Intelligent traffic information Internet acquisition and multi-source data fusion system

Abstract

The invention relates to the technical field of intelligent traffic systems and multi-source data processing, and discloses an intelligent traffic information internet acquisition and multi-source data fusion system, which comprises a semantic-driven heterogeneous facility protocol reconstruction module for analyzing an internet text and generating a dynamic acquisition protocol configuration packet adaptive to underlying perception; the anchor point data-based time delay compensation module is used for associating physical anchor point time, calculating propagation time delay and delimiting a generated backtracking time drift window; the multi-source microscopic fingerprint verification module is used for extracting physical sensor data based on a protocol packet and a time window, performing isomorphic matching on the physical sensor data and the virtual kinematics fingerprint feature template, and outputting an event authenticity verification result; and the physical feedback driven semantic evolution module is used for constructing a physical truth value feedback signal to update the model weight and correct the fuzzy geographic mapping. According to the method, the problems of internet data release lag and physical facility space-time alignment are solved, and false information filtering and system self-evolution are realized through physical fingerprint verification.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation systems and multi-source data processing technology, specifically to an intelligent transportation information Internet acquisition and multi-source data fusion system. Background Technology

[0002] With the rapid development of intelligent transportation systems, the means of traffic information perception have evolved from single physical sensor detection to a combination of physical perception and internet crowdsourced data. Existing traffic management platforms typically integrate physical sensing facilities, including video surveillance, geomagnetic coils, and microwave radar, while also capturing traffic information from internet platforms such as navigation software and social media, attempting to improve the ability to perceive traffic events through the convergence of multi-source data.

[0003] However, in practical applications, a significant semantic gap and protocol barriers exist between unstructured text data from the internet and underlying physical sensor data. Traditional physical sensing facilities typically operate in preset fixed sampling or polling modes, and their operational status is difficult to be directly driven by upper-layer internet semantic information. When internet data indicates the presence of a specific event (such as spilled material or icing) on ​​a road segment, existing systems lack a cross-layer protocol conversion mechanism to instantly translate this ambiguous semantic description into specific acquisition instructions for the underlying devices. This prevents physical facilities from adjusting sampling frequency or areas of interest in a targeted manner, leaving upper-layer applications and underlying hardware in a disconnect and hindering the realization of semantically guided proactive sensing.

[0004] Furthermore, there is often an uncertain time lag between the release time of internet traffic information and the actual occurrence time of the event. When users post traffic information on social media platforms or navigation software, it is usually some time after the event has occurred. This artificially created lag means that the timestamp of internet data cannot be directly used as the basis for retrieving physical sensor data. Existing technologies, when performing multi-source data fusion, often ignore this propagation lag and directly use the internet release time as the benchmark to retrieve real-time data from physical sensors. This misalignment of spatiotemporal benchmarks can cause the system to retrieve data within an incorrect time window, resulting in verification failures or missed reports due to the inability to match the corresponding physical features.

[0005] Meanwhile, the authenticity and accuracy of internet data are limited by the subjective judgment of the publisher, resulting in a large amount of false information, vague descriptions, and location mapping errors. Most existing fusion systems use simple weighted averaging or manual verification methods, lacking an automatic verification mechanism based on objective physical facts. The system struggles to use precise dynamic characteristics or image data collected by roadside equipment to reverse-verify the authenticity of internet text, and even less can it use verified physical truth values ​​to correct the system's internal semantic parsing model or geographic mapping relationships. This open-loop operating mode prevents the system from learning from historical false alarms, hindering the adaptive evolution of model accuracy and limiting the robustness of intelligent transportation systems in complex environments. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent transportation information internet collection and multi-source data fusion system. It aims to solve the problems in existing technologies, such as the difficulty of using unstructured semantic information from the internet to drive underlying heterogeneous facilities to actively perceive across layers, the lag in the release of internet data leading to its inability to accurately align with physical sensor data in the spatiotemporal dimension, and the lack of a closed-loop feedback mechanism based on physical objective facts, which prevents the system from effectively filtering false information and achieving adaptive model evolution.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution: an intelligent transportation information Internet collection and multi-source data fusion system, comprising: a semantically driven heterogeneous facility protocol reconstruction module, an anchor point data-based time delay compensation module, a multi-source micro-fingerprint verification module, and a physical feedback-driven semantic evolution module.

[0008] The semantically driven heterogeneous facility protocol reconstruction module is used to parse internet text data, extract output geographic location constraint information, final output event category labels, and output high-dimensional semantic feature vectors, and extract the publication timestamps of internet information. This module transforms unstructured text into structured data through semantic analysis, and generates a dynamic acquisition protocol configuration package adapted to the front-end perception and control facilities to establish communication connections between upper-layer applications and lower-layer heterogeneous devices.

[0009] The time delay compensation module based on anchor point data, according to the output geographical location constraint information, associates the physical anchor point recording time from the front-end sensing and control facilities, and combines the generated backtracking time drift window determined by propagation time delay calculation to achieve alignment of Internet data and physical sensor data in the time dimension.

[0010] The multi-source micro-fingerprint verification module, based on the generated dynamic acquisition protocol configuration packet and the generated backtracking time drift window, extracts the reconstructed physical sensor data vector from the front-end sensing and control facilities, and performs isomorphic matching between this physical sensor data vector and a virtual kinematic fingerprint feature template constructed based on the final output event category label. This module compares the physically measured data with the theoretical feature template, outputting a similarity score and the event authenticity verification result.

[0011] The physical feedback-driven semantic evolution module constructs a physical truth feedback signal in response to the event authenticity verification result, and uses the physical truth feedback signal to update the weights of the semantic parsing model and correct the fuzzy semantic geographic mapping, thereby realizing the adaptive iteration of the system's semantic understanding ability and geographic positioning accuracy.

[0012] Preferably, when generating the generated dynamic acquisition protocol configuration package, the semantically driven heterogeneous facility protocol reconstruction module matches the private protocol format of the road segment front-end sensing control facility corresponding to the output geographical location constraint information in the heterogeneous sensing device protocol feature library, uses a pre-trained protocol mapping network to convert the private protocol format into a unified middleware interface standard of the system, and then encapsulates the parameter set containing device access parameters, data decoding rules and sampling frequency requirements into the generated dynamic acquisition protocol configuration package.

[0013] Preferably, when calculating the generated backtracking time drift window, the anchor point data-based time delay compensation module calculates a propagation time delay parameter based on the difference between the physical anchor point recording time and the internet information publication timestamp; it then uses this parameter to correct the internet information publication timestamp, obtaining a corrected estimate of the actual event occurrence time; and finally, using this corrected estimate of the actual event occurrence time as the center and a preset fault tolerance buffer time parameter as the radius, it delineates the generated backtracking time drift window. The physical anchor point recording time originates from the highway toll gantry equipment or roadside detection equipment.

[0014] Preferably, when constructing the virtual kinematic fingerprint feature template, the multi-source micro-fingerprint verification module retrieves a physical parameter constraint model based on the final output event category label, generating a set containing multi-dimensional physical thresholds and logical operators. In one embodiment, for road icing events, a dynamic feature constraint function is generated, setting a joint threshold for the lateral acceleration and yaw rate from the vehicle terminal device; for road debris events, a visual feature constraint function is generated, setting a minimum motion threshold between the pixel motion vector within the spatial domain region of interest parameters of the traffic video surveillance device's image sensor and the background traffic flow.

[0015] Preferably, when outputting a similarity score, the multi-source micro-fingerprint verification module calculates the absolute difference between the measured physical feature value and the reference physical threshold in the virtual kinematic fingerprint feature template for each feature component in the reconstructed physical sensor data vector; it then uses a weighted normalization bias algorithm to solve for the normalization bias of the absolute difference relative to the reference physical threshold; and finally multiplies the value obtained by subtracting the normalization bias from 1 by the confidence weight of the feature component and the value of the logical indicator function, summing the calculation results for all feature dimensions to obtain the similarity score.

[0016] In one embodiment, the multi-source micro-fingerprint verification module is further configured to send a signal control command to the traffic signal control facility when the event authenticity verification result is true. The correction logic for the green light duration is as follows: multiply the base green light duration of the target phase in the current time period by a correction coefficient to obtain the corrected green light duration of the target phase. The correction coefficient is 1 minus the product of the traffic reduction coefficient and the event severity weight coefficient; the event severity weight coefficient is obtained by linearly mapping the similarity score.

[0017] Preferably, when constructing the physical truth feedback signal, the physical feedback-driven semantic evolution module calculates the feedback gain weight according to the following logic: calculating the absolute difference between the similarity score and the system's preset confidence threshold parameter, performing a non-linear adjustment exponential operation on the absolute difference, and multiplying it by a preset gain scaling factor. The physical truth feedback signal includes the output high-dimensional semantic feature vector, the final output event category label, the binarized truth label, and the calculated feedback gain weight.

[0018] Preferably, when updating the weights of the semantic parsing model, the physical feedback-driven semantic evolution module extracts a batch of samples containing multiple physical truth feedback signals from the high-confidence sample library; constructs a weighted loss function for physical confidence, and performs a weighted summation of the cross-entropy loss between the predicted probability by the deep neural network model and the binary truth label, wherein the weighting coefficient is the feedback gain weight; and uses the backpropagation algorithm to update the parameter set of the deep neural network model according to the loss function.

[0019] Preferably, the physical feedback-driven semantic evolution module is also used to perform dynamic adjustment of reputation weights for specific information sources. Specifically, it reads the reputation weights before the update corresponding to the virtual identity of the internet information publisher, adds a correction term to them, and uses a numerical truncation function to ensure that the result is between 0 and 1, thus obtaining the updated reputation weights. The correction term is calculated by multiplying the reputation learning rate parameter, the feedback gain weight, and the polarity mapping term determined by the binarized truth label.

[0020] Preferably, when the physical feedback-driven semantic evolution module performs the correction of the fuzzy semantic geographic mapping, if the binarized truth label is true, it obtains the verified accurate physical coordinates; calculates the spatial vector deviation between the accurate physical coordinates and the fuzzy place name entity geographic mapping center coordinate vector currently stored in the road network topology geographic information database; multiplies the spatial vector deviation by the feedback gain weight and the preset geospatial mapping learning rate, and adds it as a correction term to the mapping center coordinate vector before correction to obtain the corrected fuzzy place name entity geographic mapping center coordinate vector.

[0021] This invention provides an intelligent transportation information internet-based data collection and multi-source data fusion system. It has the following beneficial effects:

[0022] 1. This invention, through a semantically driven heterogeneous facility protocol reconstruction module, can parse unstructured text at the application layer and generate dynamic acquisition protocol configuration packages adapted to the underlying hardware. It transforms the semantic requirements of the upper layer into specific execution instructions for the underlying devices, solving the technical problem of protocol incompatibility and difficulty in linkage between application layer semantic information and heterogeneous devices at the physical layer. This enables the system to proactively retrieve physical data at specific times and spaces based on Internet intelligence.

[0023] 2. This invention utilizes a time delay compensation module based on anchor point data. By introducing the physical anchor point recording time and the Internet publishing timestamp for correlation analysis, it calculates the propagation time delay parameters and defines the backtracking time drift window. This effectively corrects various time deviations caused by the lag in user publishing behavior, ensuring that the system can retrieve physical sensor data within the correct time period, and providing an accurate spatiotemporal reference for the effective fusion of multi-source data.

[0024] 3. This invention combines a multi-source micro-fingerprint verification module with a physical feedback-driven semantic evolution module. It utilizes the objective certainty of physical sensor data to verify the authenticity of subjective semantic information and constructs a physical truth feedback signal using the verification results. This not only enables real-time identification and filtering of false traffic information but also allows for reverse updating of the weights and source reputation of the semantic parsing model and automatic correction of the geographic mapping coordinates of fuzzy place names. This achieves continuous self-optimization of the system's accuracy and robustness during long-term operation. Attached Figure Description

[0025] Figure 1 This is a system framework diagram of the present invention;

[0026] Figure 2 This is a schematic diagram of the method flow of the present invention.

[0027] Among them, 10 is the central processing platform; 20 is the communication network; 30 is the front-end perception and control facilities; 31 is the traffic video surveillance equipment; 32 is the roadside detection equipment; 33 is the highway toll gantry equipment; 34 is the vehicle-mounted terminal equipment; 40 is the traffic signal control facilities; 100 is the semantic-driven heterogeneous facility protocol reconstruction module; 200 is the time delay compensation module based on anchor point data; 300 is the multi-source micro fingerprint verification module; and 400 is the physical feedback-driven semantic evolution module. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] See attached document Figure 1 This invention provides an intelligent traffic information internet acquisition and multi-source data fusion system. Its physical architecture mainly includes a central processing platform 10, a communication network 20, front-end sensing and control facilities 30, and traffic signal control facilities 40. The central processing platform 10, as the core computing unit of the system, establishes bidirectional data connections with the front-end sensing and control facilities 30 and traffic signal control facilities 40 distributed on the road surface through the communication network 20. The communication network 20 adopts one or more combinations of fiber optic communication networks, 5G mobile communication networks, or C-V2X vehicle-to-everything (C-V2X) cooperative communication networks to carry out control command issuance and multi-source sensor data feedback.

[0030] The front-end perception and control facility 30 specifically includes traffic video surveillance equipment 31, roadside detection equipment 32, highway toll gantry equipment 33, and vehicle-mounted terminal equipment 34. Traffic video surveillance equipment 31 is deployed at key road sections and intersections, equipped with variable focal length lenses and pan-tilt controllers, for collecting images and video data streams of the road environment. Roadside detection equipment 32 includes microwave radar, millimeter-wave radar, or geomagnetic detectors, for collecting cross-sectional traffic flow and speed data. Highway toll gantry equipment 33 is installed at highway or urban expressway nodes, for collecting vehicle ETC passage records and gantry capture data. Vehicle-mounted terminal equipment 34 is installed on autonomous vehicles or intelligent logistics transport vehicles, connected to vehicle-mounted lidar, inertial measurement unit (IMU), and vehicle CAN bus interface, for acquiring vehicle micro-dynamic parameters and surrounding point cloud data.

[0031] The central processing platform 10 includes at least one processor and a memory. The memory stores computer program instructions, and the processor executes the instructions to run a system architecture comprising multiple logical sub-modules. At the logical level, the system includes: a semantically driven heterogeneous facility protocol reconstructing module 100, an anchor data-based time-delay compensation module 200, a multi-source micro-fingerprint verification module 300, and a physical feedback-driven semantic evolution module 400.

[0032] The semantically driven heterogeneous facility protocol reconstructing module 100 connects to an internet data source interface for its data input and to a control interface of the front-end sensing and control facility 30 via a communication network 20 for its data output. The semantically driven heterogeneous facility protocol reconstructing module 100 is configured to parse internet text data and generate dynamic acquisition protocol configuration instructions for the traffic video surveillance equipment 31, roadside detection equipment 32, and vehicle-mounted terminal equipment 34 based on the parsing results.

[0033] The data input terminal of the anchor-based time delay compensation module 200 is connected to the data interface of the highway toll gantry equipment 33 and the historical database. The anchor-based time delay compensation module 200 is configured to use passage records with deterministic timestamps as anchors to calculate the propagation delay of Internet data and determine the time window for data retrieval accordingly.

[0034] The multi-source micro-fingerprint verification module 300 is connected to the time delay compensation module 200 based on anchor point data and the front-end sensing and control facility 30. The multi-source micro-fingerprint verification module 300 is configured to perform feature extraction and matching operations on physical sensor data collected according to a dynamic acquisition protocol within a defined time window, and output the verification results to the traffic signal control facility 40 to execute the corresponding signal timing strategy.

[0035] The physical feedback-driven semantic evolution module 400 connects the output of the multi-source micro-fingerprint verification module 300 to the parameter update interface of the semantically driven heterogeneous facility protocol reconstruction module 100. The physical feedback-driven semantic evolution module 400 is configured to use the physical verification results as feedback signals to update the semantic parsing model weights and geospatial mapping parameters within the semantically driven heterogeneous facility protocol reconstruction module 100.

[0036] See attached document Figure 2 This invention provides a method for intelligent transportation information internet collection and multi-source data fusion. This method is implemented through logical instructions running on a central processing platform 10 and includes the following steps:

[0037] S100: Collect Internet traffic text data, parse the semantic feature vector and event category label of the Internet traffic text data, generate dynamic acquisition protocol configuration instructions for the front-end perception and control facility 30 according to the event category label, and send the dynamic acquisition protocol configuration instructions to the front-end perception and control facility 30 in the corresponding geographical area to reconstruct the data acquisition parameters.

[0038] S200, retrieve the vehicle passage records of the highway toll gantry equipment 33 as time anchor data, calculate the propagation time delay parameter of the Internet traffic text data based on the time anchor data, and determine the backtracking time drift window for retrieving historical physical sensor data according to the propagation time delay parameter;

[0039] S300, within the retrospective time drift window, extract the physical sensor data collected by the front-end perception and control facility 30 after reconstruction by the dynamic acquisition protocol configuration instruction, convert the physical sensor data into micro kinematic features, and determine whether the micro kinematic features match the preset virtual event fingerprint. If the match is successful, generate the determined event location coordinates and send the traffic signal control instruction to the traffic signal control facility 40.

[0040] S400, the matching verification result generated in step S300 is used as the physical truth feedback signal. The confidence weight model and geospatial mapping parameters used to parse semantic features in step S100 are corrected using the physical truth feedback signal to complete the iterative update of the semantic model.

[0041] The above steps will be described in detail below with reference to specific embodiments and accompanying drawings.

[0042] See attached document Figure 2 In step S100 above, the semantically driven heterogeneous facility protocol reconstruction module 100's processing of Internet text data specifically includes the following sub-steps:

[0043] S111, the semantically driven heterogeneous facility protocol reconstruction module 100 obtains raw Internet traffic text data through the application programming interface. For the obtained raw text, preprocessing operations are performed, using regular expressions to remove HTML tags and special symbols, and loading a pre-set traffic domain stop word list to filter out meaningless function words, resulting in a preprocessed text sequence to be analyzed.

[0044] S112, the preprocessed text sequence to be analyzed is input into a pre-trained deep neural network model for feature encoding. The semantically driven heterogeneous facility protocol reconstruction module 100 runs a language model based on the Transformer architecture, which maps discrete text symbols into continuous vectors through a multi-layer self-attention mechanism.

[0045] The semantic feature extraction process is represented by a semantic encoding formula, as follows:

[0046] ;

[0047] in, This represents the high-dimensional semantic feature vector of the output; This represents the encoding function of a deep neural network; This represents the preprocessed text sequence to be analyzed. This represents the set of network parameters for a pre-trained model.

[0048] S113, based on the output high-dimensional semantic feature vector, the semantically driven heterogeneous facility protocol reconstruction module 100 uses a classifier to identify specific traffic event categories. This classifier calculates the probability distribution of the current text belonging to each category in a predefined event category set. The predefined event category set includes road debris, road icing, traffic signal malfunction, abnormal congestion, and illegal lane occupation.

[0049] The semantically driven heterogeneous facility protocol reconstruction module 100 selects the traffic event category with the highest probability value as the final output event category label. .

[0050] S114, while determining the event category, the semantically driven heterogeneous facility protocol reconstruction module 100 extracts entity words with geographical attributes from the preprocessed text sequence to be analyzed using a named entity recognition algorithm. The module matches the extracted entity words with a pre-stored road network topology geographic information database, which contains the index relationship between road segment names and GIS coordinates. After matching, geographic location constraint information containing latitude and longitude ranges and road segment identifiers is generated. The extraction result of this geographic location constraint information is represented as a set of locations, as shown below:

[0051] ;

[0052] in, This indicates the output geographic location constraint information; Indicates the minimum latitude and longitude of the target area; Indicates the maximum latitude and longitude of the target area; A unique identifier for the associated road network segment.

[0053] After the semantic-driven heterogeneous facility protocol reconstruction module 100 completes semantic feature extraction and event category determination, the system enters the stage of constructing and mapping the dynamic acquisition protocol configuration package.

[0054] S121, the semantically driven heterogeneous facility protocol reconstruction module 100 accesses a pre-set event and perception mapping library based on the event category labels obtained in step S113. The event and perception mapping library stores the corresponding logical relationships between traffic event semantic categories and physical sensor observation requirements. The module locks the corresponding sensor type set and the acquisition mode identifier of each sensor in the mapping library based on the final output event category labels.

[0055] S122, based on the retrieved observation requirements, the semantically driven heterogeneous facility protocol reconstruction module 100 constructs a dynamic acquisition protocol configuration package for the front-end sensing and control facilities 30 within the target geographical area. This configuration package is a set of structured data instructions. The construction process of the dynamic acquisition protocol configuration package follows the following protocol generation formula:

[0056] ;

[0057] in, This represents the generated dynamic acquisition protocol configuration package; Indicates the protocol generation function; Indicates the event category label in the final output; This indicates the output geographic location constraint information; A hardware capability description file representing the available devices within the target area.

[0058] Furthermore, the generated dynamic acquisition protocol configuration package contains a specific set of control parameters, defined as follows:

[0059] ;

[0060] in, This represents the time-domain sampling frequency parameter of the sensor; This represents the spatial domain region of interest parameters of the image sensor; This represents the sensitivity threshold parameter for signal detection.

[0061] S123, the semantically driven heterogeneous facility protocol reconstruction module 100 assigns values ​​to the above parameters according to the specific event category logic:

[0062] First, for event categories involving abnormal vehicle microdynamics (such as road icing and vehicle skidding), the module sets the time-domain sampling frequency parameters of the sensors of the on-board terminal device 34. For high-frequency values ​​(e.g.) The data acquisition targets are specified to include triaxial acceleration and yaw rate, to ensure that the sensor can capture transient physical characteristics.

[0063] Second, for event categories involving small target recognition (such as road debris), the module defines the spatial domain region of interest parameters for the image sensor of the traffic video surveillance equipment 31. This parameter consists of a set of pixel coordinate vectors, instructing the camera to perform high-bitrate encoding only on specific lane areas in the image or to invoke the gimbal for optical zoom.

[0064] Third, for event categories involving low-speed congestion (such as congestion queuing), the module adjusts the sensitivity threshold parameters of the signal detection of the roadside detection device 32. The module lowers this parameter value, instructing the device to enter a high-sensitivity mode to detect static targets with near-zero velocity.

[0065] When the semantically driven heterogeneous facility protocol reconfiguration module 100 generates the dynamic acquisition protocol configuration package After that, the system enters the stage of issuing protocol commands and executing hardware reconfiguration.

[0066] S131, the semantically driven heterogeneous facility protocol reconstruction module 100 parses the geographic location constraint information generated in the preceding steps. The system retrieves all active front-end sensing and control facilities 30 currently located within the geographical area from the device management database to determine the target device set. The semantically driven heterogeneous facility protocol reconstructing module 100 establishes a data transmission channel using the communication network 20, and transmits the generated dynamic acquisition protocol configuration package... It is encapsulated in downlink control signaling and sent to the target device.

[0067] S132, for the traffic video surveillance device 31 in the target device set, the built-in video encoding unit of the device, after receiving the generated dynamic acquisition protocol configuration packet, parses the spatial domain region of interest parameters of the image sensor contained therein. If this parameter is valid, the video coding unit executes region-based differential coding logic, which is expressed through the quantization parameter update formula:

[0068] ;

[0069] in, Represents pixel coordinates The quantization parameter value at the location; This indicates a lower quantization parameter value, corresponding to higher image clarity; A higher quantization parameter value corresponds to a higher compression ratio; This represents the region of interest parameter in the spatial domain of the image sensor.

[0070] Through this logic, the device achieves enhanced clarity of focus for specific semantic targets.

[0071] S133, for the vehicle-mounted terminal device 34 in the target device set, its controller parses the time-domain sampling frequency parameters of the sensors in the configuration package. Subsequently, the data reading frequencies of the inertial measurement unit and the vehicle's electronic stability system are dynamically modified. The low-pass filter in normal operation is bypassed, a temporary circular buffer is established, and raw sensor data is continuously written at a frequency defined by the sensor's time-domain sampling frequency parameters. Data is uploaded via the V2X network only when the buffer data meets preset trigger conditions.

[0072] See attached document Figure 2 After the system completes the issuance of the protocol reconfiguration instruction for the front-end hardware, the logic processing flow switches in parallel to the time delay compensation module 200 based on anchor point data, and begins to execute the initial stage in step S200, namely the association and extraction of time anchor point data. This step is implemented through the following sub-steps S211 to S213.

[0073] S211, the time delay compensation module 200 based on anchor point data determines the time anchor point data source for calibration. The module reads the metadata fields of the Internet traffic text data and identifies the type attribute of the information publishing platform. If the platform type attribute is a professional logistics or commercial vehicle networking platform, the module selects the positioning trajectory and CAN bus status data of intelligent logistics vehicles as the anchor point source; if the platform type attribute is social media or map navigation, the module selects the ETC passage record or license plate recognition data generated by the highway toll gantry device 33 as the anchor point source.

[0074] The selection of the aforementioned equipment data as anchor points is based on its facility-level timestamp accuracy and location certainty. Data from the highway toll gantry equipment 33 is generated by the roadside unit, synchronized via a time synchronization server, and its geographical coordinates are fixed physical truth values; intelligent logistics data is directly collected through vehicle-mounted terminals. This type of deterministic data is used as a reference system to eliminate random delays in internet data transmission and processing.

[0075] S212, the time-delay compensation module 200 based on anchor data establishes a mapping association between the identity identifier of the internet information publisher and the entity identifier of the physical traffic participant. For data sources with real-name authentication mechanisms, the time-delay compensation module 200 based on anchor data retrieves the bound vehicle physical identifier in the road network operation database through a unique identity index. For anonymous or public data sources, the time-delay compensation module 200 based on anchor data performs association matching based on spatiotemporal probability. This association process is represented by the following entity mapping function:

[0076] ;

[0077] in, This represents the physical traffic participant entity identifier of the output; This represents an identity association mapping function; A virtual identity identifier representing an internet information publisher; Indicates the timestamp of information published on the Internet; This indicates the output geographic location constraint information.

[0078] Executing the identity association mapping function At that time, the module calculates the spatiotemporal overlap between the virtual user's historical trajectory points and the physical nodes of the road network. Specifically, the module retrieves the set of vehicles that frequently appear in the upstream road segments of the area covered by the output geographical location constraint information within a preset time threshold before the publication timestamp of the Internet information, and selects the entity with the highest trajectory similarity score as the associated object.

[0079] S213, in determining the physical identifiers of traffic participants Subsequently, the time delay compensation module 200 based on the anchor point data retrieves the most recent upstream exact passage record from the historical record database of the highway toll gantry equipment 33 or the logistics fleet's driving log, and defines the timestamp of this record as the physical anchor point record time. The extraction logic of this physical anchor point record time follows the following anchor point selection formula:

[0080] ;

[0081] in, Indicates the time recorded at the physical anchor point; Represents physical entities In the historical database The time of each passage record; Indicates the timestamp of information published on the Internet; Indicates the first Record the coordinates of the corresponding physical facilities each time; This represents the set of valid retrieval domain nodes upstream of the output geographic location constraint information, defined based on the road network topology. This indicates the operation of finding the maximum value.

[0082] Through the above steps, the module associates the event descriptions in the application layer time dimension with the records in the physical layer time dimension.

[0083] The time delay compensation module 200 based on anchor point data determines the physical anchor point recording time. After that, the processing flow enters the stage of constructing a calculation model for the propagation time delay parameters.

[0084] S221, the time delay compensation module 200 based on anchor data calculates the user's historical behavior entropy for internet information publishers to quantify the dispersion of the user's information publishing timeliness. The module extracts the user's valid reporting records from the user behavior database within a historical period, calculates the time difference between the publication time and the subsequent verification confirmation time of each record, and statistically analyzes the probability distribution of the difference within the discretized time interval. The calculation process of user historical behavior entropy follows the following behavior entropy calculation formula:

[0085] ;

[0086] in, This represents the calculated entropy of user historical behavior. This represents the total number of discretization time intervals; This indicates that the user's historical posting time lag falls on the [number]th [page number]. The probability density values ​​of each time difference interval.

[0087] S222, the propagation delay estimation model is constructed based on the anchor point data delay compensation module 200 to calculate the estimated propagation delay of the current data. This model introduces a linear regression mechanism, comprehensively considering physical travel constraints and user behavior habits. The calculation process of the estimated propagation delay follows the following propagation delay estimation formula:

[0088] ;

[0089] in, This represents the calculated estimated propagation delay; This represents the regression coefficient related to travel time; Indicates the timestamp of information published on the Internet; Indicates the time recorded at the physical anchor point; This represents the correlation coefficient of behavioral entropy. Represents the entropy of a user's historical behavior; This represents the fundamental intercept constant.

[0090] The regression coefficients , and constants The parameters are fixed and obtained by fitting a historically validated dataset with least squares data before training, and are stored in the module's parameter register.

[0091] S223, Based on the calculated estimated propagation delay, the anchor data-based delay compensation module 200 subtracts the estimated propagation delay from the internet information's publication timestamp to obtain a corrected estimate of the event's actual occurrence time. This time calibration formula is expressed as follows:

[0092] ;

[0093] in, This represents the revised estimate of the actual time of the event. Indicates the timestamp of information published on the Internet; This indicates the estimated propagation time lag.

[0094] Following the content of step S220 above, the time delay compensation module 200 based on anchor point data calculates the corrected estimate of the actual occurrence time of the event. After that, the processing flow enters the stage of determining the backtracking time drift window.

[0095] S231, the time delay compensation module 200 based on anchor point data calculates the fault tolerance buffer time parameter for expanding the retrieval range. The module then calculates the time tolerance buffer time parameter based on the output geographic location constraint information. The physical facility coordinates corresponding to the physical anchor point recording time are used to calculate the road network topology distance between them using Dijkstra's algorithm or A* algorithm. Then, combined with the historical average traffic speed and speed variation coefficient of the road segment, the time fluctuation range is estimated. The calculation process for this fault tolerance buffer time parameter follows the following buffer time calculation formula:

[0096] ;

[0097] in, This represents the calculated fault tolerance buffer time parameter; This represents the distance along the road network topology between the output geographic location constraint information and the coordinates of the physical anchor point facility; This indicates the historical average traffic speed of this road section; The coefficient of variation represents the traffic speed on that road segment; This represents the system's preset basic synchronization error constant.

[0098] S232, the time delay compensation module 200 based on anchor point data uses the corrected estimate of the actual occurrence time of the event. As a baseline time point, combined with fault tolerance buffer time parameters A backtracking time-shift window is generated for retrieving historical data. The generation process of this backtracking time-shift window follows the following window definition formula:

[0099] ;

[0100] in, This represents the generated backtracking time drift window; This represents the revised estimate of the actual time of the event. This represents the fault tolerance buffer time parameter.

[0101] S233, the time delay compensation module 200 based on anchor point data calculates the generated backtracking time drift window. The module sends a retrieval command to the historical database of the front-end sensing and control facility 30. It extracts physical sensor data fragments from the historical data whose timestamps fall within the retrospective time drift window and whose spatial coordinates conform to the output geographical location constraints. These data fragments contain the data reconstructed by the dynamic acquisition protocol configuration packet in step S130, are marked as the physical dataset to be verified, and are transmitted to the multi-source micro-fingerprint verification module 300.

[0102] See attached document Figure 2 Following step S230, after the time delay compensation module 200 based on anchor point data completes the extraction of the physical dataset to be verified, the system logic flow transfers to the multi-source microscopic fingerprint verification module 300, and begins the initial stage in step S300, namely, the generation of the virtual kinematic fingerprint feature template. This step is implemented through the following sub-steps S311 to S313.

[0103] S311, the multi-source micro fingerprint verification module 300 reads the event category label of the final output obtained in the aforementioned step S113. The module uses this as an index to access a microscopic kinematic feature library stored in non-volatile memory. This feature library stores physical parameter constraint models corresponding to various traffic events. The module retrieves the corresponding model based on the event category label in the final output, which describes the range of values ​​that the physical sensors should record when the event occurs.

[0104] S312, the multi-source micro-fingerprint verification module 300 generates a virtual kinematic fingerprint feature template based on the retrieved physical parameter constraint model. This template is a set of multi-dimensional physical thresholds and logical operators used to define the physical boundaries of valid events. The generation process of this virtual kinematic fingerprint feature template follows the fingerprint definition formula:

[0105] ;

[0106] in, Represents a virtual kinematic fingerprint feature template; This represents a set of dynamic characteristic constraint functions for vehicle terminal data, involving velocity, acceleration, and angular velocity; This represents a set of visual feature constraint functions for video surveillance data, involving image pixel intensity and optical flow gradient. It represents the logical combination relationship between various feature constraints.

[0107] S313, the multi-source micro fingerprint verification module 300 instantiates and assigns values ​​to the constraint parameters in the above formula based on the specific event category label of the final output:

[0108] First, for events categorized as road icing or vehicle skidding, the module constructs a dynamic fingerprint based on vehicle stability. The module sets a joint threshold for lateral acceleration and yaw rate to identify lateral slippage not caused by driver intent. The constraints of this dynamic fingerprint are expressed by a sideslip determination formula:

[0109] ;

[0110] in, The logical conditions for determining a sideslip event; This represents the lateral acceleration value recorded by the vehicle's onboard sensors; This indicates the preset threshold for abnormal lateral acceleration. This represents the vehicle's yaw rate recorded by the onboard sensors; This indicates the preset abnormal yaw rate threshold; This indicates the rate of change of the vehicle's steering wheel angle; This indicates the steering wheel static tolerance threshold. Represents the logical AND operation.

[0111] Second, for events categorized as road debris or abnormal obstacles, the module constructs a visual fingerprint based on the optical flow field. The module utilizes the spatial domain region of interest parameters of the image sensor defined in step S132 above. The Lucas-Kanade optical flow algorithm or the Farneback dense optical flow algorithm is used to analyze the pixel motion vectors within the region. The constraints of this visual fingerprint are expressed by the following static foreign object determination formula:

[0112] ;

[0113] in, Logical conditions for determining static obstacle events; This represents the total number of pixels within the region of interest. Represents the pixel coordinates within an image frame; This represents the spatial domain region of interest parameters of the image sensor; Represents pixels The magnitude of the optical flow velocity vector at that location; This represents the zero-velocity tolerance threshold for determining whether an object is stationary. This represents the average optical flow velocity of the background region surrounding the region of interest. This represents the minimum motion threshold of the background traffic flow. Through the above steps, the module transforms semantic concepts into a set of physical parameters that can be compared by a computer.

[0114] The multi-source micro-fingerprint verification module 300 generates a virtual kinematic fingerprint feature template. After that, the system processing flow enters the isomorphism matching operation stage of heterogeneous physical data.

[0115] S321, the multi-source micro-fingerprint verification module 300 performs spatiotemporal extraction and feature dimensionality reduction operations on the physical data. The module uses the backtracking time drift window generated in step S232. Using time indexing, the geographic location constraint information obtained in step S114 is used. For spatial indexing, the module extracts the original sensor data stream reconstructed in step S130 from the historical database. For high-frequency sampled vehicle inertial navigation data, the module extracts peak values ​​and root mean square values ​​using a sliding window algorithm; for video data based on region-of-interest coding, the module extracts the motion vector field.

[0116] The extracted and processed data is encapsulated into a reconstructed physical sensor data vector, the construction of which follows the physical feature extraction formula:

[0117] ;

[0118] in, This represents the reconstructed physical sensor data vector; Represents the feature extraction function; This represents the raw sensor data extracted within the backtracking window; This represents the generated backtracking time drift window; This indicates the generated dynamic acquisition protocol configuration package.

[0119] S322, the multi-source micro-fingerprint verification module 300 performs isomorphic matching operations to calculate the similarity score between the physical measured data and the virtual fingerprint template. The module compares each component in the reconstructed physical sensor data vector with the corresponding threshold in the virtual kinematic fingerprint feature template, and uses a weighted normalized bias algorithm to quantify the degree of matching. The calculation process of this similarity score follows the following isomorphic matching formula:

[0120] ;

[0121] in, This represents the calculated similarity score; This represents the total number of dimensions of the feature vector; Indicates the first Confidence weights of each feature component; Represents the measured physical characteristic value; Indicates the reference physical threshold; Represents a small constant; This represents a logic indicator function. It takes the value 1 when the measured value meets the logic judgment condition defined in step S313, and takes the value 0 otherwise.

[0122] S323, the multi-source micro-fingerprint verification module 300 generates the final event authenticity verification result based on the calculated similarity score. The module compares the similarity score with a preset system confidence threshold. If the score is higher than the threshold, the event is determined to be true, and a physical truth signal containing physical coordinates and event type is generated; if the score is lower than the threshold, it is determined to be false or invalid information. The verification decision logic is expressed as follows:

[0123] ;

[0124] in, This indicates the result of the event verification. This indicates the similarity score; This represents the system's preset confidence threshold parameter.

[0125] When the multi-source micro fingerprint verification module 300 outputs the event authenticity verification result When the result is true, the system enters the traffic signal linkage control phase based on the verification result.

[0126] S331, the multi-source micro fingerprint verification module 300 is based on the output geographical location constraint information. The module performs spatial addressing of associated control facilities. It accesses the road network topology database to retrieve traffic signal controllers and variable message signs located upstream of the road segment where the incident occurred. After identifying the controlled facilities, the module determines the appropriate control facility based on the final output event category label. The system matches the control level from a pre-defined event-policy mapping table. If the event category is a serious incident, a level-one blocking policy is matched; if the event category is abnormal congestion, a level-two flow control policy is matched.

[0127] S332, the multi-source micro fingerprint verification module 300 generates specific signal control command parameters according to the matched control level.

[0128] For the primary blocking strategy, the multi-source micro-fingerprint verification module 300 generates a traffic interception control command, setting the upstream intersection traffic lights to all red. For the secondary flow control strategy, the multi-source micro-fingerprint verification module 300 generates a phase adjustment command, dynamically compressing the green light time to limit traffic flow into the event segment. This green light time adjustment process follows the following signal timing correction formula:

[0129] ;

[0130] in, This indicates the corrected target phase green light duration; This indicates the base green light duration for the target phase in the current time period; This indicates the preset flow reduction factor; The weighting coefficient represents the severity of the event, which is obtained by linearly mapping the similarity score calculated in step S322 above.

[0131] S333, the multi-source micro fingerprint verification module 300 encapsulates the generated signal control command into a traffic control protocol data packet and sends it to the front-end signal controller. The structure of this data packet is defined as follows:

[0132] ;

[0133] in, This indicates the traffic control instruction package that has been issued; A unique hardware identifier representing a target controlled facility; Indicates the control mode word; This represents the specific set of control parameters, including the corrected green light duration; Indicates the effective duration of the instruction.

[0134] After receiving the instruction, the front-end signal controller suspends the local operation plan and executes the actions defined by the instruction to achieve physical control of traffic flow.

[0135] See attached document Figure 2 After the system completes the physical linkage control of traffic signals, or after determining that the event is false information in step S323, the logical processing flow enters the semantic confidence self-evolution module 400 based on physical feedback, and begins to execute the initial stage in step S400, namely the construction of the physical truth feedback signal. This step is implemented through the following sub-steps S411 to S413, which aim to transform the verification results of the physical world into mathematical supervision signals that can be understood by the semantic model.

[0136] S411, the semantic confidence self-evolution module 400 based on physical feedback performs backtracking and alignment operations on historical semantic data. The module uses the unique identifier of the event to retrieve the high-dimensional semantic feature vector generated in step S112 from the system's temporary buffer. and the event category labels of the final output generated in step S113 This process ensures that the subsequently generated feedback signals correspond precisely to the original text input that triggered the physical verification.

[0137] S412, the semantic confidence self-evolution module 400 based on physical feedback verifies the authenticity of the event output in step S323. and the similarity score calculated in step S322 Calculate the binarized truth labels and feedback gain weights used for model correction.

[0138] The module first defines the binary truth labels. If the physical verification result is true, the semantic classification of the text is confirmed to be correct, and the label is set to 1; if the physical verification result is false, the semantic classification of the text is confirmed to be incorrect, and the label is set to 0. The definition formula for this binary truth label is as follows:

[0139] ;

[0140] in, Represents the binary truth label; This indicates the result of the verification of the authenticity of the event.

[0141] Subsequently, the semantic confidence self-evolution module 400 calculates the feedback gain weight. This weight is used to quantify the degree of certainty in physical verification. The further the similarity score calculated by the physical sensor is from the system confidence threshold, the higher the confidence of the physical layer in this judgment, and the greater the weight is given to the sample in model updates; conversely, if the score is near the threshold, the weight of the sample is reduced. The calculation of this feedback gain weight follows the following confidence mapping formula:

[0142] ;

[0143] in, This represents the calculated feedback gain weight; This indicates the preset gain scaling factor; This indicates the similarity score; This represents the system's preset confidence threshold parameter; This represents the nonlinear adjustment index.

[0144] S413, the semantic confidence self-evolution module 400 based on physical feedback encapsulates the aligned semantic features, labels, and weights into a structured physical truth feedback signal. This signal constitutes the basic data unit for subsequent backpropagation updates in the neural network. The structural definition formula of this physical truth feedback signal is as follows:

[0145] ;

[0146] in, The physical truth feedback signal indicates that the construction is complete; This represents the high-dimensional semantic feature vector of the output; Indicates the event category label in the final output; Represents the binary truth label; This represents the feedback gain weight.

[0147] After the module is built, it will The samples are stored in a high-confidence sample library located on the server. This database is specifically designed to store physically validated samples for use by the semantic model to update gradients.

[0148] The physical truth feedback signal is constructed in the semantic confidence self-evolution module 400 based on physical feedback. Afterwards, the logical processing flow enters the weight update stage of the semantic parsing model. This stage is implemented through the following sub-steps S421 to S423.

[0149] S421, the semantic confidence self-evolution module 400 based on physical feedback performs online gradient updates for the deep neural network. The module extracts batches of samples containing multiple physical truth feedback signals from a high-confidence sample library, and outputs a high-dimensional semantic feature vector. The data is then re-input into the deep neural network model described in step S112 to calculate the prediction bias.

[0150] In this process, the module construction incorporates a weighted loss function based on physical confidence to ensure that the direction of model updates is constrained by the physical validation results. This weighted loss function is expressed by the following formula:

[0151] ;

[0152] in, This represents the calculated physical guidance loss value; Indicates the number of samples in the training batch; For sample index variables; Indicates the first Feedback gain weights for each sample; Indicates the first Binarized ground truth labels for each sample; The model represents the first The original predicted probability values ​​of each sample belonging to the event category label in the final output.

[0153] Based on the calculation The module uses the backpropagation algorithm to update the parameter set of the deep neural network in step S112.

[0154] S422, the semantic confidence self-evolution module 400 based on physical feedback performs dynamic adjustment of the reputation weight for a specific information source. The semantic confidence self-evolution module 400 reads the virtual identity identifier of the Internet information publisher associated in step S212. And retrieve the user's current reputation weight from the system's source reputation database. .

[0155] Based on the results of this physical verification, the module incrementally updates the user's reputation weight. If the physical verification confirms the event as true, the system increases the user's reputation value; if it confirms as false, the system deducts the user's reputation value. The update process of this reputation weight follows the following reputation iteration formula:

[0156] ;

[0157] in, This indicates the updated reputation weight; This indicates the reputation weight before the update; This represents a numerical truncation function that ensures the weight values ​​remain between 0 and 1. This represents the preset reputation learning rate parameter; Indicates the feedback gain weight; Represents the binary truth label; This represents the polarity mapping term.

[0158] S423, the semantic confidence self-evolution module 400 based on physical feedback will update the... Write back to the source reputation database.

[0159] In the subsequent execution step S113 for event category determination, this weight will be multiplied into the output layer of the classifier as a prior probability coefficient. Specifically, the system performs a dot product operation between the user's reputation weight and the probability distribution output by the classifier, making it easier for reports from high-reputation users to exceed the determination threshold, while reports from low-reputation users require stronger semantic feature support.

[0160] Through the above steps, the system implements an iterative filtering mechanism based on physical facts, enabling the semantic parsing model to dynamically evaluate the reliability of the information source based on physical feedback.

[0161] After the system completes the weight update of the semantic parsing model, the logical processing flow enters the correction stage of fuzzy semantic geographic mapping. This stage is implemented through the following sub-steps S431 to S433, which aim to use the high-precision determinism of physical space to correct the spatial mapping deviation of fuzzy geographic entities in natural language description.

[0162] S431, the semantic confidence self-evolution module 400 based on physical feedback first parses the physical truth feedback signal generated in the aforementioned step S413. Extract the binarized truth labels contained therein. Only when At this time, the semantic confidence self-evolution module 400 further retrieves the precise physical coordinates of the successful verification locked by the multi-source micro fingerprint verification module 300 in this event. These coordinates originate from the specific installation location of the highway toll gantry or roadside equipment that triggered the verification. This coordinate is marked as... The module also extracts fuzzy place name entities from the original text of the event using named entity recognition, and establishes paired samples of "fuzzy text entity - precise physical coordinates".

[0163] S432, the semantic confidence self-evolution module 400 based on physical feedback iteratively updates the pre-set road network topology geographic information database. This database stores the mapping relationship between fuzzy place name entities and their geographic coordinate center points and coverage radii. The module calculates the spatial vector deviation between the successfully verified precise physical coordinates and the currently stored mapping center point of the place name entity in the database, and uses the feedback gain weight calculated in step S412. The center point of the mapping is corrected by weighted translation. The correction process of this fuzzy geographic mapping follows the following spatial centroid iteration formula:

[0164] ;

[0165] in, This represents the coordinate vector of the geographic mapping center of the corrected fuzzy place name entity. This represents the coordinate vector of the geographic mapping center of the fuzzy place name entity before correction. This represents the preset geospatial mapping learning rate; Indicates the feedback gain weight; The precise physical coordinates indicating successful verification.

[0166] S433, the semantic confidence self-evolution module 400 based on physical feedback further corrects the spatial coverage radius of fuzzy place name entities. The module calculates and verifies the accurate physical coordinates and corrected center coordinates. The Euclidean distance between them is used to update the spatial coverage radius parameter of the place name entity. The correction logic for the spatial coverage range follows the radius update formula as follows:

[0167] ;

[0168] in, Indicates the spatial coverage radius of the corrected fuzzy place name entity; Indicates the radius update smoothing coefficient; Indicates the spatial coverage radius of the ambiguous place name entity before correction; Represents the Euclidean distance operation; The precise physical coordinates indicating successful verification; This represents the corrected geographic mapping center coordinate vector.

[0169] After the update is complete, the module writes the corrected coordinates and radius into the road network topology geographic information database. During the subsequent execution step S114 for geographic location constraint extraction, the system will use these corrected parameters to generate more accurate output geographic location constraint information. This reduces invalid search areas caused by misunderstandings of place names.

Claims

1. A smart transportation information internet acquisition and multi-source data fusion system, characterized in that, include: The semantically driven heterogeneous facility protocol reconstruction module (100) is used to parse Internet text data, extract the output geographical location constraint information, the final output event category label and the output high-dimensional semantic feature vector, and generate a dynamic acquisition protocol configuration package adapted to the front-end perception control facility (30). The time delay compensation module (200) based on anchor point data associates the physical anchor point recording time with the output geographical location constraint information and calculates the backtracking time drift window. The multi-source micro fingerprint verification module (300) extracts the reconstructed physical sensor data vector based on the dynamic acquisition protocol configuration package and the backtracking time drift window, and performs isomorphic matching between the physical sensor data vector and the virtual kinematic fingerprint feature template constructed based on the event category label of the final output, and outputs the similarity score and the event authenticity verification result. The physical feedback-driven semantic evolution module (400) constructs a physical truth feedback signal based on the event authenticity verification result, and uses the physical truth feedback signal to update the weights of the semantic parsing model and correct the fuzzy semantic geographic mapping.

2. The intelligent transportation information internet acquisition and multi-source data fusion system according to claim 1, characterized in that, When generating the generated dynamic acquisition protocol configuration package, the semantically driven heterogeneous facility protocol reconstruction module (100) performs the following operations: In the heterogeneous sensing device protocol feature library, match the private protocol format of the road segment front-end sensing control facility (30) corresponding to the output geographical location constraint information; The proprietary protocol format is converted into a unified middleware interface standard by using a pre-trained protocol mapping network. The parameter set, which includes device access parameters, data decoding rules, and sampling frequency requirements, is then encapsulated into the generated dynamic acquisition protocol configuration package.

3. The intelligent transportation information internet acquisition and multi-source data fusion system according to claim 1, characterized in that, When calculating the generated backtracking time drift window, the time delay compensation module (200) based on anchor point data executes the following calculation logic: The propagation delay parameter is calculated based on the difference between the physical anchor point recording time and the Internet information publication timestamp. The propagation delay parameter is used to correct the publication timestamp of the Internet information to obtain a corrected estimate of the actual occurrence time of the event. Centered on the corrected estimated actual event time, and with a preset fault tolerance buffer time parameter as the radius, the generated backtracking time drift window is defined. The physical anchor point recording time is derived from the highway toll gantry equipment (33) or the roadside detection equipment (32).

4. The intelligent transportation information internet acquisition and multi-source data fusion system according to claim 1, characterized in that, When constructing the virtual kinematic fingerprint feature template, the multi-source micro fingerprint verification module (300) retrieves the physical parameter constraint model based on the event category label of the final output and generates a set containing multi-dimensional physical thresholds and logical operators. The set includes: a dynamic characteristic constraint function for road icing events, wherein the dynamic characteristic constraint function sets a joint threshold of lateral acceleration and yaw rate from the on-board terminal device (34); And a visual feature constraint function for road littering events, wherein the visual feature constraint function sets the minimum motion threshold between the pixel motion vector and the background traffic flow within the spatial domain region of interest parameters of the image sensor of the traffic video surveillance device (31).

5. The intelligent transportation information internet acquisition and multi-source data fusion system according to claim 1, characterized in that, The multi-source micro fingerprint verification module (300) uses the following calculation logic when outputting similarity scores: For each feature component in the reconstructed physical sensor data vector, calculate the absolute difference between the measured physical feature value and the reference physical threshold in the virtual kinematic fingerprint feature template; The normalized deviation of the absolute difference relative to the reference physical threshold is solved using a weighted normalized deviation algorithm. The similarity score is obtained by multiplying the value obtained by subtracting the normalization bias from 1 by the confidence weight of the feature component and the value of the logical indicator function, and summing the calculation results for all feature dimensions.

6. The intelligent transportation information internet acquisition and multi-source data fusion system according to claim 5, characterized in that, The multi-source micro fingerprint verification module (300) is also used to send a signal control command to the traffic signal control facility (40) when the event authenticity verification result is true, wherein the correction of the green light duration follows the following logic: Multiply the base green light duration of the target phase in the current time period by the correction factor to obtain the corrected green light duration of the target phase. The correction coefficient is 1 minus the product of the traffic reduction coefficient and the event severity weight coefficient; wherein the event severity weight coefficient is obtained by linearly mapping the similarity score.

7. The intelligent transportation information internet acquisition and multi-source data fusion system according to claim 5, characterized in that, The physical feedback-driven semantic evolution module (400) includes calculating the feedback gain weight when constructing the physical truth feedback signal. The specific calculation logic is as follows: Calculate the absolute difference between the similarity score and the system's preset confidence threshold parameter; The absolute difference is subjected to a nonlinear adjustment exponentiation operation and multiplied by a preset gain scaling factor to obtain the feedback gain weight. The physical truth feedback signal includes the high-dimensional semantic feature vector of the output, the event category label of the final output, the binarized truth label, and the feedback gain weight.

8. The intelligent transportation information internet acquisition and multi-source data fusion system according to claim 7, characterized in that, The physical feedback-driven semantic evolution module (400) performs the following operations when updating the weights of the semantic parsing model: Extract a batch of samples containing multiple physical truth feedback signals from a high-confidence sample library; A weighted loss function for physical confidence is constructed, and the cross-entropy loss between the predicted probability by the deep neural network model and the binary ground truth label is weighted and summed, where the weighting coefficients are the feedback gain weights; the parameter set of the deep neural network model is updated using the backpropagation algorithm based on the weighted loss function for physical confidence.

9. The intelligent transportation information internet acquisition and multi-source data fusion system according to claim 7, characterized in that, The physical feedback-driven semantic evolution module (400) is also used to perform dynamic adjustment of reputation weights for specific information sources, and its calculation logic is as follows: Read the reputation weight of the internet information publisher before the update, corresponding to their virtual identity. The updated reputation weights are obtained by adding a correction term to the original reputation weights and using a numerical truncation function to ensure that the result is between 0 and 1. The correction term is the product of the reputation learning rate parameter, the feedback gain weight, and the polarity mapping term determined by the binarized truth label.

10. The intelligent transportation information internet acquisition and multi-source data fusion system according to claim 7, characterized in that, The physical feedback-driven semantic evolution module (400) performs the following computational logic when correcting the fuzzy semantic geographic mapping: When the binarized truth label is true, obtain the accurate physical coordinates of the successful verification; Calculate the spatial vector deviation between the verified precise physical coordinates and the coordinate vector of the geographic mapping center of the fuzzy place name entity currently stored in the road network topology geographic information database; The spatial vector deviation is multiplied by the feedback gain weight and the preset geospatial mapping learning rate, and then superimposed onto the original mapping center coordinate vector as a correction term to obtain the corrected fuzzy place name entity geographic mapping center coordinate vector.

Images