Traffic vertical domain model generation method and device and autonomous intelligent traffic system

Abstract

The application provides a traffic vertical domain model generation method and device and an autonomous intelligent traffic system. The generation method comprises the following steps: receiving traffic data, extracting a plurality of key features related to traffic control and a plurality of traffic conditions corresponding to each key feature from the traffic data, and a first causal graph between traffic objects corresponding to each traffic condition; creating a traffic vertical domain model, wherein the traffic vertical domain model comprises an MLM module, the MLM module identifies a second causal graph from the received text, then pairs the causes or results in the causal entities in the first and second causal graphs using a directional mask causal pair, learns the causal dependency relationship in all causal entities, and trains the traffic vertical domain model based on the traffic data. The traffic vertical domain model can be applied to an autonomous intelligent traffic system.

Description

Technical Field

[0001] This invention relates to the field of traffic control technology, and in particular to a method, apparatus, and autonomous intelligent transportation system for generating traffic vertical domain models. Background Technology

[0002] In today's rapidly globalizing and urbanizing world, the intelligent upgrading of transportation systems has become crucial for enhancing urban competitiveness. This requires addressing the complexities of traffic scenarios, the massive volume of traffic data, the heavy reliance on human intervention for traffic control and command, and the long-standing long-tail problem in the transportation industry.

[0003] Therefore, how to make autonomous intelligent transportation systems intelligent has become an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, and autonomous intelligent transportation system for generating traffic vertical domain models.

[0005] To achieve one of the aforementioned objectives, an embodiment of the present invention provides a method for generating a traffic vertical domain model for an autonomous intelligent transportation system, comprising the following steps: receiving traffic data, the traffic data including at least: traffic flow, speed, accident reports, and weather conditions; extracting several key features related to traffic control from the traffic data, as well as several traffic conditions corresponding to each key feature, and a first causal graph between traffic objects corresponding to each traffic condition; creating a traffic vertical domain model, the traffic vertical domain model including an MLM module, the MLM module identifying a second causal graph from the received text, and then using directional masking causal pairs to perform pairwise masking of causes or effects in causal entities in the first and second causal graphs, and learning the causal dependencies in all causal entities; and training the traffic vertical domain model based on the traffic data.

[0006] As a further improvement of one embodiment of the present invention, the receiving of traffic data specifically includes: receiving traffic data, then performing prediction and filling processing on missing values ​​in the traffic data, removing outliers in the traffic data, and then performing normalization processing on the traffic data.

[0007] As a further improvement of one embodiment of the present invention, the step of extracting several key features related to traffic control from the traffic data, and several traffic conditions corresponding to each key feature, and the traffic object corresponding to each traffic condition specifically includes: performing text segmentation on the traffic data to obtain several sentences and several words; using named entity recognition algorithm and dependency parsing to extract several key features related to traffic control; then, labeling the traffic data; and then using a constraint-based causal discovery algorithm to construct a first causal graph between several traffic events corresponding to each key feature and the traffic objects corresponding to each traffic event.

[0008] As a further improvement of one embodiment of the present invention, the MLM module adopts a multi-task learning architecture and a three-layer model structure of device state, spatiotemporal relationship triples, and causal entity. It also uses triple extraction to extract a second causal graph from the received text, and uses Span Cross-Entropy and relationship classification cross-entropy to predict position-sensitive spans. Furthermore, it uses a dynamic weight adjustment method based on Uncertainty Weighting to automatically balance task weights for multi-task loss fusion.

[0009] This invention also provides a device for generating a traffic vertical domain model for an autonomous intelligent transportation system, comprising the following modules: a data acquisition module for receiving traffic data, the traffic data including at least: traffic flow, speed, accident reports, and weather conditions; a data preprocessing module for extracting several key features related to traffic control from the traffic data, as well as several traffic conditions corresponding to each key feature, and a first causal graph between traffic objects corresponding to each traffic condition; a model creation module for creating a traffic vertical domain model, the traffic vertical domain model including an MLM module, the MLM module identifying a second causal graph from the received text, and then using directional masking causal pairs to perform pairwise masking of causes or effects in the causal entities in the first and second causal graphs, and learning the causal dependencies in all causal entities; and a training module for training the traffic vertical domain model based on the traffic data.

[0010] As a further improvement of one embodiment of the present invention, the data acquisition module is further configured to: receive traffic data, then perform prediction and filling processing on missing values ​​in the traffic data, remove outliers in the traffic data, and then perform normalization processing on the traffic data.

[0011] As a further improvement of one embodiment of the present invention, the data preprocessing module is further configured to: perform text segmentation on the traffic data to obtain several sentences and several words; extract several key features related to traffic control using named entity recognition algorithm and dependency parsing method; then, annotate the traffic data; and then, use a constraint-based causal discovery algorithm to construct a first causal graph between several traffic events corresponding to each key feature and traffic objects corresponding to each traffic event.

[0012] As a further improvement of one embodiment of the present invention, the MLM module adopts a multi-task learning architecture and a three-layer model structure of device state, spatiotemporal relationship triples, and causal entity. It also uses triple extraction to extract a second causal graph from the received text, and uses Span Cross-Entropy and relationship classification cross-entropy to predict position-sensitive spans. Furthermore, it uses a dynamic weight adjustment method based on Uncertainty Weighting to automatically balance task weights for multi-task loss fusion.

[0013] This invention also provides a traffic control system, comprising: a traffic vertical domain model, a data platform, and a twin map obtained by performing the above-described generation method; the traffic vertical domain model is capable of receiving real-time event data from the data platform, processing the received data, providing decision data, and then displaying it in the form of an HTML page; the twin map receives data from the HTML page.

[0014] As a further improvement of one embodiment of the present invention, the twin map receives data from the HTML page based on the WebTRC protocol.

[0015] Compared to existing technologies, the technical advantages of this invention are as follows: This invention provides a method, apparatus, and autonomous intelligent transportation system for generating a traffic vertical domain model. The generation method includes the following steps: receiving traffic data; extracting several key features related to traffic control from the traffic data, as well as several traffic conditions corresponding to each key feature, and a first causal graph between traffic objects corresponding to each traffic condition; creating a traffic vertical domain model, which includes an MLM module. The MLM module identifies a second causal graph from the received text, and then uses directional masking causal pairs to perform pairwise masking of causes or effects in the causal entities of the first and second causal graphs, and learns the causal dependencies in all causal entities; and training the traffic vertical domain model based on the traffic data. This traffic vertical domain model can be applied to an autonomous intelligent transportation system. Attached Figure Description

[0016] Figure 1This is a flowchart illustrating the method for generating a traffic vertical domain model in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the autonomous intelligent transportation system in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0019] The terms used herein, such as “above,” “over,” “below,” and “under,” indicating spatial relative position, are for illustrative purposes to describe the relationship of one unit or feature relative to another unit or feature as shown in the accompanying drawings. These terms may be intended to include different orientations of the device in use or operation other than those shown in the figures. For example, if the device in the figures is flipped, a unit described as being “below” or “under” another unit or feature would be “above” that unit or feature. Therefore, the exemplary term “below” can encompass both above and below orientations. The device may be oriented in other ways (rotated 90 degrees or otherwise), and the spatially related descriptive terms used herein will be interpreted accordingly.

[0020] Embodiment 1 of the present invention provides a method for generating a traffic vertical domain model for an autonomous intelligent transportation system, such as... Figure 1 As shown, it includes the following steps:

[0021] Step 101: Receive traffic data, which includes at least: traffic flow, speed, accident reports and weather conditions.

[0022] Here, the traffic data can be multi-source and heterogeneous, including: intersections, road sections, ancillary buildings and facilities, etc.; it can also include: traffic decision-making data, such as decision-making bodies, implementing bodies, legal and regulatory basis for decision-making and decision-making personnel, etc.; it can also include: traffic event database, such as the cause of the event, surrounding roadside perception data (vehicle speed, congestion, driving trajectory), traffic flow, speed, accident reports and weather conditions, etc.

[0023] Step 102: Extract several key features related to traffic control from the traffic data, as well as several traffic conditions corresponding to each key feature, and a first causal graph between traffic objects corresponding to each traffic condition.

[0024] Here, key features related to traffic control can include: environmental information such as time and weather; event information such as early warning events and traffic accidents; and handling information such as traffic command and traffic prompts. Traffic conditions can be: congested, smooth, etc.; traffic objects can be: traffic billboards, reversible lanes, traffic lights, broadcasts, etc.; and personnel can be: traffic police, medical personnel, firefighters, etc.

[0025] Cause-and-effect diagrams are an important analytical tool in the field of intelligent transportation. They are used to reveal the relationships between causes and effects in traffic phenomena, graphically demonstrating how different factors interact to lead to specific traffic conditions, such as traffic congestion and accidents. For example, when analyzing the causes of congestion, factors such as the number of vehicles, road construction, and traffic light settings can be considered as causal factors, and the degree of congestion as an effect. A cause-and-effect diagram clearly presents the relationships between them, helping traffic planners and managers to fully understand complex traffic problems and develop targeted solutions.

[0026] Step 103: Create a traffic vertical domain model, which includes an MLM module. The MLM module identifies the second causal graph from the received text. Then, it uses directional masking causal pairs to mask the causes or results in the causal entities in the first and second causal graphs and learns the causal dependencies in all causal entities.

[0027] MLM (Masked Language Model) is the core mechanism of pre-trained models in the field of NLP (such as BERT). Its principle is to replace parts of the text with "[MASK]", allowing the model to predict the masked words based on the context, thereby learning the semantic and grammatical relationships of the language.

[0028] Dependency parsing specifically involves analyzing the dependency relationships between words in a sentence (such as subject-verb, verb-object, attributive, etc.) and constructing a syntactic tree to reveal the structural connections between words. A causal graph is a graphical tool used to represent causal relationships, where nodes represent events or entities, and edges represent causal dependencies (such as "cause → effect").

[0029] The specific process of combining the MLM model with causal analysis:

[0030] Step 1: Identify Causal Entities – Based on Dependency Syntax or Causal Graph. Objective: Extract entity pairs (cause-effect) with causal relationships from the text. Implementation: Dependency Syntax Analysis: By analyzing the syntactic structure of sentences, identify lexical relationships with causal logic. For example, in the sentence "Straight-through congestion leads to the activation of reversible lanes," dependency parsing can determine the verb-object causal relationship between "straight-through congestion" (cause) and "reversible lane activation" (effect). Causal Graph: Pre-construct a causal knowledge graph in the transportation domain (e.g., "road construction → congestion," "accident → slow traffic," etc.), match entities in the text with the graph, and identify causal pairs. Example: The text contains "Straight-through congestion leads to reversible lanes." Using the above method, identify "straight-through congestion" as the cause and "reversible lane" (referring to reversible lane adjustment) as the effect, forming a causal entity pair.

[0031] Step 2: Targeted Masking of Causal Pairs – Paired masking of causes or effects. Objective: To force the model to learn causal dependencies and improve its understanding of causal logic by masking portions of causal entities. Implementation: Paired masking: Instead of randomly masking individual words, for identified causal entity pairs, one of the "cause" or "effect" is masked, while the other is retained as a contextual clue. For example, for the causal pair "straight-through congestion → variable," it might be processed as: masking the cause: "[MASK] produces variable," requiring the model to predict the cause "straight-through congestion" based on "variable"; or masking the result: "straight-through congestion produces [MASK]," requiring the model to predict the result "variable" based on the cause "straight-through congestion." Directionality: The masking strategy is based on the directionality of causal relationships (cause → effect), ensuring that the model learns unidirectional dependencies in causal logic, rather than simple word co-occurrence.

[0032] Traditional MLM only learns the semantic relationships of the context, but when combined with causal analysis, the model can explicitly capture the logical dependencies of "cause-effect", which is more in line with the characteristics of problems in the transportation field (such as the need to analyze the causes of congestion and the effects of measures).

[0033] Step 104: Train the traffic vertical domain model based on the traffic data.

[0034] During training, the GLM-4-9B pre-trained model framework was used, employing a single-node 8-card NVIDIA A100 80GB GPU with BFloat16 precision. A multi-module training method was adopted, consisting of three modules: traffic event sequence labeling, event level assessment, and decision tree generation. An alternating training method was used, iterating through the parameters of one module per round. During module-by-module training... The learning rate is small when fine-tuning multi-step planning. The regularization parameters are weight decay of 0.01 and label smoothing of 0.1, and the global gradient norm is constrained by gradient clipping of 1.0.

[0035] The parameter optimization process includes the learning rate ( The adjustment involves setting differentiated weight decay (0.01→0.05) for different modules, and adjusting the training batch size according to the memory usage.

[0036] After training is complete, evaluation is required, including the following steps: 1. Test set processing: manually remove duplicate and invalid data, stratify the data according to event type, and globally shuffle the data using a fixed random seed; 2. BLEU-4 metric calculation: calculate the BLEU score using the corpus_bleu function of the NLTK library; 3. ROUGE metric calculation: calculate the ROUGE metric using the rouge_scorer module of the rouge_score library.

[0037] In this embodiment, receiving traffic data specifically includes:

[0038] Traffic data is received, and then missing values ​​in the traffic data are predicted and filled, and outliers in the traffic data are removed. After that, the traffic data is normalized.

[0039] Here, the traffic data undergoes data cleaning, which involves removing duplicate data and processing erroneous data. For example, duplicate records at the same time point in traffic flow data, or erroneous data caused by sensor malfunctions, can be handled by deletion, interpolation (such as filling with the average of the preceding and following time points), or predictive filling (e.g., using algorithms such as linear interpolation).

[0040] Afterwards, the traffic data can be standardized and normalized, that is, the data can be extracted into text according to a complete event handling process, including event time, location, event type (accident, equipment damage, congestion, etc.), event factors (human, vehicle, road, environment), and handling process (traffic command, road management, law enforcement process, information release).

[0041] In this embodiment, the extraction of several key features related to traffic control from the traffic data, and several traffic conditions corresponding to each key feature, specifically includes the following traffic objects corresponding to each traffic condition:

[0042] The traffic data is segmented into several sentences and several words. Named entity recognition algorithm and dependency parsing are used to extract several key features related to traffic control. Then, the traffic data is labeled. Then, a constraint-based causal discovery algorithm is used to construct a first causal graph between several traffic events corresponding to each key feature and traffic objects corresponding to each traffic event.

[0043] Traffic objects include: equipment (e.g., traffic billboards, reversible lanes, traffic lights, and broadcasts); and personnel (e.g., traffic police, medical personnel, firefighters).

[0044] Here, the spaCy open-source tool can be used for automatic annotation, followed by manual correction and supplementation. Afterwards, [the following can be used]... The quality of the annotations was evaluated, and the annotation results were manually sampled and checked. The Kappa coefficient was used to measure the consistency among evaluators in a classification task. PA is the observed consistency rate, which refers to the proportion of samples in which evaluators reached agreement in actual observation. PE is the random expected consistency rate, which represents the probability that evaluators might reach agreement under random conditions. (1-PE) was used for standardization, constraining the results to the interval [-1, 1] to exclude the influence of random consistency.

[0045] Named Entity Recognition (NER) refers to the identification of entities with specific meanings in text, mainly including names of people, places, organizations, and proper nouns. It aims to extract meaningful nouns or phrases from text, providing a foundation for applications such as knowledge base question answering, machine translation, information retrieval, sentiment analysis, and knowledge graphs.

[0046] Dependency syntax advocates describing the syntactic relationships between words through asymmetric dependency relations, where each sentence element depends on only one parent node. First proposed by the French linguist L. Tesniere, dependency syntax describes the dependency relationships between words by analyzing sentences into dependency syntax trees, and these relationships are semantically related.

[0047] Constraint-based causal discovery algorithms, also known as the PC algorithm (Peter-Clark algorithm), are a classic Bayesian network structure learning method. The core objective is to infer causal relationships or conditional independence relationships between variables from observed data, constructing a directed acyclic graph (DAG, i.e., a Bayesian network). The PC algorithm process includes: 1. Connecting all variables to obtain a completely undirected graph; 2. Removing redundant edges through a conditional independence (d-separation) test to obtain a skeleton graph; 3. Identifying collision structures (v-structures) in the skeleton graph to determine the orientation of some edges; 4. After identifying all collision structures, attempting to determine the orientation of the remaining edges using the properties of the DAG and the Meek rule.

[0048] Here, the first challenge is dealing with a massive amount of complex and disorganized traffic data. This data may originate from multiple sources, including autonomous intelligent transportation systems, smart traffic sensors, and traffic management logs, and includes various forms such as text descriptions and voice recordings. To effectively analyze this data, the first step is text segmentation. This process is akin to breaking down a large and chaotic jigsaw puzzle into individual smaller pieces. Using advanced text processing techniques, continuous text data is divided according to certain rules, ultimately yielding several sentences and words. These sentences and words act as the basic elements of the data, laying the foundation for subsequent in-depth analysis.

[0049] Once the sentences and words are segmented, the next step is to extract key features relevant to traffic management. To achieve this, named entity recognition (NER) and dependency parsing algorithms are employed. NER acts like a precise treasure hunter, quickly identifying entities with specific meanings from numerous words and sentences. These entities include traffic locations (intersections, road sections, bridges), traffic times (peak hours, accident times), types of traffic objects (cars, buses, subways), and types of traffic events (accidents, traffic congestion, traffic control). Dependency parsing, on the other hand, acts like an experienced language analyst, examining the dependency relationships between sentence components, such as the logical connections between subject and predicate, verb and object. By combining these two methods, several key features closely related to traffic management can be accurately extracted from massive amounts of data. These key features are like hidden treasures deep within the data, containing patterns and potential problems in traffic operations.

[0050] After successfully extracting key features, the next step is to label the traffic data. Labeling is like attaching a clear label to each data element, giving them a distinct identity and attributes. Based on the previously extracted key features, each sentence, word, or data fragment is assigned a corresponding category label. For example, a sentence describing a traffic accident will be labeled "traffic accident," and further subdivided into accident types, such as collision accidents, rear-end collisions, etc.; content related to traffic congestion will be labeled "traffic congestion," with the degree of congestion and the location of the incident noted. Through meticulous labeling, the originally chaotic data becomes orderly and easy to understand, providing a high-quality data foundation for subsequent causal analysis.

[0051] Finally, a constraint-based causal discovery algorithm is employed to construct several traffic events corresponding to each key feature, as well as the traffic objects corresponding to each traffic event. This constraint-based causal discovery algorithm acts like a master logician, deeply exploring the causal relationships between key features and traffic events, and between traffic events and traffic objects, based on the correlations and constraints between data. For example, when a key feature is found to be "peak hours," the algorithm can construct several corresponding traffic events, such as "increased traffic congestion" and "increased traffic accidents." For each traffic event, the algorithm further analyzes the possible traffic objects involved; for example, in a traffic congestion event, multiple traffic objects such as cars and buses may be involved. By constructing such a causal relationship graph, the mutual influence and mechanisms of action between various factors can be intuitively observed.

[0052] In this embodiment, the MLM module adopts a multi-task learning architecture and a three-layer model structure of device state, spatiotemporal relationship triples, and causal entities. It also uses triple extraction to extract a second causal graph from the received text, and uses Span Cross-Entropy and relationship classification cross-entropy to predict position-sensitive spans. Furthermore, it uses a dynamic weight adjustment method based on Uncertainty Weighting to automatically balance task weights for multi-task loss fusion.

[0053] Improving the traditional MLM (Masked Language Model) is a core technique for training Large Language Models (LLMs). Its core idea is to randomly mask parts of the input text, forcing the model to predict the masked words based on context, thereby learning the semantic and grammatical structure of the language. The model is forced to predict causal entities in event chains, while auxiliary tasks are set up to enhance data learning in areas such as device state reasoning, spatiotemporal relation resolution, and dialogue policy consistency. Textual results are fed back to the digital twin system, ensuring that the large model can engage in conversational interaction with the digital twin.

[0054] In this scenario, the MLM model employs a multi-task learning architecture, integrating traffic domain knowledge from different dimensions through a three-layer model structure. Combined with specific task design and loss function optimization, it achieves understanding of complex traffic text. Its core framework can be broken down into the following parts:

[0055] 1. The first layer is used for the device state model. The goal is to capture the state information of traffic equipment (such as traffic lights, surveillance cameras, variable lane indicators, etc.). Examples include: "The traffic light is currently red" and "The variable lane indicator shows straight ahead." These states are the basic data of the traffic scene.

[0056] 2. The second layer is used for the spatiotemporal relation triple model. Definition: A triple refers to an "entity-relationship-entity" structure, used to represent the spatiotemporal relationship of traffic events. Components: Entity: such as road segment ("Chang'an Avenue"), time ("morning rush hour"), event ("congestion"); Relationship: such as "occurred at" (road segment-time), "caused by" (event-impact). Function: To locate traffic events in a specific spatiotemporal context. For example, "congestion occurred on Chang'an Avenue during the morning rush hour" can be represented as (Chang'an Avenue, occurred at, morning rush hour) + (Chang'an Avenue, occurred, congestion).

[0057] 3. The third layer is used for the causal entity model, continuing the logic above: identifying causal pairs (cause-effect) in the text, such as "accident (cause) → traffic disruption (effect)". Related to the first two layers: causal entities need to be combined with equipment status and spatiotemporal relationships, for example, "Traffic light malfunction on a certain road segment (equipment status) caused congestion during the morning rush hour (spatiotemporal) (causal result)".

[0058] Triple extraction task and data processing

[0059] The goal of triple extraction is to extract structured triple information from unstructured text and transform natural language into machine-processable knowledge graph elements.

[0060] The implementation method is span extraction, which locates continuous segments (spans) representing entities in the text, such as "morning rush hour" and "Chang'an Avenue congestion"; relation classification: determine the type of relationship between entities (such as spatiotemporal relationship, causal relationship).

[0061] Example: The text is "During the morning rush hour, traffic was slow on the East Third Ring Road due to a traffic accident". Extract the ternary pairs: (East Third Ring Road, time period, morning rush hour) - spatiotemporal relationship; (traffic accident, caused, slow traffic) - causal relationship.

[0062] Loss Function and Multi-Task Optimization Strategy

[0063] 1. SpanCross-Entropy is used for position-sensitive span prediction. Its function is to optimize the prediction of the start and end positions of entity spans. Principle: It treats entities as continuous intervals in the text (such as from the i-th word to the j-th word), and calculates the deviation between the predicted position and the actual position using the cross-entropy loss function, emphasizing the accurate localization of entity boundaries.

[0064] 2. Relationship Classification: Cross-entropy is used to optimize the classification of relationships between entities. Its function is to classify the relationship type of extracted entity pairs (such as "causal", "spatiotemporal", "dependent"). Example: Determine whether "traffic accident" and "slow-moving traffic" have a causal relationship, and improve the classification accuracy by using the cross-entropy loss function.

[0065] 3. UncertaintyWeighting (UHT) is used for multi-task loss fusion. Background: In multi-task learning, different tasks (such as entity extraction and relation classification) may have different importance, requiring a balance of loss weights. Core method: Automatically adjust weights based on task uncertainty: Assign higher weights to tasks with high uncertainty (such as rare relation classification) to enhance learning; Mathematical expression: The loss function can be expressed as... ,in, Dynamic calculations based on task uncertainty (e.g., measured by variance or entropy).

[0066] Input traffic text → Three-layer model extracts equipment status, spatiotemporal triples, and causal entities respectively → Triples extract task structured data → Optimize predictions for each task through Span CE and relation CE → Train the model based on UncertaintyWeighting dynamic fusion loss.

[0067] Embodiment 2 of the present invention provides a device for generating a traffic vertical domain model for an autonomous intelligent transportation system, comprising the following modules:

[0068] The data acquisition module is used to receive traffic data, which includes at least: traffic flow, speed, accident reports and weather conditions.

[0069] The data preprocessing module is used to extract several key features related to traffic control from the traffic data, as well as several traffic conditions corresponding to each key feature, and a first causal graph between traffic objects corresponding to each traffic condition.

[0070] The model creation module is used to create a traffic vertical domain model. The traffic vertical domain model includes an MLM module. The MLM module identifies a second causal graph from the received text. Then, it uses directional masking causal pairs to mask the causes or results in the causal entities in the first and second causal graphs and learns the causal dependencies in all causal entities.

[0071] The training module is used to train the traffic vertical domain model based on the traffic data.

[0072] In this embodiment, the data acquisition module is further configured to:

[0073] Traffic data is received, and then missing values ​​in the traffic data are predicted and filled, and outliers in the traffic data are removed. After that, the traffic data is normalized.

[0074] In this embodiment, the data preprocessing module is further used for:

[0075] The traffic data is segmented into several sentences and several words. Named entity recognition algorithm and dependency parsing are used to extract several key features related to traffic control. Then, the traffic data is labeled. Then, a constraint-based causal discovery algorithm is used to construct a first causal graph between several traffic events corresponding to each key feature and traffic objects corresponding to each traffic event.

[0076] In this embodiment, the MLM module adopts a multi-task learning architecture and a three-layer model structure of device state, spatiotemporal relationship triples, and causal entities. It also uses triple extraction to extract a second causal graph from the received text, and uses Span Cross-Entropy and relationship classification cross-entropy to predict position-sensitive spans. Furthermore, it uses a dynamic weight adjustment method based on Uncertainty Weighting to automatically balance task weights for multi-task loss fusion.

[0077] Embodiment 3 of the present invention provides a traffic control system, such as Figure 2 As shown, it includes: a traffic vertical domain model, a data platform, and a twin map obtained by executing the generation method in Embodiment 1.

[0078] The traffic vertical domain model can receive real-time event data from the data center, process the received data, provide decision data, and then display it in the form of an HTML page.

[0079] The twin map receives data from the HTML page.

[0080] In this embodiment, the twin map receives data from the HTML page based on the WebTRC protocol.

[0081] This traffic control system enables rapid interactive responses with geographic entities through natural language, such as quick location, spatial viewing, and event mapping, thereby improving management efficiency.

[0082] On the page, users can input commands via text dialogue in the form of a digital human. The system then calls API interfaces and matches the semantic information such as name and location returned by the API interfaces with the semantics set in the digital twin to generate functions such as location and event point mapping display.

[0083] The following processes are all completed through the interface in the digital twin system. For example, the intelligent annotation system is as follows: 1. Define the label protocol, 2. Create digital twin points, lines and surfaces, 3. Define the annotation interface through structure mapping and other methods. Through the semantic text content in step one, the event coordinates, location and other information are mapped to the interface to automatically generate digital twin annotations.

[0084] This traffic control system has the following functions:

[0085] 1. Semantic Information Reading: This involves using WebRTC on the web to efficiently and with low latency acquire audio and video streams for map transmission and interaction. WebSocket is used to establish the WebRTC connection and handle other control and non-real-time data transmission. After acquiring the data, the intelligent interactive entity uses C++ / Blueprint interpretation to read and process the data in real time, extracting tags such as name, type, and location. This involves interpreting the information returned from the web segment and parsing objects and arrays within the information using C++ / Blueprint code. This includes defining data structures, writing parsing functions, and extracting fields and arrays.

[0086] 2. Intelligent labeling: Establish a labeling system. By binding point, line, and area labels and matching them with the device types obtained from the interface, intelligent labeling is automatically completed on the map based on location information.

[0087] 3. Intelligent positioning: Through semantic reading and matching of data such as coordinate position and label system, and after transmitting it to the camera on the map via socket communication, the camera parameters are bound to the user's viewpoint.

[0088] 4. Fast response: Through a front-end dialog window, the front-end page allows users to input information such as place name and address, and the map completes a fast response by reading semantic information and intelligent positioning.

[0089] Active early warning and control specifically includes the following:

[0090] 1. Proactive data acquisition: Traffic incident information is proactively pushed using the WebSocket protocol, establishing a WebSocket connection between the client and server. After a warning is issued, the server proactively sends a message to the client.

[0091] 2. Rapid positioning of intelligent agents: Based on the acquired early warning information, the system calculates in real time and quickly calls relevant map tools to achieve rapid positioning and display of events.

[0092] 3. Large-scale model analysis and decision-making: Based on the sources of multimodal event data, features are extracted and combined with the model library. Matching decisions are made by combining event type and feature information with historical event processing procedures, and traffic control strategies are formed.

[0093] 4. Early Warning Decision Distribution: By connecting to the control protocol interface of front-end devices, data is distributed, including configuration information for traffic lights, traffic guidance signs, reversible lanes, and broadcasts. Data is pushed and distributed via the data interaction interface in the form of text, images, and voice.

[0094] 5. Early warning response: The front-end hardware devices complete the decision-making and response to events, using communication protocols such as Modbus, MQTT, and TCP to achieve remote communication between the control terminal and the hardware devices. After receiving the command, the front-end hardware devices will execute the corresponding operation, such as changing the status of traffic lights or changing lane direction.

[0095] 6. Early warning simulation: Simulation of the early warning decision-making process in the platform system. Using real-time dynamic monitoring data to drive the simulation unit of the digital twin world, the early warning response process is accurately visualized.

[0096] It should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0097] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating a traffic sub-area model for an autonomous intelligent transportation system, characterized in that, The method comprises the following steps: receiving traffic data, wherein the traffic data at least comprises traffic flow, speed, accident report and weather condition; text segmenting the traffic data to obtain a plurality of sentences and a plurality of words, extracting a plurality of key features related to traffic control by using named entity recognition algorithm and dependency syntax analysis method, then labeling the traffic data, and then constructing a first causal graph between a plurality of traffic events corresponding to each key feature and a plurality of traffic objects corresponding to each traffic event by using constraint-based causal discovery algorithm; creating a traffic vertical domain model, wherein the traffic vertical domain model comprises an MLM module, the MLM module identifies a second causal graph from the received text, then performs pair masking on causes or results in causal entities in the first and second causal graphs by using directional mask causal pairs, and learns causal dependency relationships in all causal entities; the MLM module adopts a multi-task learning architecture, adopts a device state model, a spatio-temporal relation triple model and a causal entity model three-layer model structure, extracts the second causal graph from the received text by using triple extraction, predicts position-sensitive span by using Span Cross-Entropy and relationship classification cross-entropy, and performs multi-task loss fusion based on Uncertainty Weighting automatic task weight balancing dynamic weight adjustment method; wherein the device state model aims to capture state information of traffic devices; the spatio-temporal relation triple model is an "entity-relation-entity" structure, which is used to represent the spatio-temporal correlation of traffic events; and the causal entity model is used to identify causal pairs in the text; training the traffic vertical domain model based on the traffic data.

2. The generation method of claim 1, wherein, The receiving traffic data specifically comprises: receiving traffic data, then predicting and filling missing values in the traffic data, removing abnormal values in the traffic data, and then normalizing the traffic data.

3. A device for generating a traffic sub-area model for an autonomous intelligent transportation system, characterized by The method comprises the following modules: a data acquisition module, which is used to receive traffic data, wherein the traffic data at least comprises traffic flow, speed, accident report and weather condition; a data preprocessing module, which is used to text segment the traffic data to obtain a plurality of sentences and a plurality of words, extract a plurality of key features related to traffic control by using named entity recognition algorithm and dependency syntax analysis method, then label the traffic data, and then construct a first causal graph between a plurality of traffic events corresponding to each key feature and a plurality of traffic objects corresponding to each traffic event by using constraint-based causal discovery algorithm; The model creating module is configured to create a traffic vertical domain model, which comprises an MLM module configured to identify a second cause-effect graph from the received text, and then perform pair masking on the causes or results in the cause-effect entities in the first and second cause-effect graphs using a directional mask cause-effect pair, and learn the cause-effect dependencies in all the cause-effect entities; the MLM module uses a multi-task learning architecture, and uses a device state model, a space-time relationship triple model, and a cause-effect entity model three-layer model structure, simultaneously extracts the second cause-effect graph from the received text using triple extraction, performs position-sensitive span prediction using Span Cross-Entropy and relationship classification cross-entropy, and performs multi-task loss fusion based on an Uncertainty Weighting automatic task weight balancing dynamic weight adjustment method; the device state model aims to capture the state information of the traffic device; the space-time relationship triple model is an "entity-relation-entity" structure, and is used to represent the space-time association of the traffic event; and the cause-effect entity model is used to identify the cause-effect pairs in the text. The training module is configured to train the traffic vertical domain model based on the traffic data.

4. The generating device of claim 3, wherein, The data obtaining module is further configured to: receive traffic data, then perform missing value prediction and filling processing on the traffic data, remove the abnormal values in the traffic data, and then perform normalization processing on the traffic data.

5. A traffic management system, characterized by The system comprises: a traffic vertical domain model, a data platform, and a twin map generated by the traffic vertical domain model generating method of claim 1 or 2; the traffic vertical domain model can receive real-time event data from the data platform, process the received data, and give decision data, and then display the decision data in the form of an HTML page; the twin map receives data from the HTML page.

6. The traffic management and control system according to claim 5, wherein the twin map receives data from the HTML page based on a WebTRC protocol.

Images