Intelligent road traffic risk early warning method and system based on vehicle-road cooperation

Abstract

The invention discloses an intelligent road traffic risk early warning method and system based on vehicle-road cooperation, and relates to the technical field of traffic early warning, and the method comprises the steps: unifying real-time collected road data to a UTM coordinate system through combining time alignment and coordinate conversion, and constructing standardized spatio-temporal data; based on the standardized spatio-temporal data, using a hypergraph neural network to construct a dynamic hypergraph, obtaining node risk features, and generating a comprehensive risk index; according to the risk index feature weight distribution vector, risk root cause probability distribution is obtained through intervention calculation, and a risk level and an early warning type are output; real-time road data are safely aggregated through federal learning, and the dynamic hypergraph is updated in combination with differential privacy. According to the method, the dynamic hypergraph is constructed, the hypergraph neural network is used for capturing the high-order interaction relation between the vehicles, the problem that complex space-time interaction modes between the vehicles are difficult to capture is solved, and the extraction efficiency of node risk features is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic warning, in particular to an intelligent road traffic risk warning method and system based on vehicle-road cooperation. BACKGROUND

[0002] In the field of intelligent transportation, most of the traditional road traffic risk warning methods rely on a single data source or a simple model, such as traffic flow monitoring based on fixed thresholds or risk judgment based on rules. Basically, by collecting basic data such as vehicle position and speed, and combining historical statistical rules or expert experience, the potential risks are preliminarily evaluated. In recent years, with the rise of vehicle-road cooperation technology, some studies use multi-source data fusion and machine learning models to improve the accuracy of risk identification.

[0003] The existing methods currently have some deficiencies, for example, the traditional models mostly use static graph structures or linear regression analysis, which are difficult to capture the complex spatio-temporal interaction patterns between vehicles, especially in dynamic scenes such as congestion and accidents, the extraction efficiency of node risk features is low; in addition, the existing risk indicator screening mechanism usually relies on manual threshold setting or simple statistical methods, which cannot adaptively identify key risk factors, resulting in redundancy and lag of the warning results. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an intelligent road traffic risk warning method based on vehicle-road cooperation, which solves the problems of difficulty in capturing complex spatio-temporal interaction patterns between vehicles and inability to adaptively identify key risk factors.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present application provides an intelligent road traffic risk warning method based on vehicle-road cooperation, which comprises: combining time alignment and coordinate conversion to unify real-time collected road data to the UTM coordinate system, and constructing standardized spatio-temporal data; Based on the standardized spatio-temporal data, a dynamic hypergraph is constructed using a hypergraph neural network, and node risk features are obtained to generate a comprehensive risk indicator; The comprehensive risk indicator is converted into a binary pulse sequence by generating a pulse sequence, a pulse synchronization intensity matrix is constructed, and a risk feature weight distribution vector is generated; According to the risk feature weight distribution vector, the risk root cause probability distribution is obtained by combining the hypergraph neural network, and the risk level and warning type are outputted; Based on the risk level difference, the warning type is pushed, the risk event data is generated and encrypted by data fusion and risk feature extraction, and uploaded; Real-time road data is securely aggregated by federated learning and updated with dynamic hypergraphs in combination with differential privacy.

[0007] As a preferred scheme of the intelligent road traffic risk early warning method based on vehicle-road cooperation, the real-time road data is unified to the UTM coordinate system by combining time alignment and coordinate conversion, and the standardized space-time data is constructed, and the specific steps are as follows, A unified time reference is constructed by using a high-precision clock synchronization protocol and hardware triggering, and a time tag is marked; The geographic coordinates are converted into plane rectangular coordinate positions by applying satellite positioning and inertial measurement fusion, and plane positioning data is generated; Columnar storage and spatial grid index compression are implemented, time tags and plane positioning data are associated, and standardized space-time data is constructed.

[0008] As a preferred scheme of the intelligent road traffic risk early warning method based on vehicle-road cooperation, the real-time road data is unified to the UTM coordinate system by combining time alignment and coordinate conversion, and the standardized space-time data is constructed, and the specific steps are as follows, Based on the standardized space-time data, micro-super edges are generated by Euclidean distance and behavior similarity, macro-super edges are generated by combining traffic flow density and road type, and dynamic hypergraphs are constructed; According to the dynamic hypergraph, the node space-time features are obtained by using hypergraph convolution to aggregate neighbor information, and the node risk features and embedding vectors are generated; The node risk features and embedding vectors are fused, and the comprehensive risk indicators are generated by dynamic threshold.

[0009] As a preferred scheme of the intelligent road traffic risk early warning method based on vehicle-road cooperation, the real-time road data is unified to the UTM coordinate system by combining time alignment and coordinate conversion, and the standardized space-time data is constructed, and the specific steps are as follows, The comprehensive risk indicators are converted into binary pulse sequences by generating pulse sequences, and the binary pulse trigger times are obtained according to fixed time windows, and the pulse trigger frequency is output; A dynamic competition threshold is set, and binary pulses that meet the pulse trigger frequency greater than the dynamic competition threshold are reserved as a subset, and a high-frequency index set and corresponding binary pulse trigger frequency are generated through screening; Based on the binary pulse sequence corresponding to the high-frequency index set, the arrival time difference of the binary pulse sequence is obtained and normalized, and the pulse synchronization intensity matrix is constructed; Based on the binary pulse trigger frequency and the pulse synchronization intensity matrix, the risk feature weight distribution vector is generated by using weighted fusion.

[0010] As a preferred scheme of the intelligent road traffic risk early warning method based on vehicle-road cooperation, wherein: according to the risk characteristic weight distribution vector, the risk root cause probability distribution is obtained by combining the hypergraph neural network, and the risk level and the early warning type are output, and the specific steps are as follows, The risk index causal graph is generated by taking the risk characteristic weight distribution vector as a node, combining the micro-super edge and the macro-super edge to initialize the causal edge, and generating the risk index causal graph; Based on the risk index causal graph, the counterfactual result is generated by using the hypergraph convolution to aggregate neighbor information and is normalized, and the risk root cause probability distribution is obtained; The node activation strength is strengthened using the pulse timing dependent plasticity rule in combination with the risk root cause probability distribution and the pulse neural network, and the smoothed distribution is generated by using the exponential moving average filtering to generate the smoothed distribution probability value, and the risk level and the early warning type are output.

[0011] As a preferred scheme of the intelligent road traffic risk early warning method based on vehicle-road cooperation, wherein: the risk level is differentiated to push the early warning type, the risk event is generated by data fusion and risk feature extraction, and is encrypted and uploaded, and the specific steps are as follows, The risk score is obtained by multi-feature fusion processing on the comprehensive risk score, and the early warning instruction of the matching risk level is generated by using logical judgment to determine the early warning level; The historical data and the real-time collected data are aligned using dynamic time warping, and the structured risk event data is generated using conditional generative adversarial network; Based on the structured risk event data, the data is encapsulated and encrypted and is uploaded in slices.

[0012] As a preferred scheme of the intelligent road traffic risk early warning method based on vehicle-road cooperation, wherein: the real-time road data is safely aggregated by federated learning, and the dynamic hypergraph is updated in combination with differential privacy, and the specific steps are as follows, The parameter aggregation is performed using the random mask to generate the mask parameter, and the update parameter is generated by protocol reconstruction; The update parameter is quantized by 8-bit integer, and the redundant data block is sliced by using the erasure code, and is uploaded in parallel through multiple network channels.

[0013] In a second aspect, the application provides an intelligent road traffic risk early warning system based on vehicle-road cooperation, comprising a space-time alignment module, a hypergraph modeling module, a risk screening module, a root cause analysis module, an early warning pushing module, and a federated updating module, The space-time alignment module is used to unify the real-time collected road data to the UTM coordinate system by combining time alignment and coordinate conversion, and to construct standardized space-time data; The supergraph modeling module is configured to construct a dynamic supergraph using a supergraph neural network based on standardized space-time data, and obtain node risk features to generate a comprehensive risk index. The risk screening module is configured to convert the comprehensive risk index into a binary pulse sequence by generating a pulse sequence, construct a pulse synchronization intensity matrix, and generate a risk feature weight distribution vector. The root cause analysis module is configured to obtain a risk root cause probability distribution according to the risk feature weight distribution vector in combination with the supergraph neural network, and output a risk level and a warning type. The early warning pushing module is configured to push the warning type based on the risk level difference, generate risk event data by data fusion and risk feature extraction, and encrypt and upload the risk event data. The federal update module is configured to securely aggregate real-time road data through federated learning, and update the dynamic supergraph in combination with differential privacy.

[0014] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, any step of the intelligent road traffic risk early warning method based on vehicle-road cooperation according to the first aspect of the present application is implemented.

[0015] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, any step of the intelligent road traffic risk early warning method based on vehicle-road cooperation according to the first aspect of the present application is implemented.

[0016] The present application has the following advantages: by constructing a dynamic supergraph, the supergraph neural network is used to capture the high-order interaction relationship between vehicles, the problem of being difficult to capture the complex space-time interaction mode between vehicles is solved, and the extraction efficiency of node risk features is improved; by screening high-risk indicators through biological heuristic attention and dynamic competitive threshold, the adaptive identification of risk factors is strengthened, the redundancy of early warning is reduced, and the real-time performance is enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0018] Fig. 1 The flowchart of the intelligent road traffic risk early warning method based on vehicle-road cooperation.

[0019] Fig. 2 The schematic diagram of the intelligent road traffic risk early warning system based on vehicle-road cooperation.

[0020] Fig. 3 Flow chart for bio-inspired attentional screening of high-risk indicators.

[0021] Fig. 4 Flow chart for generating risk level and warning type. DETAILED DESCRIPTION

[0022] In order to make the above objectives, features and advantages of the present application more apparent, specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0023] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details given herein, that the present application can be practiced with other than the described embodiments, and that variations from the particular embodiments described herein can be made and still be within the scope of the present application.

[0024] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments.

[0025] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides an intelligent road traffic risk warning method based on vehicle-road cooperation, comprising the following steps: S1, combine time alignment and coordinate conversion to unify real-time collected road data to UTM coordinate system, and construct standardized space-time data.

[0026] It should be noted that the real-time collected data is vehicle dynamic information, road environment state, traffic signal instruction and traffic participant behavior.

[0027] Further, a high-precision clock synchronization protocol and hardware triggering are used to construct a unified time reference and mark a time tag.

[0028] Specifically, the vehicle, roadside device and cloud are time-synchronized through the precise time protocol, the master clock node such as the roadside device periodically sends time synchronization signals, and the slave clock node such as the vehicle terminal receives and automatically calibrates the local clock after receiving, forming a unified time reference; based on the unified time reference, a deviation value (such as ±5 microseconds) is set, and the clock deviation value is periodically detected, and if it exceeds the deviation value, frequency compensation is automatically triggered; the data collection time of the vehicle sensor is synchronized through the hardware signal such as GPS second pulse, and an accurate time tag is attached to each piece of data.

[0029] The satellite positioning and inertial measurement are fused to convert the geographic coordinates into planar rectangular coordinate positions to generate planar positioning data.

[0030] Specifically, satellite observation data are acquired through a navigation satellite, and the longitude, latitude, altitude and receiver clock error in a global geographic coordinate system are solved. Vehicle acceleration and angular velocity are collected through inertial measurement fusion. Satellite observation data are detected and repaired for cycle slip and inhibited for multipath effect. Inertial measurement data are calibrated for zero offset, compensated for temperature and filtered for sensor noise. The longitude, latitude and altitude in the global geographic coordinate system are converted into a planar rectangular coordinate system through projection. The processed inertial measurement data are fused through loose coupling to generate initial planar positioning data. The velocity, position and attitude angle in the planar rectangular coordinate position are acquired through a strapdown inertial navigation method based on the initial planar positioning data and vehicle acceleration and angular velocity. When the satellite observation data are valid, the inertial measurement data and the satellite observation data are subtracted. The zero offset error of the inertial measurement data is acquired and corrected through Kalman filtering to generate planar positioning data.

[0031] Columnar storage and spatial grid index compression are implemented. Time labels and planar positioning data are associated to construct standardized spatiotemporal data.

[0032] Specifically, the planar positioning data, time labels and vehicle road ID and the like fields are structured and packaged in a columnar storage format. The continuously collected positioning data stream is segmented and cut according to a fixed time window. Each segment of data generates an independent data block. The data in the block is arranged in column order to generate a structured data file. In combination with the structured data file, a target road area is divided into uniform two-dimensional grids. Each two-dimensional grid is assigned a unique identifier. The grid size can be dynamically adjusted according to the application scenario. A space-filling curve is used to map the two-dimensional grid units to one-dimensional index values for fast conversion of planar coordinates to index values to generate a spatial grid index table. The time labels and the data in the spatial grid index table are combined into a composite key. The mapping relationship between the time stamp and the grid unit is established through a hash table to support fast retrieval according to the time range or the spatial range. Repetitive values in the columnar storage file are compressed using dictionary encoding. Floating-point coordinate data are compressed using incremental encoding. The storage volume is reduced through a general compression method to construct standardized spatiotemporal data.

[0033] It should be noted that the positioning data stream refers to the standardized positioning data generated by preprocessing the high-frequency sampled coordinate time series.

[0034] S2, based on the standardized spatiotemporal data, a hypergraph neural network is used to construct a dynamic hypergraph and acquire node risk features to generate a comprehensive risk indicator.

[0035] Further, based on the standardized spatio-temporal data, micro-superegions are generated by Euclidean distance and behavior similarity, macro-superegions are generated by traffic flow density and road type, and a dynamic supernet is constructed.

[0036] Specifically, the standardized spatio-temporal data is preprocessed by Z-score standardization, and the vehicle position, speed and other features are normalized. The vehicle trajectory data is divided into static snapshots at a fixed time interval, such as 100 milliseconds, to generate a normalized vehicle state matrix. According to the normalized vehicle state matrix, the Euclidean distance between vehicles is identified, the behavior similarity is obtained by cosine similarity based on the speed difference and acceleration difference, and if the Euclidean distance between two vehicles is less than a set distance threshold, such as 50 meters, and the behavior similarity between the two vehicles is greater than a set similarity threshold, a micro-superegion is generated, and a micro-superegion set is constructed. The road is divided into grids, and the number of vehicles and average speed in each grid are counted to obtain the grid density. If the density threshold is exceeded, it is marked as a high-density grid. The same type of road vehicles are aggregated into a superegion, and a macro-superegion set is generated. The micro-superegion set and the macro-superegion set are combined, the supernet node set is defined as the vehicle and the superegion set, the superegion weight is obtained based on the interaction intensity such as the average similarity of micro-superegions, and the supernet is updated by sliding the time window such as 100 milliseconds. A dynamic sequence is generated, and a dynamic supernet containing the high-order interaction relationship of vehicles changing over time is output.

[0037] It should be noted that the similarity threshold can be set to obtain the similarity of all vehicle pairs, and the 90th percentile of the similarity distribution is taken as the similarity threshold, for example, 0.8. The density threshold can be set based on the critical density value of urban roads, for example, 20 vehicles per kilometer. The speed difference and acceleration difference are obtained by direct measurement and satellite positioning data.

[0038] According to the dynamic supernet, the node spatio-temporal features are obtained by aggregating neighbor information using supernet convolution, and the node risk features and embedding vectors are generated.

[0039] Specifically, based on the connection relationship between the vehicle node and the hyperedge of the dynamic hypergraph, the initial weight of each hyperedge is obtained according to the interaction intensity of the micro or macro hyperedge, and the initial weight is normalized by Softmax to make the sum of all weights equal to 1, an adjacency tensor is constructed, and the hypergraph adjacency tensor is output; based on the hypergraph adjacency tensor, the numerical features such as vehicle position, speed and acceleration are extracted from the standardized space-time data and standardized, the environment data such as traffic flow density and road type are one-hot encoded, the encoded environment data and vehicle features are spliced to generate an extended node feature matrix; according to the extended node feature matrix, the vehicle node features connected by each hyperedge are aggregated by using the adjacency tensor, and the weighted sum is obtained to obtain the hyperedge level feature, the mean pooling is performed on all hyperedge level features, and the hyperedge level feature and the node feature are spliced to generate an embedding vector through a fully connected layer; the node embedding vector sequence is arranged according to time steps, and the mean value and variance in the window are obtained, the embedding vectors of adjacent nodes are weighted and summed based on the hyperedge weight of the hypergraph adjacency tensor, the spatial features of high weight interaction are highlighted, and a space-time feature matrix is generated; based on the space-time feature matrix, PCA is used to reduce the dimension of the space-time feature matrix, the risk dimension is extracted, the risk score is generated, and the space-time feature is mapped to the risk score (numerical range 0~1) through the Softmax function. The node risk feature is generated.

[0040] It should be noted that the weighted sum of the embedding vectors of adjacent nodes is to locate all adjacent nodes directly connected to the target node according to the hypergraph adjacency tensor, directly use the normalized hyperedge weight obtained in the hypergraph convolution, and take out the embedding vectors of adjacent nodes from the node embedding matrix and weighted sum; the traffic flow density is obtained by dividing the road into grids and the number of vehicles in the grid; the embedding vector is a vehicle node feature representation.

[0041] The node risk feature and the embedding vector are fused, and a comprehensive risk index is generated through a dynamic threshold.

[0042] Specifically, based on the node risk characteristics and the embedding vector, a combined feature vector is generated by merging in order; according to the combined feature vector, a sliding window statistics (such as the 90th percentile of the risk score of the last 5 minutes or 100 vehicles) is used to adjust the percentile result according to the traffic flow state (such as +0.05 in congestion and -0.03 in smooth), and a dynamic risk threshold is output; each feature value in the combined feature vector is normalized to eliminate the dimensional difference, and all normalized feature values are directly spliced to obtain a standardized combined feature vector, and the arithmetic mean of all feature values in the standardized combined feature vector is obtained to obtain an initial risk score, and the dynamic risk threshold is directly compared with the initial risk score, if the initial risk score is greater than or equal to the dynamic risk threshold, it is determined as high risk, and if the initial risk score is less than the dynamic risk threshold, it is determined as low risk; and according to a fixed numerical range, the initial risk score is divided into low risk, (0.3, 0.7] is medium risk, and (0.7, 1.0] is high risk, and a comprehensive risk index is generated.

[0043] S3, the comprehensive risk index is converted into a binary pulse sequence by generating a pulse sequence, a pulse synchronization intensity matrix is constructed, and a risk feature weight distribution vector is generated.

[0044] Further, the comprehensive risk index is converted into a binary pulse sequence by generating a pulse sequence, and the number of binary pulse triggers is obtained according to a fixed time window, and a pulse trigger frequency is output.

[0045] Specifically, according to the comprehensive risk index, the normalized processing of the comprehensive risk index of each time step is performed, if the comprehensive risk index has different physical meanings or dimensions, the feature standardization is performed first, and then the scaling is uniformly performed to the range of [0, 1], and the normalized risk index vector is output; according to the normalized risk index vector, a static threshold (such as directly setting 0.7) is used to generate a pulse firing threshold of the current time step; and using an element-by-element comparison function, each normalized risk index vector is directly compared with the pulse firing threshold, whether the normalized risk index vector is greater than or equal to the pulse firing threshold is judged, and a binary pulse sequence is generated; based on the binary pulse sequence, the pulse sequence is reshaped into a pulse sequence matrix by using a sliding window statistics, the summation in the dynamic window is obtained for each pulse in the pulse sequence matrix, a window-level pulse trigger count matrix is generated, and the pulse trigger frequency is directly output by using division.

[0046] A dynamic competition threshold is set, and binary pulses satisfying the pulse trigger frequency greater than the dynamic competition threshold are reserved as a subset, and a high-frequency index set filtered and the corresponding binary pulse trigger frequency are generated.

[0047] Specifically, the mean and the standard deviation are obtained based on the risk indicator vector, and a dynamic competition threshold is generated; each value in the risk indicator vector is compared with the dynamic competition threshold one by one, the risk indicators that meet the condition of being greater than the dynamic competition threshold are retained, the index and the value are recorded, and a high-frequency indicator set filtered and a corresponding binary pulse trigger frequency are generated.

[0048] It should be noted that the expression for generating the dynamic competition threshold is: T = μ r + β * σ r ; Wherein, T represents the dynamic competition threshold, μ r is the mean of the risk indicator vector, reflecting the overall risk level, σ r represents the standard deviation of the risk indicator vector, represents the average activity of the current risk indicator, β is the adjustment coefficient, controls the deviation of the dynamic competition threshold from the mean, balances the filtering strictness, 0 ≤ β ≤ 1, which can be set to 0.7, and r represents the risk indicator vector.

[0049] Based on the binary pulse sequence corresponding to the high-frequency indicator set, the arrival time difference of the binary pulse sequence is obtained and normalized, and a pulse synchronization strength matrix is constructed.

[0050] Specifically, based on the high-frequency indicator set, a subsequence corresponding to the high-frequency indicator set is filtered out, and the binary pulse record corresponding to each high-frequency indicator is located. The subsequence of each high-frequency indicator records whether the indicator triggers a pulse in the continuous time step, such as using the value 1 to represent triggering and 0 to represent not triggering, forming an independent time step sequence; for each pair of high-frequency indicator subsequences, their pulse trigger time points are compared respectively, the specific time step positions of all triggered pulses in each pair of subsequences are extracted, the absolute difference between time steps is obtained, and the minimum value is selected from all absolute differences as the pulse arrival time difference; initialize a square matrix corresponding to the number of high-frequency indicators, all elements are initialized to 0, fill the normalized time difference into the corresponding position of the matrix in turn, fill the fixed value 1 in the diagonal line position (i.e. the row and column corresponding to the same indicator) of the matrix, indicating that the synchronization of the indicator itself is the strongest, and combine the independent time step sequence and the pulse arrival time difference to construct the pulse synchronization strength matrix.

[0051] Based on the binary pulse trigger frequency and the pulse synchronization strength matrix, a risk feature weight distribution vector is generated using weighted fusion.

[0052] Specifically, based on the binary pulse trigger frequency and the pulse synchronization intensity matrix, the binary pulse trigger frequency is scaled to the range of [0, 1] to eliminate the dimensional difference, the maximum pulse frequency value of all high-frequency indicators is counted, the normalized value of the binary pulse trigger frequency of each high-frequency indicator is obtained, and the normalized pulse frequency vector is generated; each row and each column of the pulse synchronization intensity matrix is traversed, the elements at the diagonal line position are excluded, and the non-diagonal line pulse synchronization intensity set is output; the normalized binary pulse trigger frequency and the non-diagonal line pulse synchronization intensity set are normalized by weighted summation to generate a comprehensive weight; the synchronization strength average value of each high-frequency indicator is obtained, the normalized binary pulse trigger frequency is weighted and summed according to a fixed weight (such as 0.6 and 1-0.6), and the risk feature weight distribution vector is generated by combining the pulse frequency vector, the non-diagonal line synchronization intensity set, and the comprehensive weight.

[0053] It should be noted that the comprehensive weight is obtained by: counting the number of pulse triggers per unit time for each high-frequency indicator and normalizing it to the range of [0, 1], the higher the value, the stronger the local activity, and extracting the average synchronization strength (the average value of the normalized time difference, the higher the value, the stronger the synergy) of all high-frequency indicators, and weighting and summing the two according to a fixed ratio (such as 60% for activity and 40% for synergy) to obtain the comprehensive weight.

[0054] S4, according to the risk feature weight distribution vector, the risk root cause probability distribution is obtained by combining the hypergraph neural network, and the risk level and the warning type are output.

[0055] Further, the risk feature weight distribution vector is taken as a node, the micro-super edge and the macro-super edge are combined to initialize the causal edge, and the risk indicator causal graph is generated.

[0056] Specifically, each element in the risk feature weight distribution vector is mapped to an independent graph node, each node carries two basic attributes of indicator identifier and risk feature weight value (from the corresponding numerical value in the vector), and an initial graph structure is output; based on the initial graph structure, all micro-supersides are traversed, each superedge connects a group of nodes, for each node pair in each micro-superside, a directed or undirected causal connection edge is set according to the similarity, and the weight of the edge can be set as the correlation score of the similarity value after normalization processing; if there is a strong association (such as high similarity in behavior), a strong causal connection is initialized, indicating that the change of one indicator may directly affect another indicator, a set of local causal edges are added in the initial graph structure, forming a local causal subgraph to capture the fine-grained risk propagation path; all macro-supersides are traversed, each superedge also connects a group of nodes, the global correlation strength is obtained according to the macro feature (such as the correlation degree score obtained by statistical correlation analysis), and the correlation degree score is converted into the weight of the edge, and the corresponding directed or undirected connection is added in the initial graph, indicating the potential influence of the macro environment on the local risk indicator, a set of global causal edges are added in the initial graph structure, forming an expanded causal subgraph, covering local and global multi-scale risk correlation; the optimization dynamic threshold is set, and the graph pruning strategy is used to further optimize the graph structure, and the optimal causal connection is reserved, for the node pair with bidirectional connection, a directed causal graph is constructed according to the causal directionality judgment rule (such as time sequence to determine the direction of the edge), and the risk indicator causal graph is generated in combination with the local and global multi-scale risk correlation and the risk propagation path.

[0057] It should be noted that the optimization dynamic threshold can be obtained by sorting the correlation strength (such as similarity) of all causal edges, and selecting the 90%~95% quantile as the optimization dynamic threshold (such as 0.7~0.8); the optimal causal connection is reserved as the edge with correlation strength greater than or equal to the optimization dynamic threshold.

[0058] Based on the risk indicator causal graph, the neighbor information is aggregated using the hypergraph convolution to generate counterfactual results and normalize, and the risk root cause probability distribution is obtained.

[0059] Specifically, each edge in the risk indicator causal diagram is mapped to a hyperedge in the hypergraph. If only two nodes are connected, it is directly mapped to a normal hyperedge. If there are multiple nodes, a new hyperedge is used to include multiple nodes at the same time, indicating that they have a synergistic effect under a certain risk mode. Each hyperedge is assigned a weight to reflect the frequency of occurrence, and a hypergraph is generated. A hypergraph convolution network is used to collect all hyperedges to which each node in the hypergraph belongs. In each hyperedge, the features of all nodes (including risk feature weight values and other related attributes) are averaged to generate local aggregation features. The local aggregation features of all hyperedges are integrated by splicing to output updated node feature representations. Based on the updated node feature representations, nodes with high risk feature weights are selected as intervention targets. The risk feature weights of the intervention targets are disturbed (such as increased proportion), and new node feature representations are generated and input into the hypergraph convolution network. The risk state after intervention is extracted, and the risk state before and after intervention is compared to output a set of counterfactual results. The influence degree of the relative change proportion of each counterfactual result is obtained and normalized so that the sum is 1 to form a probability distribution, and a risk root cause probability distribution is obtained.

[0060] The risk root cause probability distribution is combined with the pulse neural network to strengthen the node activation strength using the pulse temporal dependence plasticity rule, and a smooth distribution is generated using an exponential moving average filter to generate a smooth distribution probability value, outputting a risk level and a warning type.

[0061] Specifically, the probability value corresponding to the risk root cause probability distribution is proportionally converted into the initial active energy value of the neuron (for example, probability 0.8 corresponds to active energy 80%), and the active threshold of the neuron is set (for example, 80%), when the active energy reaches or exceeds the active threshold, the neuron will be triggered to generate a pulse signal, the initial connection strength between neurons is set according to the risk indicator causal diagram (for example, the more closely connected the neurons are, the stronger the correlation between the indicators), and the initial active energy value and connection strength of each neuron in the output pulse neural network are assigned; and the active energy change of each neuron is monitored in real time, when the active energy of the neuron reaches the active threshold, a pulse signal is immediately triggered, if the high-risk neuron triggers the pulse before the low-risk neuron, the connection strength between the high-risk neuron and the low-risk neuron is enhanced, so that the low-risk neuron is more easily triggered by the high-risk neuron, if the low-risk neuron triggers the pulse before the high-risk neuron, the connection strength between the high-risk neuron and the low-risk neuron is weakened, reducing the interference of the low-risk neuron on the high-risk neuron, after each pulse trigger, the active energy of the high-risk neuron is updated, the active energy after triggering can be reduced, reducing continuous high-frequency triggering, so that the neuron corresponding to the high-risk indicator is preferentially triggered to generate a pulse and the connection strength is dynamically enhanced; the number of times each neuron triggers a pulse in a unit of time is counted to obtain the instantaneous risk probability, and a dynamic weighted average method is used to update the probability value, the weight can be adjusted according to actual needs (for example, the current instantaneous risk probability accounts for 70% and the instantaneous risk probability at the last moment accounts for 30%), and the entire risk event time window is continuously iterated and covered to generate a smoothed risk probability distribution; according to the smoothed risk probability distribution, risk level thresholds are set, for example, the highest probability indicator less than 30% is a low risk level, 30% to 60% is a medium risk level, and greater than or equal to 60% is a high risk level, according to the risk level and specific indicator combination, a warning type is selected, and the risk level and warning type are output.

[0062] It should be noted that when the risk level and warning type are output, the risk information can be transmitted to the surrounding road traffic facilities through visible light signals through a visible light communication system.

[0063] Among them, the visible light communication system includes a visible light signal receiving system and a visible light signal transmitting system. Specifically, after receiving the risk level and warning type, the risk level and warning type are transmitted to the surrounding nails through visible light signals to form an alarm on a certain nail safety distance.

[0064] It should be noted that the warning type can include low-risk no intervention, medium-risk reminding the driver to pay attention to the road conditions, high-risk suggesting starting emergency braking or reducing to a safe speed, and starting road closure or traffic control to notify rescue forces; S5, based on the risk level difference, push the warning type, generate risk event data through data fusion and risk feature extraction, and upload the data after encryption.

[0065] Further, the risk score is obtained by multi-feature fusion processing on the comprehensive risk score, and the early warning level is determined by logic judgment to generate the early warning instruction matching the risk level.

[0066] Specifically, n risk feature indicators (such as vehicle emergency braking frequency and road wetness degree) are extracted from the real-time collected multi-source data and marked as risk feature values. Each feature value is linearly normalized and mapped to the interval [0, 1]. A fixed weight is assigned to each feature. The fixed weight is obtained by historical statistics and meets the normalization requirement. The risk score is directly obtained by weighted summation (the risk score is between 0 and 1, and the larger the value, the higher the comprehensive risk). The mapping relationship between the risk score and the risk level is directly defined by logic judgment.

[0067] For example, if the risk score is greater than or equal to 0 and less than 0.3, it is determined as low risk, if the risk score is greater than or equal to 0.3 and less than 0.6, it is determined as medium risk, if the risk score is greater than or equal to 0.6 and less than 0.85, it is determined as high risk, and if the risk score is greater than or equal to 0.85 and less than or equal to 1, it is determined as extremely high risk. The mapping relationship between the risk level and the early warning instruction is defined by using conditional branching structure (such as generating no intervention instruction if it is low risk, generating instruction to remind the driver to pay attention to the road conditions if it is medium risk, generating instruction to suggest starting emergency braking or reducing to safe speed if it is high risk, and generating instruction to start road closure or traffic control and notify rescue forces if it is extremely high risk), and the early warning instruction matching the risk level is generated combined with the risk score.

[0068] The historical data and the real-time collected data are aligned using dynamic time warping, and the conditional generative adversarial network is used to generate structured risk event data.

[0069] Specifically, based on the historical data and the real-time collected data, a fixed time window length is set, the historical data and the real-time collected data are divided into discrete time segments according to the fixed time window length, if there is historical data timestamp and real-time collected data timestamp, they are directly aligned, if there is no exact match, the sample closest to the timestamp is selected as the alignment result, if higher accuracy is required, the interpolation result is obtained by linear interpolation formula for adjacent two historical data points, and the aligned data pair set is output; according to the data pair set, and based on the statistical law, the trigger condition of the risk event is predefined (for example, if the real-time occupancy rate is greater than 80% and the historical occupancy rate is less than 60%, the congestion intensification event is triggered), and all predefined rules are traversed for each time segment, if any rule condition is met, the corresponding risk event label is generated, the risk event label, timestamp and related feature value are combined as a structured record, and the structured risk event data is generated.

[0070] Based on structured risk event data, it is encapsulated and encrypted and uploaded in segments.

[0071] Specifically, data integration is used to integrate structured risk event data into a unified list, timestamps are uniformly converted into an international standard format, event types are converted into string labels such as sudden braking, characteristic values ​​are classified and stored according to numerical and enumeration types, and metadata is supplemented to add auxiliary information, including data generation device ID, version number, time range, etc., to generate standardized data packets; symmetric encryption is used to encrypt standardized data packets, and the symmetric key distributed through the encryption channel is used to make the encryption party and the decryption party key consistent, and the standardized data packets are converted into binary format, and the binary data is encrypted using the AES-128 algorithm to generate ciphertext, and the hash value of the initial data packet is obtained, which is attached to the ciphertext as a message authentication code, and the encrypted ciphertext data is output; according to the latest network transmission Maximum transmission limit, set a fixed size for each shard, obtain the total length of the ciphertext data for the shard, split it in sequence from the ciphertext header according to the fixed size, and generate multiple shards. If the last shard is less than the fixed size, retain the remaining ciphertext data as the last shard, and add metadata to the shards, attach management information to each shard, including the shard sequence number, the total number of shards, and the total ciphertext hash value, and output the shard data set; based on the shard data set, use the Hypertext Transfer Protocol Security to establish an encrypted communication channel, divide the shard set into multiple batches, upload different batches in parallel through multi-threading, and each shard is transmitted independently, supporting breakpoint resumption. After the cloud server receives the shard, it reorganizes the data according to the shard sequence number and the total number of shards, verifies the total hash value, and obtains the original risk event data through decryption.

[0072] Furthermore, when risk event data is generated and uploaded in encrypted form, the risk event data can be transmitted to surrounding vehicles and pedestrians through sound and light alarms via an alarm system based on road traffic facilities.

[0073] The road traffic facility-based warning system includes an intelligent information identification system for road traffic facilities, an intelligent information transmission system, and an audio and visual alarm system. Specifically, upon receiving risk information, the audio and visual alarm system issues an alert, reminding surrounding vehicles and pedestrians to pay attention to safety. For example, when a road stud system receives risk information, it issues an alert using high-powered LED strobe lights and an audio system, reminding people to pay attention to safety.

[0074] It should be noted that the road stud is equipped with an LED flash light, a battery, a solar cell, a sound system and an intelligent processing circuit.

[0075] S6. Securely aggregate real-time road data through federated learning and update the dynamic hypergraph in combination with differential privacy.

[0076] Further, the mask parameters are generated using random masks for parameter aggregation, and the update parameters are generated through protocol reconstruction.

[0077] Specifically, the participants locally maintain initial parameters to be uploaded, such as risk feature weights, randomly sample a noise vector from a selected noise distribution according to the current noise intensity, and add the initial parameters and the noise vector item by item to generate mask parameters; using a homomorphic encryption aggregation scheme, the participants upload the mask parameters to the aggregation server, the aggregation server performs homomorphic addition operation on the encrypted mask parameters to obtain encrypted global mask parameters, decrypts the global mask parameters through a decryption key to obtain the global mask parameters, and outputs the global mask parameters; the risk level is mapped to a noise intensity value, such as the higher the risk level, the lower the noise intensity, and the lower the risk level, the higher the noise intensity, a noise intensity adjustment rule is generated, and the current noise intensity value to be used is obtained, the participants use the noise intensity value as a distribution parameter (such as the standard deviation of a normal distribution) when generating a random noise vector, and output a dynamically adjusted noise vector; the aggregation server returns the global mask parameters to the participants, uses additive secret sharing to reconstruct the sum of the mask parameters of other participants through a protocol, and obtains the update amount of the local parameters to generate update parameters.

[0078] It should be noted that when aggregating parameters, the road facilities can interact through spikes to achieve secure data transmission and privacy protection between devices, and improve stability and data security.

[0079] The update parameters are quantized by 8-bit integers, and the erasure code is divided into redundant data blocks, which are uploaded in parallel through multiple network channels.

[0080] Specifically, all floating point values in the update parameters are traversed using uniform quantization to obtain the floating point maximum value and the floating point minimum value, and the scaling factor and zero point are obtained, the quantization value of each floating point parameter is obtained according to the scaling factor and zero point and is limited in the range of 0 to 255, the quantized floating point parameter value is packed and stored by byte (8-bit), the floating point value set is output, the Reed-Solomon code is used to divide the floating point value set into data blocks of similar size, the redundant check block is obtained based on the RS code algorithm, and the data block and the check block are encapsulated into independent data packets, each data packet contains block number and check information, and the redundant data block is output; according to the redundant data block, the multi-path transmission control protocol is used to detect the bandwidth, delay and packet loss rate of each network channel in real time, the redundant data block is distributed to multiple channels according to the channel state, each channel independently sends the allocated data block, and the server side recombines the complete data through the fragment number and the check information, if a channel loses a packet, the redundant check block is used to recover the data packet in other channels, so that all data blocks are successfully uploaded.

[0081] More preferably, in the event of an emergency, the combination of the vehicle light signal modulation system and the stud system can realize real-time information transmission and alarm push, so that every traffic participant on the road can receive the warning information in the shortest time, thereby reducing the risk of accidents and secondary accidents.

[0082] It should be noted that the vehicle light signal modulation system adopts a standard driving voltage of 12V-24V in the vehicle, uses a high-power MOS tube as the driving element, and adopts a pulse position modulation method. A small signal pulse of 3V or more is used as the modulation signal of the switching tube, and the switching tube modulates the driving electrical signal of the LED.

[0083] The embodiment also provides an intelligent road traffic risk early warning system based on vehicle-road cooperation, comprising: a space-time alignment module for aligning real-time collected road data to the UTM coordinate system by combining time alignment and coordinate conversion to construct standardized space-time data; a hypergraph modeling module for constructing a dynamic hypergraph based on the standardized space-time data using a hypergraph neural network, obtaining node risk features, and generating a comprehensive risk index; a risk screening module for converting the comprehensive risk index into a binary pulse sequence by generating a pulse sequence, constructing a pulse synchronization intensity matrix, and generating a risk feature weight distribution vector; a root cause analysis module for obtaining a risk root cause probability distribution based on the risk feature weight distribution vector and combining the hypergraph neural network to output a risk level and an early warning type; an early warning push module for pushing the early warning type based on the risk level difference, generating risk event data and encrypting uploading through data fusion and risk feature extraction; a federal update module for securely aggregating real-time road data through federated learning and updating the dynamic hypergraph based on differential privacy.

[0084] The embodiment also provides a computer device suitable for the intelligent road traffic risk early warning method based on vehicle-road cooperation, comprising: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the intelligent road traffic risk early warning method based on vehicle-road cooperation as proposed in the above embodiment.

[0085] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved by WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0086] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the intelligent road traffic risk early warning method based on vehicle-road cooperation as described in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0087] To sum up, the present application solves the problem of difficult capture of complex spatio-temporal interaction mode between vehicles by constructing a dynamic hypergraph and capturing high-order interaction relationship between vehicles by using a hypergraph neural network, and improves the extraction efficiency of node risk features. The present application also strengthens the adaptive identification of risk factors, reduces the redundancy of early warning and enhances the real-time performance by screening high-risk indicators through a biological heuristic attention and a dynamic competitive threshold.

[0088] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. An intelligent road traffic risk warning method based on vehicle-road collaboration, characterized by: include, Combine time alignment and coordinate conversion to unify real-time collected road data into the UTM coordinate system and construct standardized spatiotemporal data; Based on standardized spatiotemporal data, a dynamic hypergraph is constructed using a hypergraph neural network, and node risk characteristics are obtained to generate comprehensive risk indicators. By generating a pulse sequence, the comprehensive risk index is converted into a binary pulse sequence, a pulse synchronization intensity matrix is ​​constructed, and a risk feature weight distribution vector is generated; Based on the risk feature weight distribution vector, combined with the hypergraph neural network, the probability distribution of risk root causes is obtained, and the risk level and warning type are output; Differentiate warning types based on risk levels, generate risk event data through data fusion and risk feature extraction, and upload it in encrypted form; Real-time road data is securely aggregated through federated learning and updated on the dynamic hypergraph in combination with differential privacy.

2. The intelligent road traffic risk warning method based on vehicle-road collaboration according to claim 1, characterized in that: The method combines time alignment and coordinate conversion to unify the real-time collected road data into the UTM coordinate system and construct standardized spatiotemporal data. The specific steps are as follows: Use high-precision clock synchronization protocol and hardware triggering to build a unified time base and mark time tags; Apply satellite positioning and inertial measurement fusion to convert geographic coordinates into plane rectangular coordinates and generate plane positioning data; Implement columnar storage and spatial grid index compression, associate time tags with plane positioning data, and construct standardized spatiotemporal data.

3. The intelligent road traffic risk warning method based on vehicle-road collaboration as claimed in claim 2, characterized in that: Based on the standardized spatiotemporal data, a dynamic hypergraph is constructed using a hypergraph neural network, and node risk characteristics are obtained to generate a comprehensive risk index. The specific steps are as follows: Based on standardized spatiotemporal data, micro-hyperedges are generated through Euclidean distance and behavioral similarity, and macro-hyperedges are generated by combining traffic flow density and road type to construct a dynamic hypergraph. Based on the dynamic hypergraph, hypergraph convolution is used to aggregate neighbor information to obtain node spatiotemporal features, generating node risk features and embedding vectors; The node risk features and embedding vectors are integrated and a comprehensive risk index is generated through dynamic thresholds.

4. The intelligent road traffic risk warning method based on vehicle-road collaboration as claimed in claim 3 is characterized by: The comprehensive risk index is converted into a binary pulse sequence by generating a pulse sequence, a pulse synchronization intensity matrix is ​​constructed, and a risk feature weight distribution vector is generated. The specific steps are as follows: The comprehensive risk index is converted into a binary pulse sequence by generating a pulse sequence, and the number of binary pulse triggers is obtained according to a fixed time window, and the pulse trigger frequency is output; Set a dynamic competition threshold, retain binary pulses whose pulse trigger frequency is greater than the dynamic competition threshold as a subset, and generate a set of high-frequency indicators that pass the screening and the corresponding binary pulse trigger frequency; Based on the binary pulse sequence corresponding to the high-frequency indicator set, the arrival time difference of the binary pulse sequence is obtained and normalized to construct the pulse synchronization intensity matrix; Based on the binary pulse trigger frequency and pulse synchronization intensity matrix, weighted fusion is used to generate the risk feature weight distribution vector.

5. The intelligent road traffic risk warning method based on vehicle-road collaboration according to claim 4 is characterized in that: The risk characteristic weight distribution vector is combined with the hypergraph neural network to obtain the probability distribution of the risk root cause and output the risk level and warning type. The specific steps are as follows: Taking the risk feature weight distribution vector as the node, the causal edge is initialized by combining the micro hyperedge and macro hyperedge to generate the risk indicator causal graph; Based on the risk indicator causal graph, hypergraph convolution is used to aggregate neighbor information to generate counterfactual results and normalize them to obtain the probability distribution of risk root causes; Combining the probability distribution of risk root causes with the pulse neural network, the pulse timing-dependent plasticity rule is used to strengthen the node activation intensity, and the exponential sliding average filter is used to generate a smooth distribution, generate a smooth distribution probability value, and output the risk level and warning type.

6. The intelligent road traffic risk warning method based on vehicle-road collaboration according to claim 5 is characterized by: The differentiated push warning type based on risk level generates risk events and uploads them encrypted through data fusion and risk feature extraction. The specific steps are as follows: The risk score is obtained by integrating the features of the comprehensive risk score, and the warning level is determined by logical judgment, and warning instructions matching the risk level are generated; Use dynamic time warping to align historical data with real-time data, and use conditional generative adversarial networks to generate structured risk event data; Based on structured risk event data, it is encapsulated and encrypted and uploaded in segments.

7. The intelligent road traffic risk warning method based on vehicle-road collaboration according to claim 6, characterized in that: The method of securely aggregating real-time road data through federated learning and updating the dynamic hypergraph in combination with differential privacy is as follows: Use random masks to generate mask parameters for parameter aggregation, and generate updated parameters through protocol reconstruction; The update parameters are quantized into 8-bit integers and divided into redundant data blocks using erasure codes, which are uploaded in parallel through multiple network channels.

8. An intelligent road traffic risk warning system based on vehicle-road collaboration, based on the intelligent road traffic risk warning method based on vehicle-road collaboration according to any one of claims 1 to 7, characterized in that: Including, spatiotemporal alignment module, hypergraph modeling module, risk screening module, root cause analysis module, warning push module and federated update module, The spatiotemporal alignment module is used to unify the real-time collected road data into the UTM coordinate system by combining time alignment and coordinate conversion to construct standardized spatiotemporal data; The hypergraph modeling module is used to build a dynamic hypergraph based on standardized spatiotemporal data using a hypergraph neural network, obtain node risk characteristics, and generate comprehensive risk indicators; The risk screening module is used to convert the comprehensive risk index into a binary pulse sequence by generating a pulse sequence, construct a pulse synchronization intensity matrix, and generate a risk feature weight distribution vector; The root cause analysis module is used to obtain the probability distribution of risk root causes based on the risk feature weight distribution vector and combine it with the hypergraph neural network to output the risk level and warning type; The early warning push module is used to push differentiated warning types based on risk levels. Through data fusion and risk feature extraction, risk event data is generated and encrypted for upload; The federated update module is used to securely aggregate real-time road data through federated learning and update the dynamic hypergraph in combination with differential privacy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the intelligent road traffic risk warning method based on vehicle-road collaboration as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the intelligent road traffic risk warning method based on vehicle-road collaboration described in any one of claims 1 to 7 are implemented.

Images