicle; collecting the vehicle's basic information, passage records, characteristic information, and audit information, and updating and verifying the vehicle file with the collected information; S2, corresponding to the file fusion and preprocessing of multi-source data for each vehicle, including: accessing multi-source heterogeneous data; standardizing, cleaning and normalizing the data; using the unique vehicle identifier combined with timestamps and spatial locations, aligning and associating the data of the entrance, gantry and exit in time and space, and constructing a "trip-data block" time and space association model to describe a single complete passage; S3, extract and update multi-level structured features, which include: basic fact features directly extracted from raw data, rule features generated by comparing basic fact features with business rules, and behavioral patterns and risk features mined from historical archives through statistical analysis or machine learning; perform scenario-based feature analysis on the extracted features, and generate a risk profile of the vehicle based on the features; S4, Perform risk assessment: Invoke an atomized expert rule model to match the structured features, and invoke a pre-trained AI model to infer the structured features; obtain a pre-defined composite audit strategy through a visual orchestration method, wherein the composite audit strategy defines the combination logic between the matching results of the expert rule model and the inference results of the AI ​​model; perform logical combination operations on the matching results and / or inference results according to the composite audit strategy to generate a suspicion assessment result; dynamically distribute the audit calculation task to the real-time computing layer, near-line computing layer, or offline computing layer for execution according to the complexity, timeliness requirements, and data volume of the composite audit strategy, and dynamically allocate computing resources according to the negative computing resource allocation. S5, synthesize the electronic evidence chain and link it to the audit work order.

2. The method according to claim 1, characterized in that, In S2, when accessing multi-source heterogeneous data, the data sources of the entire domain are defined and classified for access. Multi-source heterogeneous data includes real-time streaming data sources, batch data sources and external data sources. Real-time streaming data sources include real-time data from each toll station accessed through message queues. The batch data sources include data uploaded by each toll station and gantry at regular intervals; the external data sources include cross-provincial traffic data, vehicle registration information and vehicle release information from external databases.

3. The method according to claim 2, characterized in that, In S2, a data asset catalog and metadata management system is established for the accessed multi-source heterogeneous data.

4. The method according to claim 1, characterized in that, In S2, data standardization, cleaning, and normalization include: structured data cleaning of transaction data; and unstructured data preprocessing of images / videos.

5. The method according to claim 1, characterized in that, When the same license plate is identified in different locations within a very short period of time, the spatiotemporal overlap verification technology uses the vehicle's unique identifier, timestamp, and spatial location as the core to perform spatiotemporal consistency cross-verification on multi-source data and flow data. Based on the time dimension and road network spatial dimension, the time and spatial continuity and rationality of the vehicle trajectory are verified.

6. The method according to claim 5, characterized in that, When traffic is heavy, lane coil triggering signals are introduced and video target tracking segments are used for auxiliary correction to correct mismatches between adjacent vehicle images.

7. The method according to claim 1, characterized in that, S2 also includes building a standardized thematic data layer and quality monitoring, including: generating a "vehicle-trip" star-shaped data layer from the associated and fused data, including a trip fact table, a trip detail table, a vehicle file table, and an image feature index table; and establishing full-link data quality monitoring and closed-loop feedback.

8. The method according to claim 1, characterized in that, In S3, the basic factual features include spatiotemporal facts, transaction facts, and vehicle facts; and / or the rule features include consistency features between entry and exit, compliance features related to toll policies, and labeling features based on simple rules; and / or the behavioral patterns and risk features include statistical features derived from long-term historical archives, aggregated features, and model-derived features output by AI models.

9. The method according to claim 1, characterized in that, In S3, the automated feature extraction and update engine includes a stream processing framework and a batch processing framework. The stream processing framework is used to calculate the dynamic features of vehicles in transit in real time and update the dynamic feature part in the vehicle file in real time. The batch processing framework is used to scan and calculate the full vehicle file at fixed intervals and update the periodic feature part in the vehicle file.

10. The method according to claim 1, characterized in that, In S4, each atomic rule is accompanied by metadata, which includes the rule creation and update time, historical trigger count, hit accuracy, associated toll evasion type, and applicable vehicle model range.

11. The method according to claim 10, characterized in that, S4 also includes: the expert rule model for invoking atomization further includes a rule dynamic management step: real-time monitoring of the hit rate, accuracy and resource consumption of each expert rule, and dynamic adjustment of the priority, activation status or trigger threshold of the rule according to the monitoring results; obtaining newly added evasion cases that have been manually reviewed and confirmed, and iteratively optimizing the parameters of the corresponding expert rule according to the newly added evasion cases.

12. The method according to claim 1, characterized in that, In S4, calling the pre-trained AI model to infer the structured features includes: extracting multimodal features and assembling them into feature vectors; inputting the feature vectors into the pre-trained AI model to obtain the inference results of suspected evasion; the AI ​​model can learn and provide feedback online.

13. The method according to claim 12, characterized in that, Multimodal features include spatiotemporal features, transaction behavior features, vehicle features, and related features. Historical audit work orders that have been finally confirmed by human intervention are used as labeled samples and associated with the feature vectors to form a supervised learning dataset.

14. The method according to claim 12, characterized in that, In S4, the pre-trained AI models include: a traffic trajectory anomaly prediction model, which predicts a reasonable path range based on vehicle files and real-time traffic trajectories, and marks anomalies when the actual path deviates significantly; a driving behavior clustering analysis model, which clusters all vehicle traffic behavior features and identifies abnormal groups whose behavior patterns deviate from the mainstream cluster; and an image sequence analysis model, which analyzes subtle changes in vehicles in continuous gantry-captured images to assist in identifying card-switching and license plate-changing behaviors.

15. The method according to claim 12, characterized in that, Online learning and feedback for AI models include: obtaining the final conclusions of auditors after manually reviewing the AI ​​model's prediction results as new labeled samples, automatically feeding them back to the AI ​​model training platform, and periodically using incremental data to fine-tune the AI ​​model.

16. The method according to claim 1, characterized in that, In S4, obtaining the pre-set composite audit strategy through visual orchestration includes: selecting atomized expert rules and / or AI model output results as strategy components from the model component library through the visual orchestration interface, and connecting the selected strategy components through logic gate components to generate an audit strategy flowchart; performing sandbox simulation on the generated audit strategy flowchart to output a list of simulated hit vehicles and a matching report with historical confirmed cases; and deploying the audit strategy flowchart that has passed the simulation review to the production environment as a new composite audit strategy for subsequent audit scanning tasks.

17. The method according to claim 16, characterized in that, In step S4, the dynamic distribution of audit calculation tasks includes: allocating high-frequency, high-accuracy models to the real-time streaming computing layer to provide immediate risk warnings for vehicles in transit; and allocating low-frequency, complex models to the offline batch computing layer to periodically perform in-depth scans of full or sampled vehicle files.

18. The method according to claim 1, characterized in that, In S5, the synthesis of the electronic evidence chain and its association with the audit work order includes: automatically generating a structured audit trigger command when suspected toll evasion is determined; starting the intelligent collection and scheduling engine, matching the corresponding evidence template according to the suspected toll evasion type, and retrieving multi-source heterogeneous evidence from the vehicle file according to the evidence template; standardizing and preprocessing the retrieved multi-source heterogeneous evidence, adding timestamps, spatial location tags, and metadata to each piece of evidence to generate a standardized evidence package; sorting and aligning the standardized evidence package according to millisecond-level time precision and spatial order using the vehicle's actual travel path or abnormal travel section as the time axis to generate a visual evidence trajectory map; dynamically synthesizing the sorted and aligned evidence package into an electronic evidence chain, and automatically generating an audit reason summary based on the content of the electronic evidence chain; and binding the electronic evidence chain and the audit reason summary to the audit work order.

19. The method according to claim 18, characterized in that, Before automatically binding the evidence chain and summary to the audit work order, perform an evidence chain self-check and mark the missing items and potential impacts for cases where key evidence is missing.

20. A highway vehicle passage audit system based on "one vehicle, one file," characterized in that, include: The data source layer includes toll stations, gantries, and external databases. It collects transaction records, captured images, and videos through toll stations and gantries, and collects vehicle issuance information, license plate records, and vehicle information from the transportation department through external databases. The data processing and core engine layer includes a vehicle file center, a data fusion and preprocessing module, an automated feature extraction and update engine, an intelligent audit model library, and an automated evidence chain generation module. The vehicle file center integrates external databases to establish a unique identifier for each vehicle, creates a vehicle file for each vehicle, and collects basic information, travel records, feature information, and audit information for that vehicle. It then updates and verifies the vehicle file using the collected information. The data fusion and preprocessing module accesses multi-source heterogeneous data, performs standardized cleaning and normalization processing on the data, and constructs a "trip-data block" spatiotemporal correlation model. The automated feature extraction and update engine... The new engine extracts and updates multi-level structured features, performs scenario-based feature analysis, and generates risk profiles. The intelligent audit model library includes an expert rule library based on business knowledge graphs, where the rules are atomic expert rules. The intelligent audit model library also includes a deep mining AI model based on multimodal feature learning and a composite audit strategy factory based on visual orchestration. The evidence chain automatic generation module uses the actual travel path of the vehicle as the time axis, captures the required multi-source heterogeneous evidence, performs millisecond-level spatiotemporal alignment and sorting of all evidence, automatically synthesizes an electronic evidence chain containing a visual trajectory map, and electronically binds the evidence chain to the audit work order.

21. The system according to claim 20, characterized in that, The data processing and core engine layer also includes a closed-loop feedback module, which uses the final conclusions of auditors after manually reviewing the AI ​​model prediction results as newly labeled samples and automatically feeds them back to the model training platform to periodically fine-tune the AI ​​model using incremental data.

22. The system according to claim 20, characterized in that, The system also includes an application layer, which includes real-time early warning dashboards, audit work order management, and multi-dimensional analysis reports.

Images