A multi-scene linkage vehicle abnormal state early warning method and system

Abstract

The application discloses a kind of multi-scene linkage's vehicle abnormal state early warning method and system, it is related to intelligent transportation technology field, including: each scene node in scene association event chain is mapped as cause-effect node in cause-effect diagram, and according to the dependent relationship between each cause-effect node, construct the propagation cause-effect directed graph indicating that the abnormal state activated by traffic scene identification result is propagated to associated scene;Scene evolution prediction is carried out based on the propagation cause-effect directed graph, and scene evolution prediction result is output.The application realizes the quantitative prediction of the propagation path of abnormal state to associated scene, activation probability and arrival time by constructing the propagation cause-effect directed graph, and taking the abnormal state index of root node as initial propagation intensity to calculate the activation probability and predicted arrival time of each child node along directed edge, supports proactive early warning.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method and system for early warning of abnormal vehicle conditions in multiple scenarios. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of motor vehicle ownership, road traffic safety issues are becoming increasingly prominent. Abnormal vehicle states (such as sudden deceleration, brake failure, and loss of steering control) are major causes of traffic accidents. However, the evolution of abnormal vehicle states is not an isolated event; it is often closely coupled with the specific traffic scenario (such as intersections, construction zones, and congested sections) and can propagate to surrounding areas through traffic flow interactions, triggering a wider chain of risks. Traditional vehicle warning systems typically rely solely on single-vehicle sensor data for single-state judgments, lacking deep integration of vehicle operating states with complex traffic environments. Furthermore, they struggle to capture the dynamic propagation and evolution of abnormal states across different traffic scenarios, resulting in insufficient accuracy and timeliness of warnings, failing to meet the growing demands for driving safety and proactive prevention.

[0003] Currently, existing vehicle abnormality warning technologies are mainly divided into two categories. One category is single-vehicle warning methods based on onboard sensors, such as analyzing data like vehicle speed, acceleration, and steering wheel angle on the CAN bus, and using threshold judgment or traditional machine learning models to identify abnormal driving behavior or vehicle mechanical failures. The other category is environmental warning methods based on roadside perception, which monitor traffic flow and potential hazards on the road using devices such as cameras and radar, and broadcast warning information to passing vehicles. In recent years, with the development of vehicle-road cooperative technology, some research has begun to attempt to initially combine vehicle status with the road environment, such as issuing a warning to following vehicles when a vehicle ahead brakes suddenly. In addition, a small number of works have used deep learning models to identify single traffic scenarios (such as highways or intersections) and trigger corresponding warnings according to preset rules.

[0004] However, the aforementioned existing technologies still have significant shortcomings and problems in practical applications. First, existing methods generally treat vehicle state detection and traffic scene identification as two relatively independent tasks, lacking the ability to dynamically model the deep correlation between the two, leading to false alarms or missed alarms in complex or abrupt scenarios. Second, existing technologies mainly provide reactive warnings for currently occurring abnormal states, i.e., post-event alerts, lacking the ability to predict the paths, probabilities, and arrival times of abnormal states that may propagate to other related scenarios, thus failing to achieve truly proactive pre-event warnings. Third, existing warning strategies are usually singular and fixed, such as issuing an alert only to the vehicle itself, failing to dynamically generate multi-level, multi-regional, and multi-object linkage warning schemes based on the multiple spatiotemporal ranges that abnormal states may affect. This results in a mismatch between the warning coverage and the risk level, potentially leading to insufficient warnings or excessive alerts.

[0005] Therefore, it is essential to invent a method and system for early warning of abnormal vehicle conditions that involves multiple scenarios to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for early warning of abnormal vehicle conditions in multiple scenarios, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a multi-scenario linkage vehicle abnormal state early warning method, comprising the following steps: S1: Obtain vehicle operation status data of at least one target vehicle within the target road area, as well as traffic environment perception data corresponding to the target road area; S2: Input the vehicle operating status data and the traffic environment perception data into a preset scene classification model and output the traffic scene identification result; and input the vehicle operating status data into a preset anomaly detection model and output the anomaly state index. S3: Based on the traffic scene identification result and the abnormal state index, call the preset scene association rule library to perform scene association matching and generate a scene association event chain; wherein, the scene association event chain contains at least one successor scene node; S4: Map each scene node in the scene-related event chain to a causal node in the causal graph, and assign each causal node a node attribute including spatial coordinates; construct a propagation causal directed graph based on the dependencies between the causal nodes; use the abnormal state index of the root node corresponding to the traffic scene identification result as the initial propagation intensity, and propagate step by step along the directed edge direction of the propagation causal directed graph, calculating the activation probability of each child node based on the edge weight of each directed edge; and for each child node, calculate the predicted arrival time corresponding to the child node based on the road path distance between the spatial coordinates of the child node and the spatial coordinates of the parent node and the estimated propagation speed; output the scene evolution prediction result; wherein, the scene evolution prediction result includes activation probability, propagation path and predicted arrival time. S5: Determine a multi-level linkage early warning strategy based on the scenario evolution prediction results; S6: Map the multi-level linkage early warning strategy to early warning execution instructions corresponding to each execution terminal, and distribute the early warning execution instructions to each execution terminal.

[0008] A multi-scenario linked vehicle abnormal state early warning system includes the following modules: The data acquisition module is used to acquire vehicle operation status data of at least one target vehicle within the target road area and traffic environment perception data corresponding to the target road area. The scene identification and anomaly detection module is used to input the vehicle operation status data and the traffic environment perception data into a preset scene classification model to output traffic scene identification results, and to input the vehicle operation status data into a preset anomaly detection model to output an anomaly state index. The scene association analysis module is used to perform scene association matching by calling a preset scene association rule library based on the traffic scene identification results and the abnormal state index, and generate a scene association event chain; wherein, the scene association event chain contains at least one successor scene node; The scene evolution prediction module is used to map each scene node in the scene-related event chain to a causal node in a causal graph, and assign node attributes including spatial coordinates to each causal node; construct a propagation causal directed graph based on the dependencies between the causal nodes; use the abnormal state index of the root node corresponding to the traffic scene identification result as the initial propagation intensity, and propagate step by step along the directed edge direction of the propagation causal directed graph; calculate the activation probability of each child node based on the edge weight of each directed edge; and for each child node, calculate the predicted arrival time corresponding to the child node based on the road path distance between the spatial coordinates of the child node and the spatial coordinates of the parent node and the estimated propagation speed; output the scene evolution prediction result; wherein, the scene evolution prediction result includes activation probability, propagation path, and predicted arrival time. The linkage strategy generation module is used to determine a multi-level linkage early warning strategy based on the scenario evolution prediction results; The instruction distribution module is used to map the multi-level linkage early warning strategy into early warning execution instructions corresponding to each execution terminal, and to distribute the early warning execution instructions to each execution terminal.

[0009] The technical effects and advantages of this invention are as follows:

[0010] 1. This invention achieves collaborative perception of vehicle status and traffic environment by deeply integrating vehicle operation status data with traffic environment perception data, and uses scene classification model and anomaly detection model to output traffic scene identification results and anomaly state index respectively, thereby improving the accuracy of anomaly state identification.

[0011] 2. This invention improves the recognition accuracy of complex traffic scenes by using a scene classification model based on a CNN-LSTM dual-stream network to extract spatial features from traffic environment perception data and extract temporal features from vehicle operation status data and then fuse them through attention weighting.

[0012] 3. This invention employs a semi-supervised anomaly detection model based on variational autoencoders, trains using only normal samples, performs reconstruction error analysis on the running state data, and outputs a normalized anomaly state index, thereby achieving accurate anomaly detection in the absence of a large amount of anomaly labeled data.

[0013] 4. This invention uses a pre-defined scene association rule library represented by a quadruple of "predecessor scene - triggering condition - successor scene - association strength" to perform scene association matching and generate scene association event chains, thereby realizing a structured description of the propagation relationship of abnormal states among multiple scenes.

[0014] 5. This invention constructs a causal directed graph and calculates the activation probability and predicted arrival time of each child node level by level along the directed edge using the abnormal state index of the root node as the initial propagation strength. This enables quantitative prediction of the propagation path, activation probability and arrival time of abnormal states to related scenarios, supporting proactive early warning.

[0015] 6. This invention determines a multi-level linkage early warning strategy, including early warning level, linkage scope, execution timing, and execution object, based on the scenario evolution prediction results, and maps them to early warning execution instructions for each corresponding execution terminal for distribution. This achieves accurate linkage early warning by level, region, and time period, avoiding insufficient or excessive early warning. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall method of the present invention;

[0017] Figure 2A schematic diagram illustrating the construction and evolution prediction of a causal directed graph for propagation;

[0018] Figure 3 This is a system module architecture diagram of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides, for example Figure 1 The method for early warning of abnormal vehicle status in multiple scenarios, as shown, includes the following steps: S1: Obtain vehicle operation status data of at least one target vehicle within the target road area, as well as traffic environment perception data corresponding to the target road area; In one embodiment, the specific process of step S1 is as follows: The target road area is defined by the communication coverage of roadside units (RSUs) or regional edge computing nodes. For example, the road segment corresponding to the wireless communication coverage of RSUs deployed on both sides of highways or urban expressways (based on LTE-V2X or DSRC protocols, with a coverage radius typically of 300–500 meters), or the road segment demarcated by a pre-defined geofence (a polygonal area enclosed by a set of continuous latitude and longitude coordinates), is used as the target road area to be monitored.

[0021] For at least one target vehicle that enters or is located within the target road area, its vehicle operating status data is acquired in real time via a vehicle-to-infrastructure (V2I) communication link. Specifically, the target vehicle periodically collects operating parameters from the vehicle's CAN bus via its onboard unit (OBU), which includes at least: instantaneous vehicle speed (unit: km / h) and longitudinal acceleration (unit: m / s²). 2 ), lateral acceleration (unit: m / s²) 2 The parameters include: steering wheel angle (degrees), brake master cylinder pressure (MPa), accelerator pedal opening (%), engine speed (rpm), ABS / ESP activation status, and vehicle hazard light status. The OBU encapsulates these parameters into standardized vehicle-to-infrastructure (V2I) messages (such as BSM messages defined in SAE J2735) and uploads them to the roadside unit via a wireless communication link.

[0022] Simultaneously, traffic environment perception data corresponding to the target road area is acquired. This traffic environment perception data is collected by roadside sensing devices deployed within the target road area and generated after fusion processing by a roadside computing unit. The roadside sensing devices include at least: a high-resolution traffic camera for acquiring road surface images and video streams; millimeter-wave radar for acquiring the target's position, speed, and trajectory; and a meteorological sensor for acquiring visibility, rainfall intensity, road surface temperature, and road surface slip coefficient. The roadside computing unit performs spatiotemporal alignment and target fusion on the multi-source sensing data, outputting structured traffic environment perception data. This data includes at least: lane line information, average headway for each lane, relative speed and distance between adjacent vehicles, road speed limit, road surface adhesion coefficient, and current weather conditions.

[0023] The acquired vehicle operation status data and traffic environment perception data are cleaned and timestamped to provide synchronized input data for step S2. Data cleaning includes removing outliers exceeding reasonable thresholds and filling in short-term missing continuous data (e.g., vehicle speed) using linear interpolation. Timestamping refers to interpolating both types of data to the same time reference (e.g., using the system time of the roadside unit as the reference, with a sampling frequency of 10Hz).

[0024] S2: Input the vehicle operating status data and the traffic environment perception data into a preset scene classification model and output the traffic scene identification result; and input the vehicle operating status data into a preset anomaly detection model and output the anomaly state index. The scene classification model is a multi-task learning model built on a CNN-LSTM dual-stream network. Among them, the CNN branch extracts spatial features from the traffic environment perception data and outputs scene spatial feature vectors; The LSTM branch extracts temporal features from the vehicle operating state data and outputs a state temporal feature vector. The scene space feature vector and the state temporal feature vector are fused by attention weighting to obtain a fused scene feature vector. Based on the fused scene feature vector and the preset scene category label set, the traffic scene identification result is output.

[0025] The training data for the scene classification model is constructed as follows: At least 72 hours of continuous video streams from roadside cameras and synchronous vehicle operation status data are collected from at least 10 different road areas (including urban expressways, highways, ordinary intersections, construction zones, tunnels, and ramps). Manual annotators label each frame of the image according to a predefined set of scene category labels (10 categories in total: urban road congestion, urban road smooth traffic, highway straight ahead, highway congestion, left turn at intersection, right turn at intersection, straight ahead at intersection, ramp merging, ramp exiting, and construction zone). Simultaneously, the center point coordinates of dangerous areas within the scene are labeled as the spatial location regression target. The training set, validation set, and test set are divided in an 8:1:1 ratio, with a total of no less than 50,000 labeled frames. The multi-task learning model uses a joint loss function L... total =L cls +λ·L loc Training is conducted, where L cls For cross-entropy classification loss, L loc The mean square error is used to determine the positioning loss, and λ is set to 0.5.

[0026] The anomaly detection model is a semi-supervised anomaly detection model based on variational autoencoders, and its training and inference stages are as follows: During the training phase, only normal vehicle operating state data samples are input into the variational autoencoder. The encoder maps the normal vehicle operating state data samples to the latent space to obtain the latent variable distribution. Then, the decoder reconstructs the latent variable distribution and outputs the reconstructed vehicle operating state data. The network parameters of the encoder and the decoder are optimized with the goal of minimizing the reconstruction error between the normal vehicle operating state data samples and the reconstructed vehicle operating state data, resulting in the trained variational autoencoder. During the inference phase, the vehicle operating state data obtained in step S2 is input into the trained variational autoencoder to obtain inference-reconstructed vehicle operating state data; the inference reconstruction error between the vehicle operating state data and the inference-reconstructed vehicle operating state data is calculated; the inference reconstruction error is input into a preset normalization mapping function to output the abnormal state index with a value in the interval [0, 1].

[0027] The selection criteria for the "normal vehicle operating status data sample" are: meeting all three of the following conditions simultaneously: The vehicle did not activate any ADAS warnings (FCW, LDW, etc.); The vehicle's hazard lights were not turned on; The absolute value of the vehicle's longitudinal acceleration is less than 2 m / s². 2 And the absolute value of the lateral acceleration is less than 1.5 m / s². 2 .

[0028] The data collection covered approximately 20,000 time window samples (10-second window length, 10Hz sampling frequency) covering different vehicle speed ranges (0-30km / h, 30-60km / h, 60-90km / h, and above 90km / h), different road surface adhesion coefficients (0.2-0.4 wet, 0.5-0.7 normal, and 0.8-1.0 dry), and different traffic flow densities (free flow, synchronous flow, and congested flow).

[0029] In this embodiment, the scene classification model is a multi-task learning model built based on a CNN-LSTM dual-stream network. Specifically: The CNN branch takes traffic environment perception data as input, which is a 3D tensor with dimensions (H, W, C), where H=224, W=224, and C=3, corresponding to RGB image frames captured by the vehicle-mounted camera. The CNN branch uses ResNet-18 as its backbone network, containing 5 convolutional blocks (each block consists of 2 convolutional layers with a kernel size of 3×3, a stride of 1, and padding of 1). Each convolutional block is followed by a 2×2 max-pooling layer, ultimately outputting a 2048-dimensional scene space feature vector.

[0030] LSTM Branch: The input is vehicle operating state data, which is time-series data within a time window T=10 seconds, with a sampling frequency of 10Hz. The feature dimension of each time step is D=12 (including vehicle speed, longitudinal acceleration, lateral acceleration, yaw rate, steering wheel angle, accelerator pedal opening, brake pedal opening, gear position, turn signal status, ABS activation status, ESP activation status, and GPS vehicle speed). The LSTM branch uses a two-layer LSTM with 128 hidden units in each layer, and outputs the hidden state of the last time step as a 256-dimensional state temporal feature vector.

[0031] Attention-weighted fusion: combining scene space feature vectors f s ∈R^2048 and the state-time eigenvector f t Each ∈R^256 is mapped to a common dimension d=128 through a fully connected layer to obtain f. s 'and f t '. Calculate the attention weight α = softmax(v^T·tanh(W) s ·f s '+W t ·f t ')), where W s W t ∈R^{128×128}, where v∈R^{128} are learnable parameters. The final fused feature vector f fuse =α·f s '+(1-α)·f tThe fused feature vector has a dimension of 128. This fused feature vector is then input into a fully connected layer (output dimension equal to the number of scene categories, set to K=10) and a Softmax function to output traffic scene identification results, such as "urban road congestion," "highway straight ahead," and "left turn at an intersection."

[0032] The anomaly detection model is a semi-supervised anomaly detection model based on variational autoencoders, and its training and inference stages are as follows: Model Structure: The variational autoencoder consists of an encoder and a decoder. The encoder comprises three fully connected layers with 128, 64, and 32 neurons respectively. The last layer outputs the mean μ and the logarithm of variance ln(σ) of the latent space. 2 Latent space dimension z dim =16. The decoder consists of three fully connected layers with 64, 128, and D neurons (D=12) respectively, outputting reconstructed vehicle operating state data. All hidden layers use the ReLU activation function.

[0033] Training Phase: Data samples of normal vehicle operation (i.e., operational data when the vehicle is without faults or dangerous events) are collected, totaling N=10000 time window samples. Each sample is input into the encoder to obtain the latent variable distribution q(z|x). The latent variable z is sampled using a reparameterization technique, and then reconstructed by the decoder to obtain x. recon The training loss function is the sum of the reconstruction error and the KL divergence: Loss = MSE(x, x). recon )+β·KL(q(z|x)||N(0,I)), where MSE is the mean squared error and β=0.5. With the goal of minimizing the loss, the Adam optimizer (learning rate 0.001) is used for iterative training until convergence, resulting in the trained variational autoencoder.

[0034] Inference phase: The vehicle operating status data x obtained in step S2... test Input the trained variational autoencoder to obtain inference reconstruction data x recon，test Calculate the inference reconstruction error err = MSE(x) test x recon，test The Sigmoid function is used as the normalization mapping function: Abnormal state exponent = 1 / (1+exp(-k·(err-μ)). err ))), where μ err The index is the mean of the reconstruction error of all normal samples in the training set, and k is the scaling factor (with a value of 5). The index takes values ​​in the range [0, 1], and the closer the index is to 1, the higher the degree of abnormality in the vehicle's operating state.

[0035] The scaling factor k in the normalized mapping function is determined as follows: a grid search is used on the validation set to optimize the distribution of the abnormal state index of normal samples concentrated in the interval [0, 0.3] and the distribution of the abnormal state index of artificially injected abnormal samples (simulated by changing the acceleration peak in the vehicle speed sequence or adding sensor noise) in the interval [0.7, 1]. The k value with the highest classification accuracy is selected; a typical k value is between 5 and 8. err This is the mean reconstruction error of the training set obtained after training is completed, with a typical value range of 0.05-0.15.

[0036] S3: Based on the traffic scene identification result and the abnormal state index, call the preset scene association rule library to perform scene association matching and generate a scene association event chain; wherein, the scene association event chain contains at least one successor scene node; The scene association rule base stores scene association rules, and each scene association rule is represented in the form of a quadruple of "predecessor scene-trigger condition-successor scene-association strength"; Wherein, the preceding scenario is the currently active traffic scenario; the triggering condition is that the abnormal state index exceeds the corresponding preset threshold and the spatiotemporal distance between the preceding scenario and the subsequent scenario is less than the preset association distance; the subsequent scenario is an associated traffic scenario that has a causal propagation relationship with the preceding scenario and will be affected by the abnormal state of the preceding scenario when the triggering condition is met; the association strength is a probability value determined by historical statistical data.

[0037] The construction method of the scenario association rule base is as follows: A natural driving dataset and a traffic accident database covering all scenario types in the scenario category label set for at least one year are obtained. The dataset includes vehicle operating status data, traffic environment perception data, and event chain records of actual abnormal events and their chain consequences. A causal association mining algorithm—an improved FP-Growth algorithm—is used, with a minimum support threshold of 0.05 and a minimum confidence threshold of 0.6. Frequently occurring "preceding scenario → subsequent scenario" sequence patterns are extracted from historical event chain data. For each mined pattern, the frequency of the subsequent scenario actually occurring within the next 30 seconds when the preceding scenario occurs and the triggering condition is met is calculated, and this frequency is used as the association strength. Furthermore, for rules with sparse data but clear causal relationships based on traffic engineering theory (such as "rear-end collision on highway → chain collision"), expert experience is used to supplement the data. The association strength is determined by the average score of independent scores from at least three traffic experts.

[0038] In this embodiment, the scene association rule base is stored in the vehicle's local storage or a cloud server, and is organized using a relational data table structure. Each scene association rule in the base is represented as a four-tuple of "previous scenario - trigger condition - successor scenario - association strength", as follows: Precursor Scene: The value is the currently active scene label in the traffic scene identification result output in step S2, such as "urban road congestion", "highway straight", "left turn at intersection", "ramp merging", "construction area", etc. The scene label set contains a total of K=10 preset categories.

[0039] Triggering condition: A logical AND expression, requiring both sub-conditions to be met simultaneously: Condition 1: The abnormal state index exceeds the preset threshold θ (in this embodiment, θ = 0.7, which is obtained by empirical value or offline calibration); Condition 2: The spatiotemporal distance between the preceding scenario and the subsequent scenario is less than a preset association distance. The spatiotemporal distance is calculated as follows: the estimated arrival time is obtained by dividing the Euclidean spatial distance between the location of the preceding scenario (marked with GPS coordinates and timestamps) and the predicted location of the subsequent scenario by the vehicle's current speed. If this estimated arrival time is less than a preset time threshold T... dist When the time reaches 10 seconds, the spatiotemporal distance is determined to meet the requirements.

[0040] Subsequent scenarios: These are traffic scenarios that have a causal relationship with the preceding scenarios. The causal relationship is pre-established based on traffic flow theory and accident chain analysis. For example, the subsequent scenarios of a "rear-end collision on a highway" scenario include "chain collisions," "congestion wave propagation," and "emergency lane occupancy."

[0041] Association strength: a probability value determined by historical statistical data, ranging from [0, 1]. The statistical method is as follows: in a large-scale natural driving dataset or traffic accident database, the frequency of subsequent scenarios actually occurring within a certain spatiotemporal window when the preceding scenario occurs and the triggering conditions are met is counted, and this frequency is used as the association strength.

[0042] The scenario association rule base contains the following specific rules: Rule 1: The preceding scenario is "straight on the highway". The triggering condition is that the abnormal state index is ≥0.7 and the estimated arrival time between the preceding scenario and the subsequent scenario "rampage merging" is less than 10 seconds. The subsequent scenario is "rampage merging" and the correlation strength is 0.82. Rule 2: The preceding scenario is "urban road congestion". The triggering conditions are an abnormal state index ≥ 0.7 and the estimated arrival time between the preceding scenario and the subsequent scenario "rear-end collision risk area" is less than 8 seconds. The subsequent scenario is "rear-end collision risk area" with a correlation strength of 0.76. Rule 3: The preceding scenario is "left turn at an intersection". The triggering condition is that the abnormal state index is ≥0.7 and the estimated arrival time between the preceding scenario and the subsequent scenario "side conflict zone" is less than 5 seconds. The subsequent scenario is "side conflict zone" and the association strength is 0.68. Rule 4: The preceding scenario is "construction area", the triggering condition is that the abnormal state index is ≥0.7 and the estimated arrival time between the preceding scenario and the subsequent scenario "lane reduction congestion" is less than 12 seconds, the subsequent scenario is "lane reduction congestion", and the correlation strength is 0.71; Rule 5: The preceding scenario is "exiting the ramp". The triggering condition is that the abnormal state index is ≥0.7 and the estimated arrival time between the preceding scenario and the subsequent scenario "mainline rear-end collision" is less than 6 seconds. The subsequent scenario is "mainline rear-end collision" with a correlation strength of 0.79.

[0043] The correlation strength in the above example rules is determined by a combination of historical statistical data and expert experience. The specific values ​​can be calibrated offline or adaptively adjusted online according to the traffic flow characteristics of different road areas.

[0044] The specific execution steps for the scene association matching are as follows:

[0045] Step S31: Obtain the current traffic scene identification result Sc and abnormal state index Ac output in step S2.

[0046] Step S32: In the scene association rule base, retrieve the rule subset R of all predecessor scenes equal to Sc.

[0047] Step S33: For each rule r in R, determine whether its triggering condition is met: If Ac < 0.7, the triggering condition is not met, and the rule is skipped. If Ac≥0.7, then the spatiotemporal distance between the current vehicle position and the center of the preset area of ​​effect of the subsequent scene is further calculated (the area of ​​effect of the subsequent scene is pre-stored in the rule base, for example, the area of ​​effect of the "ramp merging" scene is within 50 meters before and after the merging point of the ramp and the main road). If the estimated arrival time is less than 10 seconds, the trigger condition is met.

[0048] Step S34: Extract all rules whose triggering conditions are met, sort them from high to low according to their association strength, take the top M rules (M=3, which can be configured according to actual needs), and connect their subsequent scenes according to the spatiotemporal order of the triggering conditions to generate a scene association event chain. The event chain is represented by a directed graph structure, in the form of: Sc→(subsequent scene A, association strength 0.85)→(subsequent scene B, association strength 0.62).

[0049] Step S35: Output the event chain associated with the scenario for use in subsequent steps (such as risk warning or decision planning in S4).

[0050] Through the rule-based association matching described above, it is possible to predict the subsequent sequence of scenarios that may be affected from the current single scenario and anomaly index, thereby improving the vehicle's ability to predict continuous traffic event chains.

[0051] like Figure 2 As shown, S4: Map each scene node in the scene-related event chain to a causal node in the causal graph, and assign each causal node a node attribute including spatial coordinates; construct a propagation causal directed graph representing the propagation of the abnormal state activated by the traffic scene identification result to the related scene based on the dependency relationship between the causal nodes; use the abnormal state index of the root node corresponding to the traffic scene identification result as the initial propagation intensity, and propagate step by step along the directed edge direction of the propagation causal directed graph, and calculate the activation probability of each child node according to the edge weight of each directed edge; and for each child node, calculate the predicted arrival time corresponding to the child node according to the road path distance between the spatial coordinates of the child node and the spatial coordinates of the parent node and the estimated propagation speed; output the scene evolution prediction result; wherein, the scene evolution prediction result includes activation probability, propagation path and predicted arrival time. The method for constructing the directed graph of propagation causality is as follows: Traverse each scene node in the scene-related event chain, set each scene node as a causal node, and assign node attributes to each causal node; the node attributes include scene type label, the abnormal state index, spatial coordinates, and activation timestamp; The dependency relationships between each causal node are extracted from the scene association rule base, and each dependency relationship is mapped to a directed edge between the causal nodes; the edge weight of each directed edge is the association strength in the corresponding scene association rule; Based on each causal node and each directed edge, a propagation causal directed graph is constructed with the node corresponding to the traffic scene identification result as the root node.

[0052] The predicted arrival time in the scenario evolution prediction results is calculated as follows: For each activated child node, obtain the road path distance between the spatial coordinates of the child node and the spatial coordinates of the parent node, and determine the estimated propagation speed according to the scene type corresponding to the child node; The predicted arrival time = the activation timestamp of the parent node + (road path distance / estimated propagation speed).

[0053] In this embodiment, the method for constructing the causal directed graph is as follows: Step S41: Obtain the scene association event chain generated in step S3. The event chain consists of multiple scene nodes arranged in spatiotemporal order. Each node contains a scene type label (such as "highway straight", "ramp merging", "construction area" etc.) and the corresponding association strength.

[0054] Step S42: Traverse each scene node in the scene-related event chain, set each scene node as a causal node, and assign the following node attributes to each causal node: Scene type label: Takes a value from the preset label set output by the scene classification model; Abnormal state index: Inherited from the abnormal state index calculated in step S2; It should be noted that all nodes share this abnormal state index at the same time, but it is reserved for subsequent propagation calculations; Spatial coordinates: For the root node corresponding to the traffic scene identification result, if the current traffic scene is a point scene or a scene with a clear center (including but not limited to intersections, ramp merging points, and construction area starting points), the current GPS location projection coordinates of the target vehicle are taken as the spatial coordinates of the root node; if the current traffic scene is a linear scene or a planar scene (including but not limited to urban road congestion, highway straight-through, and tunnel sections), the spatial geometric center point coordinates along the road direction of the scene are taken as the spatial coordinates of the root node. The center point coordinates are obtained by discretizing the centerline of the road segment where the target vehicle is currently located into equidistant points and calculating the geometric mean; for child nodes, their spatial coordinates are taken as the projection coordinates of the center point of the subsequent scene's area of ​​action pre-stored in the scene association rule base. Activation timestamp: The activation timestamp of the root node is set to the current system time t0, and the activation timestamp of the child nodes is temporarily set to null, to be filled in after the predicted arrival time is calculated.

[0055] Step S43: Extract the dependency relationships between each causal node from the scene association rule base. Specifically, for any two adjacent nodes in the scene association event chain... i →Node j Search the rule base for predecessor scenarios where Node is the first instance. i The subsequent scenario is Node. j The rule is used to obtain the association strength (probability value) within that rule. This dependency is then mapped to directed edges between causal nodes, with the direction pointing from the predecessor node to the successor node. The weight w of each directed edge is... ij It equals the association strength in the corresponding rule.

[0056] Step S44: Using the node corresponding to the traffic scene identification result (i.e., the currently active scene output in step S2, which is also the starting node of the scene-related event chain) as the root node, and combining all causal nodes obtained in step S42 and all directed edges obtained in step S43, construct a directed acyclic graph (DAG), which is the propagation causal directed graph. If there are branches in the event chain (i.e., the same predecessor node corresponds to multiple successor nodes), it is represented in the graph as multiple child node branches of the root node; if the event chain is linear, the graph is a path.

[0057] The steps for generating the scene evolution prediction results are as follows: Step S45: Use the abnormal state index A0 (within the range [0, 1]) corresponding to the root node as the initial propagation strength. Propagate step by step along the directed edges of the propagation causal directed graph, calculating the activation probability of each child node. For the directed edge from parent node i to child node j in the graph, the activation probability P of child node j is... j Calculate using the following formula: P j =P i ×w ij , Among them, P i The activation probability of parent node i (P of the root node) root =A0), w ij The edge weight represents the association strength. If a child node has multiple parent nodes (cycles are not allowed in the graph, but convergence is), the maximum value or weighted sum of all incoming probabilities is taken. In this implementation, the maximum value, P, is used. j =max(P i ×w ij To avoid excessive probability inflation.

[0058] Step S46: For each activated child node, i.e., P j Child nodes with a value >0.01 are considered validly activated, based on the spatial coordinates (x, y) of that child node. j y j ) and the spatial coordinates (x) of the parent node i y i The road path distance d between ) ij =sqrt((x j -x i )^2+(y j -y i )^2), and the estimated propagation speed v prop Calculate the predicted arrival time corresponding to the child node.

[0059] Estimated propagation speed v prop Dynamically set according to scenario type and road environment parameters: For congestion propagation scenarios, vprop A value of 5 m / s (approximately 18 km / h) is used, based on empirical statistics of congestion wave speeds, and is applicable to situations where free flow transitions to congestion flow on expressways and highways; for accident cascading scenarios, v prop Take 0.5 × current vehicle speed, where the current vehicle speed is the real-time average vehicle speed on the road section corresponding to the associated scenario of this sub-node (calculated from roadside radar or vehicle-reported data). The coefficient 0.5 is determined by the statistical regression relationship between the propagation speed of the shock wave from a traffic accident and the traffic flow speed. Its value is based on the following: when the vehicle speed is between 20km / h and 120km / h, the propagation speed of the shock wave is approximately 0.4 to 0.6 times the vehicle speed. In this implementation, the middle value of 0.5 is taken. For slippery roads or low visibility weather, v prop The aforementioned baseline value is multiplied by a road surface adhesion coefficient correction factor (ranging from 0.6 to 0.9) to reflect the actual physical process of the abnormal state propagation being hindered or accelerated.

[0060] The absolute predicted arrival time of the child node is t. j =t i +Δt ij , where t i The activation timestamp of the parent node (the root node's t) root =t0), Δt ij The propagation time delay is obtained by dividing the path distance between the child node and the parent node by the estimated propagation speed, i.e., Δt. ij =Road path distance / Estimated propagation speed; If a child node has multiple parent nodes, take the smallest Δt. ij The corresponding arrival time.

[0061] Step S47: Combine the activation probability, propagation path, and predicted arrival time of each child node to generate the scene evolution prediction result. The propagation path is represented as a sequence of nodes from the root node to that child node, for example: [Highway straight → Ramp merging → Construction area]. The final output scene evolution prediction result can be organized in JSON format or tabular form, containing the following fields: child node scene label, activation probability, predicted arrival time (absolute timestamp or relative time difference), and complete propagation path.

[0062] By constructing the causal directed graph and performing step-by-step propagation calculations, the propagation effect of abnormal states in related scenarios can be quantified, providing quantitative evolutionary prediction information for downstream decision-making modules.

[0063] S5: Determine a multi-level linkage early warning strategy based on the scenario evolution prediction results; The steps for generating the multi-level linkage early warning strategy are as follows: The spatiotemporal attribute parameters and influence range parameters of each associated scenario in the scenario evolution prediction results are obtained, and the threat value of each associated scenario is calculated by combining the activation probability of each associated scenario with a preset threat evaluation function. Based on the threat level value of each associated scenario, the corresponding warning level is retrieved from the preset warning level mapping table; The linkage range is determined based on the influence range parameters of each of the associated scenarios, the execution timing is determined based on the predicted arrival time of each of the associated scenarios, and the execution object is determined based on the linkage range; the execution object includes at least: roadside variable information signs, vehicle terminals, area broadcasting equipment, and mobile communication terminals within the linkage range; The multi-level linkage early warning strategy is generated by combining the early warning level, the linkage range, the execution timing, and the execution object.

[0064] In this embodiment, the steps for generating the multi-level linkage early warning strategy are as follows: Step S51: Obtain the scene evolution prediction results output in step S4. These results include the following parameters for each associated scene (i.e., all child nodes in the propagation causal directed graph except for the root node): Scene type tag j ; Activation probability P j (Value range [0, 1]); Predicted arrival time t j (Absolute timestamp) or relative time difference Δt j ; Spatial coordinates (x) j y j ); Radius of influence R j (Based on the scenario type, the influence radius is preset: for example, the radius of influence for the "congestion propagation" scenario is 200 meters, for the "accident chain" scenario it is 100 meters, and for the "construction area" it is 50 meters. These radius values ​​are stored in the scenario association rule library in advance.)

[0065] Step S52: Combining the activation probability and spatiotemporal attribute parameters of each associated scenario, calculate the threat value (Threat) of each associated scenario using a preset threat assessment function. j The threat assessment function is designed in the following form: Threat j =P j ×(α×(1 / (1+exp(-β×(T max -Δt j ))))+γ×(Impact j )), Among them, P j The activation probability directly reflects the likelihood of this scenario occurring. Δt j To predict the arrival time (in seconds), T max A preset time decay threshold (30 seconds in this embodiment) is used, employing a Sigmoid-type time decay factor to ensure that the closer the arrival time (Δt) is... j The smaller the value, the greater the contribution to threat level; Impact j The inherent impact coefficient for the scene is predefined according to the scene type: for example, "rear-end collision on highway" is 0.9, "rampage merging conflict" is 0.7, "congestion wave" is 0.5, and "construction area" is 0.3. α, β, and γ are weighting coefficients that satisfy α + γ = 1. In this embodiment, α = 0.6, β = 0.2, and γ = 0.4.

[0066] Calculated Threat j The value ranges from [0, 1], with a larger value indicating a more serious threat.

[0067] The parameters in the threat assessment function described above are determined as follows: T max The 30-second timeframe was chosen because statistical analysis shows that warnings exceeding 30 seconds offer limited benefit to driver reaction time, and excessively long prediction windows introduce unacceptable prediction errors. Based on statistics of the time interval between the abnormality of the vehicle ahead and the accident of the vehicle behind in 200 traffic accidents, 75% of chain events occurred within 30 seconds.

[0068] α and γ: Determined using the Analytic Hierarchy Process (AHP). A judgment matrix was constructed to compare the importance of "time urgency" and "inherent impact of the scenario". Ten experts in the transportation field were invited to score the results, and α=0.6, γ=0.4, and the consistency ratio CR=0.02<0.1, which meets the consistency requirements.

[0069] β: Take 0.2, such that when Δt j From T max When the time decay factor decreases from 30 seconds to 0 seconds, it monotonically increases from approximately 0.5 to approximately 0.95, ensuring moderate sensitivity of the time factor. The selection of the β value was determined through simulation experiments: in 30 different Δt... j In the test cases, the β value that maximizes the correlation coefficient between the warning time and the actual risk time is 0.2.

[0070] Impact jThe criteria are based on historical accident fatality rates and average economic losses. For example, "rear-end collisions on highways" are 0.9 (fatality rate higher than 30%), "rampage merging conflicts" are 0.7 (fatality rate 10%~30%), "congestion waves" are 0.5 (fatality rate lower than 5%), and "construction areas" are 0.3 (mainly property damage).

[0071] Warning level mapping thresholds (0.8, 0.6, 0.3): determined based on ROC curve analysis, collecting 1000 sets of historical abnormal event samples, using the threat level value as a classifier, with the matching rate between the warning level and the actual severity level of the event (comprehensively assessed by property damage and casualties after the accident) as the optimization objective, and obtaining the optimal threshold by maximizing the Youden index.

[0072] Step S53: Based on the threat level value of each associated scenario. j The system retrieves the corresponding warning level from a pre-defined warning level mapping relationship. In this implementation, warning levels are divided into four levels: Level I (red), Level II (orange), Level III (yellow), and Level IV (blue). The mapping rules are as follows: when the threat value is greater than or equal to 0.8, it corresponds to Level I (highest) warning; when the threat value is greater than or equal to 0.6 and less than 0.8, it corresponds to Level II warning; when the threat value is greater than or equal to 0.3 and less than 0.6, it corresponds to Level III warning; and when the threat value is less than 0.3, it corresponds to Level IV (lowest) warning. If multiple related scenarios correspond to different warning levels, the highest level is taken as the overall warning level of the current system, while retaining the independent level of each scenario for tiered handling.

[0073] Step S54: Determine the scope of the linkage, the timing of execution, and the target of execution: Linkage Scope: Determined based on the influence scope parameters of each associated scenario. The influence scope of each associated scenario (in terms of (x...) j y j ( ) is the center of the circle, and the radius is R. j The union of the circular regions forms the total linkage area. Simultaneously, a buffer zone is extended 100 meters outwards from the original area. The final geometric description of the linkage range is a set of polygonal or circular regions.

[0074] Execution timing: based on the predicted arrival time t of each of the associated scenarios. j Determined. The predicted arrival time of the earliest arriving associated scenario is taken as the initial warning execution time; for each scenario, the lead time Δt... advance Set to 5 seconds (i.e., warning issuance time = t) j -5 seconds) to ensure sufficient reaction time.

[0075] Target objects: Determined based on the aforementioned linkage range. The infrastructure and terminal equipment included within the linkage range include: roadside variable message signs (VMS, placed above or to the side of the road, one every 200 meters), vehicle-mounted terminals (via vehicle-to-everything (OBU) communication), area broadcasting equipment (such as roadside loudspeakers, FM radio micro base stations), and mobile communication terminals (pushing messages to mobile phone users' apps or SMS messages via 4G / 5G). The priority for selecting target objects is as follows: first, activate VMS and vehicle-mounted terminals for targeted warnings; then, trigger area broadcasting equipment to cover traffic participants without vehicle-mounted terminals; finally, remote push notifications are sent via mobile communication terminals.

[0076] Step S55: Combine the warning level, the linkage range, the execution timing, and the execution object to generate the multi-level linkage warning strategy. The strategy is output in structured data form, such as JSON format.

[0077] Through the above steps, a multi-level linkage early warning strategy with hierarchical, regional, and time-based approaches can be dynamically generated based on the scenario evolution prediction results, thereby achieving accurate and efficient traffic risk early warning.

[0078] S6: Map the multi-level linkage early warning strategy to early warning execution instructions corresponding to each execution terminal, and distribute the early warning execution instructions to each execution terminal.

[0079] The mapping and distribution of the early warning execution command are specifically as follows: Obtain the device attribute information and communication protocol type of each execution object in the multi-level linkage early warning strategy; Based on the communication protocol type of each execution object, the warning level and execution timing are encoded into warning execution instruction formats corresponding to each execution object; among them, roadside variable information signs are encoded as display content instructions and display duration instructions; vehicle-mounted terminals are encoded as voice broadcast instructions and interface pop-up instructions; area broadcasting equipment is encoded as broadcast frequency band instructions and broadcast content instructions; and mobile communication terminals are encoded as SMS push instructions and application push instructions. The encoded warning execution command is distributed to each execution object through the corresponding communication channel.

[0080] In this embodiment, the mapping and distribution of the early warning execution command specifically refers to: Step S61: Obtain the multi-level linkage early warning strategy output in step S5, and extract the device attribute information and communication protocol type of each execution object. Device attribute information includes a unique device identifier. The communication protocol type is determined according to the classification of the execution object: roadside variable message signs use TCP / IP or NTCIP protocols; vehicle-mounted terminals use SAE J2735 message sets based on DSRC or C-V2X; area broadcasting equipment uses RDS or a custom serial port protocol; mobile communication terminals use HTTP / 2 or SMPP protocols.

[0081] Step S62: Based on the communication protocol type of each execution object, encode the warning level and execution timing in the multi-level linkage warning strategy into a warning execution instruction format corresponding to each execution object. The specific encoding method is as follows: For roadside variable message signs, the display content instruction and display duration instruction are encoded. The display content instruction includes a text string and graphic code, while the display duration instruction sets the display duration in seconds.

[0082] For vehicle-mounted terminals, the codes are coded into voice broadcast commands and interface pop-up commands. Voice broadcast commands include the broadcast text, language, volume level, and number of broadcasts; interface pop-up commands include the pop-up title, pop-up content, background color determined according to the warning level, and an automatic closing countdown.

[0083] For area broadcasting equipment, the encoding consists of a broadcast frequency band instruction and a broadcast content instruction. The broadcast frequency band instruction specifies the FM transmission frequency or the channel number for loudspeaker broadcasting; the broadcast content instruction contains an audio file URL or real-time text-to-speech text content, and for FM broadcasting, the transmission power and coverage radius must also be set.

[0084] For mobile communication terminals, the encoding is divided into SMS push commands and application push commands. SMS push commands include a list of target mobile phone numbers, SMS text content, and sending priority; application push commands include a target device token, push title, push body, additional data, and push expiration time.

[0085] Step S63: Distribute the encoded warning execution command to each execution target through the corresponding communication channel. The communication channels are allocated as follows: roadside variable message signs send via the Ethernet interface or cellular network of the roadside unit; vehicle-mounted terminals send via the C-V2XPC5 interface or the Uu interface of the cellular network; area broadcasting equipment sends via a dedicated line or cellular network to the broadcast control server; mobile communication terminals send via the operator's SMS gateway or mobile application push service. After each command is sent, a response confirmation message is received from each execution target. If no confirmation is received, the command is retried after a preset time, up to a maximum of a preset number of retries. For terminals that still fail after retrying, a failure log is recorded and supplementary notification is sent via a backup channel.

[0086] A multi-scenario linked vehicle abnormal state early warning system, such as Figure 3 As shown, it includes the following modules: The data acquisition module is used to acquire vehicle operation status data of at least one target vehicle within the target road area and traffic environment perception data corresponding to the target road area. The scene identification and anomaly detection module is used to input the vehicle operation status data and the traffic environment perception data into a preset scene classification model to output traffic scene identification results, and to input the vehicle operation status data into a preset anomaly detection model to output an anomaly state index. The scene association analysis module is used to perform scene association matching by calling a preset scene association rule library based on the traffic scene identification results and the abnormal state index, and generate a scene association event chain; wherein, the scene association event chain contains at least one successor scene node; The scene evolution prediction module is used to map each scene node in the scene-related event chain to a causal node in a causal graph, and assign node attributes including spatial coordinates to each causal node; construct a propagation causal directed graph based on the dependencies between the causal nodes; use the abnormal state index of the root node corresponding to the traffic scene identification result as the initial propagation intensity, and propagate step by step along the directed edge direction of the propagation causal directed graph; calculate the activation probability of each child node based on the edge weight of each directed edge; and for each child node, calculate the predicted arrival time corresponding to the child node based on the road path distance between the spatial coordinates of the child node and the spatial coordinates of the parent node and the estimated propagation speed; output the scene evolution prediction result; wherein, the scene evolution prediction result includes activation probability, propagation path, and predicted arrival time. The linkage strategy generation module is used to determine a multi-level linkage early warning strategy based on the scenario evolution prediction results; The instruction distribution module is used to map the multi-level linkage early warning strategy into early warning execution instructions corresponding to each execution terminal, and to distribute the early warning execution instructions to each execution terminal.

[0087] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early warning of abnormal vehicle states in multiple scenarios, characterized in that, Includes the following steps: S1: Obtain vehicle operation status data of at least one target vehicle within the target road area, as well as traffic environment perception data corresponding to the target road area; S2: Input the vehicle operating status data and the traffic environment perception data into a preset scene classification model, and output the traffic scene identification result; In addition, the vehicle operating status data is input into a preset anomaly detection model, and an anomaly status index is output. S3: Based on the traffic scene identification result and the abnormal state index, call the preset scene association rule library to perform scene association matching and generate a scene association event chain; wherein, the scene association event chain contains at least one successor scene node; S4: Map each scene node in the scene-related event chain to a causal node in the causal graph, and assign each causal node a node attribute including spatial coordinates; construct a propagation causal directed graph based on the dependencies between the causal nodes; use the abnormal state index of the root node corresponding to the traffic scene identification result as the initial propagation intensity, and propagate step by step along the directed edge direction of the propagation causal directed graph, calculating the activation probability of each child node based on the edge weight of each directed edge; and for each child node, calculate the predicted arrival time corresponding to the child node based on the road path distance between the spatial coordinates of the child node and the spatial coordinates of the parent node and the estimated propagation speed; output the scene evolution prediction result; wherein, the scene evolution prediction result includes activation probability, propagation path and predicted arrival time. S5: Determine a multi-level linkage early warning strategy based on the scenario evolution prediction results; the multi-level linkage early warning strategy includes early warning level, linkage scope, execution timing, and execution target; S6: Map the multi-level linkage early warning strategy to early warning execution instructions corresponding to each execution terminal, and distribute the early warning execution instructions to each execution terminal.

2. The method for early warning of abnormal vehicle status in multiple scenarios according to claim 1, characterized in that, The scene classification model is a multi-task learning model built on a CNN-LSTM dual-stream network. Among them, the CNN branch extracts spatial features from the traffic environment perception data and outputs scene spatial feature vectors; The LSTM branch extracts temporal features from the vehicle operating state data and outputs a state temporal feature vector. The scene space feature vector and the state temporal feature vector are fused by attention weighting to obtain a fused scene feature vector. Based on the fused scene feature vector and the preset scene category label set, the traffic scene identification result is output.

3. The method for early warning of abnormal vehicle status in multiple scenarios according to claim 1, characterized in that, The anomaly detection model is a semi-supervised anomaly detection model based on variational autoencoders, and its training and inference stages are as follows: During the training phase, only normal vehicle operating state data samples are input into the variational autoencoder. The encoder maps the normal vehicle operating state data samples to the latent space to obtain the latent variable distribution. Then, the decoder reconstructs the latent variable distribution and outputs the reconstructed vehicle operating state data. With minimizing the reconstruction error between the normal vehicle operating state data sample and the reconstructed vehicle operating state data as the training objective, the network parameters of the encoder and the decoder are optimized to obtain the trained variational autoencoder. During the inference phase, the vehicle operating state data obtained in step S2 is input into the trained variational autoencoder to obtain inference-reconstructed vehicle operating state data. Calculate the inference and reconstruction error between the vehicle operating status data and the inferred and reconstructed vehicle operating status data; The inference reconstruction error is input into a preset normalization mapping function, and the abnormal state index with a value in the interval [0, 1] is output.

4. The method for early warning of abnormal vehicle status in multiple scenarios according to claim 1, characterized in that, The scene association rule base stores scene association rules, and each scene association rule is represented in the form of a quadruple of "predecessor scene-trigger condition-successor scene-association strength"; Wherein, the preceding scenario is the currently active traffic scenario; the triggering condition is that the abnormal state index exceeds the corresponding preset threshold and the spatiotemporal distance between the preceding scenario and the subsequent scenario is less than the preset association distance; the subsequent scenario is an associated traffic scenario that has a causal propagation relationship with the preceding scenario and will be affected by the abnormal state of the preceding scenario when the triggering condition is met; the association strength is a probability value determined by historical statistical data.

5. The method for early warning of abnormal vehicle status in multiple scenarios according to claim 4, characterized in that, The method for constructing the directed graph of propagation causality is as follows: Traverse each scene node in the scene-related event chain, set each scene node as a causal node, and assign node attributes to each causal node; the node attributes include scene type label, abnormal state index, spatial coordinates, and activation timestamp; The dependency relationships between each causal node are extracted from the scene association rule base, and each dependency relationship is mapped to a directed edge between the causal nodes; the edge weight of each directed edge is the association strength in the corresponding scene association rule; Based on each causal node and each directed edge, a propagation causal directed graph is constructed with the node corresponding to the traffic scene identification result as the root node.

6. The method for early warning of abnormal vehicle status in multiple scenarios according to claim 1, characterized in that, The predicted arrival time in the scenario evolution prediction results is calculated as follows: For each activated child node, obtain the road path distance between the spatial coordinates of the child node and the spatial coordinates of the parent node, and determine the estimated propagation speed according to the scene type corresponding to the child node; The predicted arrival time = the activation timestamp of the parent node + (road path distance / estimated propagation speed).

7. The method for early warning of abnormal vehicle status in multiple scenarios according to claim 1, characterized in that, The steps for generating the multi-level linkage early warning strategy are as follows: The spatiotemporal attribute parameters and influence range parameters of each associated scenario in the scenario evolution prediction results are obtained, and the threat value of each associated scenario is calculated by combining the activation probability of each associated scenario with a preset threat evaluation function. Based on the threat level value of each associated scenario, the corresponding warning level is retrieved from the preset warning level mapping table; The linkage range is determined based on the influence range parameter of each of the associated scenarios, the execution timing is determined based on the predicted arrival time of each of the associated scenarios, and the execution object is determined based on the linkage range. The execution targets include at least: roadside variable information signs, vehicle-mounted terminals, area broadcasting equipment, and mobile communication terminals within the linkage range; The multi-level linkage early warning strategy is generated by combining the early warning level, the linkage range, the execution timing, and the execution object.

8. The method for early warning of abnormal vehicle status in multiple scenarios according to claim 1, characterized in that, The mapping and distribution of the early warning execution command are specifically as follows: Obtain the device attribute information and communication protocol type of each execution object in the multi-level linkage early warning strategy; Based on the communication protocol type of each execution object, the warning level and execution timing are encoded into warning execution instruction formats corresponding to each execution object; among them, roadside variable information signs are encoded as display content instructions and display duration instructions; vehicle-mounted terminals are encoded as voice broadcast instructions and interface pop-up instructions; area broadcasting equipment is encoded as broadcast frequency band instructions and broadcast content instructions; and mobile communication terminals are encoded as SMS push instructions and application push instructions. The encoded warning execution command is distributed to each execution object through the corresponding communication channel.

9. A multi-scenario linked vehicle abnormal state early warning system, characterized in that, Includes the following modules: The data acquisition module is used to acquire vehicle operation status data of at least one target vehicle within the target road area and traffic environment perception data corresponding to the target road area. The scene identification and anomaly detection module is used to input the vehicle operation status data and the traffic environment perception data into a preset scene classification model to output traffic scene identification results, and to input the vehicle operation status data into a preset anomaly detection model to output an anomaly state index. The scene association analysis module is used to perform scene association matching by calling a preset scene association rule library based on the traffic scene identification results and the abnormal state index, and generate a scene association event chain; wherein, the scene association event chain contains at least one successor scene node; The scene evolution prediction module is used to map each scene node in the scene-related event chain to a causal node in a causal graph, and assign node attributes including spatial coordinates to each causal node; construct a propagation causal directed graph based on the dependencies between the causal nodes; use the abnormal state index of the root node corresponding to the traffic scene identification result as the initial propagation intensity, and propagate step by step along the directed edge direction of the propagation causal directed graph; calculate the activation probability of each child node based on the edge weight of each directed edge; and for each child node, calculate the predicted arrival time corresponding to the child node based on the road path distance between the spatial coordinates of the child node and the spatial coordinates of the parent node and the estimated propagation speed; output the scene evolution prediction result; wherein, the scene evolution prediction result includes activation probability, propagation path, and predicted arrival time. The linkage strategy generation module is used to determine a multi-level linkage early warning strategy based on the scenario evolution prediction results; The instruction distribution module is used to map the multi-level linkage early warning strategy into early warning execution instructions corresponding to each execution terminal, and to distribute the early warning execution instructions to each execution terminal.

Images