Expressway data fusion management method

Abstract

The invention discloses a highway data fusion management method, and relates to the technical field of intelligent traffic management, and the method comprises the steps: collecting the global operation data of a highway in real time through a multi-source heterogeneous sensing network; performing space-time alignment and semantic annotation processing on the operation data to construct a fusion data set; inputting the feature subset into a lightweight space-time fusion model to form a multi-target decision set; executing the traffic control instruction sequence in a preset space-time window according to the multi-target decision set; and generating a self-optimization report including a model iteration path and road network health assessment based on the deviation degree of the execution feedback data and the expected optimization target. According to the expressway data fusion management method provided by the invention, collaborative optimization of passing efficiency and safety of the expressway is guaranteed.

Description

Technical Field

[0001] The present application relates to the field of intelligent traffic management technology, and in particular to a highway data fusion management method. Background Art

[0002] Existing highway management systems often rely on data from single sensors (such as loop detectors or cameras) to monitor local sections. This results in limited data dimensions and incomplete spatial and temporal coverage, making it difficult to accurately characterize dynamic traffic flow characteristics across the entire road network. Traditional data processing methods typically employ static rules or simple threshold judgments, lacking the ability to integrate heterogeneous multi-source data and susceptible to environmental noise. This leads to delayed traffic event detection and lagging control strategies, making it difficult to meet the real-time response requirements under complex road conditions.

[0003] With the development of vehicle-infrastructure collaboration (V2X) technology, some solutions have attempted to introduce edge computing devices to improve data processing efficiency. However, these solutions still face bottlenecks such as poor model generalization and insufficient decision interpretability. Existing systems, in particular, lack the ability to coordinate multi-objective optimization and dynamic resource scheduling in extreme weather or accident scenarios, making it difficult to balance multiple constraints such as traffic efficiency, safety risks, and energy consumption. Therefore, a highway data management approach that integrates multimodal perception, adaptive feature learning, and federated optimization is urgently needed to achieve accurate understanding of the entire road network and the dynamic generation of control strategies. Summary of the Invention

[0004] In order to overcome the above-mentioned deficiencies of the prior art, the present application provides a highway data fusion management method.

[0005] In a first aspect, the present application provides a highway data fusion management method, the method comprising: Real-time data collection of highway-wide operational data is collected through a multi-source heterogeneous perception network consisting of millimeter-wave radar arrays, lidar, smart cameras, and in-vehicle wireless communication equipment. This data includes vehicle trajectory point clouds, roadside unit communication logs, meteorological sensor readings, road surface status images, and emergency alarm signals. Perform spatiotemporal alignment and semantic annotation on the operational data to construct a fused dataset containing spatiotemporal stamp codes, a data quality assessment matrix, and abnormal event markers. Then, use an adaptive feature selection algorithm to extract a subset of dynamic features that are strongly correlated with traffic trends. The feature subset is input into a lightweight spatiotemporal fusion model, which generates a road section-level risk heat map and a lane-level traffic efficiency optimization plan based on a multi-head spatiotemporal attention mechanism. This forms a multi-objective decision set including variable speed limit strategies, dynamic lane allocation, and emergency resource scheduling. Execute a traffic control command sequence within a preset spatiotemporal window based on a multi-objective decision set. The command sequence includes three stages: dynamic speed limit sign updating, variable lane alignment guidance, and right-of-way allocation for rescue vehicles. Traffic flow parameters and safety indicators after command execution are monitored in real time. Based on the deviation between the execution feedback data and the expected optimization target, the parameter weights of the lightweight spatiotemporal fusion model are updated through the federated learning framework, and a self-optimization report including the model iteration path and road network health assessment is generated.

[0006] Preferably, a multi-source data conflict resolution model is constructed. When there are conflicting detection results of different sensors on the same target, confidence weights are dynamically assigned based on the device's historical accuracy and real-time signal strength. A sliding window mechanism is used to perform Kalman filtering on vehicle trajectory data to eliminate trajectory jitter noise caused by GPS drift. The contribution of each feature dimension to traffic situation prediction is calculated using the information entropy weight method, and features with a contribution higher than a preset threshold are retained to construct an orthogonal feature space. Perform adversarial enhancement on the filtered feature subset, injecting Gaussian noise and random occlusion to improve model robustness; When it is detected that the multicollinearity between feature dimensions exceeds the preset threshold, the principal component analysis dimensionality reduction operation is triggered and the variance explanation rate above the preset probability is retained.

[0007] Preferably, knowledge distillation technology is used to compress the pre-trained traffic model with tens of billions of parameters into a lightweight version suitable for roadside edge computing units, maintaining sensitivity to emergencies and decision-making accuracy; A dynamic structured pruning module is embedded in the model compression process to automatically adjust the model depth and attention head distribution based on real-time computing resources and task urgency; Build a digital twin system for highways, simulate traffic flow evolution under extreme weather conditions such as heavy rain and fog through the twin, and generate enhanced training data to improve model generalization capabilities; Implement a cache quantization strategy during the inference phase. When the key-value cache occupancy exceeds the device memory limit, a product quantization algorithm is used to compress and store the query key vector in the attention mechanism. Design a multimodal output interface, use 8-bit integer quantization to accelerate inference for real-time control instructions, and retain 32-bit floating-point calculation accuracy for post-analysis reports.

[0008] Preferably, during the dynamic speed limit update phase, the speed limit adjustment range of each road section is calculated based on the gradient change rate of the risk heat map, and a graded speed limit recommendation is sent to the vehicle via the vehicle wireless broadcast protocol; During the variable lane induction phase, a dynamic lane function conversion plan is generated based on the real-time traffic efficiency prediction model, and laser projection equipment is used to display variable lane linear markings on the road surface; During the emergency rescue dispatch phase, vehicle trajectory prediction and road network topology are combined to plan the minimum time path for rescue vehicles and dynamically adjust the signal priority along the route; The Nash equilibrium algorithm is used to optimize the weights of multi-objective decisions, balance the game relationship between traffic efficiency, safety risks and energy consumption, and generate a Pareto optimal control strategy; After the instruction is executed, the reverse verification mechanism is activated, and by comparing the deviation between the actual traffic flow parameters and the predicted values, online correction of control parameters or event backtracking analysis is triggered.

[0009] Preferably, a knowledge graph of highway events is constructed, integrating historical accident cases, traffic control rules, and emergency response plans, and providing explainable support for real-time decision-making through retrieval-enhanced generation technology; Deploy privacy computing modules on roadside edge devices to homomorphically encrypt vehicle trajectory data and record data usage authorization credentials through blockchain. When a new traffic event pattern is detected, the active learning mechanism is activated to collect sample data and trigger incremental model training. The updated model parameters are synchronized to all road network nodes through a secure multi-party computing protocol.

[0010] Preferably, entity-relationship-attribute triples are extracted from traffic accident reports, maintenance records, and traffic regulations to construct an ontology library containing road topology, event causal chains, and disposal constraints; A graph attention network is used to dynamically embed the knowledge graph to capture high-order relationships between event types, road section characteristics, and disposal measures. In the real-time decision-making process, similar historical cases are retrieved through subgraph similarity matching to generate a decision explanation report that includes the effectiveness evaluation of the disposal suggestion and the risk transmission path; Build a rule distillation channel to convert traffic management regulations into constraints of the model loss function to ensure that decisions comply with road traffic safety regulations.

[0011] Preferably, an augmented reality management interface is constructed to overlay and display the traffic flow density field, risk diffusion trend and control instruction impact range on the three-dimensional road network model; When the system detects a major safety hazard, it automatically switches to emergency control mode, initiates coordinated warning for the entire road network, and generates a risk avoidance route guidance plan.

[0012] Preferably, the optimal set of avoidance paths is calculated within a preset time based on a deep reinforcement learning algorithm, taking into account vehicle distribution density, emergency exit locations, and accessibility of rescue resources; The enhanced display interface uses thermal gradients to indicate the risk diffusion rate and impact area, and dynamically displays the priority of each lane. Through communication between vehicles and infrastructure, graded warning information is sent to affected vehicles, including guidance on evasive paths, recommended speeds, and lane change timing.

[0013] Preferably, a road network health assessment model is established to integrate the pavement damage index, equipment failure rate, and traffic flow stability index to calculate the road section health score; When the health score falls below the preset threshold, a maintenance operation plan is automatically generated and an appointment window is scheduled with the maintenance vehicle dispatch system; Embed a digital twin verification module in the maintenance plan to simulate the impact of different maintenance strategies on traffic efficiency and optimize construction timing.

[0014] Compared with the prior art, the present invention has the following characteristics and beneficial effects: First, a multimodal sensor array collects real-time operating status data and environmental parameters of the main and auxiliary power systems, achieving comprehensive awareness of the ship's power system operating status. This addresses the single-dimensional data collection issue of traditional methods, provides multi-dimensional data support for subsequent intelligent decision-making, and significantly improves the comprehensiveness and accuracy of system status monitoring. Second, the collected operating status data and environmental parameters are subjected to spatiotemporal alignment and noise suppression, constructing an input dataset consisting of spatiotemporal coding labels, a multi-source data association matrix, and abnormal feature vectors. This effectively eliminates spatiotemporal bias between sensor data and environmental noise interference. A dynamic feature selection algorithm selects the feature subset most relevant to power switching decisions, ensuring data quality and improving the efficiency of subsequent model processing. Furthermore, this feature subset is fed into a lightweight large-scale model inference engine for intelligent analysis. Based on a spatiotemporal attention mechanism, power system degradation characteristics and switching risk probability distributions are extracted, generating a decision set containing multiple optimization solutions. This overcomes the limitations of traditional threshold judgment, achieves adaptive processing capabilities for complex operating conditions, and significantly improves the scientific nature and reliability of switching decisions. The system dynamically switches between the primary and auxiliary power sources based on a multi-objective optimization decision set. A phased, refined control strategy enables smooth switching, effectively reducing the mechanical shock and electrical disturbances generated during traditional switching processes. This ensures stable operation of the ship's power system and extends its service life. Finally, a federated learning framework continuously optimizes model parameters based on switching feedback data, enabling the system to continuously adapt to changes in ship performance and environmental conditions, achieving continuous self-improvement and performance enhancement of the power switching system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of the steps of a highway data fusion management method mainly embodied in this embodiment. DETAILED DESCRIPTION

[0016] The present invention is further described in detail below with reference to the following examples.

[0017] Reference Figure 1 , a highway data fusion management method, the method comprising the following steps: S1. Real-time collection of highway-wide operational data is achieved through a multi-source heterogeneous perception network. The operational data includes vehicle trajectory point clouds, roadside unit communication logs, meteorological sensor readings, road surface status images, and emergency alarm signals. The perception network consists of a millimeter-wave radar array, lidar, smart cameras, and vehicle-mounted wireless communication equipment.

[0018] S2. Perform spatiotemporal alignment and semantic annotation on the operating data to construct a fused dataset containing spatiotemporal stamp encoding, data quality assessment matrix, and abnormal event markers. Then, use an adaptive feature selection algorithm to extract a subset of dynamic features that are strongly correlated with traffic conditions.

[0019] S3. Input the feature subset into the lightweight spatiotemporal fusion model, and generate a section-level risk heat map and lane-level traffic efficiency optimization plan based on the multi-head spatiotemporal attention mechanism, forming a multi-objective decision set including variable speed limit strategy, dynamic lane allocation and emergency resource scheduling.

[0020] S4. Execute a traffic control command sequence within a preset spatiotemporal window based on a multi-objective decision set. The command sequence includes three stages: dynamic speed limit sign updating, variable lane alignment guidance, and right-of-way allocation for rescue vehicles. Traffic flow parameters and safety indicators after command execution are monitored in real time.

[0021] S5. Based on the deviation between the execution feedback data and the expected optimization target, the parameter weights of the lightweight spatiotemporal fusion model are updated through the federated learning framework, and a self-optimization report including the model iteration path and road network health assessment is generated.

[0022] Specifically, a multimodal sensor array collects real-time operating status data and environmental parameters of the main and auxiliary power systems, enabling comprehensive awareness of the ship's power system operating status. This addresses the single-dimensional data collection issue of traditional methods, provides multi-dimensional data support for subsequent intelligent decision-making, and significantly improves the comprehensiveness and accuracy of system status monitoring. Secondly, the collected operating status data and environmental parameters are subjected to spatiotemporal alignment and noise suppression, constructing an input dataset consisting of spatiotemporal coding labels, a multi-source data association matrix, and abnormal feature vectors. This effectively eliminates spatiotemporal bias between sensor data and environmental noise interference. A dynamic feature selection algorithm selects the feature subset most relevant to power switching decisions, ensuring data quality and improving the efficiency of subsequent model processing. Furthermore, this feature subset is fed into a lightweight large-scale model inference engine for intelligent analysis. Based on a spatiotemporal attention mechanism, power system degradation characteristics and switching risk probability distributions are extracted, generating a decision set containing multiple optimization solutions. This overcomes the limitations of traditional threshold judgment, enables adaptive processing of complex operating conditions, and significantly improves the scientific nature and reliability of switching decisions. The system dynamically switches between the primary and auxiliary power sources based on a multi-objective optimization decision set. A phased, refined control strategy enables smooth switching, effectively reducing the mechanical shock and electrical disturbances generated during traditional switching processes. This ensures stable operation of the ship's power system and extends its service life. Finally, a federated learning framework continuously optimizes model parameters based on switching feedback data, enabling the system to continuously adapt to changes in ship performance and environmental conditions, achieving continuous self-improvement and performance enhancement of the power switching system.

[0023] Specifically, step S1 involves building a multi-source heterogeneous perception network, which is the core task of data collection. By deploying millimeter-wave radar arrays, lidar, smart cameras, and vehicle-mounted wireless communication equipment, real-time collection of highway operation data can be achieved. Millimeter-wave radar can operate in the 77GHz frequency band, which offers excellent penetration and resolution, enabling accurate measurement of vehicle speed and distance. Its operating principle is based on the Doppler effect. By transmitting and receiving millimeter-wave signals, the frequency changes of the echoes are analyzed to determine vehicle speed, and the round-trip signal time is used to calculate the distance between the vehicle and the radar. In actual highway scenarios, regardless of heavy traffic or significant variations in vehicle speeds, millimeter-wave radar can stably acquire vehicle motion information, providing reliable data support for subsequent traffic flow analysis and vehicle behavior monitoring.

[0024] In some embodiments, the LiDAR can use a 905nm laser beam for scanning, which can generate centimeter-level precision point cloud data of vehicle trajectories. When the LiDAR is working, it emits a laser beam into the surrounding space at an extremely fast speed. When the laser beam encounters an object such as a vehicle, it is reflected back and received by the detector. By accurately measuring the time difference between the emission and reception of the laser beam, combined with the emission angle of the laser beam, the spatial position information of the object can be calculated. Numerous position information points constitute the point cloud data. These point cloud data can accurately depict the outline, position and driving trajectory of the vehicle. Even in complex traffic environments, such as multi-lane merging and vehicles weaving in and out, the movement status of each vehicle can be clearly distinguished, providing key data for accurate perception of traffic conditions.

[0025] Smart cameras utilize advanced object detection algorithms, such as YOLOv5, to identify road conditions and anomalies in real time. Based on deep learning technology and trained on extensive image data, the YOLOv5 algorithm possesses powerful object recognition capabilities. It can quickly and accurately identify road conditions that could impact driving safety, such as accumulated water, snow, or ice, as well as abnormal events such as traffic accidents and vehicle breakdowns. Once an anomaly is detected, the smart camera rapidly transmits the relevant information to subsequent processing systems, providing a basis for timely response measures and significantly improving early warning capabilities for road safety hazards.

[0026] Vehicle-to-vehicle wireless communication devices exchange information with onboard units (OVUs) based on dedicated communication protocols. Through this communication method, vehicles can transmit real-time information such as their location, speed, direction of travel, and emergency braking, while also receiving relevant information from other vehicles and roadside units. This enables the rapid dissemination and sharing of traffic information across the entire network, enabling collaborative perception between vehicles and between vehicles and infrastructure, and providing comprehensive real-time data for intelligent traffic decision-making.

[0027] To ensure effective integration and utilization of collected data, all sensors require time synchronization and unified spatial coordinates. By employing a precise time synchronization protocol, each sensor can maintain precise time synchronization during data collection, avoiding data inconsistencies caused by time differences. The spatial coordinate system is converted to a geographic coordinate system, an internationally recognized geographic coordinate system. This provides a unified spatial reference standard for data collected by different sensors, facilitating subsequent data fusion and analysis. After data collection, the edge computing gateway standardizes the data format and converts it into an efficient columnar storage format. This format offers the advantages of compact storage and fast query speed, significantly improving data processing efficiency and facilitating the rapid retrieval and analysis of large amounts of data.

[0028] Through the above implementation, the multi-source heterogeneous perception network overcomes the limitations of traditional single-loop detectors. Traditional loop detectors can only detect the passage of vehicles at specific locations, have a limited detection range, and are easily affected by environmental factors. However, the multi-source heterogeneous perception network leverages the complementary characteristics of multiple sensors to achieve all-weather, full-section coverage monitoring of highways. While the performance of smart cameras is somewhat affected in low-light conditions, such as at night, millimeter-wave radar and lidar are not limited by light and can still accurately collect data. In inclement weather such as rain and fog, the camera's field of view is obscured. However, millimeter-wave radar and lidar, with their penetrating capabilities, can effectively compensate for the camera's limitations, ensuring stable and reliable data collection. This comprehensive and accurate data collection method provides a solid foundation for subsequent traffic management and decision-making, significantly improving the comprehensiveness and accuracy of highway operating status monitoring.

[0029] The specific step S2 includes the following sub-steps: Build a multi-source data conflict resolution model. When different sensors have conflicting detection results for the same target, dynamically assign confidence weights based on the device's historical accuracy and real-time signal strength. A sliding window mechanism is used to perform Kalman filtering on vehicle trajectory data to eliminate trajectory jitter noise caused by GPS drift. The contribution of each feature dimension to traffic situation prediction is calculated using the information entropy weight method, and features with a contribution higher than a preset threshold are retained to construct an orthogonal feature space. Perform adversarial enhancement on the filtered feature subset, injecting Gaussian noise and random occlusion to improve model robustness; When it is detected that the multicollinearity between feature dimensions exceeds the preset threshold, the principal component analysis dimensionality reduction operation is triggered and the variance explanation rate above the preset probability is retained.

[0030] Specifically, when building a multi-source data conflict resolution model, it is advisable to employ the DS evidence theory to integrate the confidence levels of multiple sensors. This theory, by comprehensively considering information from multiple evidence sources, can more accurately determine the target's status. Specifically, a sliding window is used to calculate the false alarm rate of each device over the past 30 days to assess the device's historical accuracy. In practice, a 30-day statistical period is used to record the number of false alarms for each device during this period. The false alarm rate is then calculated based on the total number of detections, serving as a quantitative indicator of historical accuracy. Real-time signal strength is quantified using the received signal strength indicator (RSSI). RSSI reflects the fluctuations in signal strength during transmission. By quantifying this, the reliability of sensor signals can be more intuitively assessed. Finally, a weighted average method is used to combine the historical accuracy and real-time signal strength to generate an optimal estimate of the target vehicle's location. This approach fully considers the device's historical performance and current signal quality, effectively resolving the issue of inconsistent detection results for the same target from different sensors and improving data accuracy and reliability.

[0031] Vehicle trajectory data is smoothed using a sliding window mechanism combined with a Kalman filter. The sliding window length is set to 500ms, and vehicle trajectory data is analyzed and processed within this time window. The Kalman filter is a commonly used optimal estimation method that recursively estimates state variables such as vehicle position, velocity, and acceleration by establishing state equations and observation equations. In highway scenarios, vehicle motion is affected by various factors, such as GPS signal drift caused by multipath, which can cause jitter in vehicle trajectories. The Kalman filter's state equation includes three-order variables: position, velocity, and acceleration, and can comprehensively describe the vehicle's motion. The observation noise covariance matrix is ​​dynamically adjusted based on the GPS signal-to-noise ratio. When the GPS signal-to-noise ratio is high, indicating good signal quality, the observation noise covariance matrix is ​​reduced accordingly, resulting in more accurate filtering results. When the GPS signal-to-noise ratio is low, the observation noise covariance matrix is ​​increased to accommodate signal uncertainty. This method effectively suppresses trajectory jitter noise caused by GPS drift, resulting in smoother and more accurate vehicle trajectories and providing reliable data for subsequent traffic flow analysis and vehicle behavior prediction.

[0032] The information entropy weighting method is a feature selection method based on information theory. The contribution of each feature to traffic situation prediction is assessed by calculating the information gain value of each feature dimension (such as traffic density, average speed, lane occupancy, etc.) in a historical dataset. The larger the information gain value, the more information the feature contains about the traffic situation and the greater its contribution to the prediction. A grid search method is used to determine a preset threshold, and features with a contribution above this threshold (e.g., 0.3) are selected. Then, the Gram-Schmidt orthogonalization method is used to construct an orthogonal feature space. Gram-Schmidt orthogonalization converts a set of linearly correlated vectors into a set of orthogonal vectors. This method reduces the correlation between processed features, avoiding the increased computational complexity and decreased model performance caused by redundant information between features. At the same time, it retains features that are important for traffic situation prediction, improving the efficiency and accuracy of subsequent model processing.

[0033] Adversarial enhancement is performed on the filtered feature subset to improve the robustness of the model. During the training phase, Gaussian noise with a standard deviation of 0.05 is injected into the feature subset, and 15% of the input image is randomly occluded. Injecting Gaussian noise simulates the noise interference found in real-world environments, allowing the model to learn to accurately recognize and process data even in noisy environments. Randomly occluding portions of the input image forces the model to learn more generalizable feature representations and avoid over-reliance on local features. This allows the model to maintain good performance in real-world applications when encountering partially missing images, noise interference, or other unseen conditions, improving its adaptability to complex environments.

[0034] When multicollinearity between feature dimensions is detected exceeding a preset threshold (for example, the Pearson correlation coefficient exceeds 0.8), principal component analysis (PCA) dimensionality reduction is triggered. PCA is a commonly used dimensionality reduction technique that converts raw data into a set of uncorrelated principal components through linear transformation. During PCA, principal components that contribute 95% of the cumulative variance are retained. This means that while reducing the data dimension, 95% of the key information in the original data is retained. PCA dimensionality reduction reduces redundant information in the data, lowers computational complexity, and speeds up model training and inference. It also avoids model overfitting caused by multicollinearity and improves model stability and generalization.

[0035] The specific step S3 includes the following sub-steps: Using knowledge distillation technology, a pre-trained traffic model with tens of billions of parameters is compressed into a lightweight version suitable for roadside edge computing units, maintaining sensitivity to emergencies and decision-making accuracy. A dynamic structured pruning module is embedded in the model compression process to automatically adjust the model depth and attention head distribution based on real-time computing resources and task urgency; Build a digital twin system for highways, simulate traffic flow evolution under extreme weather conditions such as heavy rain and fog through the twin, and generate enhanced training data to improve model generalization capabilities; Implement a cache quantization strategy during the inference phase. When the key-value cache occupancy exceeds the device memory limit, a product quantization algorithm is used to compress and store the query key vector in the attention mechanism. Design a multimodal output interface, use 8-bit integer quantization to accelerate inference for real-time control instructions, and retain 32-bit floating-point calculation accuracy for post-analysis reports.

[0036] Specifically, knowledge distillation technology is used to compress a pre-trained, multi-billion-parameter traffic model into a lightweight version suitable for roadside edge computing units (ROCs). This knowledge distillation technique is based on a teacher-student model framework. The teacher model is a large, multi-billion-parameter traffic model based on the Transformer architecture. This model is pre-trained on large-scale historical trajectory data (e.g., 1 billion trajectories) and has learned rich traffic knowledge and complex traffic patterns. The student model is a streamlined, lightweight version of the teacher model. The KL divergence loss function is used to align the output probability distributions of the teacher model in key scenarios (such as ramp merging areas and accident-prone road sections). The KL divergence loss function measures the difference between two probability distributions. By minimizing this loss function, the student model learns the teacher model's decision-making patterns and knowledge in key scenarios, significantly reducing the number of model parameters while maintaining sensitivity to unexpected events. This lightweight model can run efficiently within the limited computing resources of the RCU, enabling real-time traffic analysis and decision-making, and improving the model's practicality and deployability.

[0037] A dynamic structured pruning module is embedded in the model compression process to automatically adjust the model depth and attention head distribution based on real-time computing resources and task urgency. By monitoring the roadside unit's real-time CPU utilization (relevant information is available through the system's / proc file system), 50%-70% of the attention heads are dynamically disabled when high CPU utilization indicates limited computing resources. When disabling attention heads, attention channels sensitive to emergencies are prioritized, as these channels often play a key role in detecting and handling abnormal traffic events. For example, in emergencies such as traffic accidents or temporary road restrictions, these retained attention channels can quickly capture relevant information and respond. This approach ensures that computing resources are rationally allocated while ensuring the model's ability to handle critical events, preventing slowdowns or crashes caused by insufficient computing resources. This improves the model's adaptability and flexibility, enabling it to better adapt to diverse computing resource environments and task requirements.

[0038] A digital twin system for highways was constructed to simulate traffic flow evolution under extreme weather conditions, such as heavy rain and fog. The digital twin system was built using the SUMO simulation platform. First, road network data was imported to accurately simulate the real-world highway network structure, including road layout, number of lanes, and entrance and exit locations. Extreme weather parameters were then set, such as visibility set to less than 100 meters during heavy rain and humidity set to greater than 95% during fog. During the simulation, the system generated an enhanced dataset containing information on unusual vehicle lane changes and sudden changes in following distance. This data is difficult to obtain in practice due to the unpredictability and difficulty of collecting data in extreme weather conditions. The enhanced dataset generated by the digital twin system expanded the training sample size to three times the original data, enriching the model's training data and enabling it to learn traffic behavior patterns and patterns in a wider range of scenarios. This helps improve the model's generalization capabilities, enabling it to make more accurate predictions and decisions in a variety of complex real-world traffic conditions.

[0039] A cache quantization strategy is implemented during the inference phase. When the key-value cache usage exceeds the device memory limit, a product quantization algorithm is used to compress and store the query key vectors in the attention mechanism. This algorithm divides the 768-dimensional query key vector into eight subspaces, with each subspace having 256 cluster centers. This method compresses high-dimensional vectors into low-dimensional representations, reducing storage space to 1 / 4 of their original size. To compensate for precision loss, residual quantization is used to calculate and store the residuals of the quantized vectors, ensuring data accuracy to a certain extent. This allows for effective data compression and storage during inference when the key-value cache usage is excessive, avoiding system crashes due to insufficient memory, ensuring the normal operation of the model, and improving its efficiency and stability.

[0040] A multimodal output interface is designed, using 8-bit integer quantization to accelerate inference for real-time control commands while retaining 32-bit floating-point precision for post-event analysis reports. The multimodal output interface utilizes an acceleration engine, a high-performance deep learning inference optimization engine. 8-bit integer quantization is used for real-time control commands, such as dynamic speed limits and lane-changing instructions. During the quantization process, 1,000 sets of typical scenario data are used as a calibration set to calibrate the quantization parameters. This ensures that the quantized model maintains high accuracy while significantly improving inference speed. After quantization, inference latency is reduced from 120ms to 35ms, significantly improving the response speed of real-time control commands, enabling more timely traffic control, and effectively enhancing the real-time and accuracy of traffic management. For post-event analysis reports, 32-bit floating-point precision is retained because high-precision data is required to ensure the accuracy and reliability of analysis results during post-event processing, such as accident tracing and traffic flow analysis. This approach meets the speed requirements of real-time control while ensuring the accuracy requirements of post-event analysis.

[0041] The specific step S4 includes the following sub-steps: During the dynamic speed limit update phase, the speed limit adjustment range for each road section is calculated based on the gradient change rate of the risk heat map, and graded speed limit recommendations are sent to vehicles via the vehicle wireless broadcast protocol; During the variable lane induction phase, a dynamic lane function conversion plan is generated based on the real-time traffic efficiency prediction model, and laser projection equipment is used to display variable lane linear markings on the road surface; During the emergency rescue dispatch phase, vehicle trajectory prediction and road network topology are combined to plan the minimum time path for rescue vehicles and dynamically adjust the signal priority along the route; The Nash equilibrium algorithm is used to optimize the weights of multi-objective decisions, balance the game relationship between traffic efficiency, safety risks and energy consumption, and generate a Pareto optimal control strategy; After the instruction is executed, the reverse verification mechanism is activated, and by comparing the deviation between the actual traffic flow parameters and the predicted values, online correction of control parameters or event backtracking analysis is triggered.

[0042] Specifically, during the dynamic speed limit update phase, the speed limit adjustment range for each road section is calculated based on the gradient change rate of the risk heat map. The risk heat map is a visualization tool that directly displays the degree of traffic risk. By analyzing it, the gradient change rate of adjacent 500m sections is calculated. When the gradient exceeds 0.15 / m, it indicates a significant change in traffic risk on that section, potentially leading to increased congestion and increased accident risks. This triggers a speed limit reduction with a set step size of 10km / h. Lowering the speed limit effectively reduces vehicle speeds, minimizing conflicts and the probability of accidents. Speed ​​limit commands are transmitted to vehicles via a vehicle-to-vehicle wireless broadcast protocol, and the on-board unit uses an elliptic curve digital signature algorithm to verify the legitimacy of the command signature. This communication and verification method ensures that speed limit commands are accurately and securely transmitted to vehicles, ensuring that vehicles adjust their speeds in accordance with regulations and ensuring road safety.

[0043] During the variable lane guidance phase, a real-time traffic efficiency prediction model is used to generate dynamic lane function conversion plans. This model utilizes a long short-term memory (LSTM) network, a deep learning model specifically designed for processing time series data. LSTM is capable of effectively learning and memorizing historical patterns in traffic flow. By learning and analyzing historical traffic flow data, the LSTM model can predict lane-level traffic flow conditions for the next five minutes. When a lane's congestion index (the ratio of queue length to capacity) exceeds 0.7, indicating congestion in that lane, the laser projection device is activated. The laser projection device, with a wavelength of 650nm, projects a green arrow on the road surface to indicate the lanes vehicles can travel. The optical system, using servo motors, achieves a pointing accuracy of ±0.1°, ensuring accurate projection of the arrow on the road surface and guiding vehicles to the correct route. This variable lane guidance method dynamically adjusts lane function based on real-time traffic flow conditions, improving lane utilization efficiency, alleviating traffic congestion, and optimizing road resource allocation.

[0044] During the emergency rescue dispatch phase, vehicle trajectory predictions and road network topology are combined to plan the minimum-time path for rescue vehicles. The A algorithm is used to integrate real-time vehicle trajectory information (updated at 1Hz) to calculate the rescue path. The A algorithm is an efficient path search algorithm that incorporates heuristic functions and can quickly find the optimal path from the starting point to the end point in complex road networks. When calculating the rescue path, real-time vehicle trajectory information is taken into account, allowing for the avoidance of congested roads and ensuring that rescue vehicles reach the accident scene as quickly as possible. Simultaneously, priority signals are transmitted via optical fiber to signal controllers along the route, extending the green light phase to a minimum of 8 seconds. This ensures that rescue vehicles are given priority at intersections while driving, reducing rescue time, improving emergency rescue efficiency, and maximizing the safety of life and property.

[0045] The Nash equilibrium algorithm optimizes multi-objective decision-making weights to balance the trade-offs between traffic efficiency, safety risk, and energy consumption. This Nash equilibrium algorithm constructs a three-party game model, in which traffic efficiency is measured by maximizing vehicle-kilometers, safety risk is measured by minimizing the number of conflict points, and energy consumption is measured by optimizing fuel efficiency. The alternating direction method of multipliers (ADMM) optimizer solves the Pareto front solution set to find the optimal balance between these three objectives. This approach avoids the problem of pursuing a single objective while ignoring other aspects. It comprehensively considers multiple performance indicators of the transportation system, generates a more reasonable control strategy, achieves overall optimization of the transportation system, and improves the scientific and rational nature of traffic management.

[0046] After the command is executed, a reverse verification mechanism is initiated to evaluate the control effect by comparing the deviation between actual traffic flow parameters and predicted values. The mean absolute error (MAE) between actual traffic flow parameters (such as travel time and headway) and predicted values ​​is calculated. When the MAE exceeds 15%, it indicates a significant deviation between the actual traffic conditions and the predicted results. At this point, the online correction module based on PID control is triggered. PID control is a commonly used feedback control algorithm that adjusts control parameters such as the speed limit weight coefficient by performing proportional (P), integral (I), and differential (D) operations on the error. In this way, the control strategy can be adjusted in real time based on actual traffic conditions, making traffic control more precise, ensuring the stable operation of the traffic system, and improving the adaptability and effectiveness of traffic management.

[0047] Specifically, step S5 involves the federated learning framework using a federated averaging algorithm, with each edge node uploading model parameter increments every 24 hours. To reduce data transmission and protect data privacy, the gradient tensor is compressed using Top-k sparsification, retaining only the top 10% of significant values. The central aggregation server uses Paillier homomorphic encryption to protect parameter privacy. The parameter increments uploaded by each edge node are aggregated in an encrypted state to generate a global model version number, encoded in the format of year, month, and day. This approach ensures that the model can continuously absorb data from each edge node for optimization, achieving continuous model improvement, while also protecting data privacy and mitigating the risk of data leakage. This allows collaborative training of a globally optimized model even when different edge nodes possess large amounts of data but cannot directly share it. The self-optimization report includes two parts: the model iteration path and the road network health assessment. The model iteration path is visualized using the t-distributed stochastic neighbor embedding (t-SNE) algorithm. t-SNE is a commonly used dimensionality reduction visualization technique that maps high-dimensional model parameter space to a low-dimensional space.

[0048] In some embodiments, the highway data fusion management method may further include the following steps: Construct a knowledge graph of highway events, integrating historical accident cases, traffic control rules, and emergency response plans, and provide explainable support for real-time decision-making through retrieval-enhanced generation technology; Deploy privacy computing modules on roadside edge devices to homomorphically encrypt vehicle trajectory data and record data usage authorization credentials through blockchain. When a new traffic event pattern is detected, the active learning mechanism is activated to collect sample data and trigger incremental model training. The updated model parameters are synchronized to all road network nodes through a secure multi-party computing protocol.

[0049] Specifically, when constructing a knowledge graph for highway events, a large number of historical accident cases, traffic control rules, and emergency response plans are collected. From these historical accident cases, key information such as the time, location, type, and vehicles involved is extracted. Traffic control rules are used to identify control measures and applicable conditions for different scenarios. Emergency response plans clearly define the steps to be taken and resource allocation methods in various emergency situations. This information is integrated and, using knowledge graph technology, a network structure is constructed that encompasses various entities (such as accidents, vehicles, road sections, and control measures) and the relationships between them (e.g., an accident occurred on a certain road section, a certain control measure was used to address a certain type of accident, etc.).

[0050] Through retrieval-enhanced generation technology, when faced with real-time decision-making, the system can quickly retrieve relevant information from the knowledge graph and, combined with current traffic conditions, provide explainable support for decision-making. For example, when handling a traffic accident, the system can obtain experience in handling similar accidents, relevant traffic control rules, and applicable emergency response plans from the knowledge graph to help decision-makers understand the basis for the decision and the possible consequences. A privacy computing module is deployed on roadside edge devices, and homomorphic encryption technology is used to encrypt vehicle trajectory data to ensure data security during calculation and transmission. At the same time, blockchain technology is used to record data use authorization credentials to ensure the legitimacy and traceability of data use, thereby achieving secure data sharing and effective utilization while protecting user privacy.

[0051] When the system detects a new traffic event pattern, the active learning mechanism is triggered. This mechanism first initiates sample data collection, determining the scope, type, and frequency of data collection based on the characteristics and requirements of the event. For example, for the emerging autonomous vehicle platooning pattern, detailed data such as vehicle trajectories, speeds, spacing, and platoon structure may need to be collected. Once sufficient sample data is collected, it is used to trigger incremental training of the model. Incremental training allows the model to learn the characteristics and patterns of new event patterns without changing its existing knowledge, improving its adaptability to new situations.

[0052] Updated model parameters need to be synchronized across all nodes in the road network to ensure consistency and accuracy across the entire highway data fusion management system. This parameter synchronization is accomplished using a secure multi-party computing protocol, which enables collaborative computing and data exchange among multiple participants without disclosing their private data. This ensures that each node has timely access to updated model parameters, enabling accurate analysis and decision-making regarding new traffic event patterns. This enables rapid response and effective handling of new traffic events across the entire road network, continuously improving the intelligence and adaptability of the entire traffic management system.

[0053] In some embodiments, the highway data fusion management method may further include the following steps: Extract entity-relationship-attribute triples from traffic accident reports, maintenance records, and traffic regulations to build an ontology library that includes road topology, event causal chains, and disposal constraints; A graph attention network is used to dynamically embed the knowledge graph to capture high-order relationships between event types, road section characteristics, and disposal measures. In the real-time decision-making process, similar historical cases are retrieved through subgraph similarity matching to generate a decision explanation report that includes the effectiveness evaluation of the disposal suggestion and the risk transmission path; Build a rule distillation channel to convert traffic management regulations into constraints of the model loss function to ensure that decisions comply with road traffic safety regulations.

[0054] Specifically, when constructing the ontology, named entity recognition tools based on pre-trained language models (such as BERT) are used to accurately extract entities such as accident type, responsible party, and location from traffic accident report text. Key information such as construction sections and closure durations is retrieved through query statements from maintenance records stored in the database. Traffic regulations are parsed to identify relevant entities such as speed limits. After identifying the entities, semantic analysis and rule matching are used to determine relationships between them, such as the "action" and "accident association" relationships within an accident, and to define attributes such as accident severity. These are combined into triples to construct an ontology that encompasses road topology, event causal chains, and resolution constraints.

[0055] The Graph Attention Network (GAT) dynamically embeds the knowledge graph. It uses a multi-head attention mechanism to assign feature vectors to each node, aggregate neighbor features based on node connectivity, and utilizes gated recurrent units to memorize historical states, thereby capturing high-order correlations. During real-time decision-making, the SimRank algorithm performs subgraph similarity matching, retrieves historical similar cases, and generates a decision explanation report that includes an assessment of the effectiveness of the proposed action and the risk transmission path. Furthermore, a rule distillation pipeline is constructed to transform traffic management regulations into constraints within the model's loss function, ensuring legal and compliant decisions. This series of operations integrates information from multiple sources, providing knowledge and decision-making basis for traffic management, and enhancing the scientific and effective nature of management.

[0056] In some embodiments, the highway data fusion management method may further include the following steps: Build an augmented reality management interface to overlay the traffic flow density field, risk diffusion trend, and control instruction impact range on the 3D road network model; When the system detects a major safety hazard, it automatically switches to emergency control mode, initiates coordinated warning for the entire road network, and generates a risk avoidance route guidance plan.

[0057] Specifically, when constructing the augmented reality (AR) management interface, 3D modeling technology was used to create a high-precision highway network model, recreating the actual road geography. This model was then overlaid with traffic flow density fields, risk diffusion trends, and the impact range of control directives. The traffic flow density field was calculated based on data collected by induction coils and cameras, displaying density using color shading. Risk diffusion trends were visualized using a particle system using a predictive algorithm combined with real-time traffic event simulations. The impact range of control directives was determined and identified using a computational model. When a major safety hazard, such as a serious accident or severe weather conditions, is detected, the system automatically switches to emergency control mode, activating a coordinated warning mechanism across the entire road network. Warning messages containing hazard details were sent to vehicles and management departments via multiple devices. Simultaneously, a path planning algorithm, combined with real-time road conditions and emergency exit locations, planned optimal avoidance routes for affected vehicles and displayed within the AR interface. The AR management interface provides traffic management personnel with a visual understanding of road conditions. The warnings and route guidance provided in the emergency control mode enhance emergency response capabilities, ensure safety and road traffic flow, and improve the efficiency and safety of traffic management.

[0058] In some embodiments, the highway data fusion management method may further include the following steps: Based on the deep reinforcement learning algorithm, the optimal set of evacuation paths is calculated within a preset time, taking into account the vehicle distribution density, emergency exit locations, and accessibility of rescue resources; The enhanced display interface uses thermal gradients to indicate the risk diffusion rate and impact area, and dynamically displays the priority of each lane. Through communication between vehicles and infrastructure, graded warning information is sent to affected vehicles, including guidance on evasive paths, recommended speeds, and lane change timing.

[0059] Specifically, the system calculates the optimal set of hazard-avoiding paths based on a deep reinforcement learning algorithm. The state space, action space, and reward function are first defined. The state space encompasses information such as vehicle density, emergency exit locations, and accessibility to rescue resources. The action space uses parameterized Bezier curves to represent the sequence of path points. The reward function comprehensively considers travel time, safety distance, and fuel consumption, assigning weights to each factor to balance multiple objectives. Using the Proximal Policy Optimization (PPO) algorithm for training, the intelligent agent interacts with the environment to learn the optimal policy. After extensive training, the optimal set of hazard-avoiding paths can be calculated within a preset timeframe. The enhanced display interface uses thermal gradients to indicate the risk diffusion rate and impact area, dynamically displaying lane priority. Vehicle-to-Infrastructure (V2I) communication sends graded warning information to affected vehicles, including hazard-avoiding path guidance, recommended speed, and lane change timing. These actions provide accurate guidance to drivers in dangerous situations, optimize path planning, interface display, and information transmission, improve drivers' emergency response capabilities, reduce the probability of accidents, and ensure safe and smooth traffic on highways.

[0060] In some embodiments, the highway data fusion management method may further include the following steps: Establish a road network health assessment model that integrates pavement damage index, equipment failure rate, and traffic flow stability indicators to calculate road health scores; When the health score falls below the preset threshold, a maintenance operation plan is automatically generated and an appointment window is scheduled with the maintenance vehicle dispatch system; Embed a digital twin verification module in the maintenance plan to simulate the impact of different maintenance strategies on traffic efficiency and optimize construction timing.

[0061] Specifically, when building a road network health assessment model, a road health score is calculated by integrating the pavement damage index, equipment failure rate, and traffic flow stability indicators. The pavement damage index is analyzed using images acquired from inspection vehicles using the deep learning-based MobileNetV3 model. Equipment failure rates are calculated by collecting device operating status codes via the Simple Network Management Protocol (SNMP). Traffic flow stability is measured using the coefficient of variation of cross-sectional flow. These three indicators are weighted (e.g., 0.4 for the pavement damage index, 0.3 for the equipment failure rate, and 0.3 for the traffic flow stability index) to calculate a comprehensive score. When the health score falls below a preset threshold, a maintenance work plan is automatically generated and linked to the maintenance vehicle dispatch system to schedule a work window. The maintenance plan generation module formulates a plan based on road conditions and maintenance resources, while the dispatch system ensures that vehicles are on schedule. A digital twin verification module is embedded in the maintenance plan to simulate the impact of different maintenance strategies on traffic efficiency. Micro-simulation is used to model each vehicle, and after analyzing the results, the optimal strategy is selected to optimize the construction schedule. This series of operations enables timely identification of problems, rationally schedules maintenance, minimizes traffic impacts, extends highway life, and provides users with a better travel experience.

[0062] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A highway data fusion management method, characterized in that: The following steps are involved: Real-time collection of highway-wide operational data, including vehicle trajectory point clouds, roadside unit communication logs, meteorological sensor readings, road surface status images, and emergency alarm signals, is achieved through a multi-source heterogeneous perception network comprised of millimeter-wave radar arrays, lidar, smart cameras, and vehicle-based wireless communication equipment. Performing spatiotemporal alignment and semantic annotation on the operating data to construct a fused dataset containing spatiotemporal stamp codes, a data quality assessment matrix, and abnormal event markers, and extracting a subset of dynamic features that are strongly correlated with traffic trends through an adaptive feature selection algorithm; The feature subset is input into a lightweight spatiotemporal fusion model, and a road section-level risk heat map and lane-level traffic efficiency optimization plan are generated based on a multi-head spatiotemporal attention mechanism, forming a multi-objective decision set including variable speed limit strategy, dynamic lane allocation and emergency resource scheduling. Executing a traffic control instruction sequence within a preset spatiotemporal window based on the multi-objective decision set, the instruction sequence includes three stages: dynamic speed limit sign updating, variable lane alignment guidance, and right-of-way allocation for rescue vehicles, and monitoring traffic flow parameters and safety indicators in real time after the instruction execution; Based on the deviation between the execution feedback data and the expected optimization target, the parameter weights of the lightweight spatiotemporal fusion model are updated through the federated learning framework, and a self-optimization report including the model iteration path and road network health assessment is generated.

2. The highway data fusion management method according to claim 1 is characterized in that: The steps of performing spatiotemporal alignment and semantic annotation processing on the operating data, constructing a fusion data set including spatiotemporal stamp coding, a data quality assessment matrix, and abnormal event markers, and extracting a dynamic feature subset strongly related to the traffic situation through an adaptive feature selection algorithm are specifically as follows: Build a multi-source data conflict resolution model. When different sensors have conflicting detection results for the same target, dynamically assign confidence weights based on the device's historical accuracy and real-time signal strength. A sliding window mechanism is used to perform Kalman filtering on vehicle trajectory data to eliminate trajectory jitter noise caused by GPS drift. The contribution of each feature dimension to traffic situation prediction is calculated using the information entropy weight method, and features with a contribution higher than a preset threshold are retained to construct an orthogonal feature space. Perform adversarial enhancement on the filtered feature subset, injecting Gaussian noise and random occlusion to improve model robustness; When it is detected that the multicollinearity between feature dimensions exceeds the preset threshold, the principal component analysis dimensionality reduction operation is triggered and the variance explanation rate above the preset probability is retained.

3. A highway data fusion management method according to claim 2, characterized in that: The feature subset is input into a lightweight spatiotemporal fusion model. Based on the multi-head spatiotemporal attention mechanism, a road section-level risk heat map and a lane-level traffic efficiency optimization plan are generated. The steps of forming a multi-objective decision set including a variable speed limit strategy, dynamic lane allocation, and emergency resource scheduling are as follows: Using knowledge distillation technology, a pre-trained traffic model with tens of billions of parameters is compressed into a lightweight version suitable for roadside edge computing units, maintaining sensitivity to emergencies and decision-making accuracy. A dynamic structured pruning module is embedded in the model compression process to automatically adjust the model depth and attention head distribution based on real-time computing resources and task urgency; Build a digital twin system for highways, simulate traffic flow evolution under extreme weather conditions such as heavy rain and fog through the twin, and generate enhanced training data to improve model generalization capabilities; Implement a cache quantization strategy during the inference phase. When the key-value cache occupancy exceeds the device memory limit, a product quantization algorithm is used to compress and store the query key vector in the attention mechanism. Design a multimodal output interface, use 8-bit integer quantization to accelerate inference for real-time control instructions, and retain 32-bit floating-point calculation accuracy for post-analysis reports.

4. A highway data fusion management method according to claim 3, characterized in that: According to the multi-objective optimization decision set, a traffic control instruction sequence is executed within a preset time and space window according to the multi-objective decision set, wherein the instruction sequence includes three stages: dynamic speed limit sign update, variable lane alignment guidance, and rescue vehicle priority right-of-way allocation, and the traffic flow parameters and safety indicators after the instruction execution are monitored in real time, specifically: During the dynamic speed limit update phase, the speed limit adjustment range for each road section is calculated based on the gradient change rate of the risk heat map, and graded speed limit recommendations are sent to vehicles via the vehicle wireless broadcast protocol; During the variable lane induction phase, a dynamic lane function conversion plan is generated based on the real-time traffic efficiency prediction model, and laser projection equipment is used to display variable lane linear markings on the road surface; During the emergency rescue dispatch phase, the system combines vehicle trajectory prediction with road network topology to plan the minimum time path for rescue vehicles and dynamically adjust signal priorities along the route. The Nash equilibrium algorithm is used to optimize the weights of multi-objective decisions, balance the game relationship between traffic efficiency, safety risks and energy consumption, and generate a Pareto optimal control strategy; After the instruction is executed, the reverse verification mechanism is activated, and by comparing the deviation between the actual traffic flow parameters and the predicted values, online correction of control parameters or event backtracking analysis is triggered.

5. The highway data fusion management method according to claim 1, characterized in that: The method further comprises: Construct a knowledge graph of highway events, integrating historical accident cases, traffic control rules, and emergency response plans, and provide explainable support for real-time decision-making through retrieval-enhanced generation technology; Deploy privacy computing modules on roadside edge devices to homomorphically encrypt vehicle trajectory data and record data usage authorization credentials through blockchain. When a new traffic event pattern is detected, the active learning mechanism is activated to collect sample data and trigger incremental model training. The updated model parameters are synchronized to all road network nodes through a secure multi-party computing protocol.

6. A highway data fusion management method according to claim 5, characterized in that: The steps to build a highway event knowledge graph, integrate historical accident cases, traffic control rules, and emergency response plans, and provide explainable support for real-time decision-making through retrieval-enhanced generation technology are as follows: Extract entity-relationship-attribute triples from traffic accident reports, maintenance records, and traffic regulations to build an ontology library that includes road topology, event causal chains, and disposal constraints; A graph attention network is used to dynamically embed the knowledge graph to capture high-order relationships between event types, road section characteristics, and disposal measures. In the real-time decision-making process, similar historical cases are retrieved through subgraph similarity matching to generate a decision explanation report that includes the effectiveness evaluation of the disposal suggestion and the risk transmission path; Build a rule distillation channel to convert traffic management regulations into constraints of the model loss function to ensure that decisions comply with road traffic safety regulations.

7. The highway data fusion management method according to claim 1, characterized in that: The method further comprises: Build an augmented reality management interface to overlay the traffic flow density field, risk diffusion trend, and control instruction impact range on the 3D road network model; When the system detects a major safety hazard, it automatically switches to emergency control mode, initiates coordinated warning for the entire road network, and generates a risk avoidance route guidance plan.

8. A highway data fusion management method according to claim 7, characterized in that: When the system detects a major safety hazard, it automatically switches to emergency control mode, initiates coordinated warnings for the entire road network, and generates a risk avoidance route guidance plan. The specific steps are as follows: Based on the deep reinforcement learning algorithm, the optimal set of evacuation paths is calculated within a preset time, taking into account the vehicle distribution density, emergency exit locations, and accessibility of rescue resources. The enhanced display interface uses thermal gradients to indicate the risk diffusion rate and impact area, and dynamically displays the priority of each lane. Through communication between vehicles and infrastructure, graded warning information is sent to affected vehicles, including guidance on evasive paths, recommended speeds, and lane change timing.

9. The highway data fusion management method according to claim 1, further comprising: Establish a road network health assessment model that integrates pavement damage index, equipment failure rate, and traffic flow stability indicators to calculate road health scores; When the health score falls below the preset threshold, a maintenance operation plan is automatically generated and an appointment window is scheduled with the maintenance vehicle dispatch system; Embed a digital twin verification module in the maintenance plan to simulate the impact of different maintenance strategies on traffic efficiency and optimize construction timing.

Images