A highway multi-modal space-time risk situation awareness research and judgment method and system

Abstract

The application discloses a kind of expressway multi-modal space-time risk situation awareness and research methods and system, it is related to expressway risk research technical field, including: parallel acquisition multi-modal sensing parameter and mapping to geographical grid;By double space-time mask traffic state perception algorithm extraction traffic state tensor;By risk factor gradient coupling analytical model output initial risk potential value;By dynamic road network topology iterative reasoning function update risk potential space distribution;By multilevel risk threshold determination generates structured situation message;By global road network risk situation research platform rendering generates visual view.The application realizes the accurate perception and dynamic deduction of global risk situation of expressway by speed density gradient coupling analysis and dynamic topology iteration propagation, improves the accuracy and timeliness of risk research.

Description

Technical Field

[0001] This invention relates to the field of highway risk assessment technology, and in particular to a method and system for assessing multimodal spatiotemporal risk situation on highways. Background Technology

[0002] The highway traffic environment is complex, with traffic flow conditions influenced by multiple modal factors such as traffic density, vehicle speed distribution, weather conditions, and road surface conditions. Local disturbances can easily propagate rapidly within the road network and evolve into cascading traffic accidents or regional congestion risks. As the highway network continues to expand, operational risks exhibit characteristics such as wide spatial and temporal distribution, numerous coupling factors, and rapid evolution. This places extremely high demands on the real-time perception and accurate assessment of risk situations, necessitating a risk situation perception technology solution that can integrate multi-source heterogeneous sensor data while taking into account spatiotemporal dynamic correlations.

[0003] Existing highway risk monitoring systems mainly collect parameters such as cross-sectional flow, speed, and occupancy through fixed roadside detectors. After the collected data is aggregated to the central platform, statistical regression models or basic graph traffic flow theory are used to determine the traffic status of road sections. In addition, meteorological monitoring station data is combined to issue threshold alarms for road section risks under severe weather conditions. Some advanced systems introduce graph neural networks or convolutional long short-term memory networks to model the spatiotemporal characteristics of traffic flow under the road network topology. By training historical data, the system learns the mapping relationship from input parameters to risk levels, and finally pushes the judgment results to the graphical interface of the monitoring center for management personnel to refer to.

[0004] Existing technologies have the following drawbacks: First, existing solutions often process traffic flow parameters and meteorological environmental parameters separately, lacking quantitative analysis of the nonlinear coupling relationship between the velocity gradient field and the density gradient field, making it difficult to accurately characterize the gradual accumulation effect of risk potential energy during the transition of traffic flow from smooth to congested. Second, most existing dynamic road network risk propagation methods adopt static propagation rules under a fixed topology, failing to dynamically adjust the risk transmission weights between road network nodes based on real-time road closures, accidents blocking lanes, and other emergencies, resulting in a significant deviation between the risk propagation projection in the road network and the actual traffic situation. Summary of the Invention

[0005] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides a method and system for assessing and evaluating multimodal spatiotemporal risk situations on highways.

[0006] The technical solution adopted in this invention is a method for multimodal spatiotemporal risk situation perception and assessment of highways, comprising the following steps: S1, using multimodal sensing devices deployed on the roadside and on vehicles to collect traffic flow parameters, meteorological environmental parameters, and road surface condition parameters across the entire highway area in parallel, mapping the collected parameters to a unified geographic grid coordinate system after spatiotemporal registration, and constructing a multimodal spatiotemporal situation basic dataset; S2, inputting the multimodal spatiotemporal situation basic dataset into a dual spatiotemporal mask traffic flow state perception algorithm, applying a sliding window mask in the time dimension to capture the temporal evolution rhythm of traffic flow, and applying an adjacency graph mask in the spatial dimension to extract the spatial dependency correlation between road segments, and outputting the traffic flow state tensor of each grid unit through dual-mask interactive inference; S3, feeding the traffic flow state tensor into a risk factor gradient coupling analytical model, and analyzing the relationship between the velocity gradient field and the density gradient field along the traffic flow propagation direction. The nonlinear coupling relationship is used to weight and accumulate risk factors based on the gradient coupling strength, and the initial risk potential energy value of each grid unit is output; S4, the initial risk potential energy value is input into the dynamic road network topology iterative inference function, the adjacency weight matrix is ​​dynamically updated according to the real-time road network connectivity status, and the risk potential energy is propagated between road network nodes through topology iterative inference until the change in node risk potential energy converges to below the preset threshold, and the updated risk potential energy spatial distribution is output; S5, the updated risk potential energy spatial distribution is superimposed with the static road network attribute layer, and a risk level label is generated through multi-level risk threshold judgment, and the risk level label and the corresponding spatiotemporal coordinates and traffic flow parameters are encapsulated into a structured situation message; S6, the structured situation message is received by the full-domain road network risk situation assessment platform, and the platform parses the message and renders it on the digital twin map interface to generate a full-domain risk situation visualization view.

[0007] Furthermore, the expression for the dual spatiotemporal mask traffic flow state perception algorithm is as follows: ,in For the first Hidden state matrix of vehicle flow layers The output after masked interactive inference is the first... Hidden state matrix of vehicle flow layers It is a non-linear activation function. This represents the total number of mask types. For the first The attention weight vector of the class mask, where ⊙ is the Hadamard multiplication operator. For the first Normalized adjacency matrix under the effect of class masking For the first The learnable weight matrix corresponding to the class mask; a separate temporal mask matrix is ​​also provided. and spatial mask matrix ,Will and Through fusion function Coupled generation of joint mask tensor ,by right Dynamic modulation is performed, and the modulation relationship is as follows: ,in This is a tensor element-level modulation operator. It is the time-varying adjacency matrix after joint mask tensor modulation.

[0008] Furthermore, the expression for the risk factor gradient coupling analytical model is as follows: ,in For the first Velocity density gradient coupling strength at each grid cell For the first The velocity gradient vector at each grid cell For the first The density gradient vector at each grid cell is defined by the vector dot product operator; a separate set of risk factors is also defined. Corresponding to each meteorological and environmental parameter, based on the aforementioned coupling strength Constructing a risk weight function The initial risk potential value is the scaling factor. The calculation relationship is as follows ,in For the first The spatial integration region encompassed by each grid cell For the area of ​​a microelement, The total number of risk factors. For the first The measured values ​​of each risk factor. , .

[0009] Furthermore, the expression for the dynamic road network topology iterative inference function is: ,in For the first Node at the next iteration The risk potential value, For the first Node at the next iteration The risk potential value, This is the iteration step size factor. For nodes The set of adjacent nodes, and Both are adjacency sets Node index in For nodes With nodes Real-time network impedance distance between them For nodes With nodes Real-time network impedance distance between them Here, is the distance decay coefficient, is the product operator, exp is the natural exponential function, and the iteration continues until... The process terminates when the value falls below the preset threshold.

[0010] Furthermore, the multimodal sensing device includes a millimeter-wave radar deployed on a gantry along the highway, an optical video acquisition unit deployed on the roadside, and a piezoelectric dynamic weighing sensor deployed inside the road surface; the detection range of the millimeter-wave radar is set to 150 meters to 250 meters, the sampling frame rate of the optical video acquisition unit is set to 15 frames / second to 30 frames / second, and the sampling frequency of the piezoelectric dynamic weighing sensor is set to 500 Hz to 1000 Hz; the grid cell size of the geographic grid coordinate system is set to a 500-meter × 500-meter square grid; the time window width of the sliding window mask is... The interval is set to 5 to 15 minutes, and the window sliding step is set to 1 to 3 minutes; the spatial adjacency order of the adjacency graph mask is set to 1st to 3rd order adjacency; the multi-level risk threshold includes three levels: low risk threshold, medium risk threshold, and high risk threshold. The low risk threshold is the 30th percentile value of the risk potential value distribution, the medium risk threshold is the 60th percentile value, and the high risk threshold is the 90th percentile value; the structured situation report is encapsulated in JSON format, including timestamp field, spatial coordinate field, risk level field, and traffic flow parameter field.

[0011] Further, S2 includes the following sub-steps: S21, extracting the time series components from the multimodal spatiotemporal situational baseline dataset, performing segmentation and truncation operations on the time series components using the sliding window mask, and inputting the truncated sub-sequences of each time period into a gated recurrent unit to extract temporal latent features; S22, extracting the spatial distribution components from the multimodal spatiotemporal situational baseline dataset, masking the information transmission channels between non-adjacent nodes using the adjacency graph mask, retaining only the spatial feature aggregation path between adjacent nodes, and performing graph convolution operations to extract spatial latent features; S23, performing a concatenation operation on the feature dimension of the temporal latent features and the spatial latent features, mapping the concatenated joint feature tensor to a preset latent space dimension through a linear projection layer to generate a temporal-spatial fusion feature; S24, performing alternating iterative interactive reasoning of the temporal-spatial fusion feature using the dual mask, alternately applying the time mask and the spatial mask in each iteration until the number of iterations reaches a preset upper limit, and outputting the traffic flow state tensor.

[0012] Further, S3 includes the following sub-steps: S31, calculating the partial derivatives of the velocity component and density component along the flow direction of each grid cell in the traffic flow state tensor, and organizing the partial derivatives of the velocity component into a velocity gradient field and the partial derivatives of the density component into a density gradient field; S32, performing a vector dot product operation on the velocity gradient field and the density gradient field at the same grid cell, using the dot product result as the gradient coupling strength of that grid cell, and traversing all grid cells to generate a gradient coupling strength distribution map; S33, substituting the coupling strength values ​​of each grid cell in the gradient coupling strength distribution map into a nonlinear mapping function, and generating risk factor weight coefficients for each grid cell after transformation by the nonlinear mapping function; S34, multiplying the risk factor weight coefficients of each grid cell by the measured values ​​of various risk factors of the corresponding grid cell, summing the multiplications, and outputting the initial risk potential value of each grid cell.

[0013] Further, S4 includes the following sub-steps: S41, constructing a dynamic adjacency matrix based on the real-time road network connectivity status, setting the adjacency elements corresponding to road segments closed due to traffic control or accidents to zero, and assigning the adjacency elements corresponding to normally passable road segments the reciprocal of the real-time road network impedance; S42, using the initial risk potential energy value as the initial risk potential energy value of each node, and using the dynamic adjacency matrix as the weight base, spreading the current risk potential energy value of each node to its adjacent nodes according to the weight ratio; S43, in each iteration step, accumulating the spread potential energy received by each node from all adjacent nodes, and then replacing the current risk potential energy value of the node with the accumulated result to complete the synchronous update of all nodes; S44, calculating the maximum value of the absolute difference between the risk potential energy values ​​of all nodes between the current iteration step and the previous iteration step. If the maximum value is higher than the preset threshold, S42 to S43 are repeated. If the maximum value is lower than or equal to the preset threshold, the iteration is terminated and the updated risk potential energy spatial distribution is output.

[0014] Further, S5 includes the following sub-steps: S51, performing a weighted superposition operation on the risk potential energy values ​​of each grid unit in the updated risk potential energy spatial distribution with the road curvature attributes and longitudinal slope attributes in the pre-stored static road network attribute layer to generate a composite risk potential energy value; S52, comparing the composite risk potential energy value with the low-risk threshold, medium-risk threshold, and high-risk threshold respectively, and determining that each grid unit belongs to one of the low-risk, medium-risk, or high-risk levels based on the comparison results; S53, filling the determined risk level label, the current time stamp, the central geographic coordinates of each grid unit, and the corresponding speed parameters, density parameters, and flow parameters into the fields of the preset message template; S54, serializing the filled content into the structured situation message according to JSON syntax rules, and pushing the structured situation message to the full-domain road network risk situation assessment platform through the message queue protocol.

[0015] A multimodal spatiotemporal risk situation awareness and assessment system for highways is provided. This system is applied to a multimodal spatiotemporal risk situation awareness and assessment method for highways, including: a multimodal sensor parameter spatiotemporal registration and grid mapping unit, a dual spatiotemporal mask traffic flow state tensor inference unit, a velocity density gradient coupled risk potential energy analysis unit, a dynamic topology iteration risk potential energy update unit, a multi-level risk situation structured message generation unit, and a digital twin map risk situation visualization rendering unit.

[0016] The multimodal sensing parameter spatiotemporal registration and grid mapping unit is used to map traffic flow parameters, meteorological environment parameters and road surface condition parameters collected by multi-source sensors to the geographic grid coordinate system after spatiotemporal alignment. The output end is connected to the input end of the dual spatiotemporal mask vehicle flow state tensor inference unit.

[0017] The dual spatiotemporal masking traffic flow state tensor inference unit is used to apply a mask in the dual dimensions of temporal window and spatial adjacency and perform multi-layer interactive inference on the implicit state of traffic flow to generate a traffic flow state tensor. The output end is connected to the input end of the velocity density gradient coupled risk potential energy analysis unit.

[0018] The velocity-density gradient coupled risk potential energy analysis unit is used to analyze the initial risk potential energy value of each grid cell by weighting and accumulating various risk factors based on the dot product coupling strength of the velocity gradient field and the density gradient field. The output end is connected to the input end of the dynamic topology iteration risk potential energy update unit.

[0019] The dynamic topology iteration risk potential energy update unit is used to dynamically update the adjacency weight matrix according to the real-time road network connectivity status and iteratively propagate the risk potential energy between road network nodes until convergence to output the updated risk potential energy spatial distribution. The output end is connected to the input end of the multi-level risk situation structured message generation unit.

[0020] The multi-level risk situation structured message generation unit is used to overlay the spatial distribution of risk potential energy with static road network attributes and generate risk level labels through multi-level threshold determination. The labels, spatiotemporal coordinates, and traffic flow parameters are then encapsulated into structured situation messages. The output end is connected to the input end of the digital twin map risk situation visualization rendering unit.

[0021] The digital twin map risk situation visualization rendering unit is used to receive and parse the structured situation message, mark the risk level of each grid cell with different color levels on the digital twin map interface, and output the marked map view to the display terminal.

[0022] The present invention has the following beneficial effects:

[0023] By collecting multimodal parameters of traffic flow, meteorological environment, and road surface conditions in parallel and mapping them to a unified geographic grid, this approach addresses the lack of spatiotemporal collaborative registration of multi-source heterogeneous data in existing technologies, providing a high-precision basic dataset for risk analysis. Utilizing a dual spatiotemporal masking traffic flow state perception algorithm, it simultaneously extracts the rhythm of traffic flow evolution and spatial dependencies between road segments in both temporal sliding windows and spatial adjacency dimensions, overcoming the shortcomings of traditional methods in adequately characterizing the spatiotemporal coupling features of traffic flow. Furthermore, by using a risk factor gradient coupling analytical model along the traffic flow propagation direction, it analyzes the nonlinear dot product coupling relationship between the velocity gradient field and the density gradient field, and based on this, it applies risk factors to various risk analysis methods. The risk factors are weighted and accumulated, solving the problem that existing technologies neglect the coupling effect of speed density gradient and thus cannot quantify the gradual accumulation process of risk. The dynamic road network topology iterative inference function dynamically updates the adjacency weights and iteratively propagates the risk potential energy based on the real-time road network connectivity status, solving the drawback of the existing static topology propagation rules being unable to adapt to sudden changes in road conditions, which leads to inaccurate risk prediction. The whole-domain road network risk situation assessment platform renders structured risk level labels and spatiotemporal coordinates on the digital twin map interface, enabling managers to intuitively obtain the overall picture of risk distribution across the entire domain, improving the accuracy and timeliness of highway risk situation perception. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0025] Figure 2 This is a flowchart of method step S2 of the present invention;

[0026] Figure 3 This is a flowchart of method step S3 of the present invention;

[0027] Figure 4 This is a flowchart of method step S4 of the present invention;

[0028] Figure 5 This is a flowchart of step S5 of the method of the present invention;

[0029] Figure 6 This is a diagram showing the system unit composition of the present invention. Detailed Implementation

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] like Figure 1 As shown, a method for assessing and evaluating multimodal spatiotemporal risk situation on highways includes the following steps: S1, using multimodal sensing devices deployed on the roadside and on vehicles to collect traffic flow parameters, meteorological environmental parameters, and road surface condition parameters across the entire highway area in parallel, mapping the collected parameters to a unified geographic grid coordinate system after spatiotemporal registration, and constructing a multimodal spatiotemporal situation basic dataset; S2, inputting the multimodal spatiotemporal situation basic dataset into a dual spatiotemporal masking traffic flow state perception algorithm, applying a sliding window mask in the time dimension to capture the temporal evolution rhythm of traffic flow, and applying an adjacency graph mask in the spatial dimension to extract the spatial dependency correlation between road segments, and outputting the traffic flow state tensor of each grid unit through dual-mask interactive inference; S3, feeding the traffic flow state tensor into a risk factor gradient coupling analytical model, and analyzing the nonlinear coupling between the velocity gradient field and the density gradient field along the traffic flow propagation direction. The relationship is established, and the risk factors are weighted and accumulated according to the gradient coupling strength to output the initial risk potential energy value of each grid unit; S4, the initial risk potential energy value is input into the dynamic road network topology iterative inference function, and the adjacency weight matrix is ​​dynamically updated according to the real-time road network connectivity status. The risk potential energy is propagated between road network nodes through topology iterative inference until the change in node risk potential energy converges to below the preset threshold, and the updated risk potential energy spatial distribution is output; S5, the updated risk potential energy spatial distribution is superimposed with the static road network attribute layer, and a risk level label is generated through multi-level risk threshold judgment. The risk level label and the corresponding spatiotemporal coordinates and traffic flow parameters are encapsulated into a structured situation message; S6, the structured situation message is received by the full-domain road network risk situation assessment platform. The platform parses the message and renders it on the digital twin map interface to generate a full-domain risk situation visualization view.

[0032] When implementing step S1: A gantry is installed every 500 to 800 meters along the highway, equipped with a millimeter-wave radar with a detection range of 150 to 250 meters, a horizontal angular resolution of 1.2 degrees, and a vertical angular resolution of 2.4 degrees. It outputs traffic flow, instantaneous speed, and vehicle type classification data, with a sampling period of 100 milliseconds. An optical video acquisition unit is installed every 200 meters along the roadside, with a sampling frame rate of 25 frames per second and an image resolution of 1920 pixels × 1080 pixels. Vehicle targets are extracted and their trajectories are tracked using background subtraction. A set of piezoelectric dynamic weighing sensors is buried every kilometer inside the road surface. The sensor, with a sampling frequency set to 800 Hz, outputs axle load and vehicle speed data. The aforementioned multimodal sensor data is transmitted to an edge computing node via optical fiber. The edge computing node has an embedded GPS timing module, and each sensor data is stamped with a unified timestamp, with a time synchronization accuracy of 1 millisecond. The edge computing node divides the acquired data into 500m × 500m geographic grid units, maps the data to the corresponding grid according to the latitude and longitude coordinates of the installation location of each sensor device, and uses bilinear interpolation to perform spatial registration for the data at the grid boundary. Finally, a multimodal spatiotemporal situational awareness dataset covering the entire 200km to 300km section of the road is generated.

[0033] In step S2: The edge computing node reads the time series components from the multimodal spatiotemporal situational awareness dataset. The time window width is set to 10 minutes, and the window sliding step size is set to 2 minutes. Within each window, the velocity, density, and flow time series data are segmented and truncated. Each segment is set to 60 sampling points. Each segment subsequence is input into a gated recurrent unit. The hidden layer dimension is set to 128 dimensions, and the temporal latent features are output. Simultaneously, the edge computing node extracts the spatial distribution components. The spatial adjacency order is set to second-order adjacency, meaning that each grid only interacts with adjacent grids within a 1500-meter radius. The information transmission channels of non-adjacent nodes are masked, and the data is preserved. The adjacent paths are subjected to graph convolution operations, with 3 layers and 64-dimensional output for each layer, outputting spatial latent features. Edge computing nodes concatenate the temporal and spatial latent features in terms of feature dimensions, resulting in a 192-dimensional joint feature, which is then mapped to a 256-dimensional latent space via a linear projection layer. Alternating iterative interactive inference is then performed, with temporal and spatial masks applied alternately. In each iteration, the temporal mask covers the data for 3 minutes before and after the current time window, while the spatial mask covers nodes outside the second-order adjacency. The number of iterations is set to 8. After each iteration, the joint features are processed by an activation function, ultimately outputting a 256-dimensional traffic flow state tensor.

[0034] In step S3: The edge computing node extracts the velocity component of each grid cell from the traffic flow state tensor along the traffic flow direction. It calculates the partial derivative of this velocity component along the flow direction, with a difference step size of 500 meters. The partial derivative values ​​of all grid cells are then processed to form the velocity gradient field. The edge computing node also extracts the density component from the same traffic flow state tensor. It calculates the partial derivative of this density component along the flow direction, with a difference step size of 500 meters. The partial derivative values ​​of all grid cells are then processed to form the density gradient field. At each grid cell, the edge computing node performs a vector dot product operation on the velocity gradient field value and the density gradient field value. The sign of the dot product result indicates whether the velocity and density change in the same or opposite direction. The absolute value... To characterize the coupling strength, a gradient coupling strength distribution map is generated by traversing all grid cells. The coupling strength values ​​range from negative 1.5 to positive 1.5. The edge computing nodes substitute the coupling strength values ​​of each grid cell into a nonlinear mapping function. This function is in the form of exponential decay, with a scale adjustment coefficient of 0.8. After conversion, risk factor weight coefficients are generated for each grid cell, ranging from 0.1 to 1.0. The edge computing nodes multiply the risk factor weight coefficients by the measured values ​​of temperature, rainfall intensity, and road friction coefficient of the corresponding grid cell, respectively. The product results are accumulated grid by grid, and the initial risk potential energy value of each grid cell is output after accumulation. The initial risk potential energy value ranges from 0 to 100.

[0035] During step S4: Edge computing nodes obtain the road network control instruction set from the traffic control center in real time. Based on the accident-closed and construction-controlled road sections marked in the instruction set, a dynamic adjacency matrix is ​​constructed. The matrix dimension is equal to the total number of grid cells. For closed and controlled road sections, the corresponding adjacent elements are set to 0. For normally open road sections, the corresponding adjacent elements are assigned the reciprocal of the real-time road network impedance. The road network impedance is calculated by dividing the average travel speed of the road section by the road section length. The average travel speed is taken from the average radar detection value within the previous 5 minutes. The edge computing nodes use the initial risk potential energy value as the initial risk potential energy value of each node. Using the dynamic adjacency matrix as the weight base, each node diffuses its own risk potential energy to adjacent nodes according to the weight ratio. Thirty percent; In each iteration step, each node receives the diffusion potential energy from all neighboring nodes, sums it up, and adds it to seventy percent of the node's original risk potential energy to obtain the updated risk potential energy value of the node. All nodes synchronously perform the above update operation; The iteration step size factor is 0.5, and the distance decay coefficient is 0.3. After each iteration, the maximum value of the absolute difference of the risk potential energy of all nodes between the current iteration step and the previous iteration step is calculated. The preset convergence threshold is 0.001. When the maximum value is higher than 0.001, the iteration is repeated. When the maximum value is lower than or equal to 0.001, the iteration is terminated. The number of iterations is usually between 15 and 25. The updated risk potential energy spatial distribution is output.

[0036] When implementing step S5: The edge computing node reads the pre-stored static road network attribute layer, which includes the road curvature radius value and longitudinal slope percentage value of each grid cell. The curvature radius value ranges from 200 meters to 2000 meters, and the longitudinal slope percentage value ranges from -5% to +5%. The edge computing node multiplies the risk potential energy value of each grid cell in the updated risk potential energy spatial distribution by a curvature correction factor of 0.3, and then multiplies it by a slope correction factor of 0.2. The multipliers are then added to the original risk potential energy value to generate a composite risk potential energy value. The edge computing node statistically analyzes the distribution of composite risk potential energy values ​​of all grid cells, calculates the 30th percentile of the distribution as the low-risk threshold, the 60th percentile as the medium-risk threshold, and the 90th percentile as the high-risk threshold, and performs operations on the composite risk potential energy value of each grid cell with the three thresholds respectively. Numerical comparisons are used to classify areas as low-risk (below the low-risk threshold), medium-risk (between the low-risk and medium-risk thresholds), high-risk (above the high-risk threshold), and medium-high-risk (between the medium-risk and high-risk thresholds). Edge computing nodes input the determined risk level label along with the current UTC timestamp, the latitude and longitude coordinates of each grid cell center, average speed, average density, and average traffic flow into a JSON message template. The JSON message template includes four primary fields: timestamp, coordinates, risk level, and traffic flow. Each primary field contains several secondary subfields. After inputting the data, the JSON object is serialized into a string-based situational message, which is then pushed to the full-domain road network risk situation assessment platform via the MQTT protocol. The MQTT service quality level is set to level 1.

[0037] During step S5: The full-domain road network risk assessment platform is deployed on a central server cluster, which includes 3 application servers and 1 database server. The servers are interconnected via 10 Gigabit Ethernet. The platform continuously listens to the MQTT protocol port (port number 1883). After receiving structured situational messages pushed by edge computing nodes, it calls a JSON parsing library to perform deserialization on the messages, extracting the risk level labels and latitude / longitude coordinates of each grid unit. Upon startup, the platform pre-loads a digital twin map base for the entire highway. This map base is built based on OpenStreetMap vector data, including vector layers such as lane lines, guardrails, gantries, and toll stations. The map coordinate system uses WGS84 latitude / longitude projection, with a scaling level set to dynamically adapt from 15 to 18 levels. The platform then uses the solution... The obtained latitude and longitude coordinates are used to locate the positions of each grid cell on the digital twin map interface. High-risk grid cells are marked with red (RGB values ​​set to 255, 0, 0), medium-high risk grid cells with orange (RGB values ​​set to 255, 165, 0), medium risk grid cells with yellow (RGB values ​​set to 255, 255, 0), and low risk grid cells with green (RGB values ​​set to 0, 255, 0). The size of each grid cell is set to 500 meters × 500 meters, the transparency is set to 0.7, and the edges of the cells are outlined with a white border with a width of 2 pixels. After marking, the complete map view is output to the display terminal of the management center at a refresh rate of 30 frames per second. The display terminal resolution is 3840 pixels × 2160 pixels.

[0038] Preferably, the expression for the dual spatiotemporal mask traffic flow state perception algorithm is: ,in For the first Hidden state matrix of vehicle flow layers The output after masked interactive inference is the first... Hidden state matrix of vehicle flow layers It is a non-linear activation function. This represents the total number of mask types. For the first The attention weight vector of the class mask, where ⊙ is the Hadamard multiplication operator. For the first Normalized adjacency matrix under the effect of class masking For the first The learnable weight matrix corresponding to the class mask; a separate temporal mask matrix is ​​also provided. and spatial mask matrix ,Will and Through fusion function Coupled generation of joint mask tensor ,by right Dynamic modulation is performed, and the modulation relationship is as follows: ,in This is a tensor element-level modulation operator. It is the time-varying adjacency matrix after joint mask tensor modulation.

[0039] Specifically, the dual spatiotemporal masking traffic flow state perception algorithm is built on a graph neural network message passing framework. It abstracts the highway network as a topological graph composed of nodes and edges, where each grid cell corresponds to a graph node. The edge weights between nodes are determined based on actual road connectivity and geographical distance. The propagation of traffic flow state between adjacent nodes follows a diffusion process. Therefore, spatial dimension feature aggregation is achieved by multiplying the adjacency matrix and the hidden state matrix, followed by multiplication by a learnable weight matrix to complete the linear transformation of the feature space. The total number of mask types, K, is set to 4, corresponding to four categories: forward temporal mask, reverse temporal mask, first-order spatial adjacency mask, and second-order spatial adjacency mask. Each type of mask independently acts on the adjacency matrix to generate different normalized adjacency matrices. The attention weight vector is a 4-dimensional vector, with each component value learned by the network's self-attention mechanism. Its value stabilizes between 0.15 and 0.45 after training convergence. The Hadamard quadrature operation applies the attention weights element-wise. Feature recalibration is achieved by multiplying the aggregation result. The ReLU function is chosen as the activation function σ to introduce nonlinear expressive power. The dimension of the time mask matrix is ​​set to 60 x 60, corresponding to 60 time segments divided into 2-minute sliding steps within 10 minutes. The dimension of the spatial mask matrix is ​​set to N x N, where N equals the total number of grid cells in the road network. For a highway with a length of 200 kilometers, the total number of grid cells N equals 800. The fusion function Φ uses bilinear interpolation to upgrade the time mask matrix and the spatial mask matrix to the same tensor dimension, then performs element-wise addition and compresses them to the 0 to 1 interval using the sigmoid function to generate a joint mask tensor with a dimension of N x N x 60. This tensor is used to dynamically modulate the normalized adjacency matrix in the first formula. The modulation operator is tensor element-wise multiplication, which makes each element of the adjacency matrix change dynamically with different time segments, thereby giving the algorithm the ability to dynamically capture the spatiotemporal coupling features of traffic flow.

[0040] Preferably, the expression for the risk factor gradient coupling analytical model is: ,in For the first Velocity density gradient coupling strength at each grid cell For the first The velocity gradient vector at each grid cell For the first The density gradient vector at each grid cell is defined by the vector dot product operator; a separate set of risk factors is also defined. Corresponding to each meteorological and environmental parameter, based on the aforementioned coupling strength Constructing a risk weight function The initial risk potential value is the scaling factor. The calculation relationship is as follows ,in For the first The spatial integration region encompassed by each grid cell For the area of ​​a microelement, The total number of risk factors. For the first The measured values ​​of each risk factor. , .

[0041] Specifically, the risk factor gradient coupling analytical model is based on the constitutive relationship between velocity potential and density field in fluid mechanics. In highway traffic flow, the velocity gradient represents the spatial rate of change of vehicle flow acceleration or deceleration, while the density gradient represents the spatial rate of change of vehicle flow aggregation or sparsity. The physical meaning of their dot product operation is to quantify the consistency between the direction of velocity change and the direction of density change. When the velocity decreases along the flow direction while the density increases along the flow direction, a negative dot product indicates that the vehicle flow is in a compressed state, and the risk accumulation trend is increasing. When the dot product is positive, it indicates that the vehicle flow is in an expanded state, and the risk tends to dissipate. The differential operator calculates the partial derivative with a step size of 500 meters. Both the velocity gradient and the density gradient are discretized and calculated grid by grid after being calculated with this step size. The dot product result is directly used as the gradient coupling strength output. The second formula is based on the spatial weighted integration of risk factors. The risk factor set includes temperature, rainfall intensity, and road conditions. The surface friction coefficient consists of three factors, each with a clearly defined measured value. The risk weighting function adopts a negative exponential square form, with a scale adjustment coefficient of 0.8. This function reaches a maximum of 1.0 when the coupling strength is zero, and monotonically decreases to near zero as the absolute value of the coupling strength increases. Its physical meaning is that the greater the coupling strength, the more severe the compression or expansion of the traffic flow, and the lower the risk weight should be, while the actual risk should be higher. Therefore, this weighting function is used as a decay factor to weight the risk factors and then sum them. Then, an area integration operation is performed within the spatial integration area of ​​the grid cell. The integration area is a 500m by 500m square area, and the area micro-element is 1 square meter. The summation operation accumulates the measured values ​​of the three types of risk factors. The integration operation aggregates the weighted risk values ​​of each point in space into the overall initial risk potential value of the grid cell, with a value range between 0 and 100.

[0042] Preferably, the expression for the dynamic road network topology iterative inference function is: ,in For the first Node at the next iteration The risk potential value, For the first Node at the next iteration The risk potential value, This is the iteration step size factor. For nodes The set of adjacent nodes, and Both are adjacency sets Node index in For nodes With nodes Real-time network impedance distance between them For nodes With nodes Real-time network impedance distance between them Here, is the distance decay coefficient, is the product operator, exp is the natural exponential function, and the iteration continues until... The process terminates when the value falls below the preset threshold.

[0043] Specifically, the dynamic road network topology iterative inference function is based on the numerical solution of the heat diffusion equation on a discrete graph. The propagation of risk potential energy between road network nodes follows anisotropic diffusion, meaning that risk potential energy flows from high-value nodes along the road network edge to low-value nodes. The flow rate is controlled by a negative exponential function of the road network impedance distance between nodes. The derivation process starts from the continuous form of the diffusion equation, expressing the spatial Laplace operator in a discretized form on the graph as a weighted sum of the potential energy differences between adjacent nodes. Then, the explicit Euler forward difference method is used to discretize the time partial derivative. The value of the iteration step size factor is constrained by the numerical stability condition. In the road network topology graph, a value of 0.5 ensures the convergence of the iteration process. The distance attenuation coefficient is taken as the negative first power of 0.3 km, and its value range is set between 0.2 and 0.5 based on the risk influence radius of highway traffic flow. This coefficient is multiplied by the road network impedance distance between nodes and then exponentially reduced. When the impedance distance is... The exponent value is 0.74 at 1 km and 0.22 at 5 km, ensuring that nearby nodes receive a larger diffusion weight and distant nodes receive a smaller diffusion weight. The normalized denominator is summed over the weights of all adjacent nodes, making the total diffusion weight from each node to its adjacent nodes equal to 1, satisfying the mass conservation constraint of risk potential energy during propagation. In each iteration step, each node receives diffusion potential energy from all adjacent nodes. This diffusion amount is equal to the difference between the current potential energy of the adjacent nodes and the current potential energy of the node itself, multiplied by the normalized diffusion weight, and then multiplied by the iteration step size factor. The potential energy increment of the node is obtained by accumulating the diffusion amounts of each adjacent node. The increment is added to the original potential energy of the node to complete one iteration update of the node. The preset convergence threshold is 0.001. The iteration terminates when the maximum value of the absolute difference of potential energy between all nodes in two iterations is lower than this threshold. The number of iterations is usually between 15 and 25.

[0044] Preferably, the multimodal sensing device includes a millimeter-wave radar deployed on a gantry along the highway, an optical video acquisition unit deployed on the roadside, and a piezoelectric dynamic weighing sensor deployed inside the road surface; the detection range of the millimeter-wave radar is set to 150 meters to 250 meters, the sampling frame rate of the optical video acquisition unit is set to 15 frames / second to 30 frames / second, and the sampling frequency of the piezoelectric dynamic weighing sensor is set to 500 Hz to 1000 Hz; the grid cell size of the geographic grid coordinate system is set to a 500-meter × 500-meter square grid; and the time window width of the sliding window mask is specified. The timeout is set to 5 to 15 minutes, and the window sliding step is set to 1 to 3 minutes; the spatial adjacency order of the adjacency graph mask is set to 1st to 3rd order adjacency; the multi-level risk threshold includes three levels: low risk threshold, medium risk threshold, and high risk threshold. The low risk threshold is the 30th percentile value of the risk potential value distribution, the medium risk threshold is the 60th percentile value, and the high risk threshold is the 90th percentile value; the structured situation report is encapsulated in JSON format, including timestamp field, spatial coordinate field, risk level field, and traffic flow parameter field.

[0045] Preferred, such as Figure 2 As shown, S2 includes the following sub-steps: S21, extracting the time series components from the multimodal spatiotemporal situational baseline dataset, performing segmentation and truncation operations on the time series components using the sliding window mask, and inputting the truncated subsequences of each time period into a gated recurrent unit to extract temporal latent features; S22, extracting the spatial distribution components from the multimodal spatiotemporal situational baseline dataset, masking the information transmission channels between non-adjacent nodes using the adjacency graph mask, retaining only the spatial feature aggregation path between adjacent nodes, and performing graph convolution operations to extract spatial latent features; S23, performing a concatenation operation on the feature dimension of the temporal latent features and the spatial latent features, mapping the concatenated joint feature tensor to a preset latent space dimension through a linear projection layer to generate a temporal-spatial fusion feature; S24, performing alternating iterative interactive reasoning of the temporal-spatial fusion feature using the dual mask, alternately applying the time mask and the spatial mask in each iteration until the number of iterations reaches a preset upper limit, and outputting the traffic flow state tensor.

[0046] Specifically, in step S21, the edge computing node extracts three time-series components—velocity, density, and flow—from the multimodal spatiotemporal situational dataset. Each component has a sampling interval of 100 milliseconds, and each component includes 6000 sampling points within a 10-minute time window. The sliding window step size is set to 2 minutes, allowing adjacent windows to overlap by 8 minutes to maintain the continuity of temporal features. A segmentation operation divides the data within each window into 100 segments of 60 sampling points each. Each segment is input into a gated recurrent unit (ROU), with the hidden layer dimension set to 128. The bias terms of the reset and update gates are initialized to 0.1. Forward propagation extracts the temporal latent features of each segment. In step S22, the edge computing node extracts the spatial distribution component from the dataset. The spatial adjacency order is set to 2, meaning each grid cell only establishes adjacency with grids within a 1000-meter radius. In step S23, temporal and spatial latent features are concatenated along their feature dimensions, resulting in a 192-dimensional joint feature. This joint feature is then mapped to a 256-dimensional latent space via a linear projection layer. The weight matrix of the linear projection layer is uniformly initialized using Xavier. In step S24, temporal and spatial masks are applied alternately. In each iteration, the temporal mask covers the data segments before and after the current time, while the spatial mask covers all nodes except those in the second-order adjacency. The number of iterations is preset to 8. After each iteration, the ReLU activation function is applied, and the output traffic flow state tensor has a dimension of 800 x 256.

[0047] Preferred, such as Figure 3 As shown, step S3 includes the following sub-steps: S31, calculate the partial derivatives of the velocity component and density component along the flow direction of each grid cell in the traffic flow state tensor, and organize the partial derivatives of the velocity component into a velocity gradient field and the partial derivatives of the density component into a density gradient field; S32, perform a vector dot product operation on the velocity gradient field and the density gradient field at the same grid cell, and use the dot product result as the gradient coupling strength of the grid cell, and traverse all grid cells to generate a gradient coupling strength distribution map; S33, substitute the coupling strength value of each grid cell in the gradient coupling strength distribution map into a nonlinear mapping function, and generate the risk factor weight coefficient of each grid cell after transformation by the nonlinear mapping function; S34, multiply the risk factor weight coefficient of each grid cell by the measured values ​​of various risk factors of the corresponding grid cell, and output the initial risk potential value of each grid cell after accumulating the multiplications.

[0048] Specifically, in step S31, the edge computing nodes extract the velocity and density components of each grid cell from the traffic flow state tensor along the highway's driving direction. The velocity component is measured in kilometers per hour, and the density component is measured in vehicles per kilometer. First-order partial derivatives are calculated for both the velocity and density components along the flow direction. The difference step size is set to 500 meters based on the grid size. The partial derivatives are calculated using a central difference scheme, which divides the difference between the values ​​at 250 meters downstream and 250 meters upstream of the current grid by 500 meters. The velocity partial derivatives of all grid cells are then used to obtain the velocity gradient field, and the density partial derivatives of all grid cells are used to obtain the density gradient field. The gradient values ​​are on the order of vehicles per square kilometer per hour. In step S32, the scalar values ​​of the velocity gradient field and the density gradient field are multiplied at the same grid cell. The sign of the product reflects the direction of velocity change and density change. The direction of coupling is determined by the relative orientation of the coupling elements. The absolute value reflects the strength of the coupling effect. A gradient coupling strength distribution map is generated by traversing all 800 grid cells. The coupling strength values ​​are statistically within the range of -1.2 to +1.2. In step S33, the coupling strength values ​​of each grid cell are substituted into a nonlinear mapping function in the form of exponential decay. The scale adjustment coefficient is set to 0.8. When the coupling strength is zero, the mapping output value is 1.0. When the absolute value of the coupling strength is 1.2, the mapping output value decays to 0.24, generating the risk factor weight coefficient for each grid cell. In step S34, the weight coefficient is multiplied by the measured values ​​of temperature, rainfall intensity, and road friction coefficient of each grid cell. The product is accumulated grid by grid. The measured temperature value is in degrees Celsius, the rainfall intensity value is in millimeters per hour, and the road friction coefficient is dimensionless. After accumulation, the initial risk potential energy value is output.

[0049] Preferred, such as Figure 4 As shown, S4 includes the following sub-steps: S41, constructing a dynamic adjacency matrix based on the real-time road network connectivity status, setting the adjacency elements corresponding to road segments closed due to traffic control or accidents to zero, and assigning the adjacency elements corresponding to normally passable road segments the reciprocal of the real-time road network impedance; S42, using the initial risk potential energy value as the initial risk potential energy value of each node, and using the dynamic adjacency matrix as the weight base, spreading the current risk potential energy value of each node to its adjacent nodes according to the weight ratio; S43, in each iteration step, accumulating the spread potential energy received by each node from all adjacent nodes, and then replacing the current risk potential energy value of the node with the accumulated result to complete the synchronous update of all nodes; S44, calculating the maximum value of the absolute difference of the risk potential energy values ​​of all nodes between the current iteration step and the previous iteration step. If the maximum value is higher than the preset threshold, S42 to S43 are repeated. If the maximum value is lower than or equal to the preset threshold, the iteration is terminated and the updated risk potential energy spatial distribution is output.

[0050] Specifically, in step S41, edge computing nodes obtain control command sets in real time from the traffic control center via dedicated links. These commands include the accident chainage range, information on closed lanes due to construction, and estimated recovery time. Based on this information, the corresponding grid cell numbers for the accident and construction sections are identified. All rows and columns corresponding to these numbers in the dynamic adjacency matrix are set to 0. For normally passable road sections, the corresponding adjacent elements are assigned the reciprocal of the real-time road network impedance for that section. The road network impedance is obtained by dividing the length of the road section by the average travel speed of the vehicles traveling on that section in the previous 5 minutes. The length is in kilometers, the speed is in kilometers per hour, and the impedance is in hours per kilometer. In step S42, the initial risk potential energy value is used as the initial potential energy value for each node in the 0th iteration, with the dynamic adjacency matrix as the diffusion weight base. Each node distributes 30% of its current risk potential energy to all neighboring nodes according to the weight ratio. In step S43, each node sums up the potential energy diffused from all neighboring nodes, and then adds the sum to 70% of the node's original potential energy. The result is used as the potential energy value of the node in the new round of iteration. All nodes complete this update process synchronously. In step S44, after each iteration, the absolute difference of the risk potential energy of all 800 nodes between the current iteration step and the previous iteration step is calculated. The maximum value is taken and compared with the preset convergence threshold of 0.001. If the maximum value is greater than 0.001, S42 to S43 are repeated. If the maximum value is less than or equal to 0.001, the iteration process is terminated immediately, and the updated risk potential energy spatial distribution is output.

[0051] Preferred, such as Figure 5 As shown, step S5 includes the following sub-steps: S51, performing a weighted superposition operation on the risk potential energy values ​​of each grid unit in the updated risk potential energy spatial distribution with the road curvature attributes and longitudinal slope attributes in the pre-stored static road network attribute layer to generate a composite risk potential energy value; S52, comparing the composite risk potential energy value with the low-risk threshold, medium-risk threshold, and high-risk threshold respectively, and determining that each grid unit belongs to one of the low-risk, medium-risk, or high-risk levels based on the comparison results; S53, filling the determined risk level label, the current time stamp, the central geographic coordinates of each grid unit, and the corresponding speed parameters, density parameters, and flow parameters into the fields of the preset message template; S54, serializing the filled content into the structured situation message according to JSON syntax rules, and pushing the structured situation message to the global road network risk situation assessment platform through the message queue protocol.

[0052] Specifically, in step S51, the edge computing node reads the static road network attribute layer. This attribute layer is pre-stored offline on the edge computing node's local solid-state drive, including the road curvature radius value of each grid cell in meters, and the longitudinal slope percentage value. The curvature radius value ranges from 200 meters to 2000 meters, and the longitudinal slope percentage value ranges from -5% to +5%. The updated risk potential energy value is added to the inverse of the curvature radius multiplied by 0.3, and then added to the absolute value of the longitudinal slope multiplied by 0.2 to generate a composite risk potential energy value. In step S52, the edge computing node sorts the composite risk potential energy values ​​of all 800 grid cells in ascending order, and calculates the 240th value after sorting as the low-risk threshold. The 480th value is used as the medium-risk threshold, and the 720th value is used as the high-risk threshold. The composite risk potential value of each grid unit is compared with the above three thresholds to determine the risk level of each grid. In step S53, the risk level label, the current UTC timestamp, the latitude and longitude coordinates of the center of each grid unit, the average speed, the average density, and the average flow are filled into the corresponding fields of the JSON message template. In step S54, the filled content is serialized into a status message in string form according to the JSON syntax rules. During the serialization process, floating-point numbers are retained to two decimal places. The message is pushed to the whole-domain road network risk status assessment platform through the MQTT protocol. The keep-alive interval of the MQTT session is set to 60 seconds.

[0053] like Figure 6 As shown, a multimodal spatiotemporal risk situation awareness and assessment system for highways is applied to a multimodal spatiotemporal risk situation awareness and assessment method for highways. The system includes: a multimodal sensor parameter spatiotemporal registration and grid mapping unit, a dual spatiotemporal mask traffic flow state tensor inference unit, a velocity density gradient coupled risk potential energy analysis unit, a dynamic topology iteration risk potential energy update unit, a multi-level risk situation structured message generation unit, and a digital twin map risk situation visualization rendering unit.

[0054] The multimodal sensing parameter spatiotemporal registration and grid mapping unit is used to map traffic flow parameters, meteorological environment parameters, and road surface state parameters collected by multi-source sensors to the geographic grid coordinate system after spatiotemporal alignment. Its output is connected to the input of the dual spatiotemporal masking vehicle flow state tensor inference unit. The dual spatiotemporal masking vehicle flow state tensor inference unit applies a mask in both temporal and spatial adjacency dimensions and performs multi-layer interactive inference on the implicit state of the vehicle flow to generate a vehicle flow state tensor. Its output is connected to the input of the velocity density gradient coupling risk potential energy analysis unit. The velocity density gradient coupling risk potential energy analysis unit is used to weight and accumulate various risk factors based on the dot product coupling strength of the velocity gradient field and the density gradient field to analyze the initial risk potential energy value of each grid cell. Its output is connected to the input of the dynamic topology iteration risk potential energy update unit. The topology iterative risk potential energy update unit is used to dynamically update the adjacency weight matrix based on the real-time road network connectivity status and iteratively propagate the risk potential energy among road network nodes until convergence, outputting the updated risk potential energy spatial distribution. The output end is connected to the input end of the multi-level risk situation structured message generation unit. The multi-level risk situation structured message generation unit is used to overlay the risk potential energy spatial distribution with static road network attributes, generate risk level labels through multi-level threshold judgment, and encapsulate the labels, spatiotemporal coordinates, and traffic flow parameters into a structured situation message. The output end is connected to the input end of the digital twin map risk situation visualization rendering unit. The digital twin map risk situation visualization rendering unit is used to receive and parse the structured situation message, mark the risk level of each grid unit on the digital twin map interface with different color levels, and output the marked map view to the display terminal.

[0055] A method and system for assessing multimodal spatiotemporal risk situation on highways is proposed. By constructing a unified geographic grid coordinate system, multimodal sensing parameters are collected in parallel and spatiotemporally registered, fully exploring the spatiotemporal collaborative information between traffic flow, weather, and road surface conditions, avoiding the one-sidedness caused by independent judgment of a single parameter. A dual spatiotemporal masking mechanism applies constraints simultaneously in the time and spatial dimensions, ensuring that the extraction process of the traffic flow state tensor takes into account both the temporal evolution rhythm and the road network topology dependency, significantly improving the accuracy of traffic flow state characterization. A gradient-coupled analytical model directly analyzes the nonlinear coupling strength between the velocity field and density field along the traffic flow propagation direction, and accordingly weights and accumulates each risk factor, making the quantification results of risk potential energy highly consistent with the actual operation mechanism of traffic flow. Dynamic topological iterative reasoning dynamically adjusts the risk transmission weights based on the real-time road network condition, enabling risk diffusion simulation to respond to emergencies.

[0056] This invention incorporates velocity gradient and density gradient into the same analytical framework for dot product coupling calculation through a gradient coupling analytical model. This solves the shortcomings of existing technologies that process traffic flow parameters and meteorological environmental parameters separately, and cannot quantitatively characterize the gradual accumulation effect of risks. At the same time, this invention adopts a dynamic adjacency matrix update strategy driven by real-time road network connectivity status. In each iteration, the risk transmission weights between nodes are adjusted in real time based on road closure or accident road occupancy information. This ensures that the risk propagation simulation is always consistent with the actual road network operation status, effectively overcoming the problem of large deviations between the risk simulation results and the actual situation under static topology propagation rules.

[0057] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for multimodal spatiotemporal risk situational awareness and assessment of highways, characterized in that, Includes the following steps: S1 collects traffic flow parameters, meteorological environmental parameters, and road surface condition parameters across the entire highway area through multimodal sensing devices deployed on the roadside and on vehicles. The collected parameters are then mapped to a unified geographic grid coordinate system after spatiotemporal registration to construct a multimodal spatiotemporal situational dataset. S2, input the multimodal spatiotemporal situational basic dataset into the dual spatiotemporal mask traffic flow state perception algorithm, apply a sliding window mask in the time dimension to capture the temporal evolution rhythm of traffic flow, apply an adjacency graph mask in the spatial dimension to extract the spatial dependency association between road segments, and output the traffic flow state tensor of each grid unit through dual mask interactive reasoning. S3, feed the traffic flow state tensor into the risk factor gradient coupling analytical model, analyze the nonlinear coupling relationship between the velocity gradient field and the density gradient field along the traffic flow propagation direction, and weight and accumulate the risk factors according to the gradient coupling strength to output the initial risk potential value of each grid cell. S4. Input the initial risk potential energy value into the dynamic road network topology iterative inference function, dynamically update the adjacency weight matrix according to the real-time road network connectivity status, and propagate the risk potential energy between road network nodes through topology iterative inference until the change in node risk potential energy converges to below the preset threshold, and output the updated risk potential energy spatial distribution. S5, the updated risk potential energy spatial distribution is superimposed with the static road network attribute layer, and a risk level label is generated through multi-level risk threshold determination. The risk level label and the corresponding spatiotemporal coordinates and traffic flow parameters are encapsulated into a structured situational message. S6, the structured situational message is received by the whole-domain road network risk situation assessment platform. The platform parses the message and renders it on the digital twin map interface to generate a whole-domain risk situation visualization view.

2. The method for multimodal spatiotemporal risk situational awareness and assessment of highways according to claim 1, characterized in that, The expression for the dual spatiotemporal mask traffic flow state perception algorithm is as follows: ,in For the first Hidden state matrix of vehicle flow layers The output after masked interactive inference is the first... Hidden state matrix of vehicle flow layers It is a non-linear activation function. This represents the total number of mask types. For the first The attention weight vector of the class mask, where ⊙ is the Hadamard multiplication operator. For the first Normalized adjacency matrix under the effect of class masking For the first The learnable weight matrix corresponding to the class mask; a separate temporal mask matrix is ​​also provided. and spatial mask matrix ,Will and Through fusion function Coupled generation of joint mask tensor ,by right Dynamic modulation is performed, and the modulation relationship is as follows: ,in This is a tensor element-level modulation operator. It is the time-varying adjacency matrix after joint mask tensor modulation.

3. The method for multimodal spatiotemporal risk situational awareness and assessment of highways according to claim 1, characterized in that, The expression for the risk factor gradient coupling analytical model is as follows: ,in For the first Velocity density gradient coupling strength at each grid cell For the first The velocity gradient vector at each grid cell For the first The density gradient vector at each grid cell is defined by the vector dot product operator; a separate set of risk factors is also defined. Corresponding to each meteorological and environmental parameter, based on the aforementioned coupling strength Constructing a risk weight function The initial risk potential value is the scaling factor. The calculation relationship is as follows ,in For the first The spatial integration region encompassed by each grid cell For the area of ​​a microelement, The total number of risk factors. For the first The measured values ​​of each risk factor. , .

4. The method for multimodal spatiotemporal risk situational awareness and assessment of highways according to claim 1, characterized in that, The expression for the dynamic road network topology iterative inference function is: ,in For the first Node at the next iteration The risk potential value, For the first Node at the next iteration The risk potential value, This is the iteration step size factor. For nodes The set of adjacent nodes, and Both are adjacency sets Node index in For nodes With nodes Real-time network impedance distance between them For nodes With nodes Real-time network impedance distance between them Here, is the distance decay coefficient, is the product operator, exp is the natural exponential function, and the iteration continues until... The process terminates when the value falls below the preset threshold.

5. The method for multimodal spatiotemporal risk situational awareness and assessment of highways according to claim 1, characterized in that, The multimodal sensing device includes a millimeter-wave radar deployed on a gantry along the highway, an optical video acquisition unit deployed on the roadside, and a piezoelectric dynamic weighing sensor deployed inside the road surface. The detection range of the millimeter-wave radar is set to 150 to 250 meters, the sampling frame rate of the optical video acquisition unit is set to 15 to 30 frames per second, and the sampling frequency of the piezoelectric dynamic weighing sensor is set to 500 Hz to 1000 Hz. The grid cell size of the geographic grid coordinate system is set to a 500-meter × 500-meter square grid. The time window width of the sliding window mask is set to... The window sliding step size is set to 1 minute to 3 minutes, with a time interval of 5 to 15 minutes. The spatial adjacency order of the adjacency graph mask is set to 1st to 3rd order adjacency. The multi-level risk threshold includes three levels: low risk threshold, medium risk threshold, and high risk threshold. The low risk threshold is the 30th percentile value of the risk potential value distribution, the medium risk threshold is the 60th percentile value, and the high risk threshold is the 90th percentile value. The structured situational message is encapsulated in JSON format and includes a timestamp field, a spatial coordinate field, a risk level field, and a traffic flow parameter field.

6. The method for multimodal spatiotemporal risk situational awareness and assessment of highways according to claim 1, characterized in that, S2 includes: The time series components in the multimodal spatiotemporal situational awareness dataset are extracted, and the time series components are segmented and truncated using the sliding window mask. The truncated subsequences of each time period are then input into a gated recurrent unit to extract temporal latent features. The spatial distribution components in the multimodal spatiotemporal situational awareness dataset are extracted, and the information transmission channels between non-adjacent nodes are masked using the adjacency graph mask. Only the spatial feature aggregation paths between adjacent nodes are retained, and graph convolution operations are performed to extract spatial latent features. The temporal latent features and the spatial latent features are concatenated along the feature dimension. The concatenated joint feature tensor is then mapped to a preset latent space dimension via a linear projection layer to generate a temporal-spatial fusion feature. The temporal-spatial fusion feature is then subjected to alternating iterative interactive reasoning using the dual masks. In each iteration, the temporal mask and the spatial mask are applied alternately until the number of iterations reaches a preset upper limit, and the traffic flow state tensor is output.

7. The method for multimodal spatiotemporal risk situational awareness and assessment of highways according to claim 1, characterized in that, S3 includes: Calculate the partial derivatives of the velocity component and the density component along the flow direction of each grid cell in the traffic flow state tensor. Organize the partial derivatives of the velocity component into a velocity gradient field and the partial derivatives of the density component into a density gradient field. Perform a vector dot product operation on the velocity gradient field and the density gradient field at the same grid cell, and use the dot product result as the gradient coupling strength of that grid cell. Traverse all grid cells to generate a gradient coupling strength distribution map. The coupling strength value of each grid cell in the gradient coupling strength distribution map is substituted into the nonlinear mapping function, and the risk factor weight coefficient of each grid cell is generated after transformation by the nonlinear mapping function. The risk factor weight coefficient of each grid cell is multiplied by the measured value of each type of risk factor of the corresponding grid cell, and the multipliers are accumulated and then the initial risk potential value of each grid cell is output.

8. The method for multimodal spatiotemporal risk situational awareness and assessment of highways according to claim 1, characterized in that, S4 includes the following steps: S41. Construct a dynamic adjacency matrix based on the real-time road network connectivity status. Set the adjacency elements corresponding to road segments that are closed due to traffic control or accidents to zero, and assign the adjacency elements corresponding to road segments that are open to traffic to the reciprocal of the real-time road network impedance. S42, the initial risk potential value is used as the initial risk potential value of each node, and the current risk potential value of each node is propagated to its neighboring nodes according to the weight ratio using the dynamic adjacency matrix as the weight basis. S43, In each iteration step, the diffusion potential energy received by each node from all neighboring nodes is accumulated, and the accumulated result is used to replace the current risk potential energy value of the node to complete the synchronous update of all nodes. S44, calculate the maximum absolute difference of all node risk potential energy values ​​between the current iteration step and the previous iteration step. If the maximum value is higher than the preset threshold, repeat S42 to S43. If the maximum value is lower than or equal to the preset threshold, terminate the iteration and output the updated risk potential energy spatial distribution.

9. The method for multimodal spatiotemporal risk situational awareness and assessment of highways according to claim 1, characterized in that, S5 includes: The risk potential energy values ​​of each grid cell in the updated risk potential energy spatial distribution are weighted and superimposed with the road curvature attributes and longitudinal slope attributes in the pre-stored static road network attribute layer to generate composite risk potential energy values. The composite risk potential value is compared with the low-risk threshold, medium-risk threshold, and high-risk threshold respectively. Based on the comparison results, each grid cell is determined to belong to one of the low-risk, medium-risk, or high-risk levels. The determined risk level label, along with the current timestamp, the center geographic coordinates of each grid cell, and the corresponding speed, density, and flow parameters, are filled into the fields of a preset message template. The filled content is serialized into the structured situation message according to JSON syntax rules, and the structured situation message is pushed to the global road network risk situation assessment platform through a message queue protocol.

10. A multimodal spatiotemporal risk situational awareness and assessment system for highways, characterized in that, This system is applied to the multimodal spatiotemporal risk situational awareness and assessment method for highways as described in claim 1, comprising: The multimodal sensing parameter spatiotemporal registration and grid mapping unit is used to map traffic flow parameters, meteorological environment parameters and road surface condition parameters collected by multi-source sensors to the geographic grid coordinate system after spatiotemporal alignment, and the output end is connected to the input end of the dual spatiotemporal mask vehicle flow state tensor inference unit. The dual spatiotemporal masking traffic flow state tensor inference unit is used to apply a mask in the dual dimensions of temporal window and spatial adjacency and perform multi-layer interactive inference on the implicit state of traffic flow to generate a traffic flow state tensor. The output end is connected to the input end of the velocity density gradient coupled risk potential energy analysis unit. The velocity-density gradient coupled risk potential energy analysis unit is used to analyze the initial risk potential energy value of each grid cell by weighting and accumulating various risk factors based on the dot product coupling strength of the velocity gradient field and the density gradient field. The output end is connected to the input end of the dynamic topology iteration risk potential energy update unit. The dynamic topology iterative risk potential energy update unit is used to dynamically update the adjacency weight matrix based on the real-time road network connectivity status and iteratively propagate the risk potential energy between road network nodes until convergence to output the updated risk potential energy spatial distribution. The output end is connected to the input end of the multi-level risk situation structured message generation unit. The multi-level risk situation structured message generation unit is used to overlay the spatial distribution of risk potential energy with static road network attributes and generate risk level labels through multi-level threshold determination. The labels, spatiotemporal coordinates, and traffic flow parameters are then encapsulated into structured situation messages. The output end is connected to the input end of the digital twin map risk situation visualization rendering unit. The digital twin map risk situation visualization rendering unit is used to receive and parse the structured situation message, mark the risk level of each grid unit on the digital twin map interface with different color levels, and output the marked map view to the display terminal.

Images