A traffic large model training method, a traffic intelligent scheduling method and equipment

Abstract

This application provides a method for training a large-scale traffic model, a method for intelligent traffic scheduling, and an apparatus, relating to the field of intelligent transportation technology. The method includes constructing a traffic knowledge graph using road networks, equipment, traffic regulations, environmental states, and traffic events as nodes, and causal / spatiotemporal relationships between nodes as edges. The traffic knowledge graph is then transformed into differentiable graph embedding vectors. At each edge node, the graph embedding vectors are used to enhance the features of training samples, resulting in enhanced samples with graph logic constraints. Based on the entities in the enhanced samples, the actual evolution paths of traffic events corresponding to the enhanced samples are retrieved in the traffic knowledge graph. Based on the enhanced samples, the large-scale traffic model is locally trained, and the local parameters of the large-scale traffic model are updated based on the deviation loss between the traffic event evolution paths generated by the large-scale model and the actual evolution paths. This invention improves the accuracy and interpretability of the inference results.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and more specifically, to a method for training a large traffic model, a method for intelligent traffic scheduling, and a device. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of motor vehicle ownership, Intelligent Transportation Systems (ITS) have emerged, with the core objective of achieving accurate perception, efficient scheduling, and intelligent management of traffic flow.

[0003] In existing intelligent transportation systems, a retrieval-enhanced generation model is typically adopted, which involves identifying traffic events using a general large model and then retrieving the subsequent evolution path of the events based on a static knowledge graph.

[0004] However, in this approach, the knowledge graph is only used as an external reference, which results in the evolution path of the large model generation lacking internal logical constraints. It is difficult to internalize traffic regulations and accident evolution patterns, and it is very easy to produce illusory results that violate physical common sense or regulations, thus reducing the reliability of the output results of the large model. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a large traffic model training method, a traffic intelligent scheduling method and device, so as to improve the accuracy of the inference results of large models specifically for the traffic field and reduce illusions.

[0006] Firstly, a method for training a large-scale traffic model is provided, including: A traffic knowledge graph is constructed using road network, equipment, traffic regulations, environmental conditions, and traffic events as nodes, and causal / spatiotemporal relationships between nodes as edges; the traffic knowledge graph is then transformed into a differentiable graph embedding vector. At each edge node, the training samples are augmented with graph embedding vectors to obtain augmented samples with graph logic constraints. Based on the entities in the augmented samples, the actual evolution path of the traffic event corresponding to the augmented sample is retrieved in the traffic knowledge graph. Based on augmented samples, the traffic model is locally trained, and the local parameters of the traffic model are updated based on the deviation loss between the traffic event evolution path generated by the large model and the actual evolution path.

[0007] Optionally, feature enhancement of the training samples can be performed using graph embedding vectors to obtain enhanced samples with graph logic constraints, including: Based on the raw data of the training samples, key entities are identified and extracted, and the key entities are mapped to corresponding nodes in the traffic knowledge graph; key entities include at least one of road network, equipment, traffic regulations, environmental status and traffic events; By utilizing the relationships between nodes in the graph, auxiliary feature vectors associated with key entities can be retrieved; The auxiliary feature vector is fused with the original feature vector of the training sample to obtain an enhanced sample with graph logic constraints.

[0008] Optionally, the entities based on the augmented samples can retrieve the actual evolution paths of traffic events corresponding to the augmented samples in the traffic knowledge graph, including: The road network nodes obtained from the mapping in the enhanced samples are used as the starting point for traversal. By using graph attention mechanism to perform multi-hop search along the associated edges of the traffic knowledge graph, the sequence of successor nodes that matches the preset regulatory constraints and event evolution logic is selected; and the actual evolution path is constructed based on the road network nodes and the sequence of successor nodes.

[0009] Optionally, the method also includes: Upload the local parameters of each edge node to the cloud for global aggregation and optimization; Iteratively perform local training and global optimization until convergence to obtain a large-scale traffic model with logical reasoning capabilities in the traffic domain.

[0010] Secondly, a traffic intelligent scheduling method is provided, including: Acquire multi-source sensor data within the target area and preprocess the multi-source sensor data to obtain structured event vectors; then use the graph embedding vectors of the traffic knowledge graph to enhance the features of the structured event vectors. Based on the traffic model trained by the first method, reasoning is performed on the enhanced structured event vectors to output the target event, event evolution path and evolution result under the constraints of the traffic knowledge graph. Based on the target event, its evolution path, and the outcome, dispatch instructions are generated for each traffic control terminal within the target area and sent to each traffic control terminal.

[0011] Optionally, the structured event vector includes spatiotemporal coordinates, event type identifiers, and confidence scores; before generating scheduling instructions, the method further includes: Based on the spatiotemporal coordinates, event type identifiers, and confidence scores of structured event vectors, real-time dynamic features related to the target event are extracted. These real-time dynamic features include the probability of event occurrence, the radius of influence, and the expected duration. Based on the probability of an event occurring, the radius of its impact, and the expected duration, the risk level of the target event is calculated using a pre-set risk assessment model. The priority of handling target events is determined based on the risk level, and the priority is used as a constraint for generating subsequent scheduling instructions.

[0012] Optionally, based on the target event, its event evolution path, and the evolution result, the dispatch instructions for each traffic control terminal within the target area are generated, including: We analyze the key node sequence in the evolution path of traffic incidents and extract the impact characteristics of the evolution path on traffic flow, road network capacity and potential secondary risks at different time steps. Based on the priority of handling target events, a multi-objective optimization function is constructed with the optimization objectives of minimizing congestion duration, maximizing the probability of safe passage, and minimizing resource consumption. The optimal intervention strategy sequence is planned according to the impact characteristics. The optimal intervention strategy sequence is mapped to a set of executable instructions for each traffic control terminal within the target area. The set of executable instructions includes at least signal timing adjustment schemes, lane control instructions, variable message sign guidance information, and emergency vehicle guidance routes.

[0013] Optionally, multi-source sensor data within the target area is acquired, and the multi-source sensor data is preprocessed to obtain a structured event vector, including: Acquire fixed sensing data and mobile sensing data within the target area; Semantic alignment and fusion are performed on fixed sensing data and mobile sensing data to obtain an initial structured event vector; A consistency check is performed on several initial structured event vectors within the same spatiotemporal window to obtain the final structured event vector.

[0014] Optionally, the method also includes: Obtain environmental feedback data from each traffic control terminal when executing the current dispatch command; The cumulative reward value of the current strategy is calculated based on environmental feedback data and a pre-built composite reward function; the composite reward function includes a traffic efficiency gain term, a safety risk reduction term, and an energy consumption control term. The cumulative reward value is used as the target signal for reinforcement learning, and the parameters of the large traffic model are updated using the policy gradient algorithm.

[0015] Thirdly, an electronic device is provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements either the first aspect or the second aspect of the method.

[0016] This invention provides a method for training a large-scale traffic model, a method for intelligent traffic scheduling, and an apparatus. It constructs a traffic knowledge graph using road networks, equipment, traffic regulations, environmental states, and traffic events as nodes, and causal / spatiotemporal relationships between nodes as edges. The traffic knowledge graph is then transformed into differentiable graph embedding vectors. At each edge node, the graph embedding vectors are used to enhance the features of training samples, resulting in enhanced samples with graph logic constraints. Based on the entities in the enhanced samples, the actual evolution paths of traffic events corresponding to the enhanced samples are retrieved in the traffic knowledge graph. Based on the enhanced samples, the large-scale traffic model is locally trained, and the local parameters of the large-scale traffic model are updated based on the deviation loss between the traffic event evolution paths generated by the large model and the actual evolution paths. This invention transforms traffic domain knowledge into the model's intrinsic capabilities, significantly suppressing the illusion problem when large models handle complex traffic scenarios, and improving the accuracy and interpretability of the inference results.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a training method for a large traffic model provided in an embodiment of the present invention is shown; Figure 2 An example diagram of the traffic knowledge graph provided in an embodiment of the present invention is shown; Figure 3 A flowchart illustrating a traffic intelligent scheduling method provided in an embodiment of the present invention is shown; Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] This invention provides a method for training a large traffic model, such as... Figure 1 As shown, the method includes the following steps: Step S101: Construct a traffic knowledge graph using road network, equipment, traffic regulations, environmental status, and traffic events as nodes, and causal / spatiotemporal relationships between nodes as edges; and transform the traffic knowledge graph into a differentiable graph embedding vector.

[0022] In one example, the road network node is a road segment A, the device node is camera 01 (which may also include radar sensors, road condition sensors, mobile sensing devices, etc.), the traffic regulation node is the speed limit of 60km / h, the environmental status node is rainstorm weather, and the traffic event node is a rear-end collision event. Figure 2 The image shows an example of a traffic knowledge graph built based on each node.

[0023] In Figure 2 In the diagram, road segment A (road network node) and camera 01 (device node) are connected through the relationship of "deployed at" or "covered", indicating that the device is the physical entry point for monitoring data of this road segment.

[0024] The rear-end collision event (traffic event node) is marked as occurring on road segment A, reflecting the spatial location constraint of the event.

[0025] Heavy rain (environmental state node) serves as the initial trigger, pointing to an implicit intermediate state (such as "slippery road surface") through the "cause" relationship, and then to a rear-end collision event (traffic event node) through the "initiate" relationship. This constitutes a complete causal path from environmental anomaly to the occurrence of an accident.

[0026] The speed limit of 60 km / h (traffic regulation node) is applied to road segment A through a "constraint" relationship, defining the safe driving threshold for that segment. When the actual vehicle speed (implied in the data) exceeds this threshold and is accompanied by heavy rain, the probability of risk increases significantly, thereby accelerating the evolution into a rear-end collision.

[0027] This graph is not simply a list of nodes, but rather a dynamic reasoning network. It clearly demonstrates how traffic events (rear-end collisions) are triggered at specific spatiotemporal nodes (road segment A). This structured representation forms the basis for subsequently transforming the graph into differentiable graph embedding vectors, enabling the model to learn these complex nonlinear causal relationships, thereby achieving prediction and attribution analysis of long-tailed accidents.

[0028] It should be understood that graph embedding can be implemented using various algorithms such as TransE, RotatE, or graph neural networks, as long as it can achieve a faithful mapping from symbolic knowledge to a continuous vector space.

[0029] In one feasible implementation, converting a traffic knowledge graph into graph embedding vectors using a graph neural network includes the following steps: Step S101A: Node feature initialization and encoding.

[0030] Semantic encoding is performed on each node (road network, equipment, traffic regulations, environmental status, and traffic events) in the traffic knowledge graph.

[0031] Specifically, pre-trained natural language processing models (such as BERT) or one-hot encoding are used to transform the text descriptions of nodes (such as "Road Segment A", "Speed ​​Limit 60km / h", "Heavy Rain") into initial dense vector representations; for numerical attribute nodes (such as device ID, timestamp), normalized numerical vector representations are used to form the initial feature matrix of the nodes. .

[0032] Step S101B: Construct the adjacency matrix and edge weight mapping.

[0033] Construct the adjacency matrix of the graph based on the causal / spatiotemporal relationships defined in the graph. Different weights or learnable parameters can be assigned to different types of edges: For spatiotemporal relationships (such as "deployed at" or "located in"), deterministic weights are calculated based on spatial distance or time interval; For causal relationships (such as "cause" or "trigger"), learnable attention weight parameters are introduced to reflect the differences in importance of different causal paths to downstream inference tasks, thus obtaining a weighted adjacency matrix. .

[0034] Step S101C: Perform multi-layer graph neural network aggregation propagation.

[0035] Utilizing graph neural network architectures (such as GraphConvolutionalNetwork or GraphAttentionNetwork), in Feature aggregation and updating are performed in the layered network.

[0036] In the In the layer, the feature representation of each node is updated by aggregating the features of its neighboring nodes and the attribute information of the edges: (1); in, For activation function, For the first The learnable weight matrix of the layer; Indicates the first Feature matrix of layer nodes; Represents the weighted adjacency matrix The degree matrix.

[0037] This process enables each node to capture its local topology information and higher-order semantic relationships (for example, through multi-hop propagation, the "rear-end collision" node can indirectly perceive the impact of "heavy rain" and "speed limit of 60km / h").

[0038] Step S101D: Generate differentiable graph embedding vectors.

[0039] Will pass The final node feature representation obtained after layer propagation This serves as the embedding vector for the differentiable graph of the traffic knowledge graph.

[0040] This vector preserves the structured logical constraints (causal / spatiotemporal relationships) and semantic information of the graph, and the entire transformation process (from node input to output vector) is end-to-end differentiable, supporting the direct optimization of graph construction parameters or loss functions for subsequent tasks through the backpropagation algorithm.

[0041] Step S102: At each edge node, the training samples are enhanced using graph embedding vectors to obtain enhanced samples with graph logic constraints. Based on the entities in the enhanced samples, the actual evolution paths of traffic events corresponding to the enhanced samples are retrieved in the traffic knowledge graph.

[0042] In this embodiment, the edge node can be a computing device deployed in a roadside equipment room, a regional control center, or a mobile sensing vehicle.

[0043] Specifically, edge nodes use graph embedding vectors to enhance the features of locally acquired raw training samples (such as sensor time-series data, video structured descriptions, etc.).

[0044] Feature enhancement injects causal logic from the knowledge graph into the input vector explicitly. For example, it merges the original observation of "slippery road surface" with the association vector of "reduced friction coefficient" in the knowledge graph, forming enhanced samples with richer semantics and more complete logic. This ensures that the model converges based on the physical laws of the traffic domain rather than statistical coincidence, thereby effectively suppressing illusions and improving generalization ability.

[0045] At the same time, based on the entities identified in the enhanced samples, retrieval or reasoning is performed in the traffic knowledge graph to obtain the real evolution path corresponding to the traffic event.

[0046] This real evolutionary path represents a standard evolutionary process that conforms to traffic regulations and common sense in physics. It will serve as a supervisory signal for subsequent model training, used to correct any erroneous associations that the model may generate.

[0047] Step S103: Based on the augmented samples, perform local training on the large traffic model, and update the local parameters of the large traffic model based on the deviation loss between the traffic event evolution path generated by the large model and the actual evolution path.

[0048] Specifically, after receiving augmented samples, the model attempts to generate an evolutionary path prediction for the traffic event. It then calculates the distance metric, or bias loss, between the model-generated predicted path vector sequence and the retrieved actual evolutionary path vector sequence.

[0049] In one example, the mean squared error can be used to calculate this deviation loss: (2); in, This represents the total number of time steps. Indicates the current time step The predicted path vector sequence; Indicates the current time step The sequence of actual evolutionary path vectors.

[0050] Then the loss is calculated using the backpropagation algorithm. Relative to model parameters gradient ( The calculated gradient is used to update the model parameters. .

[0051] This bias loss reflects the degree of deviation between the model's current reasoning logic and the objective laws of the traffic field. For example, if the model predicts "rainy weather" and then directly follows "normal driving," while the actual path requires "slowing down," the distance between the two in the vector space will be large, resulting in a high loss value.

[0052] By minimizing this bias loss, the backpropagation algorithm can precisely adjust the model parameters, forcing the model's internal representation to gradually align with the logical manifold defined by the traffic knowledge graph.

[0053] This training method effectively transforms traffic domain knowledge into the model's intrinsic capabilities, significantly suppressing the illusion problem of large models when dealing with complex traffic scenarios, and improving the accuracy and interpretability of inference results.

[0054] Based on the above embodiments, feature enhancement of training samples is performed using graph embedding vectors to obtain enhanced samples with graph logic constraints, including: Step S102A1: Based on the original data of the training samples, identify and extract key entities, and map the key entities to the corresponding nodes in the traffic knowledge graph.

[0055] Key entities include at least one of the following: road network, equipment, traffic regulations, environmental conditions, and traffic incidents.

[0056] In this embodiment, the training samples are structured vectors obtained by preprocessing the collected raw data. For example, video data and image data can be converted into visual feature vectors and image feature vectors by an encoder such as ResNet, and text data can be converted into text feature vectors by a language model such as BERT.

[0057] Raw data includes, for example, video data (such as surveillance video clips), image data (such as traffic scene snapshots), and text data (such as traffic flow, traffic trajectory, and environmental conditions).

[0058] In one feasible implementation, the BERT model (an existing model) is used as an extractor to extract keywords or phrases with clear traffic semantics from these raw data. For example, words such as "rainy day," "sudden braking," and "rear-end collision" can be extracted from an accident description.

[0059] Subsequently, by calculating the semantic similarity between the extracted entity and each node in the graph, the node with the highest matching degree is selected to complete the mapping.

[0060] Step S102A2: Utilize the relationships between nodes in the graph to retrieve auxiliary feature vectors associated with key entities.

[0061] Specifically, taking each mapped node as the center, based on the adjacency matrix of the graph or a pre-computed graph embedding representation, neighboring nodes or path nodes with strong correlations are retrieved. The graph embedding vectors corresponding to these retrieved nodes are the auxiliary feature vectors.

[0062] Unlike existing technologies that rely solely on statistical co-occurrence relationships to obtain context vectors, the auxiliary feature vectors in this embodiment are strictly constrained by the physical laws and regulatory logic of the graph definition. For example, if the central node is "rainy day", the retrieved auxiliary vectors will not only include weather-related features such as "low visibility", but will also retrieve feature vectors that conform to vehicle dynamics principles, such as "reduced road friction coefficient" and "extended braking distance", along the causal edges.

[0063] This retrieval mechanism ensures that the introduced external knowledge is interpretable domain prior, rather than a false relevance driven by data, thus effectively preventing the model from learning erroneous patterns that violate common sense.

[0064] Step S102A3: Fuse the auxiliary feature vector with the original feature vector of the training sample to obtain the enhanced sample with graph logic constraints.

[0065] Finally, the retrieved auxiliary feature vectors are mathematically fused with the original feature vectors of the training samples. The fusion can be achieved through weighted summation, concatenation followed by a nonlinear transformation, or dynamic aggregation based on an attention mechanism.

[0066] Regardless of the specific algorithm used, the core purpose is to use the graph logic information contained in the auxiliary feature vector to correct or complete the position of the original feature vector in the semantic space.

[0067] To illustrate this process more clearly, we will use a specific rainy day accident scenario as an example below.

[0068] In this scenario, assume that the original training samples describe the situation of "the car in front braking while driving in the rain".

[0069] After initial encoding, its original feature vector for Each dimension represents abstract semantics such as environment, speed, operation, and risk.

[0070] At this point, although the vector contains information about "rain" and "brake", it does not explicitly express the dangerous consequences resulting from the combination of the two, and the activation value of the risk dimension is only 0.2.

[0071] During step S102A2, auxiliary feature vectors such as "reduced friction coefficient" and "increased braking distance" that are strongly correlated with "rainy weather" and "braking" were retrieved based on the map. .

[0072] For example, using attention mechanisms (3) The auxiliary vector The high-risk semantics inherent in the vector are injected into the original vector. Represents attention weights; the fused enhanced sample vector Become .

[0073] It can be observed that the values ​​of dimensions representing physical properties (such as friction and braking) and risk level increased significantly (from 0.4 to 0.7 and 0.8).

[0074] This indicates that the semantic representation of the original samples has been successfully guided by the graph logic towards a direction consistent with physical reality. Through this explicit vector fusion, the model no longer receives vague sensory data during the training phase, but rather structured knowledge embedded with causal chains. This significantly reduces the probability of the model generating illusions such as "it can stop normally even in the rain," thereby improving the credibility and security of the inference results.

[0075] Based on the above embodiments, the retrieval of the actual evolution path of traffic events corresponding to the enhanced samples in the traffic knowledge graph for each entity based on the enhanced samples includes: Step S102B1: Use the road network nodes mapped from the enhanced samples as the starting point for traversal.

[0076] Step S102B2: Utilize graph attention mechanism to perform multi-hop search along the associated edges of the traffic knowledge graph to select the successor node sequence that matches the preset regulatory constraints and event evolution logic; and construct the real evolution path based on the road network nodes and the successor node sequence.

[0077] In this process, the graph attention mechanism acts as a dynamic router, which calculates the transition probability of successor nodes not only based on the static connection strength between nodes, but also in combination with the contextual semantics of the current node.

[0078] For example, when the current node is "rainy day", although there may be multiple related nodes such as "car wash" and "rainbow" in the graph, the attention mechanism will give higher attention weight to "slippery road surface" or "reduced friction coefficient" based on the potential context of "traffic accident", thereby guiding the search direction to extend along a path that conforms to the physical causal law.

[0079] More importantly, this embodiment introduces a dual verification mechanism during the search process, namely, the joint filtering of preset regulatory constraints and event evolution logic.

[0080] In practice, traffic regulations (such as "no red light" and "no lane changing over solid lines") are encoded as hard masks or high-penalty items on the graph. In each step of the multi-hop search, if a candidate successor node violates the current regulatory constraints (e.g., transitioning from a "running a red light" node to a "normal passage" node), the transition probability of that edge will be forcibly set to zero or subject to a very large negative penalty, so that the illegal path is automatically pruned during the search process.

[0081] This design ensures that the generated real evolutionary paths are not only statistically frequently co-occurring, but also legally and physically compliant.

[0082] Finally, the actual evolution path is constructed based on the road network nodes and the selected successor node sequence. This path can be represented as an ordered sequence of node IDs or converted into a corresponding graph embedding vector sequence, so that it can be directly used in subsequent steps to measure the distance between the path and the model prediction path in the vector space.

[0083] This double-verified path serves as a supervisory signal, accurately correcting potential cognitive biases in large models and forcing them to learn correct reasoning patterns that conform to objective laws and legal norms during training. This fundamentally improves the safety and reliability of the model in complex traffic scenarios.

[0084] Based on the above embodiments, the method further includes: Step S104: Upload the local parameters of each edge node to the cloud for global aggregation and optimization.

[0085] In this embodiment, these local parameters can be full copies of the model weights, or gradient increments or difference parameters relative to the previous round of the global model.

[0086] This step, by uploading only model parameters, physically cuts off the risk path of sensitive traffic data leakage, ensuring strict protection of data privacy for each region or operating entity, while also significantly reducing bandwidth pressure on wide area network transmission.

[0087] It should be understood that the upload operation is not continuous in real time, but is controlled by preset triggering conditions. For example, the system will only trigger a parameter upload request when the local training of the edge node reaches a predetermined number of iterations, the cumulative number of processed samples exceeds a threshold, or the performance improvement of the local model on the validation set meets specific indicators.

[0088] In addition, before uploading, edge nodes will perform homomorphic encryption or security masking on the parameters to prevent man-in-the-middle attacks or parameter reverse engineering attacks during transmission, further strengthening the security boundary of the system.

[0089] After receiving local parameters from multiple edge nodes, the cloud will execute a global aggregation optimization algorithm.

[0090] Specifically, the cloud server acts as an aggregator and coordinator, employing a weighted average strategy to merge the parameters of each node.

[0091] Taking the federated averaging algorithm as an example, the update of the global model parameters follows this logic: the new global model parameters are equal to the weighted sum of the local model parameters of each edge node, where the weight of each node is proportional to the number of samples used for its local training. This means that edge nodes with more high-quality training data have a greater say in the global model update, thus ensuring that the global model can accurately reflect the statistical patterns of mainstream traffic scenarios.

[0092] Meanwhile, to prevent individual nodes with extremely unique data distributions (such as nodes that only cover a single tunnel scenario) from causing negative drift to the global model, anomaly detection and pruning mechanisms can be introduced into the aggregation algorithm to automatically remove parameter updates with excessive deviation.

[0093] Through this precise mathematical aggregation, the global model generated in the cloud not only absorbs the refined features learned by each edge node in specific scenarios, but also integrates diverse knowledge across regions, thus possessing stronger generalization ability and robustness than any single node.

[0094] Step S105: Iteratively perform local training and global optimization until convergence, to obtain a large traffic model with logical reasoning capabilities in the traffic domain.

[0095] When the global loss decreases by less than a preset small threshold (e.g., 0.001) over several consecutive rounds, or when the model's accuracy on the critical traffic event identification task reaches a predetermined target value, the model is considered to have converged.

[0096] In addition, to prevent training from getting stuck in an infinite loop, a maximum number of iterations is usually set as a fallback condition. Once the convergence condition is met, the cloud locks the current global model parameters and releases it as the final traffic model.

[0097] To illustrate the advantages of the above-mentioned collaborative mechanism more intuitively, the following example demonstrates the advantages using a cross-regional large-scale traffic model training scenario.

[0098] Suppose a city's traffic management system connects to three different types of edge computing nodes: the city center, highways, and remote suburbs. The city center nodes have a large amount of data but complex congestion patterns; the highway nodes have fast data flow and obvious accident characteristics; and the suburban nodes have sparse data but contain a large amount of interference from atypical agricultural vehicles.

[0099] In the traditional centralized training model, if all raw data is gathered in the cloud, it not only faces huge risks of privacy leakage and network transmission bottlenecks, but also the heterogeneity of data can easily lead to poor model performance in certain scenarios.

[0100] In the collaborative architecture of this embodiment, each of the three nodes uses its local graph embedding vector to enhance the samples and complete local training, and only uploads the refined model gradient.

[0101] Through weighted aggregation, the cloud-based model retains the detailed understanding of complex congestion in the central urban area, integrates the accident response logic of highway scenarios, and also takes into account the special interference suppression capabilities of suburban areas. After multiple rounds of iteration of "local augmentation training - cloud aggregation optimization," the final global model not only outperforms the individually trained local model in various benchmark tests, but also demonstrates logical reasoning capabilities that conform to traffic regulations and common sense when faced with unprecedented complex traffic events, truly achieving a leap from data silos to collective intelligence.

[0102] Based on the same inventive concept, a traffic intelligent scheduling method is provided, such as... Figure 3 As shown, the method includes the following steps: Step S301: Acquire multi-source sensor data within the target area and preprocess the multi-source sensor data to obtain structured event vectors; then use the graph embedding vectors of the traffic knowledge graph to enhance the features of the structured event vectors.

[0103] Specifically, the target area can be a single intersection, a section of highway, a jurisdiction, or even the entire urban road network, and its boundaries can be dynamically defined according to actual control needs.

[0104] Multi-source sensing data encompasses data collected by fixed sensing devices (such as roadside cameras, millimeter-wave radar, lidar, and weather stations) as well as real-time streaming data transmitted back by mobile sensing units (such as connected vehicles, drones, and law enforcement recorders).

[0105] Because these raw data differ significantly in modality, frequency, and semantic granularity, directly inputting them into the model would lead to inference confusion. Therefore, preprocessing is necessary to generate structured event vectors.

[0106] This structured event vector is a standardized semantic representation format that includes not only the spatiotemporal coordinates of the event, event type identifiers (such as "traffic accident", "abnormal parking", "slippery road surface") and confidence scores, but also environmental context information.

[0107] This vectorized representation eliminates the semantic gap between heterogeneous data, providing a unified, standardized, and high-quality input interface for subsequent large-scale model inference.

[0108] The specific preprocessing mechanism will be further detailed in subsequent embodiments.

[0109] Step S302: Based on the trained traffic model, reason about the enhanced structured event vectors to output the target event, event evolution path and evolution result under the constraints of the traffic knowledge graph.

[0110] In this embodiment, the traffic model used is the model trained by the method described in the above embodiment, whose internal parameters have incorporated traffic regulations, road network topology, and event evolution logic.

[0111] Among them, the target event is the accurate identification of the current situation; the event evolution path is the trajectory of event development deduced by the model based on the embedded knowledge graph logic, such as the complete causal chain from "rainfall" to "reduction of road friction coefficient", and then to "extension of braking distance" and "increased risk of rear-end collision"; the evolution result is the probabilistic prediction of the final state (such as "severe congestion" or "secondary accident") that may occur within a certain time window in the future.

[0112] The evolution path output in this embodiment has undergone explicit constraints of graph logic. Each step of the deduction conforms to the physical laws and traffic regulations in the field of transportation, thus giving the reasoning results strong interpretability and credibility.

[0113] Step S303: Based on the target event, the event evolution path and evolution result of the target event, generate dispatch instructions for each traffic control terminal within the target area and send them to each traffic control terminal.

[0114] This embodiment intervenes in advance to address future risk points predicted along the evolutionary path.

[0115] For example, when the evolution path shows "current minor scratch → expected to cause back queue overflow in 5 minutes → leading to main road congestion", the generated dispatch instructions will not only include rescue guidance at the accident site, but also preventive traffic control of traffic lights at upstream intersections, remote diversion guidance information for variable message signs, and dynamic opening instructions for emergency lanes.

[0116] The term "traffic control terminal" here is a broad concept, including but not limited to traffic signal controllers, variable message signs, lane indicators, ramp controllers, broadcasting systems, and mobile push services.

[0117] This embodiment achieves an end-to-end closed loop from cognitive reasoning to physical control by mapping abstract evolution paths to specific, executable device control parameters. This ensures the scientific rigor and foresight of the scheduling strategy, effectively preventing secondary disasters or congestion spread caused by reaction delays, and truly leveraging the intelligent decision-making value of the large-scale traffic model in complex scenarios.

[0118] Based on the above embodiments, before generating scheduling instructions, the method further includes: Step S304: Extract real-time dynamic features related to the target event based on the spatiotemporal coordinates, event type identifiers, and confidence scores of the structured event vector.

[0119] Real-time dynamic characteristics include the probability of event occurrence, the radius of influence, and the expected duration.

[0120] Among them, the probability of an event is a Bayesian estimate that combines the confidence score and the historical prior probability. For example, when the confidence score is high but the historical occurrence rate of this type of event is extremely low during this period, the system will appropriately adjust the probability value to suppress false alarms. The radius of influence is the equivalent spillover distance dynamically calculated based on the specific road segment located by spatiotemporal coordinates, combined with the current road network topology, real-time traffic density, and road grade, rather than a fixed geometric radius; The expected duration is obtained by analyzing the distribution of historical response records corresponding to the event type identifier and performing regression prediction in combination with the current availability of emergency rescue resources.

[0121] This precise mapping from semantic space to physical parameter space enables the transformation of abstract event descriptions into measurable risk factors, laying a data foundation for subsequent quantitative assessments.

[0122] Step S305: Based on the probability of the event, the radius of influence, and the expected duration, calculate the risk level of the target event using a preset risk assessment model.

[0123] In this embodiment, the risk assessment model adopts a weighted multidimensional evaluation algorithm, and its core calculation formula can be expressed as: (4); in, Assess risk level; , , These represent the normalized probability of the event, the radius of the affected area, and the expected duration, respectively. , , These are dynamic weighting coefficients.

[0124] It is important to note that these weighting coefficients are not fixed empirical values, but rather adaptively adjusted based on the current traffic control objectives and environmental context. For example, during morning and evening rush hours, the weight of the radius of influence is adjusted because the external costs of congestion are extremely high. The weighting will be automatically increased; however, at night or in severe weather conditions, safety becomes the primary consideration, with the probability of an event and its potential consequences taking precedence. Then it will take the lead.

[0125] Through this dynamic weighting mechanism, the risk assessment model can accurately reflect the true risk level of traffic incidents under different spatiotemporal backgrounds, avoiding the adaptability bias of a single threshold judgment in different scenarios. The calculated risk score... They will be mapped to a preset risk level range, such as being divided into four levels: low risk, medium risk, high risk, and extremely high risk, with each level corresponding to a clear numerical boundary.

[0126] Step S306: Determine the handling priority of the target event based on the risk level, and use the handling priority as a constraint condition for generating subsequent scheduling instructions.

[0127] Specifically, risk levels are linked to the allocation of scheduling resources and response time requirements.

[0128] For example, for events deemed extremely high-risk, their handling priority is set to the highest level (P0). The system will forcibly lock the optimal emergency resource channel and increase the weight of the "safe passage probability" objective in the subsequent multi-objective optimization function, while allowing for the breaking of conventional energy consumption or efficiency limits. For low-risk events, their handling priority is lower (P3), and they will only respond when high-priority tasks are met and resources are redundant. Furthermore, the optimization will focus more on the balance between passage efficiency and energy consumption.

[0129] This design, which transforms subjective priorities into objective optimization constraints, ensures that dispatch decisions are both rationally holistic and capable of responding quickly to emergencies in accordance with human intuition.

[0130] Based on the above embodiments, the dispatch instructions generated for each traffic control terminal within the target area, based on the target event, the event evolution path of the target event, and the evolution result, include: Step S303A: Analyze the key node sequence in the traffic event evolution path and extract the impact characteristics of the evolution path on traffic flow, road network capacity and potential secondary risks at different time steps.

[0131] In this embodiment, semantic parsing is first performed on key nodes on the path to identify variable nodes that characterize the traffic flow operation status, such as "queue length exceeds the threshold", "road segment saturation reaches 0.9" or "average vehicle speed drops below 20km / h".

[0132] Subsequently, these state variables are mapped onto a future timeline, forming a series of impact feature vectors with timestamps.

[0133] It should be understood that the granularity of the time step can be flexibly set according to actual control needs. For example, it can be set to 1 minute in the expressway scenario and 5 minutes in the urban arterial road scenario, in order to adapt to the rate of change of traffic flow in different scenarios.

[0134] Step S303B: Combining the handling priority of the target event, construct a multi-objective optimization function with the optimization objectives of minimizing congestion duration, maximizing safe passage probability, and minimizing resource consumption, and plan the optimal intervention strategy sequence based on the impact characteristics.

[0135] After obtaining the dynamic impact characteristics of future time periods (such as congestion duration, safe passage probability, and scheduling resource consumption), these are used as input parameters and substituted into the preset multi-objective optimization model.

[0136] The core of this optimization function lies in balancing the complex game-like relationship between efficiency, safety, and cost. Specifically, the objective function J can be expressed as: (5); in, For the predicted duration of congestion, To ensure safe passage probability, This is to manage resource consumption (such as the number of police officers deployed, the number of times equipment is started and stopped). , , The weighting coefficients for each item are dynamically adjusted based on the priority of disposal.

[0137] It should be noted that, , , All values ​​are normalized.

[0138] The solution process can utilize genetic algorithms, particle swarm optimization, or reinforcement learning policy networks to find the global optimum among possible combinations of interventions. These are all techniques well-known to those skilled in the art and will not be elaborated upon here.

[0139] This optimal intervention strategy sequence describes the macro-control actions that should be taken at various future time points, such as "t+0 to t+10 minutes: upstream intersection traffic interception; t+10 to t+20 minutes: remote diversion and guidance; after t+20 minutes: restoration of normal timing".

[0140] Through this mathematical multi-objective optimization process, the present invention transforms abstract scheduling experience into a computable and verifiable optimal solution, effectively overcoming the subjectivity and limitations of human decision-making in multi-objective trade-offs.

[0141] Step S303C: Map the optimal intervention strategy sequence to an executable instruction set for each traffic control terminal within the target area.

[0142] The executable instruction set includes at least signal timing adjustment schemes, lane control instructions, variable message sign guidance information, and emergency vehicle guidance routes.

[0143] Since the optimized output strategy sequence is usually a macro-level control intent, it must undergo protocol conversion before it can be executed by physical devices. This embodiment pre-defines a strategy-instruction mapping knowledge base, which defines the correspondence between various macro-level strategies and underlying device control parameters.

[0144] For example, when the strategy sequence includes the action of "intercepting traffic at the upstream intersection", the current timing scheme of the traffic signal at that intersection will be automatically retrieved, and a specific traffic signal timing adjustment instruction of "increasing the cycle length by 15 seconds and reducing the green light ratio in the east-west direction by 10%" will be generated according to the preset interception intensity level. When the strategy includes "remote diversion guidance", it will filter out the variable message signs of key upstream nodes based on the geographical location of the affected road section, and generate text guidance information such as "Accident ahead, please detour via [road name]" and the corresponding release time. When the strategy involves emergency rescue, the system will also plan an emergency vehicle guidance route that avoids congestion nodes based on real-time traffic conditions and road network topology, and send it to the ambulance or fire truck terminal through the vehicle-road cooperative interface.

[0145] Furthermore, lane control commands can also cover operations such as tidal lane switching, temporary opening of bus lanes, or closure of ramps. Through this standardized mapping mechanism, a seamless connection is achieved from logical decision-making at the cognitive reasoning layer to equipment control at the physical execution layer, ensuring the accurate implementation and automated execution of dispatch commands.

[0146] Based on the above embodiments, acquiring multi-source sensor data within the target area and preprocessing the multi-source sensor data to obtain a structured event vector includes: Step S301A: Acquire fixed sensing data and mobile sensing data within the target area.

[0147] In this embodiment, these data can be collected using a dynamic sampling frequency, such as 10-20 Hz.

[0148] Specifically, fixed sensing data mainly comes from equipment such as visual cameras, millimeter-wave radar, lidar and weather sensors deployed on the roadside. Its characteristics are fixed location and stable sampling frequency, but it is easily affected by environmental obstruction or bad weather. Mobile sensing data comes from mobile terminals such as connected vehicles, drones and law enforcement recorders. Its characteristics are flexible perspective and dynamic blind spot filling, but it has problems such as positioning drift, data sparsity and communication delay.

[0149] This embodiment uses a high-speed communication network to aggregate these two types of data in real time, forming a holographic perception data pool covering the target area.

[0150] Step S301B: Perform semantic alignment and fusion on the fixed sensing data and the mobile sensing data to obtain the initial structured event vector.

[0151] Because the original data have huge differences in modality, coordinate system and semantic granularity, direct splicing cannot be effectively utilized by the model.

[0152] This embodiment employs a semantic alignment mechanism based on a collaborative adaptation algorithm.

[0153] Specifically, the basic weight and real-time reliability of each sensing node are first evaluated. The basic weight is determined by the device type; for example, the visual recognition capability of a fixed camera is usually given a higher weight, while mobile terminals are given a lower weight due to fluctuations in accuracy.

[0154] Real-time reliability is dynamically calculated based on the current environment. For example, under backlight or heavy rain conditions, the reliability score of a camera will drop significantly, while radar remains highly reliable because it is not affected by light.

[0155] Subsequently, the two are combined using a preset adaptation coefficient to calculate the comprehensive weight of each sensing point at the current moment.

[0156] Node data with high overall weight will be prioritized as semantic benchmarks, while node data with low weight will be used as auxiliary verification or suppressed.

[0157] Based on this, the filtered high-reliability data is mapped to a unified semantic space, and lightweight semantic labels such as "abnormal deceleration", "vehicle stagnation" and "low visibility" are added, and precise spatiotemporal stamps are added, thereby generating an initial structured event vector containing preliminary semantic information.

[0158] This dynamic weight-based fusion strategy enables the system to maintain the stability of input data quality even when sensor performance changes with the environment, effectively avoiding perception collapse caused by the failure of a single sensor.

[0159] Step S301C: Perform a consistency check on several initial structured event vectors within the same spatiotemporal window to obtain the final structured event vector.

[0160] In real-world scenarios, multiple sensors may simultaneously observe the same event, but due to factors such as perspective, error, or noise, the generated initial vectors may exhibit surface conflicts.

[0161] Therefore, this embodiment sets strict spatiotemporal window constraints (e.g., time deviation less than 500 milliseconds, spatial deviation less than 10 meters), and only performs correlation analysis on vectors falling within the same window.

[0162] Within the window, a pre-defined semantic anchor template library is invoked for pattern matching. The semantic anchor template library stores multimodal feature combinations of various typical traffic events. For example, the "congestion event" template requires that three conditions be met simultaneously: "visual traffic stagnation," "radar low-speed density," and "trajectory clustering."

[0163] When the matching degree between the input initial vector group and a certain template exceeds a preset threshold, the system determines that the data consistency is high and generates a high-confidence final structured event vector accordingly. If the matching degree is low or there is a logical contradiction (such as visual display being smooth but radar display being high-density), the anomaly handling mechanism is triggered to resolve the conflict through voting or weighted averaging, or directly mark it as a low-confidence sample for subsequent manual review.

[0164] The final output structured event vector is a standardized data packet that includes event type identifier, confidence score, semantic label set, spatiotemporal center coordinates, and compressed feature embedding vector.

[0165] These vectors, which have undergone rigorous consistency testing, completely eliminate noise and ambiguity in the original data, providing a clean, reliable, and semantically context-rich input for the large-scale traffic model, fundamentally improving the robustness and accuracy of the model's inference.

[0166] Based on the above embodiments, the method further includes: Step S307: Obtain environmental feedback data from each traffic control terminal when executing the current dispatch instruction.

[0167] Environmental feedback data focuses on post-event assessment and long-term statistics. The data collected mainly includes key performance indicators such as actual travel time, accident rate, and equipment energy consumption.

[0168] The actual travel time can be calculated from the time difference of passing vehicles using roadside geomagnetic detectors, video checkpoints, or floating car GPS trajectories, and is used to quantify the real impact of dispatch instructions on traffic flow efficiency; the accident rate not only includes records of collision accidents that have occurred, but should also cover the frequency of traffic conflict events such as sudden braking and abnormal lane changes identified through video analysis, serving as a proxy indicator of safety risk; and equipment energy consumption is directly collected from the smart meter readings of terminals such as traffic lights and guidance screens, reflecting the resource consumption cost of the dispatch strategy.

[0169] It should be understood that the collection of these feedback data is lagging and cumulative. They are usually aggregated and statistically analyzed in fixed time windows (such as 5 minutes or 15 minutes) to eliminate instantaneous noise interference and ensure the stability and representativeness of the feedback signal.

[0170] Step S308: Calculate the cumulative reward value of the current strategy based on environmental feedback data and a pre-built composite reward function; the composite reward function includes a traffic efficiency gain term, a risk reduction term, an energy consumption control term, and a deduction error term.

[0171] In this embodiment, a multi-dimensional composite reward function R is constructed, and its mathematical expression is as follows: (6); in, The total reward value, To improve traffic efficiency. The percentage decrease in risk level As a percentage reduction in energy consumption, For dynamic extrapolation error, , , , The reward weight is dynamically adjusted based on the control objectives.

[0172] It should be noted that, , , , All values ​​are normalized.

[0173] Specifically, , , These three parameters are all normalized percentages calculated using the logic of (baseline value - current value) / baseline value. The baseline value is the value from the previous time point. For example, the total energy consumption of the previous cycle can be used as the baseline.

[0174] Specifically, dynamic inference error The calculation is as follows: (7); in, For the inference error (target) ), The number of times the large model is run; To provide feedback on the actual evolution path of events, This is the derivation path for the large model. For the first The weights of each simulation are dynamically adjusted according to the simulation accuracy requirements.

[0175] After calculating the immediate reward for a single step or a single window, the reward values ​​from multiple windows need to be summed to obtain the cumulative reward value.

[0176] This process typically employs a cumulative calculation method with a discount factor, meaning that the cumulative reward at the current moment depends not only on the immediate feedback but also on the discounted sum of expected future rewards.

[0177] The underlying mechanism of this design lies in the fact that traffic scheduling is a typical sequential decision-making process. A current instruction (such as upstream intersection closure) may cause local delays in the short term (negative immediate reward), but in the long run, it can prevent downstream bottleneck overflow and thus improve the efficiency of the entire network (positive long-term reward). By calculating cumulative rewards, the model can learn to tolerate short-term local sacrifices for the sake of long-term global optimum, effectively overcoming the limitations of short-sighted greedy strategies.

[0178] Step S309: Use the accumulated reward value as the target signal for reinforcement learning, and update the parameters of the traffic model using the policy gradient algorithm.

[0179] In practice, policy gradient algorithms such as PPO (Proximal Policy Optimization) or A3C (Asynchronous Advantage Actor-Critic) use the cumulative reward value as a scalar feedback signal to calculate the probability gradient of generating that reward value under the current policy.

[0180] If a certain sequence of scheduling instructions results in a high cumulative reward, the algorithm will adjust the model parameters in the direction of increasing the probability of generating that sequence of instructions; conversely, if it results in a low cumulative reward or even a penalty, the parameters will be adjusted in the direction of decreasing the probability.

[0181] These reinforcement learning algorithms are all existing technologies and will not be elaborated upon here.

[0182] Through thousands of such "execution-feedback-update" iterative cycles, the traffic big model gradually internalizes the optimal decision-making strategy in complex traffic environments.

[0183] Through this complete reinforcement learning loop, the present invention endows the traffic intelligent agent with true autonomous evolution capabilities, enabling it to continuously learn from the real environment after deployment, constantly adapt to changes in road network structure, traffic flow pattern migration and control target adjustment, and form long-term technological competitive advantage and life cycle value.

[0184] Based on the same technical concept, embodiments of the present invention also provide an electronic device, such as... Figure 4 As shown, it includes a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404.

[0185] Memory 403 is used to store computer programs; When the processor 401 executes the program stored in the memory 403, it implements the steps of the above method embodiment.

[0186] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0187] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0188] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0189] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0190] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0191] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for training a large-scale traffic model, characterized in that, include: A traffic knowledge graph is constructed using road networks, equipment, traffic regulations, environmental conditions, and traffic events as nodes, and causal / spatiotemporal relationships between nodes as edges. The traffic knowledge graph is then transformed into a differentiable graph embedding vector. At each edge node, the training samples are enhanced using the graph embedding vector to obtain enhanced samples with graph logic constraints. Based on the entities of the enhanced samples, the actual evolution path of the traffic event corresponding to the enhanced samples is retrieved in the traffic knowledge graph. Based on the enhanced samples, the traffic model is locally trained, and the local parameters of the traffic model are updated based on the deviation loss between the traffic event evolution path generated by the large model and the actual evolution path.

2. The method according to claim 1, characterized in that, The step of using the graph embedding vector to enhance the features of the training samples to obtain enhanced samples with graph logic constraints includes: Based on the raw data of the training samples, key entities are identified and extracted, and the key entities are mapped to corresponding nodes in the traffic knowledge graph; the key entities include at least one of road network, equipment, traffic regulations, environmental status and traffic events; By utilizing the relationships between nodes in the graph, auxiliary feature vectors associated with the key entity are retrieved; The auxiliary feature vector is fused with the original feature vector of the training sample to obtain an enhanced sample with graph logic constraints.

3. The method according to claim 2, characterized in that, Based on the enhanced samples, each entity retrieves the actual evolution path of the traffic event corresponding to the enhanced sample in the traffic knowledge graph, including: The road network nodes mapped from the enhanced samples are used as the starting point for traversal; The graph attention mechanism is used to perform multi-hop search along the associated edges of the traffic knowledge graph to select the successor node sequence that matches the preset regulatory constraints and event evolution logic; and the actual evolution path is constructed based on the road network nodes and the successor node sequence.

4. The method according to claim 1, characterized in that, The method further includes: Upload the local parameters of each edge node to the cloud for global aggregation and optimization; Iteratively perform local training and global optimization until convergence to obtain a large-scale traffic model with logical reasoning capabilities in the traffic domain.

5. A traffic intelligent scheduling method, characterized in that, include: Acquire multi-source sensor data within the target area, preprocess the multi-source sensor data to obtain structured event vectors, and use graph embedding vectors of traffic knowledge graph to enhance the features of the structured event vectors; Based on the traffic big model trained by the method described in any one of claims 1-4, reasoning is performed on the enhanced structured event vector to output the target event, event evolution path and evolution result under the constraints of the traffic knowledge graph; Based on the target event, its evolution path, and the result of the evolution, dispatch instructions for each traffic control terminal within the target area are generated and sent to each traffic control terminal.

6. The method according to claim 5, characterized in that, The structured event vector includes spatiotemporal coordinates, event type identifiers, and confidence scores; before generating scheduling instructions, the method further includes: Based on the spatiotemporal coordinates, event type identifiers, and confidence scores of the structured event vector, real-time dynamic features related to the target event are extracted. These real-time dynamic features include the probability of event occurrence, the radius of influence, and the expected duration. Based on the probability of the event occurring, the radius of its impact, and the expected duration, the risk level of the target event is calculated using a preset risk assessment model. The handling priority of the target event is determined based on the risk level, and the handling priority is used as a constraint condition for the generation of subsequent scheduling instructions.

7. The method according to claim 6, characterized in that, The generation of dispatch instructions for each traffic control terminal within the target area based on the target event, the event evolution path of the target event, and the evolution result includes: The key node sequence in the evolution path of the traffic event is analyzed, and the impact characteristics of the evolution path on traffic flow, road network capacity and potential secondary risks at different time steps are extracted. Based on the handling priority of the target events, a multi-objective optimization function is constructed with the optimization objectives of minimizing congestion duration, maximizing safe passage probability, and minimizing resource consumption. The optimal intervention strategy sequence is then planned for the impact characteristics. The optimal intervention strategy sequence is mapped to a set of executable instructions for each traffic control terminal within the target area. The set of executable instructions includes at least traffic light timing adjustment schemes, lane control instructions, variable message sign guidance information, and emergency vehicle guidance routes.

8. The method according to claim 5, characterized in that, The process of acquiring multi-source sensor data within the target area and preprocessing the multi-source sensor data to obtain a structured event vector includes: Acquire fixed sensing data and mobile sensing data within the target area; The fixed sensing data and mobile sensing data are semantically aligned and fused to obtain an initial structured event vector; A consistency check is performed on several initial structured event vectors within the same spatiotemporal window to obtain the final structured event vector.

9. The training method according to claim 5, characterized in that, The method further includes: Obtain environmental feedback data from each traffic control terminal when executing the current dispatch command; The cumulative reward value of the current strategy is calculated based on the environmental feedback data and the pre-constructed composite reward function; the composite reward function includes a traffic efficiency gain term, a safety risk reduction term, and an energy consumption control term. The accumulated reward value is used as the target signal for reinforcement learning, and the parameters of the traffic model are updated using the policy gradient algorithm.

10. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the method described in any one of claims 1-4 or 5-9.

Images