A traffic travel feature mining system based on multi-source data fusion

Abstract

The application relates to the technical field of big data mining and data processing, and particularly discloses a traffic travel feature mining system based on multi-source data fusion, which constructs a closed-loop feedback framework of data fusion-feature mining-virtual simulation verification-strategy optimization, realizes dynamic extraction of traffic travel features, effect verification in a high-fidelity virtual simulation environment, self-adaptive calibration of model parameters, and automatic generation of management strategies. The application constructs a virtual twin environment which can interact with a real traffic system, significantly improves the accuracy, timeliness and decision support value for actual traffic management of the mined features.

Description

Technical Field

[0001] This invention belongs to the field of big data mining and data processing technology, specifically relating to a traffic travel feature mining system based on multi-source data fusion. Background Technology

[0002] In the fields of smart cities and intelligent transportation systems, data-driven decision support has become a core means to improve urban operational efficiency and management. The main direction of current technological development is to integrate massive amounts of data generated by various sensors, mobile terminals, and business systems to perceive, analyze, and predict urban traffic conditions.

[0003] Traffic feature mining, as an important branch of intelligent transportation systems, aims to extract valuable information from multi-source, heterogeneous traffic data, such as travel patterns, spatiotemporal distribution patterns, and traffic demand hotspots. This technology provides crucial data support for traffic planning, road network optimization, public transportation scheduling, and travel services through in-depth analysis of these features.

[0004] Existing technologies typically rely on offline analysis of historical data or real-time computation based on pre-defined models to mine travel features. However, existing methods face significant challenges: the mining process is disconnected from the dynamic feedback mechanisms of real traffic systems, and there is a lack of a closed-loop mechanism for effectiveness verification in high-fidelity virtual microscopic traffic simulation environments. This makes it difficult to effectively verify the extracted features in complex real-world traffic environments. Furthermore, due to the lack of closed-loop interaction with real-time traffic conditions and control strategies, the feature models cannot be dynamically calibrated and iteratively optimized based on system performance. This lack of closed-loop interaction between the virtual verification environment and real-time control strategies significantly reduces the accuracy, timeliness, and guiding value of feature mining results for practical traffic management, becoming a technical bottleneck restricting the development of intelligent transportation systems towards refinement and adaptability. Summary of the Invention

[0005] The purpose of this invention is to provide a traffic feature mining system based on multi-source data fusion, so as to solve the problem that the existing technology lacks a closed-loop verification and optimization mechanism that combines virtual and real data in the traffic feature mining process, resulting in insufficient feature accuracy, timeliness and practical guidance value.

[0006] The technical solution of this invention is a traffic travel feature mining system based on multi-source data fusion. This system realizes dynamic extraction of travel features, real-time verification in a virtual traffic environment, and continuous optimization by constructing a closed-loop feedback architecture that includes a data fusion layer, a feature mining layer, a virtual simulation verification layer, and a strategy optimization layer.

[0007] The system includes a data fusion module, a feature mining module, a virtual simulation verification module, a strategy optimization module, and a closed-loop control module.

[0008] The data fusion module is used to access and process real-time and historical traffic data from multiple heterogeneous data sources. This module includes a data access submodule, a preprocessing submodule, and a spatiotemporal alignment submodule. The data access submodule synchronously receives real-time traffic data from fixed traffic detectors, floating cars, mobile signaling, public transport card swipe records, and internet map services through a standardized application programming interface. The preprocessing submodule cleans the received raw data, specifically by removing outliers that significantly exceed physical thresholds, filling in short-term data gaps caused by signal loss, and resampling data from different sampling frequencies to unify the time reference. The spatiotemporal alignment submodule uses a spatial matching algorithm based on a geographic information system to map traffic status data from different sources onto a unified road network topology and assigns a unified timestamp and spatial segment identifier to each data record.

[0009] The feature mining module connects to the data fusion module to extract multi-dimensional traffic and travel features based on the fused multi-source data. This module includes a pattern recognition submodule, a hotspot analysis submodule, and a demand prediction submodule. The pattern recognition submodule uses a density-based clustering algorithm to perform spatiotemporal clustering analysis on the preprocessed individual travel trajectory data to identify common typical travel patterns such as commuting, leisure, and business travel, and calculates the spatiotemporal distribution center, coverage area, and frequency of occurrence for each pattern.

[0010] The hotspot analysis submodule uses traffic flow, speed, and density data for each road segment within a specific time period. Applying a kernel density estimation algorithm, it dynamically identifies traffic congestion hotspots, accident-prone areas, and areas with high travel demand, and calculates the intensity index and spatiotemporal evolution trend of each hotspot. The demand forecasting submodule constructs a deep learning model integrating long short-term memory networks and attention mechanisms. This model takes historical travel demand sequences, real-time traffic conditions, weather information, and holiday factors as input, and outputs a travel demand matrix between different traffic zones for the next 1 to 24 hours.

[0011] The virtual simulation verification module, connected to the feature mining module, is used to construct a high-fidelity virtual micro-traffic simulation environment to verify the effectiveness of extracted travel features. This module includes a virtual simulation modeling submodule, a feature injection submodule, and an effectiveness evaluation submodule. The virtual simulation modeling submodule establishes a virtual micro-traffic simulation model that can reproduce real traffic flow characteristics based on the actual urban road network structure, traffic signal control scheme, and driver behavior parameters. The feature injection submodule dynamically loads the travel pattern distribution, demand hotspots, and predicted demand matrix output by the feature mining module as input parameters and initial conditions into the virtual micro-traffic simulation environment.

[0012] During the simulation, the effect evaluation submodule collects traffic flow data generated by the simulation system in real time, including average speed of road segments, travel time, queue length and intersection delay. It then compares and analyzes these simulation indicators with the actual observed traffic indicators for the corresponding time periods and road segments obtained through the data fusion module, and calculates the root mean square error, mean absolute percentage error and Pearson correlation coefficient between the two to quantitatively evaluate the fidelity and effectiveness of the extracted features in the virtual micro-traffic simulation environment.

[0013] The strategy optimization module connects to the virtual simulation verification module and generates optimized suggestions for traffic management strategies based on the verification results, driving the calibration of the feature model in reverse. This module includes a strategy generation submodule and a model calibration submodule. The strategy generation submodule receives the error analysis report output by the effect evaluation submodule. When the deviation between the simulation indicators and the actual observed indicators for a specific area exceeds a preset threshold, this submodule automatically generates traffic signal timing adjustment schemes, variable lane control schemes, or path guidance suggestion schemes for that area based on a preset optimization objective function. The model calibration submodule dynamically adjusts the model parameters within the feature mining module based on the virtual simulation verification results. Specifically, this submodule uses the difference between the simulation indicators and the actual observed indicators as part of the loss function, fine-tuning the weight parameters of the deep learning network in the demand prediction submodule or adjusting the distance threshold of the clustering algorithm in the pattern recognition submodule through backpropagation, so that the results of the next round of feature extraction can more accurately reflect the dynamic characteristics of the real traffic system.

[0014] The closed-loop control module, acting as the system's central scheduling unit, connects to the data fusion module, feature mining module, virtual simulation verification module, and strategy optimization module. It coordinates the execution timing and data flow of each module, enabling closed-loop operation and continuous optimization of the system across virtual and real data. This module triggers a complete closed-loop process at a preset fixed time period, such as every 15 minutes. Within each period, the closed-loop control module first instructs the data fusion module to collect and process the latest multi-source data; then, it instructs the feature mining module to perform feature extraction based on the latest data; next, it instructs the virtual simulation verification module to load the newly extracted features for simulation verification; finally, it instructs the strategy optimization module to generate strategy suggestions based on the verification results and trigger model calibration. The calibrated feature model will be used to process new data in the next period, thus forming a continuously iterative and self-optimizing closed loop.

[0015] Furthermore, the spatiotemporal alignment submodule in the data fusion module employs a map-matching process based on road network topology for its spatial matching algorithm. This process matches the latitude and longitude coordinates of floating cars or mobile signaling points to the most likely road segment sequence by calculating their projected distances to candidate road segments and using a hidden Markov model that considers road connectivity, thus ensuring an accurate correspondence between the travel trajectory and the road network structure.

[0016] Furthermore, the pattern recognition submodule within the feature mining module employs a noisy density-based clustering method. This method, by defining core objects, directly density-reachable concepts connected to density, can automatically discover clusters of arbitrary shapes from noisy trajectory point data, effectively identifying irregularly shaped travel pattern regions.

[0017] Furthermore, the virtual simulation modeling submodule in the virtual simulation verification module includes driver behavior parameters such as expected speed distribution, safe headway, lane change decision threshold, and response characteristics to traffic signals. These parameters are calibrated using historical actual traffic flow data to ensure that the simulated vehicle behavior can reflect the real habits of local drivers.

[0018] Furthermore, the strategy generation submodule in the strategy optimization module has a multi-objective weighted sum as its objective function. The specific objectives include minimizing the total travel time in the region, maximizing the capacity of key intersections, and balancing the traffic distribution of the road network. The weights of each objective can be dynamically configured according to the real-time control priorities of the management department.

[0019] Furthermore, the closed-loop control module also integrates a visual interactive interface, which displays the data fusion status, the currently extracted core travel features, the comparison results of simulation verification, strategy optimization suggestions, and the historical trend curve of model calibration in real time, providing system administrators with a global situational awareness and manual intervention entry point.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. This invention constructs a complete technical closed loop encompassing data fusion, feature mining, simulation verification, and strategy optimization, fundamentally changing the open-loop, offline operation mode of traditional feature mining technology. The system provides a high-fidelity testing ground for abstract feature data through a virtual simulation verification module, enabling the effectiveness of each feature extraction result to be verified in a near-realistic environment. This resolves the fundamental contradiction of feature models being detached from actual dynamic feedback for verification, significantly improving the reliability and credibility of the mined features.

[0022] 2. This invention, through the design of a strategy optimization module and a closed-loop control module, realizes an automatic feedback chain from feature verification to model calibration and strategy generation. Based on simulation verification results, the system can automatically diagnose deviations in the feature model and trigger parameter calibration, enabling the feature mining process to have continuous learning and adaptive optimization capabilities. Simultaneously, the targeted traffic management strategies generated based on the verification conclusions directly transform the value of feature mining into executable control commands, forming a complete data application chain of perception-analysis-decision-optimization, greatly enhancing the system's real-time support for actual traffic management.

[0023] 3. This invention employs a rigorous spatiotemporal alignment and preprocessing process at the multi-source data fusion level, ensuring the usability and consistency of heterogeneous data under a unified benchmark, thus laying a solid data foundation for subsequent accurate feature mining. At the feature extraction level, it comprehensively utilizes multiple algorithms such as clustering, hotspot analysis, and deep learning prediction to achieve a three-dimensional feature characterization from micro-level individual patterns to macro-level regional needs. The entire system architecture is clearly hierarchical, with well-defined and closely coordinated functions for each module. Through a periodic closed-loop operation mechanism, it can continuously track the dynamic changes of the transportation system, providing smart cities with a self-evolving, refined solution for transportation feature mining and decision support.

[0024] 4. This invention constructs a high-fidelity virtual microscopic traffic simulation environment, realizing digital twin modeling and virtual-real interaction verification of traffic systems. It provides a safe, controllable, and reproducible virtual experimental field for traffic feature mining and management strategy testing, which helps to promote the development of intelligent transportation systems towards digitalization, virtualization, and intelligence. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution architecture of the traffic travel feature mining system based on multi-source data fusion proposed in this invention;

[0026] Figure 2 This is a schematic diagram of the core principle framework of closed-loop feedback and self-optimization from multi-source data to management strategies in this invention;

[0027] Figure 3 This is a logical flowchart of the data fusion and feature mining stages in this invention;

[0028] Figure 4 This is a logical flow diagram of the simulation verification and strategy optimization stages in this invention;

[0029] Figure 5 This is a schematic diagram illustrating the multi-level interaction relationship between the closed-loop control module and the execution timing and data flow of each module in this invention. Detailed Implementation

[0030] Example 1: The overall technical architecture of the traffic travel feature mining system based on multi-source data fusion proposed in this invention is shown in the attached figure. Figures 1 to 5 As shown in the figure. This system constructs a virtual-real closed-loop feedback system consisting of a data fusion module, a feature mining module, a virtual simulation verification module, a strategy optimization module, and a closed-loop control module. This enables dynamic perception, accurate extraction, high-fidelity verification, and continuous optimization of urban traffic characteristics. The entire system's operational logic follows the appendix. Figure 2 The framework illustrates a closed-loop feedback and self-optimization principle from multi-source data input to management strategy output, and employs supplementary methods within each stage. Figure 3 With appendix Figure 4 The data fusion and feature mining logic processes, simulation verification and strategy optimization logic processes, are described in detail separately. The collaborative scheduling and timing control between the modules are handled by the appendix. Figure 5 The closed-loop control module shown completes the task in a unified and coordinated manner.

[0031] First, the data fusion module, as the system's first-level processing unit, undertakes the core task of transforming raw traffic information from multiple heterogeneous data sources into a structured, standardized, and spatiotemporally consistent high-quality dataset. This module comprises three functional sub-modules: a data access sub-module, a preprocessing sub-module, and a spatiotemporal alignment sub-module. The data access sub-module synchronously receives real-time and historical data streams from five main data sources through a set of predefined and securely certified standard application programming interfaces (APIs). These include, but are not limited to, cross-sectional flow and speed data uploaded by fixed traffic detectors, vehicle location and driving status sequences transmitted from floating car terminals, anonymized signaling location records provided by mobile communication operators, card transaction logs submitted by public transportation operators, and real-time traffic indexes and event information published by internet map service platforms. All accessed data carries original timestamps and spatial coordinates or road segment identifiers, but due to different sources, there are differences in format, sampling frequency, and semantic granularity.

[0032] The preprocessing submodule follows the data access submodule, performing a triple cleaning operation on the raw data. The first step is outlier removal, setting threshold boundaries based on physical constraints, such as a vehicle instantaneous speed not exceeding 120 km / h and an average road segment travel time not less than 0.8 times the free-flow time. Data points exceeding these reasonable ranges are marked as invalid and removed. The second step is missing value imputation. For short-term data gaps (usually no more than 5 minutes) caused by communication interruptions or equipment failures, linear interpolation or a weighted imputation strategy based on the historical average of adjacent time periods is used for repair. For long-term missing data, it is marked as an unusable interval to avoid introducing false information. The third step is time benchmark unification. Data streams with different sampling frequencies (e.g., fixed detectors report every 30 seconds, while floating cars report every 10 seconds) are resampled and unified to the system-defined benchmark time granularity (e.g., one time slice per minute) to ensure time alignment for subsequent analysis.

[0033] The spatiotemporal alignment submodule is a crucial component of the data fusion module. Its core task is to map preprocessed multi-source data onto a unified urban road network topology model, forming a fused dataset with strict spatiotemporal consistency. This submodule first loads standard road network vector data provided by the geographic information system platform. This data contains structured information such as all road centerlines, intersection nodes, number of lanes, speed limits, and directional attributes, and assigns a unique global identifier to each road segment. Subsequently, for floating car trajectory points or mobile signaling positioning points with latitude and longitude coordinates, a map-matching algorithm based on the road network topology is executed. This algorithm comprehensively considers the geometric projection distance of candidate road segments, road connectivity constraints, and the motion continuity between trajectory points, and uses a hidden Markov model to calculate the most probable path. Specifically, for each trajectory point sequence to be matched, the algorithm first filters out all candidate road segments within a 200-meter radius centered on it. Then, based on the matching result of the previous moment and the observation probability of the current point, it recursively calculates the joint probability of each candidate path. Finally, it selects the path with the highest probability as the matching result for the trajectory and assigns each trajectory point its corresponding road segment identifier and relative position along the road segment. For fixed detector data or bus card swipe records that already have road segment identifiers, spatial alignment is directly performed using the road segment identifiers. All aligned data are assigned a unified timestamp (accurate to the second) and spatial road segment identifier, forming a structured fusion data table for use by the feature mining module.

[0034] The feature mining module receives the fused dataset from the data fusion module and performs multi-dimensional, multi-level travel feature extraction tasks. This module consists of a pattern recognition submodule, a hotspot analysis submodule, and a demand prediction submodule, corresponding to three analysis levels: micro-level individual behavior, meso-level regional status, and macro-level demand trends, respectively.

[0035] The pattern recognition submodule focuses on identifying common typical travel patterns from individual travel trajectories. Its input is a set of spatiotemporally aligned individual trajectories, where each trajectory consists of a series of time-ordered road segment visits. This submodule employs a noisy density-based clustering method to perform spatiotemporal clustering of the trajectory data. The algorithm first treats each trajectory as an object in a high-dimensional feature space, whose feature vector is composed of elements such as origin road segment, destination road segment, departure time, travel duration, and key nodes passed through. The algorithm defines two core parameters: neighborhood radius. And the minimum number of neighborhood points MinPts. For any trajectory object p, if its If a neighborhood contains at least MinPts other trajectory objects, then p is called a core object. If object q is located within the core object p... Within a neighborhood, q is said to be directly density-reachable from p; if a series of core objects connect p and q, then p and q are said to be density-connected. All density-connected objects constitute a cluster. This method can effectively handle noise points in trajectory data (such as occasional detours or data errors) and automatically discover cluster structures of arbitrary shapes, thereby identifying typical travel patterns such as commuting (e.g., from residential area A to business area B during morning rush hour), leisure (e.g., from the city center to park C on a weekend afternoon), and business (e.g., shuttling between multiple points within Financial Street D on a weekday morning). For each identified pattern cluster, the system calculates its spatiotemporal distribution center (i.e., the weighted average position of the origin and destination of all trajectories within the cluster), coverage area (defined by a 95% confidence ellipse), and occurrence frequency (the number of times the pattern occurs per unit time), forming a pattern feature profile.

[0036] The hotspot analysis submodule focuses on dynamic hotspot identification at the road network level. Its input consists of traffic state indicators for each road segment within a continuous time slice in the fused dataset, including flow rate, average speed, and density. This submodule employs a kernel density estimation algorithm to construct a dynamic heat field on a two-dimensional road network plane. Specifically, for each time slice t, the system traverses all road segments i, using their geometric center as the kernel function center, and weighted by road segment length and traffic state intensity (e.g., congestion index = free-flow speed / actual speed), superimposing a Gaussian kernel function to generate a heat distribution surface for the entire road network. Subsequently, by setting a heat threshold (e.g., taking the 90th percentile of the heat values ​​for the entire road network), local maxima regions are extracted as hotspot candidates. Further combined with temporal continuity analysis, if a candidate region is identified as a hotspot in more than three consecutive time slices, it is confirmed as a valid hotspot and classified as a congestion hotspot, an accident-prone hotspot, or a high-demand hotspot (judged based on dominant indicators). The system simultaneously calculates the intensity index (i.e., peak heat value), spatial range (measured by the area enclosed by contour lines), and evolution trend (such as the drift speed of the hotspot center and the rate of intensity growth) of each hotspot, forming a hotspot feature map.

[0037] The demand forecasting submodule is responsible for quantitatively predicting travel demand for future periods. It constructs a deep learning model that integrates a long short-term memory network and an attention mechanism. The input layer of this model receives four types of features: a historical travel demand matrix (OD demand for the same period in the past 7 days), a real-time traffic state vector (average speed and congestion index for each zone), weather codes (uniquely hot-coded variables such as sunny, rainy, and snowy), and holiday markers (0 or 1). The main body of the model consists of two stacked layers of long short-term memory units, used to capture the long-term dependencies and short-term fluctuations in the demand sequence. In the decoding stage, an attention mechanism is introduced to dynamically weight the hidden states of historical time steps, enabling the model to focus on the period most influential on the current prediction. The output layer is a fully connected layer that generates a travel demand matrix for the next 1 to 24 hours, granularized in 15-minute increments. Indicates time From traffic zones To partition The model predicts travel volume. During the training phase, it uses mean squared error as the loss function and updates parameters via backpropagation with the Adam optimizer.

[0038] The virtual simulation verification module receives three types of feature outputs from the feature mining module—travel pattern distribution, hotspot area maps, and demand prediction matrices—and verifies their effectiveness in a high-fidelity virtual micro-traffic simulation environment (i.e., a traffic digital twin environment). This module includes a virtual simulation modeling submodule, a feature injection submodule, and an effectiveness evaluation submodule.

[0039] The virtual simulation modeling submodule constructs a high-fidelity virtual traffic world based on actual urban road network data. This model employs a two-layer car-following / lane-changing behavioral framework, with the car-following model utilizing an intelligent driver model. Its core formula is as follows:

[0040] ;

[0041] in, Let be the acceleration of the vehicle at time t. For maximum acceleration, At the current speed, For the desired speed, For acceleration index, This is the actual distance between the front of the vehicle. To account for the speed difference with the vehicle in front, The expression for the desired front-end spacing is:

[0042] ;

[0043] in, For minimum safe distance, For safe headway, For comfort deceleration. All driver behavior parameters (such as...) Mean and standard deviation The distribution of traffic flow and lane-changing decision thresholds are calibrated using historical actual traffic flow data. Genetic algorithms or Bayesian optimization methods are used to minimize the difference between simulated and measured traffic flow, ensuring that simulated vehicle behavior can truly reflect local driving habits.

[0044] The feature injection submodule transforms the feature mining results into input parameters for the virtual microscopic traffic simulation model. Specifically, the travel mode distribution is used to generate travel plans in the simulation: the system generates virtual travelers proportionally based on the occurrence frequency and OD distribution of each mode, and assigns them corresponding departure times, origins, destinations, and route preferences. The hotspot map is used to initialize event disturbances in the simulation: if a region is identified as a high-risk accident hotspot, low-speed obstacle vehicles are randomly injected into that region; if it is a congestion hotspot, demand intensity is increased at the upstream entrance. The demand forecast matrix directly serves as the simulation OD demand matrix, driving the simulation operation within a specified forecast period.

[0045] During the simulation, the performance evaluation submodule collects traffic flow indicators from the simulation output at 1-minute intervals, including average speed, travel time, queue length, and average intersection delay for each road segment. Simultaneously, it obtains actual observation indicators for the same time period and road segment from the data fusion module. The system performs a road segment-by-road segment and time slice-by-time comparative analysis of these two types of indicators, calculating three core evaluation indicators: root mean square error (RMSE), mean absolute percentage error (MASE), and Pearson correlation coefficient. RMSE measures the absolute deviation between the simulation and measured values; MASE reflects the relative error level; and Pearson correlation coefficient assesses the consistency of their trends. If two of the three indicators for a certain area exceed preset thresholds (e.g., RMSE > 15 km / h, MASE > 20%, Pearson correlation coefficient < 0.7), the feature extraction for that area is deemed to have a significant deviation, requiring the triggering of a calibration mechanism.

[0046] The strategy optimization module performs both strategy generation and model calibration tasks based on the evaluation results of the virtual simulation verification module. This module includes a strategy generation submodule and a model calibration submodule.

[0047] After receiving the deviation report from the effect evaluation submodule, the strategy generation submodule automatically generates traffic management optimization suggestions for areas where deviations exceed limits. Its core is a multi-objective optimization problem, with the objective function defined as:

[0048] ;

[0049] in, Total travel time within the region. To improve the capacity of key intersections, The coefficient of variation of road network traffic (measures traffic balance). , , The weights are configurable, with initial values ​​of 0.5, 0.3, and 0.2, and can be dynamically adjusted according to the real-time priorities of the management department. Optimization variables include signal timing schemes (cycle, green light ratio, phase difference), variable lane direction switching instructions, and path guidance information. The solver employs an improved particle swarm optimization algorithm to search for Pareto optimal solutions while satisfying signal safety constraints (such as minimum green light time and yellow light transition), and outputs the top three feasible solutions for decision-making reference.

[0050] The model calibration submodule is responsible for fine-tuning the parameters of the internal model of the feature mining module. Its mechanism is as follows: it adjusts the differences between simulation and measured metrics (such as velocity deviation). This serves as additional supervisory signal, backpropagated to the feature mining model. For the demand prediction submodule, the system will... A composite loss function is constructed by correlating it with the forecasting error of demand. ,in The adjustment coefficient is set to 0.1, and the weights of the Long Short-Term Memory network are updated using gradient descent. For the pattern recognition submodule, if a certain travel pattern causes excessive congestion on a specific road segment in simulation but does not occur in actual testing, it indicates that the coverage of the pattern is too large or the frequency of occurrence is too high. The system will automatically reduce its cluster radius. Alternatively, its frequency weights can be reduced. The calibrated model parameters are saved and used in the next round of feature mining.

[0051] The closed-loop control module, acting as the system's central nervous system, coordinates the orderly execution of the aforementioned four modules according to a preset 15-minute cycle. (See attached...) Figure 5As shown, this module maintains a state machine containing four main states: data acquisition, feature extraction, simulation verification, and strategy optimization. At the beginning of each cycle, the module first enters the data acquisition state, sending instructions to the data fusion module to process multi-source data within the latest 15-minute window. After data fusion is completed and a confirmation signal is returned, it switches to the feature extraction state, triggering the feature mining module to run. After feature mining is completed, it enters the simulation verification state, starting the virtual simulation verification module to load new features and run the simulation. After the simulation ends, it enters the strategy optimization state, performing strategy generation and model calibration. The calibrated model takes effect immediately and is used for data processing in the next cycle. In addition, the closed-loop control module also integrates a visual interactive interface, supporting administrators to conduct real-time monitoring and manual intervention in a virtual micro-traffic simulation environment, realizing collaborative decision-making between humans, machines, and the virtual and real environments. It also renders fused data heatmaps, Sankey diagrams of travel patterns, hotspot spatiotemporal evolution animations, simulation-measurement comparison curves, strategy suggestion cards, and historical line graphs of model parameter calibration in real time, supporting administrators to perform manual review, weight adjustment, or emergency intervention.

[0052] Through the aforementioned closed-loop mechanism, the system achieves a complete chain from data perception to strategy output, and possesses the ability to continuously self-correct and optimize. Each cycle not only produces current travel characteristics and management suggestions, but also continuously corrects the feature model through simulation verification feedback, making it increasingly closer to the dynamic laws of real traffic systems, thereby providing highly timely, accurate, and reliable decision support for smart city traffic management.

[0053] Example 2: Building upon Example 1, this example enhances the virtual simulation verification module, further improving its ability to simulate special traffic events and the realism of the virtual environment. Specifically, in addition to the basic driver behavior model, the virtual simulation modeling submodule adds a sudden event response submodel. This submodel is specifically designed to simulate the impact of non-steady-state events such as traffic accidents, road construction, and large-scale events on traffic flow. When the feature injection submodule detects high-intensity accident hotspots or abnormal demand hotspots output by the hotspot analysis submodule, the virtual simulation modeling submodule automatically activates this submodel.

[0054] The emergency response sub-model comprises three stages: event detection, impact range estimation, and driver rerouting. In the event detection stage, the system determines whether an event constitutes a valid event based on the hotspot intensity index and duration. If the congestion index of a road segment exceeds the threshold of 3.0 (free-flow speed / actual speed > 3) for two consecutive time slots, and the hotspot range remains stable, it is classified as a traffic accident event. In the impact range estimation stage, the model uses a diffusion equation based on the event location, road grade, and real-time traffic flow to estimate the upstream queue length and downstream speed decay range of the event. Specifically, the queue length... Estimated by the following formula:

[0055] ;

[0056] in, The inflow rate (vehicles / hour) of the road segment before the incident occurred. To estimate the clearance time (in hours). The congestion density is denoted as vehicles / km / lane. During the driver rerouting phase, simulated vehicles within the affected area dynamically replan their routes based on their path preferences and real-time road network conditions after detecting congestion ahead. The rerouting algorithm employs an improved A* algorithm, whose cost function includes not only distance and time but also a congestion penalty factor, which increases non-linearly with the road segment congestion index.

[0057] By introducing this enhanced model, the virtual simulation verification module can more realistically reproduce the traffic flow evolution process under emergencies, thereby providing a more rigorous test of the feature mining module's performance under abnormal conditions. For example, if the demand forecasting submodule fails to accurately predict the instantaneous surge in demand after a large concert, severe queue overflow will occur in the simulation, while actual test data may show that temporary traffic control alleviates congestion. In this case, the effect evaluation submodule will identify significant deviations, and the strategy optimization module will not only generate emergency signal timing schemes but also trigger the demand forecasting model to learn special event factors, such as explicitly adding large-scale event indicators and event scale levels to the input features, and performing transfer learning through historical similar event data to improve the robustness of future predictions for similar events.

[0058] Furthermore, this embodiment incorporates an event priority scheduling mechanism in the closed-loop control module. When a major emergency is detected, the system can temporarily shorten the closed-loop cycle to 5 minutes, prioritizing the processing of data and features in the event-related area to ensure rapid response of management strategies. Simultaneously, the visual interface highlights the event area and pushes early warning information to the traffic command center. This enhancement mechanism significantly improves the system's adaptability and decision-making value in abnormal traffic scenarios, further strengthening the practicality and resilience of the closed-loop feedback.

Claims

1. A traffic travel feature mining system based on multi-source data fusion, characterized in that, include: The data fusion module is used to access and process real-time and historical traffic data from multiple heterogeneous data sources; The feature mining module is connected to the data fusion module and is used to extract multi-dimensional traffic and travel features based on the fused multi-source data. The virtual simulation verification module, connected to the feature mining module, is used to build and run a high-fidelity virtual microscopic traffic simulation environment to form a digital twin of the traffic system and to verify the effectiveness of the extracted travel features. The strategy optimization module, connected to the virtual simulation verification module, is used to generate optimization suggestions for traffic management strategies based on the verification results and drive the calibration of the feature model in reverse. The closed-loop control module is connected to the data fusion module, the feature mining module, the virtual simulation verification module and the strategy optimization module respectively. It is used to coordinate the execution timing and data flow of each module to realize the closed-loop operation and continuous interaction of the system between real data and virtual micro-traffic simulation environment. The feature mining module includes a pattern recognition submodule, a hotspot analysis submodule, and a demand prediction submodule. The pattern recognition submodule uses a density-based clustering algorithm to perform spatiotemporal clustering analysis on the preprocessed individual travel trajectory data to identify typical travel patterns and calculate the spatiotemporal distribution center, coverage area and occurrence frequency of each pattern. The hotspot analysis submodule uses traffic flow, speed and density data of each road segment in the road network within a specific time period, and applies kernel density estimation algorithm to dynamically identify traffic congestion hotspots, accident-prone areas and areas with high travel demand, and calculates the intensity index and spatiotemporal evolution trend of each hotspot. The demand prediction submodule constructs a deep learning model that integrates long short-term memory network and attention mechanism. The model takes historical travel demand sequence, real-time traffic status, weather information and holiday factors as input, and outputs a travel demand matrix between different traffic zones in the next 1 hour to 24 hours. The virtual simulation verification module includes a virtual simulation modeling submodule, a feature injection submodule, and an effect evaluation submodule. The virtual simulation modeling submodule constructs a high-fidelity virtual microscopic traffic simulation model based on the actual urban road network structure, traffic signal control scheme, and driver behavior parameters, serving as a digital twin environment for the real traffic system. The feature injection submodule uses the travel pattern distribution, demand hotspots, and predicted demand matrix output by the feature mining module as input parameters and initial conditions for the virtual micro-traffic simulation model, and dynamically loads them into the virtual micro-traffic simulation environment. During the simulation operation, the effect evaluation submodule collects traffic flow data generated by the simulation system in real time, and compares and analyzes these simulation indicators with the actual observed traffic indicators of the corresponding time period and road segment obtained by the data fusion module. It calculates the root mean square error, mean absolute percentage error and Pearson correlation coefficient between the two to quantitatively evaluate the degree of restoration and effectiveness of the extracted features in the virtual micro traffic simulation environment. The strategy optimization module includes a strategy generation submodule and a model calibration submodule; The strategy generation submodule receives the error analysis report output by the effect evaluation submodule. When the deviation between the simulation index and the actual observation index of the region exceeds a preset threshold, it automatically generates a traffic signal timing adjustment scheme, variable lane control scheme or path guidance suggestion scheme for the region based on the preset optimization objective function. The model calibration submodule dynamically adjusts the model parameters within the feature mining module based on the virtual simulation verification results. Specifically, it uses the difference between the simulation indicators and the actual observed indicators as part of the loss function, and fine-tunes the weight parameters of the deep learning network in the demand prediction submodule or adjusts the distance threshold of the clustering algorithm in the pattern recognition submodule through the backpropagation algorithm.

2. The traffic travel feature mining system based on multi-source data fusion according to claim 1, characterized in that, The data fusion module includes a data access submodule, a preprocessing submodule, and a spatiotemporal alignment submodule. The data access submodule synchronously receives real-time traffic data from fixed traffic detectors, floating cars, mobile signaling, public transportation card swipe records, and Internet map services through a standardized application programming interface. The preprocessing submodule cleans the received raw data, including removing outliers that exceed physical thresholds, filling in short-term data gaps caused by signal loss, and resampling data at different sampling frequencies to unify the time base. The spatiotemporal alignment submodule uses a spatial matching algorithm based on geographic information systems to map traffic status data from different sources onto a unified road network topology, and assigns a unified timestamp and spatial segment identifier to each data record.

3. The traffic travel feature mining system based on multi-source data fusion according to claim 2, characterized in that, The closed-loop control module triggers a complete closed-loop process according to a preset fixed time period. In each period, the data fusion module is instructed to collect and process the latest multi-source data. Then, the feature mining module is instructed to perform feature extraction based on the latest data. Next, the virtual simulation verification module is instructed to load the newly extracted features for virtual simulation verification. Finally, the strategy optimization module is instructed to generate strategy suggestions based on the verification results and trigger model calibration.

4. The traffic travel feature mining system based on multi-source data fusion according to claim 3, characterized in that, The spatial matching algorithm of the spatiotemporal alignment submodule adopts a trajectory matching process based on road network topology. This process matches the sequence of latitude and longitude coordinates of floating cars or mobile signaling points to the most likely road segment sequence by calculating their projected distances to candidate road segments and using a hidden Markov model that considers road connectivity.

5. A traffic travel feature mining system based on multi-source data fusion according to claim 4, characterized in that, The density-based clustering algorithm used in the pattern recognition submodule is a noisy density-based clustering method. This method automatically discovers clusters of arbitrary shapes from noisy trajectory point data by defining core objects, direct density reachability and density-connected concepts.

6. A traffic travel feature mining system based on multi-source data fusion according to claim 5, characterized in that, The driver behavior parameters of the virtual simulation modeling submodule include expected speed distribution, safe headway, lane change decision threshold, and response characteristics to traffic signals. These parameters are calibrated using historical actual traffic flow data to ensure that vehicle behavior in the virtual micro-traffic simulation environment can realistically reflect driving characteristics in the real world.

7. A traffic travel feature mining system based on multi-source data fusion according to claim 6, characterized in that, The optimization objective function of the strategy generation submodule is a multi-objective weighted sum. The specific objectives include minimizing the total travel time in the region, maximizing the capacity of key intersections, and balancing the traffic distribution of the road network. The weights of each objective are dynamically configured according to the real-time control priority of the management department.

Images