Big data based traffic accident risk assessment system

Abstract

This invention relates to the field of traffic management technology, specifically a traffic accident risk assessment system based on big data. The system includes: a feature acquisition module, a topology construction module, a feature extraction module, a hazard discovery module, and a risk assessment module. In this invention, a regional traffic state matrix is ​​constructed by multi-source fusion using vehicle trajectory speed fluctuation features and vehicle counts. Topology weight parameters are obtained based on the physical attenuation characteristics of the spatial straight-line distance between road network connection nodes, and spatial weighting is used to construct a spatiotemporal topology connectivity tensor. Internal multi-dimensional local peak point coordinates and spatial topology change rate parameters are analyzed to extract abnormal coordinates and construct a multi-dimensional risk candidate vector. The degree of deviation from safety thresholds is compared to extract the numbers of over-limit road nodes and construct a set of hazard nodes. Historical accident frequencies are fused with the current node hazard status to generate risk level labels. Multi-source dynamic data is deeply mined and global spatiotemporal correlation mapping is strengthened to lock in risk areas and improve the accuracy of traffic safety assessment.

Description

Technical Field

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

[0002] The field of traffic management technology covers core aspects such as road traffic flow monitoring, traffic signal coordination and control, traffic violation recording, and traffic safety situation analysis. This field mainly collects road physical characteristics, vehicle trajectory and environmental meteorological parameters through front-end sensing hardware, and transmits the collected elements to the central command platform through communication lines, thereby supervising, planning and coordinating the operation of the entire road network.

[0003] Traditional traffic accident risk assessment systems refer to systems that quantify the probability of road safety events such as motor vehicle collisions occurring at road network intersections and designated road sections. They involve reviewing historical traffic accident files to extract the latitude and longitude coordinates, time points, and types of vehicles involved in the accidents; combining this with road construction drawings to obtain the horizontal curve radius and longitudinal slope values ​​of the road sections; then retrieving the absolute number of vehicles and average driving speed indicators of the cross-section obtained by microwave radar speed measuring instruments; and finally, manually substituting the above specific discrete characteristic values ​​into the Poisson regression equation or negative binomial regression equation to carry out mathematical formula calculations, thereby obtaining the expected frequency value of accidents occurring within a specific road area over a given time span.

[0004] Traditional traffic accident risk assessment systems rely on historical records to extract accident coordinates and vehicle types, combined with maps to obtain road curve radii and slope values, and speed cameras to obtain traffic flow and average speed indicators. The system then manually substitutes these discrete feature values ​​into regression equations for calculation. This operational mode, which depends on manually collecting discrete static data and quantifying it using fixed mathematical models, is unable to capture the dynamic fluctuations in complex road networks. It also severs the topological connections between multi-source traffic data in the spatiotemporal plane, causing risk predictions to deviate from actual road conditions and failing to adapt to dynamic traffic situations. Summary of the Invention

[0005] To address the technical problems of traditional traffic accident risk assessment systems, which rely on manually collecting discrete static data and inputting it into a fixed mathematical model, making it difficult to capture the dynamic fluctuations in complex road networks and severing the topological connections between multi-source traffic data in the spatiotemporal plane, thus causing risk predictions to deviate from actual road conditions and failing to adapt to dynamic traffic situations, this invention provides a traffic accident risk assessment system based on big data.

[0006] On the one hand, a traffic accident risk assessment system based on big data is provided, which includes: The feature acquisition module acquires vehicle count and vehicle trajectory data based on intersection sensors and a big data platform, extracts speed fluctuation features within the vehicle trajectory data, and combines them with vehicle counts to perform multi-source feature fusion. The features are then combined according to spatial arrangement to construct a regional traffic state matrix. The topology construction module, based on the regional traffic state matrix, obtains topology weight parameters according to the physical attenuation characteristics of the spatial straight-line distance values ​​of road network connection nodes, performs spatial weighting processing on the elements in the regional traffic state matrix, and constructs a spatiotemporal topology connectivity tensor. The feature extraction module analyzes the coordinates of local peak points and spatial topological change rate parameters in multiple dimensions within the spatiotemporal topological connectivity tensor, uses support vector machine for joint classification, extracts abnormal coordinate parameters, and constructs multidimensional risk candidate vectors. The hidden danger discovery module compares the deviation of the multidimensional risk candidate vector from the preset safety judgment threshold, extracts the node numbers of the over-limit roads, and constructs a set of accident hidden danger nodes; The risk assessment module, based on the set of accident hazard nodes, uses a random forest algorithm to analyze the frequency of historical accidents, and performs feature fusion with the hazard status of the current node to generate a risk level label.

[0007] As a further aspect of the present invention, the regional traffic state matrix includes road segment saturation, lane time occupancy rate, and traffic delay index; the spatiotemporal topological connectivity tensor includes road segment impedance, network connectivity, and spatial reachability; the multidimensional risk candidate vector includes collision time, post-intrusion time, and velocity variance; the accident hazard node set includes nodes with excessive historical accident frequency, nodes with abnormal line-of-sight parameters, and nodes with excessive radius of curvature; and the risk level label specifically includes high-risk level, medium-risk level, and low-risk level.

[0008] As a further aspect of the present invention, the feature acquisition module includes: The trajectory feature analysis submodule acquires vehicle count and vehicle trajectory data based on intersection sensors and big data platform, extracts the historical driving speed sequence of a single vehicle in the vehicle trajectory data, calculates the discrete standard deviation of the historical driving speed sequence of a single vehicle within a set global observation time window, and generates speed fluctuation feature quantity. The multi-source feature fusion submodule calls the speed fluctuation feature quantity, converts the speed fluctuation feature quantity and vehicle count into a multi-dimensional data feature tensor of the same dimension, assigns corresponding weight coefficients to the multi-dimensional data feature tensor of the same dimension, and then performs weighted fusion to obtain the traffic feature fusion value. The state array construction submodule, for the traffic feature fusion values, performs horizontal row and column topological splicing and combination of the traffic feature fusion values ​​according to the relative arrangement order of the two-dimensional absolute coordinates of the data monitoring nodes in the global spatial reference coordinate system, to establish a regional traffic state matrix.

[0009] As a further aspect of the present invention, the topology construction module includes: The distance feature extraction submodule extracts the two-dimensional spatial positioning coordinates of the corresponding road network connection nodes based on the regional traffic state matrix, calculates the absolute Euclidean space straight-line span values ​​between the road network connection nodes, and generates a node straight-line distance array. The decay weight calculation submodule calls the node linear distance array, converts the matrix element values ​​in the node linear distance array into their reciprocals, sets the natural logarithm base as a constant and performs natural exponential decay scalar operation with the corresponding reciprocal values ​​to obtain the topological weight decay coefficient. The spatiotemporal tensor construction submodule performs a spatially weighted, element-wise multiplication scalar operation on the topological weight decay coefficient and the corresponding values ​​of the corresponding mapped coordinates in the regional traffic state matrix. It then continuously performs three-dimensional topological structure expansion in the order of the timestamp sequence to construct a spatiotemporal topological connectivity tensor.

[0010] As a further aspect of the present invention, the feature extraction module includes: The parameter extraction submodule extracts the spatial positioning coordinates of the gradient flip point within the spatiotemporal topological connectivity tensor, extracts the displacement derivative datasets between adjacent slices, merges and calculates the distribution variance of the displacement derivative dataset, and concatenates the spatial positioning coordinates and distribution variance row by row to generate a peak discrete feature set. The projection classification submodule calls the peak discrete feature set, substitutes the peak discrete feature set into the inner product kernel function of the support vector machine to perform spatial dimensionality increase, compares the mapping output algebraic value with the segmentation judgment benchmark threshold, extracts the corresponding array elements whose mapping output algebraic value is lower than the segmentation judgment benchmark threshold, and obtains the abnormal coordinate parameters. The sequence combination submodule performs one-dimensional rearrangement of the abnormal coordinate parameters according to the preset communication node encoding rules to construct a rearrangement queue. A timestamp identifier segment is pushed into the end of the rearrangement queue, and the transposed sequence element format is a single-column multi-row matrix to construct a multi-dimensional risk candidate vector.

[0011] As a further aspect of the present invention, the operation of substituting the peak discrete feature set into the inner product kernel function of the support vector machine to perform spatial dimension increase is to use the Gaussian radial basis kernel function as the inner product kernel function, calculate the high-dimensional Euclidean space absolute distance between any two feature data vectors in the peak discrete feature set, and perform nonlinear transformation mapping processing on the high-dimensional Euclidean space absolute distance value in combination with a preset constant-level exponential decay coefficient, and output a nonlinear mapping relationship matrix containing multiple row and column elements as the mapping output algebraic value.

[0012] As a further aspect of the present invention, the hidden danger detection module includes: The deviation calculation submodule calls the multidimensional risk candidate vector, calculates the difference between each element in the multidimensional risk candidate vector and the preset safety judgment benchmark threshold of the same dimension, and obtains the absolute value parameter of the difference. The absolute value parameter of the difference is normalized by the extreme value mapping proportional transformation algorithm to generate the multidimensional vector deviation degree. The node extraction submodule, for the multidimensional vector deviation, checks the corresponding three-dimensional tensor matrix index positions that are greater than the preset tolerance within the multidimensional vector deviation, converts the three-dimensional tensor matrix index positions into a real-world topological logical sequence according to the preset coordinate transformation rules, extracts the attribute data of the real-world topological logical sequence, and extracts the over-limit road node number. The hidden danger construction submodule, based on the over-limit road node number, matches the spatial coordinate parameter attached to the over-limit road node number, and performs one-way pointer merging and connection processing on the over-limit road node number and the spatial coordinate parameter to construct a set of accident hidden danger nodes.

[0013] As a further aspect of the present invention, the preset safety judgment benchmark threshold is determined by retrieving the element risk benchmark values ​​from a multidimensional risk sample dataset within a historical time period and calculating the sum of the arithmetic mean and standard deviation values.

[0014] As a further aspect of the present invention, the risk assessment module includes: The frequency analysis submodule extracts the internal frequency parameters of the preset sample test dataset based on the set of accident hazard nodes, establishes decision calculation nodes based on the sequence within the set of accident hazard nodes, substitutes the internal frequency parameters into the decision calculation nodes to calculate the Gini impurity and calculate the mean, and generates the accident frequency distribution coefficient. The state fusion submodule obtains environmental change parameters within the set of accident hazard nodes based on the accident frequency distribution coefficient. After normalizing the environmental change parameters, it performs a multiplication-addition fusion operation with the accident frequency distribution coefficient to generate a hazard state feature vector. The risk level labeling submodule calls the hazard status feature vector, performs inner product calculation between the hazard status feature vector and the preset risk assessment matrix to obtain the mapping algebraic value, performs segmented matching and truncation of the mapping algebraic value according to the preset interval judgment threshold, and generates risk level labels.

[0015] As a further aspect of the present invention, the step of calculating the inner product of the hidden danger state feature vector and the preset risk assessment matrix to obtain the mapping algebraic value refers to performing sequence slicing on the hidden danger state feature vector according to the dimension of the preset risk assessment matrix to generate state feature slices, multiplying the elements in the state feature slices with the elements at the corresponding positions of the preset risk assessment matrix to generate basic product values, and performing cumulative summation on all basic product values ​​to output the mapping algebraic value. The preset interval determination threshold is determined by extracting the standard deviation and mean of the distribution of the mapped algebraic values ​​of the preset sample test dataset to divide the numerical boundary.

[0016] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: By acquiring speed fluctuation characteristics within vehicle trajectories and combining them with vehicle counts for multi-source fusion to construct a regional traffic state matrix, and by obtaining topological weight parameters based on the physical attenuation characteristics of spatial straight-line distances between road network connection nodes and performing spatial weighting processing to construct a spatiotemporal topological connectivity tensor, the system analyzes the coordinates of multi-dimensional local peak points and spatial topological change rate parameters to extract abnormal coordinates and construct multi-dimensional risk candidate vectors. By comparing the deviation of candidate vectors from safety thresholds, the system extracts the numbers of over-limit road nodes and constructs a set of hidden danger nodes. By fusing historical accident frequencies with the current node hidden danger status, the system generates risk level labels, deeply mines multi-source dynamic data, strengthens global spatiotemporal correlation mapping, accurately identifies risk areas in the road network, and improves the accuracy of traffic safety assessment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0018] Figure 1 This is a system schematic diagram of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the feature acquisition module in this invention; Figure 4 This is a flowchart of the topology construction module in this invention; Figure 5 This is a flowchart of the feature extraction module in this invention; Figure 6 This is a flowchart of the hidden danger detection module in this invention; Figure 7 This is a flowchart of the risk assessment module in this invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] This invention provides a traffic accident risk assessment system based on big data, such as... Figure 1-2 The diagram shown illustrates a traffic accident risk assessment system based on big data. This system includes: The feature acquisition module acquires vehicle count and vehicle trajectory data based on intersection sensors and a big data platform, extracts speed fluctuation features within the vehicle trajectory data, and combines them with vehicle counts to perform multi-source feature fusion. The features are then combined according to spatial arrangement to construct a regional traffic state matrix. The topology construction module, based on the regional traffic state matrix, obtains topology weight parameters according to the physical attenuation characteristics of the spatial straight-line distance values ​​of road network connection nodes, performs spatial weighting processing on the elements in the regional traffic state matrix, and constructs a spatiotemporal topology connectivity tensor. The feature extraction module analyzes the coordinates of local peak points and the spatial topological change rate parameters in multiple dimensions within the spatiotemporal topological connectivity tensor, uses support vector machine for joint classification, extracts anomaly coordinate parameters, and constructs multidimensional risk candidate vectors. The hazard discovery module compares the deviation of multi-dimensional risk candidate vectors from preset safety judgment thresholds, extracts the node numbers of over-limit roads, and constructs a set of accident hazard nodes. The risk assessment module, based on a set of accident hazard nodes, uses a random forest algorithm to analyze the frequency of historical accidents and fuses the features with the hazard status of the current node to generate a risk level label.

[0022] The regional traffic status matrix includes road segment saturation, lane time occupancy rate, and traffic delay index. The spatiotemporal topological connectivity tensor includes road segment impedance, network connectivity, and spatial accessibility. The multidimensional risk candidate vector includes collision time, post-intrusion time, and velocity variance. The accident hazard node set includes nodes with excessive historical accident frequency, nodes with abnormal line-of-sight parameters, and nodes with excessive curvature radius. The risk level labels are specifically high-risk, medium-risk, and low-risk.

[0023] Specifically, such as Figure 2 , 3 As shown, the feature acquisition module includes: The trajectory feature analysis submodule acquires vehicle count and vehicle trajectory data based on intersection sensors and big data platform, extracts the historical driving speed sequence of a single vehicle in the vehicle trajectory data, calculates the discrete standard deviation of the historical driving speed sequence of a single vehicle within a set global observation time window, and generates speed fluctuation feature quantity. The trajectory feature analysis submodule calls the data access interface of the 77 GHz multi-transmitter, multi-receiver millimeter-wave radar sensor deployed at the intersection front end through the interconnection communication extension bus of the underlying hardware peripheral components. The internal read probe component continuously reads the original vehicle motion status message protocol data frames at a fixed sampling frequency of 10 Hz. The parity check logic gate removes lost data frames and abnormal parity check misalignment data frames at the communication layer according to the message protocol header. The interception component extracts the single-vehicle historical driving speed sequence data field encapsulated in the valid data frame, parses the 32-bit single-precision floating-point numerical parameters in the field that conform to the IEEE 754 standard, and the memory allocator pushes the parsed single-vehicle historical driving speed sequence data into the first-in-first-out (FIFO) mode level 1 dynamic random access memory queue in an ascending timestamp sequence. The register read component extracts the global observation time window configuration parameter pre-frozen in non-volatile memory, sets the specific value of this parameter to 60 seconds, and the high-frequency timing controller component triggers a high-level data full-load interrupt signal once every 60 seconds. The discrete measurement logic unit captures the interrupt signal and extracts 600 instantaneous vehicle speed values ​​accumulated in the memory queue. The floating-point accumulator configured in this logic unit summarizes all 600 instantaneous vehicle speed values ​​and divides them by a constant 600 to obtain the arithmetic mean. Importing underlying test case data as input parameters, the first instantaneous speed value is extracted as 12.5 m / s and the second as 13.2 m / s. After summing all 600 sample values, the accumulator performs a division operation to obtain an arithmetic mean of 12.8 m / s. The multiplication logic gate receives the difference between each instantaneous speed value output by the subtractor and the corresponding arithmetic mean value. This difference is multiplied by itself to generate a square value parameter. After accumulating all 600 square value parameters, it is divided by a constant 599. Finally, the square root of the quotient obtained through Newton's iterative square root operation core is used to generate the discrete standard deviation parameter. Substituting the first velocity value into the subtractor yields a difference of -0.3 m / s. This difference is then squared using a multiplication logic gate, resulting in a value of 0.09. The hardware calculation logic continuously accumulates and takes the square root to obtain a discrete standard deviation of 1.5 m / s. The feature encapsulation unit packages this 1.5 m / s into a general feature description structure and pushes it into the output buffer register queue as a velocity fluctuation feature.

[0024] The multi-source feature fusion submodule calls the speed fluctuation feature quantity, and converts the speed fluctuation feature quantity and vehicle count into a multi-dimensional data feature tensor of the same dimension. After assigning corresponding weight coefficients to the multi-dimensional data feature tensor of the same dimension, it performs weighted fusion to obtain the traffic feature fusion value. The multi-source feature fusion submodule retrieves the speed fluctuation feature data packet, and the external bus interface component synchronously connects to the traffic management center's cloud-based big data cluster platform interface to parse the vehicle count integer parameter field within a 5-minute fixed time slice in the distributed log table. The feature dimension alignment unit calls the perceptron feature space dimensionality-upgrading network architecture component embedded in the submodule. This network architecture component is physically mapped within a dedicated neural network acceleration chip. The network includes one input node layer that receives the underlying parameters, two consecutive hidden feature mapping operation layers, each containing 64 operation nodes, and one output mapping dimension expansion layer containing 128 operation nodes. The input node layer directly reads the 1-dimensional speed fluctuation feature and the 1-dimensional vehicle count integer parameter. The multiply-accumulate operator inside the hidden feature mapping operation layer performs linear combination mapping calculations on the input parameters through a pre-set fully connected weight numerical matrix. The threshold activation unit uses a modified linear unit activation function to perform non-linear truncation and filtering of negative elements on the linear calculation results. The network output mapping dimension extension layer generates a velocity feature tensor parameter with 128 elements and a vehicle count feature tensor parameter with the same 128 elements based on the final-level operation results. Through the aforementioned dimension alignment mapping hardware component, the velocity fluctuation feature and the vehicle count are transformed into multidimensional data feature tensors of the same dimension. The Hadamard product operator reads the two generated 128-dimensional feature tensor parameters. The arithmetic logic unit array assigns corresponding weight coefficients to the floating-point elements at the same dimension index position within the two feature tensor parameters according to the underlying vector processing parallel instruction set, and then synchronously performs a weighted fusion operation. Inputting local test sandbox verification example data, the specific value latched at the first-dimensional index position of the speed feature tensor parameter is 2.5, and the specific value latched at the first-dimensional index position of the vehicle count feature tensor parameter is 40.0. The fusion operator first assigns the corresponding weight coefficient of 0.8 to the speed feature tensor parameter to convert it to 2.0, and assigns the corresponding weight coefficient of 0.2 to the vehicle count feature tensor parameter to convert it to 8.0. Then, the multiplier multiplies 2.0 and 8.0 to obtain the traffic feature weighted fusion product value of 16.0. The internal traversal engine drives the arithmetic logic unit array to execute the same weighted fusion instruction in parallel on all 128 elements of the entire dimension. The output register buffers the 1-dimensional feature tensor parameter containing the traffic feature fusion values ​​of 128 product elements.

[0025] The state array construction submodule, for traffic feature fusion values, combines the traffic feature fusion values ​​in the horizontal plane row and column topology splicing according to the relative arrangement order of the two-dimensional absolute coordinates of the data monitoring nodes in the global spatial reference coordinate system to establish a regional traffic state matrix. The array initialization component within the state array construction submodule allocates a 2D grid data storage area in memory. The coordinate transformation parser obtains the latitude and longitude coordinates of the monitoring nodes within the global spatial reference coordinate system issued by the external reference geographic information system. The scale conversion unit projects the latitude and longitude coordinates into 2D absolute coordinate values ​​under the Gauss-Krüger projection plane according to a fixed mapping scale of 1:10000. The topology sorting engine reads the list of monitoring nodes in the global space and rearranges them in ascending order of memory pointers based on the X-axis values ​​of the 2D absolute coordinate values. If the X-axis values ​​are the same, they are sorted in ascending order based on the Y-axis values, generating a relative arrangement order table of the 2D absolute coordinates of the monitoring nodes. The spatial grid mapping logic gate reads the relative arrangement sequence table and establishes the initial empty array frame of the regional traffic state matrix with a grid dimension of 5 by 5. The address offset calculation unit maps the sequence number of the monitoring node in the arrangement sequence table to the corresponding matrix row and column index coordinate position, executes the horizontal row and column topology splicing combination allocation instruction, and directly copies the 1D feature tensor parameter of the traffic feature fusion value passed from the previous stage to the data node storage block at the corresponding row and column index coordinate position in the matrix. The memory block is called to copy the debugging example data parameter, and the converted 2D absolute coordinate parameter of a certain monitoring node is read. The horizontal X value is 5000 meters and the vertical Y value is 8000 meters. The topology sorting engine marks the node as the 7th sequence number according to its numerical arrangement. The address offset calculation unit allocates and maps the traffic feature fusion value corresponding to the 7th sequence number by dividing the number 7 by the number of matrix columns 5 and taking the remainder and quotient logic, and writes it into the memory address segment at the intersection of the 2nd row and 3rd column in the initial empty array frame. The data bus traverses and fills all 25 grid memory address segments according to the above allocation logic. The matrix encapsulation component extracts the starting address of the fully filled memory region and the data block length parameter, encapsulates and backs it up as a regional traffic status matrix storage object and registers it to the global variable address mapping table.

[0026] Specifically, such as Figure 2 , 4 As shown, the topology building module includes: The distance feature extraction submodule extracts the two-dimensional spatial positioning coordinates of the corresponding road network connection nodes based on the regional traffic state matrix, calculates the absolute Euclidean space straight-line span between the road network connection nodes, and generates a node straight-line distance array. The distance feature extraction submodule is equipped with a spatial coordinate resolver component, which reads the regional traffic state matrix storage object structure residing in memory. The graph structure extraction logic gate traverses the 25 grid memory address segments within the matrix, extracts the corresponding road network connection node identifier code encapsulated in the header of each data block, and sends a query payload containing this identifier code to the local 2D geographic information model cache. It matches and returns the corresponding 2D spatial positioning coordinate parameters, whose data structure includes floating-point X-coordinate and Y-coordinate values. The Euclidean distance measurement core operator receives two sets of 2D spatial positioning coordinate parameters corresponding to any two road network connection nodes. The subtraction logic gate within this core operator performs a subtraction operation on the X-coordinate values ​​and a subtraction operation on the Y-coordinate values ​​of the two sets of coordinate parameters, respectively. The multiplication logic gate multiplies these two independent differences by themselves to generate two squared values. The adder component accumulates these two squared values ​​to obtain the combined absolute spatial distance squared value. The square root operation component calls the underlying floating-point operation hardware instructions to perform a square root operation on the comprehensive value to extract the absolute Euclidean space linear span parameter. The array generation unit establishes an empty matrix of the same dimension based on the sequence index of each node in the original matrix, and fills the extracted span parameter into the corresponding cross-index coordinate cell to generate the node linear distance array parameter. Substituting the spatial ranging calculation example parameters, the X-coordinate value of the two-dimensional spatial positioning coordinate parameter corresponding to node sequence 1 is read as 100.0 meters and the Y-coordinate value is read as 200.0 meters. The X-coordinate value of the two-dimensional spatial positioning coordinate parameter corresponding to node sequence 2 is read as 130.0 meters and the Y-coordinate value is read as 240.0 meters. The subtraction logic gate calculates the difference in X coordinate as 30.0 meters and the difference in Y coordinate as 40.0 meters. The multiplication logic gates generate square values ​​of 900.0 and 1600.0 respectively. The adder components sum up to 2500.0. The square root operation component takes the square root to obtain the absolute Euclidean space linear span between nodes, which is 50.0 meters. The array is then filled into the first row and second column of the 2D empty matrix to complete the array construction.

[0027] The decay weight calculation submodule calls the node linear distance array, converts the matrix element values ​​in the node linear distance array into their reciprocals, sets the natural logarithm base as a constant and performs natural exponential decay scalar operation with the corresponding reciprocal values ​​to obtain the topological weight decay coefficient. The floating-point divider logic array inside the attenuation weight calculation submodule extracts the node linear distance array parameter storage block passed from the previous stage from the cache stack. The internal polling component reads the floating-point absolute Euclidean space linear span numerical parameter elements contained in the array storage block one by one. The floating-point divider logic array takes the constant 1.0 as the dividend input value and the read span numerical parameter elements as the divisor input value, performs high-precision floating-point division operation, generates the reciprocal parameter of the corresponding array element value and temporarily resides in the level 1 cache register. The logarithmic calculation logic gate calls the system kernel's underlying mathematical function library to extract the natural logarithm base parameter as the base constant for exponential operations. The exponential operation core component obtains this natural logarithm base parameter, converts the reciprocal parameter residing in the cache register into double-precision floating-point format, and then uses it as the exponent parameter for the natural exponential decay scalar operation, which is then fed into the exponential operation logic. The exponential decay arithmetic unit calculates the specified exponent value of the natural logarithm base parameter using Taylor series expansion hardware acceleration instructions. The truncation unit truncates this result value to a topology weight decay coefficient parameter retaining 4 decimal places and overwrites it back to the corresponding coordinate position in the original memory block. Substituting the decay arithmetic example data for hardware logic deduction, the polling component reads the span parameter of the distance from the first row and second column coordinate position in the array as 50.0 meters. The floating-point divider logic array divides the constant 1.0 by 50.0 to obtain the reciprocal parameter of 0.02. The logarithmic calculation logic gate extracts the natural logarithm base parameter constant value of approximately 2.718. The exponential operation core component receives the operation control instruction and converts the reciprocal parameter 0.02 into a negative value of -0.02 in its hardware sign bit logic, inputting it into the exponential term register. Then, it calculates the natural exponential decay scalar value of the natural logarithm base parameter 2.718 to the power of -0.02. The exponential operation core component outputs a calculation result of approximately 0.9802. The truncation unit encapsulates 0.9802 as the extracted topological weight decay coefficient parameter for backup. Based on the above parallel instructions, the reciprocal conversion and decay parameter generation and replacement mechanism of all matrix elements are completed.

[0028] The spatiotemporal tensor construction submodule performs a spatial weighted element-wise multiplication scalar operation on the topological weight decay coefficient and the corresponding values ​​of the corresponding mapping coordinates in the regional traffic state matrix. It then continuously performs three-dimensional topological structure expansion in the order of the timestamp sequence to construct a spatiotemporal topological connectivity tensor. The spatiotemporal tensor construction submodule is equipped with a parallel weighted multiplication logic processor component. This processor component synchronously reads feature matrix data blocks containing topological weight attenuation coefficient parameters and state value data blocks containing regional traffic state matrix parameters from the cache memory. The memory address mapping unit inside the processor component establishes a corresponding mapping coordinate association index table between the two data blocks. Based on the association index table, the corresponding storage node positions are located one by one according to the row and column 2D spatial orientation. The multiplication arithmetic logic gate performs a spatial weighted element-wise multiplication scalar operation on the value units located under each set of corresponding mapping coordinates, calculates the traffic state weighted value parameters at each spatial grid point, and writes them into a temporary 2D slice register block. The tensor sequence expansion controller listens to the global system clock component and collects a snapshot of the state data generated in the temporary 2D slice register block every 5 minutes. The 3D stitching engine allocates a contiguous 3D memory addressing space in the dynamic random access memory and continuously pushes the acquired 2D slice register block data into the 3D memory addressing space in ascending order of the timestamp sequence along the Z-axis time dimension to perform 3D topology expansion. It is set to collect a total of 12 timestamp slice data to form a complete tensor block, generating a 3D spatiotemporal topology connectivity tensor parameter. The hardware weighted operation test example parameter description is loaded. The memory address mapping unit is located at the same coordinate position in the 2nd row and 3rd column. The traffic feature fusion value parameter in the regional traffic state matrix is ​​read as 16.0. At the same time, the topology weight attenuation coefficient parameter is read as 0.9802. The multiplication arithmetic logic gate directly multiplies 16.0 and 0.9802 to obtain the spatially weighted traffic state value parameter 15.6832, which is locked in the 2D slice. The tensor sequence expansion controller stacks and merges 12 sets of 2D weighted array values, labeled as slice 1 to slice 12, along the depth pointer at 5-minute intervals. The 3D stitching engine finally outputs a spatiotemporal topological connectivity tensor parameter data packet object containing 25 spatial plane nodes and a depth of 12 dimensions.

[0029] Specifically, such as Figure 2 , 5 As shown, the feature extraction module includes: The parameter extraction submodule extracts the spatial location coordinates at the gradient flip point within the spatiotemporal topological connectivity tensor, extracts the displacement derivative datasets between adjacent slices, merges and calculates the distribution variance of the displacement derivative dataset, and concatenates the spatial location coordinates and distribution variance row by row to generate a peak discrete feature set. The parameter extraction submodule contains a 3D convolutional kernel gradient detector component. This detector component is equipped with a 3×3×3 Sobel differential operator. The operator performs a full 3D sliding traversal calculation with a step size of 1 within the spatiotemporal topological connectivity tensor parameter. The first-order partial derivative calculation logic gate captures the numerical derivative parameters of each spatial coordinate node within the tensor in the three dimensions of XYZ. The extremum detection comparator compares the sign bit data of adjacent coordinate derivative numerical parameters point by point. Once a sign reversal occurs, an interrupt record is triggered. The coordinate truncation unit extracts the spatially located coordinate parameter at the gradient reversal point of the trigger mark and caches it in the stack space. The slice difference analyzer retrieves two adjacent slice plane data storage blocks on the Z-axis of the time dimension at the trigger position. It performs a difference operation on the values ​​at the same 2D coordinate positions to extract a set of displacement derivative data parameters containing 5 elements. The variance calculation logic processor calculates the arithmetic mean baseline value parameter of each element within the set. The square difference accumulator calculates the sum of the squares of the differences between each element and the arithmetic mean baseline value parameter, and then divides it by the total number of elements (5) to obtain the distribution variance value of the set parameters. The splicing and merging module performs row-by-row pointer linking and splicing combination of the spatial positioning coordinate parameters and the distribution variance value in the memory linked list to establish a 1D data chain of peak discrete feature set. Input parameter calculation extracts examples for verification and analysis. The extreme value detection comparator detects a gradient flip at the position of the 1st row, 2nd column, and 3rd layer depth within the tensor when the derivative sign changes from positive to negative. The coordinate extraction unit records the spatial positioning coordinate parameters as 1 horizontal, 2 vertical, and 3 depth. The slice difference analyzer extracts the derived values ​​at the same position from adjacent slices and calculates the difference to obtain the displacement derivative data set parameters. Specific elements include 2.0, 4.0, 6.0, 8.0, and 10.0. The variance calculation logic processor calculates the sum of the above five floating-point values, 30.0, and divides it by 5 to obtain the arithmetic mean baseline value parameter 6.0. The square difference accumulator calculates the squares of each difference and accumulates them to 40.0. Dividing by 5, it obtains the distribution variance value of 8.0. The concatenation and merging module packages the coordinate labels 1, 2, and 3 with the variance value 8.0 and writes them into the peak discrete feature set 1D data chain data segment.

[0030] The projection classification submodule calls the peak discrete feature set, substitutes the peak discrete feature set into the inner product kernel function of the support vector machine to perform spatial dimensionality increase, compares the mapping output algebraic value with the segmentation judgment benchmark threshold, extracts the corresponding array elements whose mapping output algebraic value is lower than the segmentation judgment benchmark threshold, and obtains the abnormal coordinate parameters. The projection classification submodule incorporates a feature dimensionality upscaling hardware accelerator component. It reads the 1D data chain segment of the peak discrete feature set from the previous stage. The upscaling network unit uses the system kernel's pre-built Gaussian radial basis function calculation instruction set to calculate the high-dimensional Euclidean distance between any two feature data vectors within the peak discrete feature set. The nonlinear mapping transformation component performs a power operation on this absolute distance parameter based on a preset constant-level exponential decay coefficient, outputting a nonlinear mapping matrix storage block containing multiple row and column elements as the mapping output algebraic value. The system internally configures the corresponding arithmetic logic operation array using the formula: ; Calculate the algebraic value of the above mapping output. Wherein, The output represents the algebraic value of the mapping. This represents the total number of feature elements within the peak discrete feature set. The sequence index subscript representing the feature element. The first peak discrete feature set represents the first peak. Each feature element Representing the The baseline reference value corresponding to each feature element Representing the The projection weight coefficients corresponding to each feature element This represents the bias constant of the spatial mapping. An example parameter from algebraic operations is introduced for system demonstration, setting the total number of characteristic elements within the set. Set the sequence index subscript for each feature element. The corresponding specific numerical parameters Call the corresponding baseline reference value parameter preset in the reference library. Pre-defined projection weight coefficients Pre-adjusted spatial mapping constant bias At the physical execution level, the hardware components perform the algebraic derivation calculation process by directly calling the above formula: ; ; ; Finally, the mapping output algebraic value is generated. The value is 0.64. The comparison trigger then extracts the preset segmentation judgment benchmark threshold parameter in the register and sets it to 0.5. The hardware comparator performs a numerical comparison and finds that 0.64 is not lower than 0.5, so it judges that the value is out of bounds and directly discards the data pointer. If the mapping output algebraic value generated by another data stream measurement node is 0.4 after being calculated by the above formula, and is lower than the judgment threshold of 0.5, the data interceptor will forcibly intercept the memory segment of the corresponding array element, seal the output as an abnormal coordinate parameter sequence, and write it to the back-end memory for subsequent risk assessment.

[0031] The sequence combination submodule performs one-dimensional rearrangement of abnormal coordinate parameters according to the preset communication node encoding rules, constructs a rearrangement queue, pushes a timestamp identifier segment at the end of the rearrangement queue, transposes the sequence element format into a single-column multi-row matrix, and constructs a multi-dimensional risk candidate vector. The array recombination controller integrated within the sequence combination submodule is connected to the output endpoint of the abnormal coordinate parameter sequence. The encoding rule parsing engine reads the preset communication node encoding rule parameter file in the fixed storage unit. This encoding rule parameter file defines the pre- and post-order sorting logic conventions for the hardware address identification number and the corresponding parameter identifier bits. The 1D rearrangement logic unit extracts the embedded data bit identifier fragments of each abnormal coordinate parameter according to this logic convention. In the continuous linear dynamic random access memory space, it performs a 1D rearrangement pointer movement operation on the above-mentioned disordered parameter elements, and constructs a rearranged queue parameter linked list object with the head and tail connected according to the convention. The system's underlying real-time clock generator intercepts the current hardware crystal oscillator clock count value and converts it into a 64-bit integer timestamp identifier segment parameter data block conforming to the ISO 8601 standard. The stack pointer operator addresses and locates the end of the tail memory address of the aforementioned rearranged queue parameter linked list object, and uses a push instruction to directly push the 64-bit integer timestamp identifier segment parameter data block and connect it to the tail position of the linked list. The transpose transformation operation component reads the memory structure size of the entire single-row 1D linked list sequence including the timestamp. By modifying the matrix shape reset function instruction through the row and column step index parameters, it forcibly repositions the sequence element format, presenting it as a single-column, multi-row matrix arrangement from a 3D addressing perspective. The memory encapsulation module then packages the reshaped storage blocks into a multi-dimensional risk candidate vector feature structure set for output. Substituting the storage pointer control operation example data, the abnormal coordinate parameters extract internal data segments such as 0.4, 0.3, and 0.2. The 1D rearrangement logic unit rearranges these segments into a linear queue parameter of 0.2, 0.3, and 0.4 based on the encoding size order. The system's underlying real-time clock generator captures a timestamp identifier segment parameter with a numerical parameter of 1700000000. The stack pointer operator appends it to the end to form a 1-row, 4-column structure parameter with values ​​of 0.2, 0.3, 0.4, and 1700000000. The transpose deformation operation component modifies the read step index to 1, directly stretching and mapping the structure into a 4-row, 1-column morphological parameter structure matrix. Finally, an independent multidimensional risk candidate vector feature structure set is constructed and stored.

[0032] Specifically, such as Figure 2 , 6 As shown, the hazard detection module includes: The deviation calculation submodule calls the multidimensional risk candidate vector, calculates the difference between each element in the multidimensional risk candidate vector and the preset safety judgment benchmark threshold in the same dimension, and obtains the absolute value parameter of the difference. The absolute value parameter of the difference is normalized by the extreme value mapping proportional transformation algorithm to generate the multidimensional vector deviation. The deviation calculation submodule includes a memory data read / accessor component to call the previously generated multidimensional risk candidate vector feature structure set. The threshold calculation component connects to the historical sample database cluster server and retrieves a list of multidimensional risk sample datasets spanning the past 30 days. The data field extractor extracts the historical element risk benchmark numerical parameters from the list. The statistical analysis arithmetic processor calculates the arithmetic mean of all historical numerical parameters and obtains the standard deviation of the corresponding discrete features. The addition logic gate adds the arithmetic mean and standard deviation to obtain the sum, which is then fixed as the same-dimensional preset safety judgment benchmark threshold constant and latched into a fast register. The vector difference core processor reads the numerical parameters of each element in the multidimensional risk candidate vector, executes a subtraction operation command with the same-dimensional preset safety judgment benchmark threshold constant, and the absolute value function module converts the sign bit of the difference to positive to generate a set of absolute value parameters. The extreme value scanning probe traverses the above set to find the maximum and minimum extreme value parameters. The normalization proportional converter uses an extreme value mapping proportional conversion algorithm to forcibly linearly compress all the absolute value parameters of the difference to a floating-point value range of 0 to 1, generating the final multi-dimensional vector deviation array output data block. The parameter calculation module is called to perform a calculation example to sort out the execution logic. The threshold calculation component extracts the risk benchmark numerical parameters of 100 historical elements from the database list and substitutes them into the statistical analysis arithmetic processor. The summation and division by 100 calculate the arithmetic mean parameter as 5.0, and the square root is taken to calculate the standard deviation parameter as 1.5. The addition logic gate adds 5.0 and 1.5 to obtain the same dimension preset safety judgment benchmark threshold constant fixed as 6.5. The vector difference core processor reads the numerical parameter of a certain element in the multi-dimensional risk candidate vector as 4.0, and the difference is taken to obtain -2.5. The absolute value function module removes the negative sign and extracts the absolute value parameter of the difference as 2.5. The extreme value scanning probe locks the maximum extreme value parameter as 5.0 and the minimum extreme value parameter as 0.0 within the set. The normalized scaling converter subtracts 0.0 from the parameter 2.5 and divides it by the difference between the maximum extreme value 5.0 and 0.0 to calculate the scaling value of 0.5 and fills it into the deviation array to generate a standardized multidimensional vector deviation value.

[0033] The node extraction submodule, for the multidimensional vector deviation, checks the corresponding three-dimensional tensor matrix index positions that are greater than the preset tolerance within the multidimensional vector deviation, converts the three-dimensional tensor matrix index positions into the actual topological logical sequence according to the preset coordinate transformation rules, extracts the attribute data of the actual topological logical sequence, and extracts the node number of the over-limit road. The node extraction submodule is equipped with a spatial location tracking and locator component to receive the transmitted multidimensional vector deviation array numerical blocks. The internal data decoding engine strips the header file of the associated multidimensional spatial structure mapping parameters bound to the numerical blocks. Simultaneously, the tolerance loader retrieves a preset tolerance constant residing in a register. This preset tolerance is pre-calibrated based on the sum of the arithmetic mean of the deviations extracted by the system during a historical normal traffic flow window without accidents and twice the discrete standard deviation, serving as a physical hard threshold benchmark for filtering out conventional network fluctuation noise. The matrix pointer detector sequentially backtracks through the 3D matrix memory block reserved by the system, checking the 3D tensor matrix index coordinates corresponding to the feature values ​​within the multidimensional vector deviation array that exceed the preset tolerance. The coordinate transformation processor loads a preset coordinate transformation rule algorithm from the geographic information mapping matching dictionary parameter library. This rule algorithm has built-in linear interpolation transformation logic code and a local geographic offset correction parameter table. The field attribute mapper substitutes the obtained original abstract 3D tensor matrix index position coordinate values ​​into the above transformation logic to perform address permutation operations, completely transforming and outputting a field topological logical sequence structured parameter model that matches the real road network point sorting code. The attribute truncation component reads the information tree nodes embedded in this structured parameter model and performs a breadth-first traversal search. By comparing the characteristics of the data packet header flags, it forcibly truncates the set of field topological logical sequence attribute data parameters associated with the corresponding grid cell. The string matcher accurately parses the key field of the over-limit road node number, named using a specific integer or character combination specification, from this set of attribute data parameters and pushes it to the subsequent cache area. The spatial sequence addressing algorithm parameter calculation test process verification mechanism is initiated. The matrix pointer detector captures an abnormal multidimensional vector deviation greater than the preset tolerance, corresponding to a spatial node parameter recorded in the memory offset address with an X-axis of 5, a Y-axis of 10, and a Z-axis of 2. The coordinate transformation processor retrieves the baseline offset constant configuration within the preset coordinate transformation rules, setting it to X-axis +100 and Y-axis +200. The original index parameters are added together to calculate the actual coordinate projection parameters as X-axis 105 and Y-axis 210. Based on this mapping value, the corresponding real-world topological logical sequence is retrieved from the dictionary as the 3rd intersection of the southern section of the main urban road. The attribute extraction component extracts its associated information data segment, and the string matcher matches and extracts the out-of-limit road node number identifier parameter value of 105210992 generated under a specific naming rule and outputs it separately.

[0034] The hidden danger construction submodule, based on the over-limit road node number, matches the spatial coordinate parameter attached to the over-limit road node number, and performs one-way pointer merging and connection processing on the over-limit road node number and the spatial coordinate parameter to construct a set of accident hidden danger nodes; The hazard construction submodule is equipped with a doubly linked list construction engine component. The probe reading interface retrieves the passed-in over-limit road node number identifier parameter string from the pre-order buffer. The geospatial database connection driver constructs a standardized structured query language access instruction code block based on this identifier parameter and sends it to the remote or local map server query engine terminal. The matching and association retrieval mechanism retrieves the response data packet carrying a dataset of spatial coordinate parameters, including two-dimensional latitude and longitude floating-point values ​​bound to the over-limit road node number. The pointer merging operation logic unit instantiates a new structure storage node object in the memory heap area, fills the over-limit road node number data value into the front data slot of the object, and fills the corresponding parsed two-dimensional spatial coordinate parameter set into the back data slot of the object. The system utilizes a low-level memory management interface to allocate memory addresses for a singly linked list. An address reference assignment command is used to directly attach the head address pointer of the backend data slot to the offset region at the tail of the structure memory of the frontend data slot, completing the singly linked pointer merging and connection mechanism. The collection object manager appends the combined structure object nodes created in each loop to the full linked list pool manager. Finally, these are encapsulated, packaged, named, and stored as an independent dataset of accident hazard node parameters in non-volatile storage. The system lists node binding operations and calculates parameter combinations to reconstruct the system's execution state. The probe reading interface captures and retrieves the specific text value of the preceding over-limit road node number identifier parameter, which is 105210992. The geospatial database connection driver retrieves the corresponding longitude spatial coordinate parameter of 113.5 degrees and latitude spatial coordinate parameter of 23.6 degrees from the backend server. The pointer merging operation logic unit allocates a memory block at physical memory address 1000 to store the number 105210992, and stores the coordinate combination parameters 113.5 and 23.6 at physical memory address 1048. Then, it forcibly writes the physical memory address number 1048 into the control bits at the end of the memory block 1000 via an instruction to implement a reference, completing the underlying chaining operation. The collection object manager iteratively gathers and packages all such spatial mapping structure objects containing potential hazard characteristics to form a parameter pool for the overall set of accident hazard nodes.

[0035] Specifically, such as Figure 2 , 7 As shown, the risk assessment module includes: The frequency analysis submodule extracts the internal frequency parameters of the preset sample test dataset based on the set of accident hazard nodes, establishes decision calculation nodes based on the sequence within the set of accident hazard nodes, substitutes the internal frequency parameters into the decision calculation nodes to calculate the Gini impurity and calculate the mean, and generates the accident frequency distribution coefficient. The frequency analysis submodule's built-in data iteration scanning probe traverses and calls the overall parameter pool file of the accident hazard node set residing in the storage medium. The sample library interactive probe loads the preset sample test dataset parameter table from the built-in driver, and extracts the internal frequency parameter integer values ​​of each node in the test set according to the index number, based on past statistical periods. The decision tree construction engine instantiates and builds a decision computing node object cluster containing a multi-level branch logical operation tree in memory space for the sorted node sequence within the accident hazard node set. The impurity measurement logic extracts the internal frequency parameter integer value of a certain computing node, and calculates the Gini impurity parameter under the distribution of that category's proportion, which is calculated by subtracting the sum of the squares of all frequency proportion distributions from the constant 1. The accumulator traverses all decision computing node object clusters and collects various Gini impurity parameter values. Then, it triggers the division execution unit to divide by the total number of computing nodes to calculate and obtain the mean value of the Gini impurity parameter. The arithmetic processing unit performs a simple constant scaling correction on the above mean parameter value to generate a usable accident frequency distribution coefficient feature quantity output.

[0036] Table 1: Statistical Distribution of Testing Frequency for Accident Hazard Nodes

[0037] As shown in Table 1, the relevant internal frequency parameters of nodes 1 and 2 in the preset sample test data are extracted. For node 1, the internal frequency parameters are extracted with a category distribution ratio of 0.8 and 0.2, and their squares are calculated to be 0.64 and 0.04 respectively. The sum is 0.68, and the Gini impurity is obtained by subtracting using a constant of 1, which is 0.32. For node 2, the category distribution ratio is 0.5 and 0.5, the sum of squares is 0.50, and the Gini impurity is obtained by subtracting, which is 0.50. The accumulator accumulates the impurity value to 0.82 and divides it by the total number of nodes, 2, to obtain the Gini mean of 0.41. The scaling constant is set to 1, and the output value of the accident frequency distribution coefficient feature parameter is 0.41.

[0038] The state fusion submodule obtains environmental change parameters within the set of accident hazard nodes based on the accident frequency distribution coefficient. After normalizing the environmental change parameters, it performs a multiplication-addition fusion operation with the accident frequency distribution coefficient to generate a hazard state feature vector. The state fusion submodule integrates a parameter alignment fusion operation array component to read the accident frequency distribution coefficient feature parameter value file transmitted from the frequency analysis submodule. The meteorological and road condition environment probe interface module concurrently accesses the vehicle-road cooperative infrastructure server cache pool, extracting environmental variation parameter value packets containing rainfall and road friction coefficients at the corresponding coordinate locations based on the point information attached to the accident hazard node set. The normalization mapping converter loads the limit scaling algorithm model to perform a minimum-maximum linear transformation calculation on the environmental variation parameters, forcibly compressing them into a standard size range of 0 to 1 for normalization. The fusion multiply-add operator extracts the normalized environmental variation floating-point numbers and the input accident frequency distribution coefficient parameter floating-point values, calling the system's underlying tensor multiply-add operation instructions to perform a cross-multiply-add fusion operation mechanism involving the product of two data vectors at the same-dimensional node indices and the sum of constants. The vector encapsulation component writes the results into a 1D or 2D linear continuous memory address band according to node number order and adds data boundary verification markers, thereby outputting independent and regular hazard state feature vector matrix units for subsequent extraction. The following is a detailed explanation of the parameter data introduced in the fusion measurement process: The environmental probe interface module captures the original rainfall value of 40.0 mm from the environmental change parameter package of a certain node area. The normalization mapping converter is configured with a maximum tolerance range of 100.0 mm and a minimum of 0.0 mm for rainfall. Through division, the corresponding normalized floating-point value is obtained as 0.4. The fusion multiply-add operator reads the accident frequency distribution coefficient parameter value obtained from the pre-calculation, which is 0.41. The multiplier inside the operator first multiplies 0.4 and 0.41 to obtain the primary product parameter 0.164. In the hardware constant register, the configuration constant 1.0 is called to perform the addition operation to obtain the feature value 1.164. The vector encapsulation component packages the sequence containing multiple similar calculation result data values ​​such as 1.164 into a standard set of hidden danger status feature vector parameter datasets according to storage blocks.

[0039] The risk level labeling submodule calls the hazard status feature vector, performs inner product calculation between the hazard status feature vector and the preset risk assessment matrix to obtain the mapping algebraic value, performs segmented matching and truncation of the mapping algebraic value according to the preset interval judgment threshold, and generates risk level labels. The level calibration submodule calls the hazard state feature vector parameter dataset matrix unit block residing in the high-speed cache. The matrix inner product hardware calculation core calls the preset risk assessment matrix parameter architecture form pre-stored in read-only memory to perform the corresponding inner product operation to obtain the mapping algebraic value. The inner product execution flow controller determines the specific dimensions contained in the preset risk assessment matrix and guides the slice segmenter to perform memory segmentation operation to extract state feature slice array elements of equal length dimension from the hazard state feature vector. The multiplication arithmetic logic unit extracts data elements from the state feature slices and the preset risk assessment matrix at the same address offset index, performs parallel multiplication calculation to generate a series of basic product value parameter arrays. The accumulator core adds all the basic product values ​​in the array to obtain a single scalar mapping algebraic value output. The statistical calculation unit pulls a massive amount of historical mapping algebraic values ​​from the preset sample test dataset and calculates the distribution standard deviation and distribution mean values ​​of the set. The threshold setter defines different interval ranges by superimposing and subtracting the standard deviation with the mean as the center, and generates a preset interval judgment threshold parameter table. The tag assigner compares the real-time mapping algebraic value with the threshold parameter table to locate the interval range, extracts the identification code data attached to the interval, and finally encapsulates and replaces it with risk level tag information in text type for sending and display.

[0040] Table 2: Reference Table for Risk Level Certification Parameter Testing

[0041] As shown in Table 2, the inner product hardware calculation core reads the feature slices within the vector containing the numerical elements 2.0 and 3.0. The corresponding numerical elements in the risk assessment matrix are 4.0 and 5.0. The multiplier multiplies 2.0 by 4.0 to obtain the basic product value 8.0, and multiplies 3.0 by 5.0 to obtain 15.0. The accumulator sums the results to calculate the measured mapped algebraic value of 23.0. The statistical calculation unit tests the historical mean parameter, which is set to 20.0, and the distribution standard deviation parameter is set to 5.0. The threshold setter establishes the interval preset interval judgment threshold with a lower limit of 15.0 and an upper limit of 25.0, marking it as a medium-level code. The label assigner determines that 23.0 falls within the interval, thus matching and extracting the medium-level code and converting it into a medium-risk level label parameter packet for output to the alarm system.

[0042] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A traffic accident risk assessment system based on big data, characterized in that, The system includes: The feature acquisition module acquires vehicle count and vehicle trajectory data based on intersection sensors and a big data platform, extracts speed fluctuation features within the vehicle trajectory data, and combines them with vehicle counts to perform multi-source feature fusion. The features are then combined according to spatial arrangement to construct a regional traffic state matrix. The topology construction module, based on the regional traffic state matrix, obtains topology weight parameters according to the physical attenuation characteristics of the spatial straight-line distance values ​​of road network connection nodes, performs spatial weighting processing on the elements in the regional traffic state matrix, and constructs a spatiotemporal topology connectivity tensor. The feature extraction module analyzes the coordinates of local peak points and spatial topological change rate parameters in multiple dimensions within the spatiotemporal topological connectivity tensor, uses support vector machine for joint classification, extracts abnormal coordinate parameters, and constructs multidimensional risk candidate vectors. The hidden danger discovery module compares the deviation of the multidimensional risk candidate vector from the preset safety judgment threshold, extracts the node numbers of the over-limit roads, and constructs a set of accident hidden danger nodes; The risk assessment module, based on the set of accident hazard nodes, uses a random forest algorithm to analyze the frequency of historical accidents, and performs feature fusion with the hazard status of the current node to generate a risk level label.

2. The traffic accident risk assessment system based on big data according to claim 1, characterized in that, The regional traffic state matrix includes road segment saturation, lane time occupancy rate, and traffic delay index; the spatiotemporal topological connectivity tensor includes road segment impedance, network connectivity, and spatial reachability; the multidimensional risk candidate vector includes collision time, post-intrusion time, and velocity variance; the accident hazard node set includes nodes with excessive historical accident frequency, nodes with abnormal line-of-sight parameters, and nodes with excessive curvature radius; and the risk level label specifically includes high-risk level, medium-risk level, and low-risk level.

3. The traffic accident risk assessment system based on big data according to claim 1, characterized in that, The feature acquisition module includes: The trajectory feature analysis submodule acquires vehicle count and vehicle trajectory data based on intersection sensors and big data platform, extracts the historical driving speed sequence of a single vehicle in the vehicle trajectory data, calculates the discrete standard deviation of the historical driving speed sequence of a single vehicle within a set global observation time window, and generates speed fluctuation feature quantity. The multi-source feature fusion submodule calls the speed fluctuation feature quantity, converts the speed fluctuation feature quantity and vehicle count into a multi-dimensional data feature tensor of the same dimension, assigns corresponding weight coefficients to the multi-dimensional data feature tensor of the same dimension, and then performs weighted fusion to obtain the traffic feature fusion value. The state array construction submodule, for the traffic feature fusion values, performs horizontal row and column topological splicing and combination of the traffic feature fusion values ​​according to the relative arrangement order of the two-dimensional absolute coordinates of the data monitoring nodes in the global spatial reference coordinate system, to establish a regional traffic state matrix.

4. The traffic accident risk assessment system based on big data according to claim 1, characterized in that, The topology construction module includes: The distance feature extraction submodule extracts the two-dimensional spatial positioning coordinates of the corresponding road network connection nodes based on the regional traffic state matrix, calculates the absolute Euclidean space straight-line span values ​​between the road network connection nodes, and generates a node straight-line distance array. The decay weight calculation submodule calls the node linear distance array, converts the matrix element values ​​in the node linear distance array into their reciprocals, sets the natural logarithm base as a constant and performs natural exponential decay scalar operation with the corresponding reciprocal values ​​to obtain the topological weight decay coefficient. The spatiotemporal tensor construction submodule performs a spatially weighted, element-wise multiplication scalar operation on the topological weight decay coefficient and the corresponding values ​​of the corresponding mapped coordinates in the regional traffic state matrix. It then continuously performs three-dimensional topological structure expansion in the order of the timestamp sequence to construct a spatiotemporal topological connectivity tensor.

5. The traffic accident risk assessment system based on big data according to claim 1, characterized in that, The feature extraction module includes: The parameter extraction submodule extracts the spatial positioning coordinates of the gradient flip point within the spatiotemporal topological connectivity tensor, extracts the displacement derivative datasets between adjacent slices, merges and calculates the distribution variance of the displacement derivative dataset, and concatenates the spatial positioning coordinates and distribution variance row by row to generate a peak discrete feature set. The projection classification submodule calls the peak discrete feature set, substitutes the peak discrete feature set into the inner product kernel function of the support vector machine to perform spatial dimensionality increase, compares the mapping output algebraic value with the segmentation judgment benchmark threshold, extracts the corresponding array elements whose mapping output algebraic value is lower than the segmentation judgment benchmark threshold, and obtains the abnormal coordinate parameters. The sequence combination submodule performs one-dimensional rearrangement of the abnormal coordinate parameters according to the preset communication node encoding rules to construct a rearrangement queue. A timestamp identifier segment is pushed into the end of the rearrangement queue, and the transposed sequence element format is a single-column multi-row matrix to construct a multi-dimensional risk candidate vector.

6. The traffic accident risk assessment system based on big data according to claim 5, characterized in that, The operation of substituting the peak discrete feature set into the inner product kernel function of the support vector machine to perform spatial dimension increase involves using the Gaussian radial basis kernel function as the inner product kernel function, calculating the high-dimensional Euclidean space absolute distance between any two feature data vectors in the peak discrete feature set, and performing nonlinear transformation mapping on the high-dimensional Euclidean space absolute distance value in combination with a preset constant-level exponential decay coefficient, outputting a nonlinear mapping relationship matrix containing multiple row and column elements as the mapping output algebraic value.

7. The traffic accident risk assessment system based on big data according to claim 1, characterized in that, The hazard detection module includes: The deviation calculation submodule calls the multidimensional risk candidate vector, calculates the difference between each element in the multidimensional risk candidate vector and the preset safety judgment benchmark threshold of the same dimension, and obtains the absolute value parameter of the difference. The absolute value parameter of the difference is normalized by the extreme value mapping proportional transformation algorithm to generate the multidimensional vector deviation degree. The node extraction submodule, for the multidimensional vector deviation, checks the corresponding three-dimensional tensor matrix index positions that are greater than the preset tolerance within the multidimensional vector deviation, converts the three-dimensional tensor matrix index positions into a real-world topological logical sequence according to the preset coordinate transformation rules, extracts the attribute data of the real-world topological logical sequence, and extracts the over-limit road node number. The hidden danger construction submodule, based on the over-limit road node number, matches the spatial coordinate parameter attached to the over-limit road node number, and performs one-way pointer merging and connection processing on the over-limit road node number and the spatial coordinate parameter to construct a set of accident hidden danger nodes.

8. The traffic accident risk assessment system based on big data according to claim 7, characterized in that, The preset safety judgment benchmark threshold is determined by retrieving the element risk benchmark values ​​from the multidimensional risk sample dataset within a historical time period and calculating the sum of the arithmetic mean and standard deviation values.

9. The traffic accident risk assessment system based on big data according to claim 1, characterized in that, The risk assessment module includes: The frequency analysis submodule extracts the internal frequency parameters of the preset sample test dataset based on the set of accident hazard nodes, establishes decision calculation nodes based on the sequence within the set of accident hazard nodes, substitutes the internal frequency parameters into the decision calculation nodes to calculate the Gini impurity and calculate the mean, and generates the accident frequency distribution coefficient. The state fusion submodule obtains environmental change parameters within the set of accident hazard nodes based on the accident frequency distribution coefficient. After normalizing the environmental change parameters, it performs a multiplication-addition fusion operation with the accident frequency distribution coefficient to generate a hazard state feature vector. The risk level labeling submodule calls the hazard status feature vector, performs inner product calculation between the hazard status feature vector and the preset risk assessment matrix to obtain the mapping algebraic value, performs segmented matching and truncation of the mapping algebraic value according to the preset interval judgment threshold, and generates risk level labels.

10. The traffic accident risk assessment system based on big data according to claim 9, characterized in that, The step of calculating the mapping algebraic value by performing an inner product calculation between the hidden danger state feature vector and the preset risk assessment matrix refers to performing sequence slicing on the hidden danger state feature vector according to the dimension of the preset risk assessment matrix to generate state feature slices, multiplying the elements in the state feature slices with the elements at the corresponding positions in the preset risk assessment matrix to generate basic product values, and performing cumulative summation on all basic product values ​​to output the mapping algebraic value. The preset interval determination threshold is determined by extracting the standard deviation and mean of the distribution of the mapped algebraic values ​​of the preset sample test dataset to divide the numerical boundary.

Images