A highway accident prediction system based on big data

Abstract

The application relates to the technical field of traffic safety prediction, in particular to a highway accident prediction system based on big data, which comprises the following modules: a multi-dimensional traffic flow perception module for generating a traffic flow state tensor, a dynamic graph mapping module for constructing a road network topology structure graph and generating a node feature vector, a risk potential calculation module for calculating a risk transmission weight matrix, and an accident probability deduction module for generating an accident prediction heat map by using time sequence evolution analysis. In the application, the adaptive graph attention mechanism is used to calculate the risk propagation weight between adjacent nodes and cross-region nodes, nonlinear decoupling and deep fusion analysis of multi-source risk factors such as environmental interference, road geometric shape and vehicle abnormal behavior are realized, the precursors of traffic accidents are accurately captured in a millisecond time window, and the accident evolution path is dynamically deduced, so that the technical problems of prediction lag and weak generalization ability of traditional linear models in complex and changeable traffic scenes are completely solved.

Description

Technical Field

[0001] This invention relates to the field of traffic safety prediction technology, and in particular to a highway accident prediction system based on big data. Background Technology

[0002] The field of traffic safety prediction technology encompasses a comprehensive range of technologies that utilize information technology to monitor, analyze, and assess potential risks in road traffic conditions. Traditional highway accident prediction systems typically employ induction coils or fixed-point cameras deployed along the roadside to collect traffic flow and speed data. They then use linear regression models or fixed-threshold algorithms to statistically analyze historical accident frequencies, perform simple physical attribution analysis based on the road segment's geometric parameters, and ultimately generate static risk level reports.

[0003] Existing technologies rely solely on linear extrapolation of historical statistical data from a single dimension, neglecting the strong nonlinear coupling between the physical properties of the road environment, real-time weather changes, and the microscopic driving trajectories of vehicles. Furthermore, they lack in-depth analysis of the risk transmission mechanism across regions under complex road network topologies. Consequently, when faced with sudden weather disturbances or instantaneous traffic flow pulses, the system cannot capture the dynamic evolution of risk factors in real time, leading to significant time lags and high false alarm rates in the prediction results. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a highway accident prediction system based on big data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a highway accident prediction system based on big data, comprising: The multi-dimensional traffic flow perception module acquires vehicle trajectory coordinates and road surface physical state data, extracts vehicle speed vectors and road friction coefficients, performs spatiotemporal alignment calculations of multi-source data based on unified timestamps and spatial coordinates, and fuses the vehicle speed vectors and road friction coefficients to construct a traffic flow state tensor. The dynamic graph mapping module divides roads into discrete road segment nodes based on road network geographic information, constructs a road network topology graph based on the physical connection relationship between road segments and the actual traffic flow guidance, and maps the traffic flow state tensor to the corresponding graph nodes to generate node feature vectors. The risk potential energy calculation module aggregates the node feature vectors of adjacent nodes using graph convolution operation logic, and quantitatively calculates the dynamic mutual influence intensity between nodes based on the spatial distance decay function and the traffic flow direction correlation between nodes, generating a risk transmission weight matrix. The accident probability extrapolation module inputs the risk transmission weight matrix into a time-varying recurrent neural network for deep time-series evolution analysis. It combines extreme value theory to set dynamic thresholds, calculates the mutation probability value of each discrete road segment node at the current time step, and generates an accident prediction heatmap.

[0006] As a further aspect of the present invention, the specific function of the multi-dimensional traffic flow sensing module is as follows: The heterogeneous spatiotemporal alignment submodule acquires the vehicle's driving trajectory coordinates and the road surface physical state data. Based on the atomic clock timing reference, it performs time reference unification processing on various heterogeneous data. Using a preset coordinate transformation matrix, it maps sensor data from different perspectives to a standard geographic coordinate system. It then performs the multi-source data spatiotemporal alignment operation to generate a spatiotemporal synchronization reference array. The parameter parsing and splicing submodule extracts the dynamic displacement sequence and static road surface parameters from the spatiotemporal synchronization reference array. Based on the continuous time frame displacement difference logic, it accurately calculates the vehicle velocity vector in three-dimensional space. Combined with the filtering and noise reduction algorithm, it smooths the static road surface parameters and purifies them to obtain the road surface friction coefficient. The vehicle velocity vector and the road surface friction coefficient are orthogonally concatenated to construct the traffic flow state tensor.

[0007] As a further aspect of the present invention, the specific function of the dynamic map mapping module is as follows: The topology node segmentation module obtains the road geometric features extracted from the road network geographic information, divides the continuous road space into discrete road segment nodes according to the preset distance tolerance and intersection coordinates, analyzes the physical connection relationship between the discrete road segment nodes and the actual traffic flow guidance, and constructs the road network topology structure map. The tensor feature fusion submodule obtains the road network topology diagram and the traffic flow state tensor, calculates the spatial boundary coordinates corresponding to each discrete road segment node, accurately associates and maps the data slices of the corresponding region in the traffic flow state tensor to the corresponding map nodes based on the spatial boundary coordinates, performs high-dimensional feature aggregation operation, and generates the node feature vector.

[0008] As a further aspect of the present invention, the specific function of the risk potential energy calculation module is as follows: The graph domain convolution aggregation submodule obtains the node feature vectors, configures a graph convolution kernel operator based on Chebyshev multinomials, aggregates the node feature vectors of adjacent nodes using the graph convolution operation logic, extracts spatial dependency features within the local road network, and outputs a neighborhood-aware feature matrix. The potential energy weight assessment submodule extracts the neighborhood perception feature matrix, quantifies the static spatial impedance between nodes based on the spatial distance attenuation function, quantitatively calculates the dynamic mutual influence intensity between nodes by combining the traffic flow direction correlation between nodes, and performs cross-validation and weighted summation on the static spatial impedance and the dynamic mutual influence intensity to generate the risk transmission weight matrix.

[0009] As a further aspect of the present invention, the specific function of the accident probability deduction module is as follows: The deep temporal inference submodule obtains the risk transmission weight matrix, inputs the risk transmission weight matrix into the network architecture composed of gated recurrent units to perform the deep temporal evolution analysis, extracts the cumulative evolution law of traffic risk in the time dimension, and outputs the temporal risk evolution feature vector. The extreme value heat map generation submodule analyzes the time-series risk evolution feature vector, fits the risk tail data with the extreme value theory of the generalized Pareto distribution, sets a dynamic threshold, calculates the mutation probability value of each discrete road segment node at the current time step, performs two-dimensional kernel density estimation and color mapping interpolation based on the spatial distribution coordinates of the mutation probability value, and generates the accident prediction heat map.

[0010] As a further aspect of the present invention, the process of performing the multi-source data spatiotemporal alignment operation specifically includes: The timestamp information of the header of the raw data packet transmitted by the heterogeneous sensors is obtained, the relative drift error between the timestamp information and the global standard clock of the network is calculated, the Kalman filter equation is used to smooth and compensate the relative drift error, and the accurate time tag under the unified clock reference is output. Extract the three-dimensional relative coordinates of various heterogeneous sensor data in the local sensor coordinate system, call the multi-view camera calibration parameters to establish a rotation and translation extrinsic parameter matrix, use the rotation and translation extrinsic parameter matrix to reproject the three-dimensional relative coordinates to the globally unified world geodetic coordinate system, fuse the precise time label, and generate the spatiotemporal synchronization reference array.

[0011] As a further aspect of the present invention, the process of constructing the road network topology diagram specifically includes: The road network geographic information contains the center lines of each road segment and lane connectivity data. The center lines of each road segment are traversed to identify intersections and divergence points in the road topology skeleton. Based on the intersections and divergence points, the continuously distributed road surface geometry model is truncated to generate multiple independent node entities. Extract the latitude and longitude coordinates of the boundaries at both ends of each node entity, evaluate the Euclidean distance and traffic accessibility index between node entities, and if the Euclidean distance is lower than the preset distance tolerance and the traffic accessibility index exhibits unidirectional or bidirectional conduction attributes, then establish connection edges between the corresponding node entities, combine all node entities and connection edges, and construct the road network topology diagram.

[0012] As a further aspect of the present invention, the process of quantitatively calculating the intensity of the dynamic mutual influence between nodes specifically includes: The spatial physical distance between nodes and the sequence of node transfers within a continuous time period are obtained. The spatial physical distance is then nonlinearly mapped and transformed according to the Gaussian decay kernel function to output the basic potential energy decay benchmark. Extract the node transfer train sequence, use Pearson correlation coefficient logic to calculate the dynamic coordinated volatility between the source node traffic trend and the target node traffic trend, substitute the basic potential energy decay benchmark and the dynamic coordinated volatility into the preset potential energy coupling formula for product operation, and obtain the dynamic mutual influence intensity. The specific formula for potential energy coupling is as follows: ; in, This represents the intensity of the dynamic interaction. Represents the system calibration coefficient. Represents the natural constant. Represents the spatial attenuation damping coefficient. This represents the spatial physical distance between nodes. This represents the dynamic co-variance volatility.

[0013] As a further aspect of the present invention, the process of setting the dynamic threshold specifically includes: Obtain the extreme value sample set of the time series risk evolution feature vector within the historical observation period, and use the maximum likelihood estimation method to approximate the distribution tail features of the extreme value sample set, and calculate the shape parameter and scale parameter of the generalized Pareto distribution function. Based on shape and scale parameters, a tail probability density function for the corresponding road segment region is constructed. A confidence boundary is set, and the inverse function quantile of the tail probability density function under the confidence boundary condition is solved. The inverse function quantile is then configured as a dynamic threshold under the current road network condition.

[0014] As a further aspect of the present invention, the process of performing the deep temporal evolution analysis specifically includes: Extract the historical discrete slices that are continuously distributed along the time axis from the risk transmission weight matrix, input the historical discrete slices into the update gate component and reset gate component inside the gating network, and use the reset gate component to control the fusion ratio of historical state information and current input data to generate candidate hidden states; The update gate component is used to evaluate the contribution weight of the candidate hidden state to the long-term traffic evolution trend. Based on the contribution weight, selective forgetting and information overwriting operations are performed on the historical memory unit. The comprehensive evolution tensor is output through multi-level iterative superposition. The comprehensive evolution tensor is then subjected to dimensionality reduction and flattening to output the temporal risk evolution feature vector.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by utilizing an adaptive graph attention mechanism to calculate the risk propagation weights between adjacent nodes and cross-regional nodes, nonlinear decoupling and deep fusion analysis of multi-source risk factors such as environmental interference, road geometry, and abnormal vehicle behavior are achieved. This enables the accurate capture of traffic accident precursors within a millisecond time window and the dynamic deduction of accident evolution paths, thus completely solving the technical problems of prediction lag and weak generalization ability of traditional linear models in complex and variable traffic scenarios. Attached Figure Description

[0016] Figure 1 This is the overall flowchart of the highway accident prediction system based on big data of the present invention; Figure 2 This is a flowchart of the multidimensional traffic flow perception and state tensor construction process of the present invention. Figure 3 This is a flowchart of the dynamic graph mapping and feature vector generation process of the present invention; Figure 4 This is a flowchart of the risk potential energy calculation and weight matrix generation process of the present invention; Figure 5 This is a flowchart of the accident probability extrapolation and prediction heatmap generation process of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0018] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0019] Please see Figure 1 and Figure 2 This invention provides a technical solution: a highway accident prediction system based on big data, comprising: The multi-dimensional traffic flow perception module acquires vehicle trajectory coordinates and road surface physical state data, extracts vehicle speed vectors and road friction coefficients, performs spatiotemporal alignment calculations of multi-source data based on unified timestamps and spatial coordinates, and fuses vehicle speed vectors and road friction coefficients to construct a traffic flow state tensor. The specific functions of the multi-dimensional traffic flow perception module are as follows: The heterogeneous spatiotemporal alignment submodule acquires vehicle trajectory coordinates and road surface physical state data, performs time reference unification processing on various heterogeneous data based on atomic clock timing reference, maps sensor data from different perspectives to the standard geographic coordinate system using a preset coordinate transformation matrix, performs multi-source data spatiotemporal alignment operation, and generates a spatiotemporal synchronization reference array. The parameter parsing and splicing submodule extracts the dynamic displacement sequence and static road surface parameters from the spatiotemporal synchronization reference array. Based on the continuous time frame displacement difference logic, it accurately calculates the vehicle velocity vector in three-dimensional space. Combined with the filtering and noise reduction algorithm, it smooths the static road surface parameters and purifies them to obtain the road surface friction coefficient. The vehicle velocity vector and the road surface friction coefficient are orthogonally concatenated to construct the traffic flow state tensor. The process of performing spatiotemporal alignment operations on multi-source data specifically includes: The timestamp information of the header of the raw data packets transmitted by heterogeneous sensors is obtained, the relative drift error between the timestamp information and the global standard clock of the network is calculated, the Kalman filter equation is used to smooth and compensate the relative drift error, and the accurate time tag under the unified clock reference is output. Extract the three-dimensional relative coordinates of various heterogeneous sensor data in the local sensor coordinate system, call the multi-view camera calibration parameters to establish a rotation and translation extrinsic parameter matrix, use the rotation and translation extrinsic parameter matrix to reproject the three-dimensional relative coordinates to the globally unified world geodetic coordinate system, integrate precise time labels, and generate a spatiotemporal synchronization reference array.

[0020] The multi-dimensional traffic flow perception module retrieves real-time data streams from the 77GHz millimeter-wave radar array and the piezoelectric thin-film sensor array deployed along the route to obtain the three-dimensional coordinate set of the vehicle's trajectory and the physical state data of the road surface stress distribution. The high-speed cache unit inside the multi-dimensional traffic flow perception module continuously receives raw data packets from the aforementioned heterogeneous sensing devices at a sampling frequency of 100Hz, parsing the raw data packet payload to extract the instantaneous velocity vector components of the vehicle in the X and Y axes. The piezoelectric thin-film sensor array synchronously feeds back state parameters including road surface moisture content, ice thickness, and roughness. Based on the slip ratio calculation logic when the vehicle tires contact the road surface, the multi-dimensional traffic flow perception module extracts the specific road surface friction coefficient for the corresponding road segment. Utilizing its internal field-programmable gate array (FPGA) main control chip, the multi-dimensional traffic flow perception module performs matrix alignment operations on multi-source data under the same spatiotemporal reference, based on the absolutely unified timestamp distributed by the Network Time Protocol (NAT) and the calibrated WGS84 globally universal spatial coordinates. The multidimensional traffic flow perception module concatenates the aligned vehicle three-dimensional velocity vector data with the road surface friction coefficient in a high-dimensional feature, constructs a traffic flow state tensor containing three dimensions: batch size, road segment length, and number of feature channels, and pushes this tensor into the next level memory pool to await its use.

[0021] The heterogeneous spatiotemporal alignment submodule acquires the raw hexadecimal encoding of vehicle trajectory coordinates and road surface physical state data via a direct fiber optic connection. It extracts the pulse timing reference from the high-precision rubidium atomic clock built into the device and performs nanosecond-level time reference unification processing on various heterogeneous data uploaded by millimeter-wave radar and piezoelectric thin-film sensors, eliminating timestamp offsets caused by hardware response delays. The submodule reads a preset 4x4 homogeneous coordinate transformation matrix from non-volatile memory and precisely maps the vehicle tracking sensor data from the local polar coordinate system of the millimeter-wave radar to the standard WGS84 geodetic coordinate system through matrix multiplication. Finally, the submodule allocates a three-dimensional mapping space in the video memory and performs spatiotemporal alignment operations on multi-source data in ascending time order with spatial grid indices, generating a spatiotemporal synchronization reference array containing four dimensions: time, longitude, latitude, and altitude. Experimental data show that when the time synchronization error parameter is set to less than 5 milliseconds, the success rate of multi-source data fusion alignment reaches 98.7%, which is 14.2% higher than that of conventional systems without atomic clock reference calibration.

[0022] The parameter analysis and stitching submodule reads the dynamic displacement three-dimensional floating-point sequence and static road surface stress, humidity, and other physical parameters encapsulated within the spatiotemporal synchronization reference array via a direct memory access mechanism. Based on the displacement difference logic between two consecutive time frames with a 0.01-second interval, the submodule subtracts the coordinate values ​​of the two frames and divides by the time difference to accurately calculate the instantaneous velocity vector of the vehicle in three-dimensional space. To address the high-frequency noise transmitted from the piezoelectric film sensor, the submodule uses a one-dimensional Gaussian filter noise reduction algorithm to smooth the static road surface parameters, setting the Gaussian kernel window size to 5 and the standard deviation to 1.2, eliminating outliers exceeding 3 times the standard deviation. The submodule then substitutes the smoothed stress and humidity data into a preset road surface adhesion coefficient lookup table for matching and purification to obtain the accurate road surface friction coefficient under the current road conditions. This parameter's value strictly ranges between 0.1 and 0.9. In the tensor operation unit, the parameter analysis splicing submodule performs orthogonal concatenation splicing operations on the vehicle velocity vector matrix with a dimension of N x 3 and the road surface friction coefficient matrix with a dimension of N x 1 to construct a comprehensive traffic flow state tensor with an overall dimension of N x 4.

[0023] The heterogeneous spatiotemporal alignment submodule acquires the initial timestamp information contained in the header of the raw data packets transmitted by heterogeneous sensors. It then calculates the microsecond-level relative drift error between this timestamp information and the global standard clock of the IEEE 1588 precision time protocol network deployed in the system core layer. The submodule uses a Kalman filter equation to smoothly track and continuously compensate for this relative drift error sequence. The system execution process of the Kalman filter equation is as follows: the optimal error estimate from the previous moment is extracted as the prior state prediction for the current moment, and the prior estimate covariance is calculated; subsequently, the actual time drift measurement value of the currently arriving data packet is extracted, and the measurement residual is calculated; the prior estimate covariance and the measurement noise covariance are set to 0.05 and substituted into the calculation to obtain the Kalman gain matrix, which dynamically converges to around 0.3; the Kalman gain is used to correct the prior state prediction error, and finally, the updated optimal error estimate is output. The heterogeneous spatiotemporal alignment submodule uses this optimal error estimate as a compensation term and superimposes it onto the original timestamp, outputting a precise time stamp under the globally unified clock reference of the fully aligned network.

[0024] The heterogeneous spatiotemporal alignment submodule extracts the 3D relative coordinates (X, Y, Z) captured by various heterogeneous sensor data, such as multi-view cameras and roadside radars, within their own local device coordinate systems. It then calls upon the Zhang calibration parameters of the multi-view cameras stored in memory to extract the camera's internal focal length and distortion coefficients, and establishes a rotation-translation extrinsic parameter matrix containing a 3x3 rotation matrix and a 3x1 translation vector. Utilizing a tensor multiplication core, the submodule left-multiplies the column vector carrying the local 3D relative coordinates by this rotation-translation extrinsic parameter matrix, performing a linear coordinate transformation to reproject the 3D relative coordinates onto the globally unified WGS84 world geodetic coordinate system. Finally, the submodule performs row-level alignment and primary key fusion between the reprojected global spatial coordinate matrix and the precise time stamps generated by prior smoothing tracing, generating a highly cohesive spatiotemporal synchronization reference array data table in the in-memory database. This ensures that each sensor record has a unique and absolutely accurate global spatiotemporal index.

[0025] The aforementioned 77GHz millimeter-wave radar array refers to a collection of frequency-modulated continuous wave radar devices operating in the 77 GHz frequency band, used for all-weather, high-precision detection of vehicle distance, speed, and angular distribution information.

[0026] Please see Figure 1 and Figure 3 The dynamic map mapping module divides roads into discrete road segment nodes based on road network geographic information, constructs a road network topology map based on the physical connection relationship between road segments and the actual traffic flow guidance, maps the traffic flow state tensor to the corresponding map nodes, and generates node feature vectors. The specific functions of the dynamic graph mapping module are as follows: The topology node segmentation module obtains the road geometric features extracted from the road network geographic information, divides the continuous road space into discrete road segment nodes according to the preset distance tolerance and intersection coordinates, analyzes the physical connection relationship and actual traffic flow guidance between discrete road segment nodes, and constructs a road network topology structure map. The tensor feature fusion submodule obtains the road network topology diagram and traffic flow state tensor, calculates the spatial boundary coordinates of each discrete road segment node, accurately associates and maps the data slices of the corresponding region in the traffic flow state tensor to the corresponding map nodes based on the spatial boundary coordinates, performs high-dimensional feature aggregation operation, and generates node feature vectors. The process of constructing a road network topology diagram specifically includes: The system acquires the centerline and lane connectivity data of each road segment from the road network geographic information, traverses the centerline of each road segment to identify intersections and divergence points in the road topology skeleton, and performs truncation processing on the continuously distributed road surface geometry model based on the intersections and divergence points to generate multiple independent node entities. Extract the latitude and longitude coordinates of the boundaries at both ends of each node entity, evaluate the Euclidean distance and traffic accessibility index between node entities, and if the Euclidean distance is lower than the preset distance tolerance and the traffic accessibility index shows unidirectional or bidirectional conduction attributes, then establish connection edges between the corresponding node entities, combine all node entities and connection edges, and construct a road network topology diagram.

[0027] The dynamic map mapping module retrieves Shapefile format road network geographic information files published by the geographic information system server and divides the continuous main highway into numerous discrete road segment nodes according to a preset fixed spatial step of 50 meters. Based on the spatial physical topological connections between adjacent discrete road segment nodes and the one-way or two-way traffic guidance attributes indicated by actual road markings, the module constructs a road network topology map composed of node entities and connecting edges in the graph database. The module then activates the spatial association mapping engine, accurately associating and cropping the traffic flow state tensor (containing velocity vectors and friction coefficients) generated by the previous module to the corresponding map node attribute fields based on the latitude and longitude bounding boxes of each discrete road segment node, generating a high-density node feature vector with a length of 128 dimensions.

[0028] The topology node segmentation module parses the incoming Shapefile road network geographic information file, extracting the road geometric shape features and curvature parameters contained in the polyline layer. Based on a preset 50-meter distance tolerance parameter, and combined with the latitude and longitude coordinates of intersections and ramp entrances / exits marked in the layer, the module equidistantly truncates and geometrically divides the continuous physical road space along the road centerline, generating independent discrete road segment node sets. The module traverses these discrete road segment node sets, reads the road level and traffic direction fields from the road network attribute table, analyzes whether there are actual physical paving connections between adjacent discrete road segment nodes, and strictly constructs an adjacency matrix of the road network topology structure that conforms to the actual traffic flow logic, based on the actual traffic flow guidance of the lane lines. In the matrix, an element of 1 represents connectivity, and 0 represents obstruction.

[0029] The tensor feature fusion submodule reads the completed road network topology map data model from the graph database and simultaneously retrieves the traffic flow state tensor from the cache. It calculates the coordinates of the minimum bounding rectangle spatial boundary covered by each discrete road segment node in the graph. Using the R-Tree spatial indexing algorithm, it performs a fast spatial intersection query on the global coordinate set contained in the traffic flow state tensor based on these spatial boundary coordinates. It extracts tensor data slices that completely fall within the spatial boundary of the node and accurately maps them to the corresponding graph node objects. The tensor feature fusion submodule performs high-dimensional feature aggregation operations on the tensor slice arrays collected within each node object. The aggregation process involves applying max pooling and average pooling operators along the time and channel dimensions to extract extreme and mean features within local regions. The pooling results are then input into a perceptron containing two fully connected layers for nonlinear dimensionality upscaling, ultimately generating a 128-dimensional node feature vector representing the comprehensive traffic situation of the node.

[0030] The tensor feature fusion submodule acquires the centerline coordinate sequences of each road segment and the connectivity logic data at the lane level from the road network geographic information data package. It then iterates through the centerline coordinate sequences of each road segment, calculating the curvature abrupt changes in the centerline vector and the spatial intersections with other centerlines to identify all intersections and grade-separated breakpoints in the road topology skeleton network that possess diverging or merging characteristics. Using these intersections and breakpoints as cutting elements, the tensor feature fusion submodule performs spatial polygon truncation processing on the continuously distributed three-dimensional geometric polygon model of the road surface, decomposing long-distance roads and generating multiple node entity data records with independent spatial boundary polygons.

[0031] The topology node slicing module extracts the latitude and longitude coordinates of the boundary endpoints of the spatial polygons generated by the truncation process. Using the spherical distance calculation formula, it accurately assesses the Euclidean distance between the endpoints of any two adjacent node entities and retrieves the road network accessibility rules table to evaluate the traffic accessibility index between them. The internal logic of the topology node slicing module is as follows: if the Euclidean distance between the endpoints of two node entities is lower than the preset 5-meter distance tolerance limit, and their traffic accessibility index shows a unidirectional conduction value of 1 or a bidirectional conduction attribute value of 2 in the rule table, then the two nodes are determined to be physically and logically connected, and a connection edge with directional attributes is created between the corresponding node entity objects in the graph database. After completing the global traversal, the topology node slicing module combines all node entity records and the established connection edge records to solidify and output a complete road network topology diagram. Experimental data shows that when the distance tolerance parameter is set to 5 meters, the topology node matching accuracy reaches 99.2%, which is 8.5% higher than the traditional method based on fixed road network map segmentation.

[0032] The R-Tree spatial indexing algorithm mentioned above refers to a tree-like data structure based on spatial bounding rectangles. By dividing and wrapping multidimensional spatial data according to the minimum bounding rectangle, it achieves efficient intersection and containment queries of spatial data.

[0033] Please see Figure 1 and Figure 4 The risk potential energy calculation module uses graph convolution operation logic to aggregate the node feature vectors of adjacent nodes, and quantitatively calculates the dynamic mutual influence intensity between nodes based on the spatial distance decay function and the correlation of traffic flow direction between nodes, generating a risk transmission weight matrix. The specific functions of the risk potential energy calculation module are as follows: The graph domain convolution aggregation submodule obtains node feature vectors, configures graph convolution kernel operators based on Chebyshev multinomials, aggregates node feature vectors of adjacent nodes using graph convolution operation logic, extracts spatial dependency features within the local road network, and outputs a neighborhood-aware feature matrix. The potential energy weight assessment submodule extracts the neighborhood perception feature matrix, quantifies the static spatial impedance between nodes based on the spatial distance attenuation function, and quantitatively calculates the dynamic mutual influence intensity between nodes by combining the traffic flow direction correlation between nodes. The static spatial impedance and dynamic mutual influence intensity are cross-validated and weighted summed to generate the risk transmission weight matrix. The process of quantitatively calculating the intensity of dynamic interactions between nodes specifically includes: The spatial physical distance between nodes and the sequence of node transfers within a continuous time period are obtained. The spatial physical distance is nonlinearly mapped and transformed based on the Gaussian decay kernel function, and the basic potential energy decay benchmark is output. Extract the node transfer train sequence, use Pearson correlation coefficient logic to calculate the dynamic coordinated volatility between the source node traffic trend and the target node traffic trend, substitute the basic potential energy decay benchmark and the dynamic coordinated volatility into the preset potential energy coupling formula for product operation, and obtain the dynamic mutual influence intensity. The specific formula for potential energy coupling is as follows: ; in, This represents the intensity of dynamic mutual influence. Represents the system calibration coefficient. Represents the natural constant. Represents the spatial attenuation damping coefficient. Represents the spatial physical distance between nodes. This represents dynamic co-variance volatility.

[0034] The risk potential energy calculation module inputs the high-dimensional node feature vectors generated by each node into the built-in graph neural network accelerator. Using graph convolution logic, it aggregates the node feature vectors of adjacent nodes in the adjacency matrix of the topology, extracting the spatial dependencies within the local road network. Based on the nonlinear decay function of physical spatial distance and combined with historical traffic flow correlation parameters between nodes, the module quantitatively calculates the intensity of dynamic mutual influence between any two nodes in the network topology. The module then combines and normalizes the numerical values ​​of the dynamic mutual influence intensity among all network nodes, generating an asymmetric NxN-dimensional risk transmission weight matrix. This matrix intuitively quantifies the transmission probability and potential energy distribution of traffic accident risks between nodes in different road segments.

[0035] The graph convolution aggregation submodule acquires the feature vector matrix of all graph nodes with dimension N x 128 via the memory bus and configures a graph convolution kernel operator based on Chebyshev polynomial expansion within the deep learning framework. The submodule then uses this graph convolution operation logic to perform spatial feature aggregation. The system operation process is as follows: First, the adjacency matrix of the road network topology is extracted and its degree matrix is ​​calculated, resulting in a normalized Laplacian matrix. The Laplacian matrix is ​​then subjected to eigenvalue decomposition, and its corresponding eigenvector matrix is ​​extracted as a Fourier transform basis. The Chebyshev polynomial order of the graph convolution is configured to be 3 to capture the local road network topology information within a 3-hop range around the central node. The submodule multiplies the node feature vector matrix with the Chebyshev polynomial basis matrix and superimposes a learnable parameter matrix mapped by a nonlinear activation function. Through the forward propagation operation of the convolutional layer of the graph, the features of the node itself and the features of the neighboring nodes within 3 hops are weighted and fused according to the degree of connection of the topological structure, which fully extracts the spatial impedance and traffic flow dependence features in the local road network. The final output is a neighborhood perception feature matrix with the same dimension of N times 128.

[0036] The potential energy weighting evaluation submodule extracts the updated neighborhood perception feature matrix from the output port of the graph domain convolution aggregation submodule. Based on the built-in Gaussian spatial distance decay function, the submodule exponentially maps and quantifies the actual Euclidean distance between nodes, outputting a static spatial impedance parameter that represents the greater the physical obstacle as the distance increases. The submodule extracts the vehicle transfer time series between nodes from the network traffic monitoring unit statistics and uses the Pearson correlation coefficient algorithm to quantitatively calculate the dynamic mutual influence intensity between the source node and the target node in the traffic flow fluctuation trend. The submodule performs element-wise multiplication cross-validation and adaptive weighted summation on the calculated static spatial impedance matrix and dynamic mutual influence intensity matrix, setting the static impedance weight to 0.4 and the dynamic influence weight to 0.6, ultimately generating a risk transmission weight matrix of dimension N by N, used to accurately represent the risk diffusion probability.

[0037] The potential energy weighting assessment submodule obtains the actual spatial physical distance data between any two related nodes through the road network topology database query interface. Simultaneously, it retrieves the absolute number of vehicle transfers within the past 24 hours from the traffic statistics database, divided into 96 consecutive time periods of 15 minutes each. Based on a Gaussian decay kernel function equation with parameters set to a mean of 0 and a variance of 500, the submodule performs a nonlinear mapping transformation on the obtained continuous spatial physical distance values. This compresses the numerical differences over long distances and smoothly transitions them within the 0-1 range. The final output is a baseline potential energy attenuation matrix characterizing the basic spatial physical barrier effect.

[0038] The potential energy weighting assessment submodule extracts the 96-dimensional node transfer train sequence vectors corresponding to each of the two nodes. It then uses Pearson correlation coefficient logic to perform statistical analysis on these two sequences: calculating the covariance values ​​of the two sequences and multiplying their respective standard deviations. Dividing the covariance value by the standard deviation product accurately calculates the dynamic co-variance volatility between the source node's flow trend and the target node's flow trend, with the result strictly falling within the floating-point range of -1 to 1. The potential energy weighting assessment submodule then substitutes the previously calculated baseline potential energy decay value and this dynamic co-variance volatility, along with the system's preset environmental coefficients, into a preset potential energy coupling formula for tensor multiplication, obtaining a dynamic mutual influence strength index characterizing the efficiency of risk potential energy transfer.

[0039] Table 1. Risk Potential Energy Parameter Configuration and Calculation Results

[0040] As shown in Table 1, the system extracted the distance parameter and the final potential energy intensity calculation results for different node combinations in the actual road network. The closer the road segment and the higher the consistency of the traffic flow trend, the greater the value of its influence intensity.

[0041] The specific configuration of the potential energy coupling formula called by the core computing power within the potential energy weight evaluation submodule is as follows: ; in, This represents the strength of the dynamic interaction between the source node and the target node. The system calibration coefficients are used to adjust the reference amplitude of the overall potential energy. Represents the base of the natural constant. The spatial attenuation damping coefficient is used to control the attenuation rate of risk transmission over distance. This represents the spatial physical distance between the source node and the target node. It represents the dynamic co-variability between the traffic trends of the source node and the traffic trends of the target node.

[0042] The potential energy weighting assessment submodule has a built-in logical mapping relationship that supports the configuration of specific environmental coefficients. The system calibration coefficient is determined by the historical accident rate baseline of the monitoring system. In this embodiment, the system calibration coefficient is set to 1.5 based on the statistical average of accidents on urban main roads. The spatial attenuation damping coefficient characterizes the road's physical conditions' ability to impede risk transmission. It is dynamically assigned based on road surface smoothness and the number of lanes. In this road network environment, sensor feedback indicates a smooth road surface and a two-way six-lane configuration, so the spatial attenuation damping coefficient is set to 0.002.

[0043] The numerical logic processor of the potential energy weighting assessment submodule directly substitutes the previously obtained system calibration coefficient of 1.5, spatial attenuation damping coefficient of 0.002, spatial physical distance between nodes of 120 meters, and dynamic cooperative volatility of 0.92 into the above potential energy coupling formula to perform floating-point multiplication calculation. The calculation logic is as follows: First, multiply the product term by -0.002 using the base of the natural constant and perform an exponential operation to obtain the distance attenuation term of 0.786. Then, multiply this attenuation term by 1.5 and 0.92. The final dynamic mutual influence strength result is calculated to be 1.08. This result shows that under the condition of consistent spatial distance and traffic flow fluctuation, the risk transmission from the source node to the target node has a strong potential energy penetration force. The advantage of the formula is that by introducing an exponential attenuation term that is strictly correlated with the physical distance and a cooperative volatility reflecting real-time traffic flow, the static topology map has the ability to perceive dynamic and time-varying risk transmission. Experimental data show that after introducing the co-variability parameter and performing the formula calculation, the lead time for cross-road accidents increased from 5 minutes to 12 minutes.

[0044] The Chebyshev polynomial mentioned above refers to a mathematical approximation method based on orthogonal polynomial sequences. Here, it is used to perform polynomial expansion on the graph Laplacian matrix to reduce the computational complexity of eigenvalue decomposition of the graph convolution kernel and to achieve local spatial feature extraction.

[0045] Please see Figure 1 and Figure 5 The accident probability deduction module inputs the risk transmission weight matrix into the time-varying recurrent neural network for deep time-series evolution analysis, combines extreme value theory to set dynamic thresholds, calculates the mutation probability value of each discrete road segment node at the current time step, and generates an accident prediction heatmap. The specific functions of the accident probability simulation module are as follows: The deep temporal inference submodule obtains the risk transmission weight matrix, inputs the risk transmission weight matrix into the network architecture composed of gated recurrent units for deep temporal evolution analysis, extracts the cumulative evolution law of traffic risks in the time dimension, and outputs the temporal risk evolution feature vector. The extreme value heat generation submodule analyzes the time-series risk evolution feature vector, fits the risk tail data with the extreme value theory of the generalized Pareto distribution, sets a dynamic threshold, calculates the mutation probability value of each discrete road segment node at the current time step, performs two-dimensional kernel density estimation and color mapping interpolation based on the spatial distribution coordinates of the mutation probability value, and generates an accident prediction heat map. The process of setting dynamic thresholds specifically includes: The extreme value sample set of the time series risk evolution feature vector within the historical observation period is obtained. The maximum likelihood estimation method is used to approximate the distribution tail features of the extreme value sample set. The shape parameter and scale parameter of the generalized Pareto distribution function are calculated. Based on shape and scale parameters, construct the tail probability density function for the corresponding road segment region, set confidence boundaries, solve the inverse function quantile of the tail probability density function under the confidence boundary conditions, and configure the inverse function quantile as the dynamic threshold under the current road network state; The process of conducting in-depth temporal evolution analysis specifically includes: Extract historical discrete slices that are continuously distributed along the time axis from the risk transmission weight matrix, input the historical discrete slices into the update gate component and reset gate component inside the gating network, and use the reset gate component to control the fusion ratio of historical state information and current input data to generate candidate hidden states; The contribution weight of candidate hidden states to long-term traffic evolution trends is evaluated using the update gate component. Based on the contribution weight, selective forgetting and information overwriting operations are performed on historical memory units. The comprehensive evolution tensor is output through multi-level iterative superposition. The comprehensive evolution tensor is then reduced in dimension and flattened to output the temporal risk evolution feature vector.

[0046] The accident probability inference module utilizes a gated recurrent unit framework to receive time-series slices of the N×N risk transmission weight matrix passed from the previous stage. These slices are then input into a time-varying recurrent neural network (RNN) to perform multiple rounds of forward propagation, conducting in-depth temporal evolution analysis of traffic flow characteristics. The module extracts hidden layer features from the RNN's terminal output and statistically fits the extreme tail region of the risk features using the extreme value theory of the generalized Pareto distribution, setting a dynamic threshold suitable for the current time step. The module defines the integral area where the predicted risk features exceed the dynamic threshold as the mutation probability value for each discrete road segment node. Subsequently, based on the latitude and longitude spatial distribution coordinates of each node in the real road network, it calls the graphics rendering engine to perform two-dimensional kernel density estimation and color space mapping interpolation operations, generating an intuitive accident prediction heatmap on the electronic map layer displaying high-risk areas.

[0047] The deep temporal derivation submodule reads a sequence of risk propagation weight matrices stacked in time slices from the cache. Using this sequence as a continuous input source (with a length set to the past 60 time steps), it injects this sequence into a network architecture composed of cascaded gated recurrent units with multiple hidden layers of 256 dimensions to perform deep temporal evolution analysis. During the network's temporal unfolding computation, the learnable weight matrices within the network capture minute gradient changes between consecutive time slices, extracting the nonlinear cumulative evolution of traffic congestion or accident risks from latent to eruptive over time. The deep temporal derivation submodule extracts the output layer activation vector from the last time step, serving as an N-dimensional (256) temporal risk evolution feature vector that fully integrates historical temporal context information.

[0048] The extreme value heat generation submodule reads the time-series risk evolution feature vector input from the deep time-series extrapolation submodule and compresses it into a one-dimensional risk score scalar set using a tensor dimensionality reduction operator. The extreme value heat generation submodule calls the generalized Pareto distribution theory library deployed on the statistical analysis computing platform and uses the extreme value limit theorem to fit the distribution pattern of the risk tail anomaly data exceeding the initial extreme value baseline in the first 5% interval of the risk score scalar set. The extreme value heat generation submodule uses the fitted parameter distribution curve to set a dynamic threshold, converting the absolute values ​​of risk score scalars exceeding this dynamic threshold into mutation probability values ​​within the 0-1 interval. The extreme value heat generation submodule extracts the absolute geographical coordinates (longitude and latitude) of each discrete road segment node, uses the calculated mutation probability values ​​as weights, and initiates a two-dimensional kernel density estimation module in the spatial domain for probability density smoothing. After the calculation is completed, the extreme value heat generation submodule uses the built-in RGB color gradient mapping table to map high probability density areas to red pixels and low probability density areas to blue pixels through interpolation, outputting accident prediction heat map layer data with a resolution of 1920 x 1080.

[0049] The extreme value heatmap generation submodule performs a full table scan of the historical database to obtain all time-series risk evolution feature vectors within the most recent 30-day observation period. Utilizing a peak exceedance threshold extraction algorithm, this submodule filters out the top extreme data points in the feature vectors whose values ​​exceed the 95th percentile of the overall sample, constructing a dedicated extreme value sample set for abnormal road conditions. This extreme value sample set is then input into the parameter optimizer, where the maximum likelihood estimation method is used to construct the maximum likelihood logarithm function. An optimization algorithm is then initiated to iteratively approximate the tail features of the extreme value sample set's distribution using multiple rounds of gradient descent. After iterative calculations until the loss function converges, the submodule obtains two core parameters of the generalized Pareto distribution function that maximizes the likelihood function value: the shape parameter controlling the tail thickness and the scale parameter controlling the spread range.

[0050] The extreme value thermal generation submodule substitutes the acquired shape parameter 0.15 and scale parameter 2.4 into the standard model library to construct the tail probability density function curve of the corresponding road segment area, exhibiting long-tail distribution characteristics. Based on system reliability requirements, the extreme value thermal generation submodule rigidly sets the confidence boundary to a high-pressure warning level of 0.99. The submodule then calls the numerical integration inverse operator in memory and uses Newton's iteration method to solve for the precise value of the inverse function quantile of the tail probability density function accumulated to the 0.99 confidence boundary condition. The submodule extracts this inverse function quantile value and configures it as a dynamic threshold for intercepting high-risk situations under the current road network conditions. Experimental data shows that after setting the confidence boundary to 0.99 and adopting dynamic threshold configuration, the system's false negative rate for severe rear-end collisions decreased from 12.4% to 1.8%.

[0051] Table 2. Dynamic Extreme Value Threshold Parameter Fitting Calculation Table

[0052] As shown in Table 2, the system extracts extreme value samples at different quantiles and applies maximum likelihood estimation to calculate shape and scale parameters. The final dynamic threshold result data is used as a hard comparison benchmark for subsequent probability determination.

[0053] The deep temporal inference submodule reads from the cache and extracts 120 historical discrete weight matrix slices continuously distributed along the time axis over the past 60 minutes from the risk propagation weight matrix based on timestamps. This chronologically ordered historical discrete slice data stream is then input frame-by-frame into a gated recurrent unit (GRU) containing an encapsulated gating mechanism. Within this GRU, the hidden layer state from the previous time step and the current input slice are simultaneously fed into an update gate component and a reset gate component, which contain fully connected layers and activation functions. The deep temporal inference submodule performs a dot product operation using the gating variable (between 0 and 1) output by the reset gate component, precisely controlling the fusion ratio of the historical state information matrix and the current input data matrix in the mixed feature space with hardware-level precision. Subsequently, a candidate hidden state matrix incorporating short-term fluctuation features is generated using the hyperbolic tangent activation function.

[0054] The deep time-series inference submodule then utilizes the gating vector output by the update gate component to evaluate the relative contribution weights of candidate hidden states to characterizing long-term traffic evolution trends such as the spread of tidal flows during morning and evening peak hours within the tensor computation core. The memory controller within the deep time-series inference submodule performs element-wise selective forgetting and new information overwriting operations on the original historical memory units based on this contribution weight vector, where a weight approaching 0 represents forgetting and a weight approaching 1 represents overwriting, thereby updating the final hidden state at the current moment. Through serial iterative computation via a multi-level recurrent network with 120 time steps, the deep time-series inference submodule superimposes and outputs a comprehensive evolution tensor that integrates the entire time-series cycle. The deep time-series inference submodule pushes this multi-dimensional comprehensive evolution tensor into a flattening layer, performing matrix dimensionality reduction and one-dimensional flattening to remove redundant structured holes, ultimately outputting a one-dimensional dense array-form temporal risk evolution feature vector containing compact and dense features.

[0055] The aforementioned generalized Pareto distribution refers to a continuous probability distribution function used in extreme value theory to model the tail probability distribution of random variables that exceed a specific threshold. It is mainly used to assess the probability of occurrence and risk limits of extreme abnormal events.

[0056] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A highway accident prediction system based on big data, characterized in that, The system includes: The multi-dimensional traffic flow perception module acquires vehicle trajectory coordinates and road surface physical state data, extracts vehicle speed vectors and road friction coefficients, performs spatiotemporal alignment calculations of multi-source data based on unified timestamps and spatial coordinates, and fuses the vehicle speed vectors and road friction coefficients to construct a traffic flow state tensor. The dynamic graph mapping module divides roads into discrete road segment nodes based on road network geographic information, constructs a road network topology graph based on the physical connection relationship between road segments and the actual traffic flow guidance, and maps the traffic flow state tensor to the corresponding graph nodes to generate node feature vectors. The risk potential energy calculation module aggregates the node feature vectors of adjacent nodes using graph convolution operation logic, and quantitatively calculates the dynamic mutual influence intensity between nodes based on the spatial distance decay function and the traffic flow direction correlation between nodes, generating a risk transmission weight matrix. The accident probability extrapolation module inputs the risk transmission weight matrix into a time-varying recurrent neural network for deep time-series evolution analysis. It combines extreme value theory to set dynamic thresholds, calculates the mutation probability value of each discrete road segment node at the current time step, and generates an accident prediction heatmap.

2. The highway accident prediction system based on big data according to claim 1, characterized in that, The specific functions of the multi-dimensional traffic flow perception module are as follows: The heterogeneous spatiotemporal alignment submodule acquires the vehicle's driving trajectory coordinates and the road surface physical state data. Based on the atomic clock timing reference, it performs time reference unification processing on various heterogeneous data. Using a preset coordinate transformation matrix, it maps sensor data from different perspectives to a standard geographic coordinate system. It then performs the multi-source data spatiotemporal alignment operation to generate a spatiotemporal synchronization reference array. The parameter parsing and splicing submodule extracts the dynamic displacement sequence and static road surface parameters from the spatiotemporal synchronization reference array. Based on the continuous time frame displacement difference logic, it accurately calculates the vehicle velocity vector in three-dimensional space. Combined with the filtering and noise reduction algorithm, it smooths the static road surface parameters and purifies them to obtain the road surface friction coefficient. The vehicle velocity vector and the road surface friction coefficient are orthogonally concatenated to construct the traffic flow state tensor.

3. The highway accident prediction system based on big data according to claim 1, characterized in that, The specific functions of the dynamic graph mapping module are as follows: The topology node segmentation module obtains the road geometric features extracted from the road network geographic information, divides the continuous road space into discrete road segment nodes according to the preset distance tolerance and intersection coordinates, analyzes the physical connection relationship between the discrete road segment nodes and the actual traffic flow guidance, and constructs the road network topology structure map. The tensor feature fusion submodule obtains the road network topology diagram and the traffic flow state tensor, calculates the spatial boundary coordinates corresponding to each discrete road segment node, accurately associates and maps the data slices of the corresponding region in the traffic flow state tensor to the corresponding map nodes based on the spatial boundary coordinates, performs high-dimensional feature aggregation operation, and generates the node feature vector.

4. The highway accident prediction system based on big data according to claim 1, characterized in that, The specific function of the risk potential energy calculation module is as follows: The graph domain convolution aggregation submodule obtains the node feature vectors, configures a graph convolution kernel operator based on Chebyshev multinomials, aggregates the node feature vectors of adjacent nodes using the graph convolution operation logic, extracts spatial dependency features within the local road network, and outputs a neighborhood-aware feature matrix. The potential energy weight assessment submodule extracts the neighborhood perception feature matrix, quantifies the static spatial impedance between nodes based on the spatial distance attenuation function, quantitatively calculates the dynamic mutual influence intensity between nodes by combining the traffic flow direction correlation between nodes, and performs cross-validation and weighted summation on the static spatial impedance and the dynamic mutual influence intensity to generate the risk transmission weight matrix.

5. The highway accident prediction system based on big data according to claim 1, characterized in that, The specific functions of the accident probability deduction module are as follows: The deep temporal inference submodule obtains the risk transmission weight matrix, inputs the risk transmission weight matrix into the network architecture composed of gated recurrent units to perform the deep temporal evolution analysis, extracts the cumulative evolution law of traffic risk in the time dimension, and outputs the temporal risk evolution feature vector. The extreme value heat generation submodule analyzes the time-series risk evolution feature vector, fits the risk tail data with the extreme value theory of the generalized Pareto distribution, sets a dynamic threshold, calculates the mutation probability value of each discrete road segment node at the current time step, performs two-dimensional kernel density estimation and color mapping interpolation based on the spatial distribution coordinates of the mutation probability value, and generates the accident prediction heat map.

6. The highway accident prediction system based on big data according to claim 2, characterized in that, The process of performing the multi-source data spatiotemporal alignment operation specifically includes: The timestamp information of the header of the raw data packet transmitted by the heterogeneous sensors is obtained, the relative drift error between the timestamp information and the global standard clock of the network is calculated, the Kalman filter equation is used to smooth and compensate the relative drift error, and the accurate time tag under the unified clock reference is output. Extract the three-dimensional relative coordinates of various heterogeneous sensor data in the local sensor coordinate system, call the multi-view camera calibration parameters to establish a rotation and translation extrinsic parameter matrix, use the rotation and translation extrinsic parameter matrix to reproject the three-dimensional relative coordinates to the globally unified world geodetic coordinate system, fuse the precise time label, and generate the spatiotemporal synchronization reference array.

7. The highway accident prediction system based on big data according to claim 3, characterized in that, The process of constructing the road network topology diagram specifically includes: The road network geographic information contains the center lines of each road segment and lane connectivity data. The center lines of each road segment are traversed to identify intersections and divergence points in the road topology skeleton. Based on the intersections and divergence points, the continuously distributed road surface geometry model is truncated to generate multiple independent node entities. Extract the latitude and longitude coordinates of the boundaries at both ends of each node entity, evaluate the Euclidean distance and traffic accessibility index between node entities, and if the Euclidean distance is lower than the preset distance tolerance and the traffic accessibility index exhibits unidirectional or bidirectional conduction attributes, then establish connection edges between the corresponding node entities, combine all node entities and connection edges, and construct the road network topology diagram.

8. The highway accident prediction system based on big data according to claim 4, characterized in that, The process of quantitatively calculating the intensity of the dynamic interaction between nodes specifically includes: The spatial physical distance between nodes and the sequence of node transfers within a continuous time period are obtained. The spatial physical distance is then nonlinearly mapped and transformed according to the Gaussian decay kernel function to output the basic potential energy decay benchmark. Extract the node transfer train sequence, use Pearson correlation coefficient logic to calculate the dynamic coordinated volatility between the source node traffic trend and the target node traffic trend, substitute the basic potential energy decay benchmark and the dynamic coordinated volatility into the preset potential energy coupling formula for product operation, and obtain the dynamic mutual influence intensity. The specific formula for potential energy coupling is as follows: ； in, This represents the intensity of the dynamic interaction. Represents the system calibration coefficient. Represents the natural constant. Represents the spatial attenuation damping coefficient. This represents the spatial physical distance between nodes. This represents the dynamic co-variance volatility.

9. The highway accident prediction system based on big data according to claim 5, characterized in that, The process of setting dynamic thresholds specifically includes: Obtain the extreme value sample set of the time series risk evolution feature vector within the historical observation period, and use the maximum likelihood estimation method to approximate the distribution tail features of the extreme value sample set, and calculate the shape parameter and scale parameter of the generalized Pareto distribution function. Based on shape and scale parameters, a tail probability density function for the corresponding road segment region is constructed. A confidence boundary is set, and the inverse function quantile of the tail probability density function under the confidence boundary condition is solved. The inverse function quantile is then configured as a dynamic threshold under the current road network condition.

10. The highway accident prediction system based on big data according to claim 5, characterized in that, The process of performing the aforementioned deep temporal evolution analysis specifically includes: Extract the historical discrete slices that are continuously distributed along the time axis from the risk transmission weight matrix, input the historical discrete slices into the update gate component and reset gate component inside the gating network, and use the reset gate component to control the fusion ratio of historical state information and current input data to generate candidate hidden states; The update gate component is used to evaluate the contribution weight of the candidate hidden state to the long-term traffic evolution trend. Based on the contribution weight, selective forgetting and information overwriting operations are performed on the historical memory unit. The comprehensive evolution tensor is output through multi-level iterative superposition. The comprehensive evolution tensor is then subjected to dimensionality reduction and flattening to output the temporal risk evolution feature vector.

Images