Rapid response processing system for transportation business based on large language model technology

Abstract

The present invention belongs to the field of artificial intelligence technology, and specifically relates to a transportation business rapid response processing system based on a large language model technology, the system comprising: a multi-source transportation data fusion unit, used to integrate multi-source transportation data from different data sources, using tensor fusion technology to express the fusion of multi-source transportation data as a high-order tensor, and obtain dimensionality reduction key features; a transportation knowledge graph construction unit, used to construct a transportation knowledge graph based on a hypergraph based on dimensionality reduction key features; a business impact and processing unit, used to receive user requests, use large language model technology to encode user requests into high-dimensional semantic vectors, calculate the cross-modal semantic similarity between user requests and nodes and relationships in the transportation knowledge graph, and send data corresponding to the node for the user request based on the posterior probability. The present invention realizes intelligent processing and efficient response to complex business scenarios in the transportation industry.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a rapid response processing system for transportation business based on large language model technology. Background Art

[0002] With the acceleration of modern urbanization, transportation systems are becoming increasingly complex. Especially in large cities and transportation hubs, the rapid growth of traffic flow has made traffic management and scheduling increasingly difficult. In order to cope with the increasing traffic pressure, many existing technologies have proposed solutions for intelligent transportation systems (ITS), using various technical means to monitor, analyze and optimize traffic flow in real time. However, existing intelligent transportation systems still have many problems when facing complex traffic conditions, especially in dealing with emergencies, responding quickly to user needs, and efficiently processing multi-source heterogeneous traffic data. Based on this background, how to develop a more intelligent, efficient and fast-responding transportation business processing system has become a major challenge in the current technology field. Existing intelligent transportation systems mainly rely on traditional data collection and analysis technologies. These systems usually collect traffic data in real time through devices such as cameras, radars, GPS, sensors, etc. arranged on the road, and analyze traffic flow based on predefined rules or statistical models. For example, many existing systems judge traffic flow based on camera data, infer traffic congestion based on vehicle GPS trajectories, or use historical traffic data for short-term traffic forecasts. Although these systems have improved traffic management to a certain extent, their main problem is that they have limited ability to integrate and process multi-source heterogeneous data, and are unable to cope with the complexity and variability of traffic data.

[0003] For example, in traditional traffic data processing, data from different sensors often have different formats, dimensions, and spatiotemporal characteristics. GPS data provides vehicle location information, while road sensors may only monitor the speed of vehicles passing through a certain section of road, cameras record traffic images, and weather data may affect the safety and speed of vehicle driving. These data sources are diverse and the information is complex. It is difficult for traditional traffic systems to effectively integrate them into a unified data representation, resulting in incomplete decision-making basis. In addition, the data processing method in traditional systems is usually linear, and it is impossible to capture the complex dependencies between nodes in the traffic network. For example, the congestion of a road depends not only on its own traffic, but also on adjacent roads, traffic signals, and other factors. In addition, the existing technology also faces a serious problem, that is, how to deal with emergencies in the traffic system. Traffic accidents, bad weather, road construction and other emergencies often lead to traffic congestion on local roads, and the impact of these emergencies may quickly expand to the entire traffic network. It is difficult for traditional systems to make effective response measures in a timely manner. Existing systems usually rely on preset emergency handling rules or simple traffic models to deal with emergencies, but the design of these rules and models is often inflexible and cannot dynamically adapt to complex actual traffic conditions. Therefore, when emergencies occur, the response speed and processing effect of traditional systems often cannot meet actual needs, resulting in inefficient traffic management and even further exacerbating traffic congestion. In addition, existing intelligent transportation systems also have great limitations in user interaction. Most systems cannot effectively process users' natural language requests. Users need to query and operate through fixed-format instructions or options, which greatly limits the availability and flexibility of the system. With the development of big data and artificial intelligence technologies, especially the rise of large language models, how to apply large language models to transportation systems so that the system can understand and process complex natural language queries raised by users has become a new research direction. However, the exploration of existing technologies in this regard is still very limited. Most systems lack the ability to deeply understand and analyze user requests, resulting in poor interactive experience of the system and the inability to fully tap the potential of intelligent transportation systems. Summary of the invention

[0004] In view of this, the main purpose of the present invention is to provide a rapid response processing system for transportation business based on large language model technology. The present invention realizes intelligent processing and efficient response to complex business scenarios in the transportation industry.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A rapid response processing system for transportation business based on large language model technology, the system includes: a multi-source transportation data fusion unit, which is used to integrate multi-source transportation data from different data sources, use tensor fusion technology to express the fusion of multi-source transportation data as a high-order tensor, extract features from the fused high-order tensor, use a method combining tensor decomposition and kernel principal component analysis to extract key features and reduce dimensions to obtain reduced-dimensional key features; a transportation knowledge graph construction unit, which is used to construct a transportation knowledge graph based on a hypergraph based on the reduced-dimensional key features; a business impact and processing unit, which is used to receive user requests, use large language model technology to encode user requests into high-dimensional semantic vectors, calculate the cross-modal semantic similarity between user requests and nodes and relationships in the transportation knowledge graph, find the most relevant nodes, build a probabilistic graph model, reason about the most relevant nodes, obtain the posterior probability of the nodes in each transportation knowledge graph under a given user request, and send the data corresponding to the node for the user request based on the posterior probability.

[0007] Furthermore, the multi-source traffic data includes: real-time traffic flow, vehicle GPS trajectory, road infrastructure data and weather information; the road infrastructure data includes: road length, road width and road type; the weather information includes: temperature, precipitation, wind speed and wind direction; high-order tensor Calculated by the following formula:

[0008] ;

[0009] in, is the number of data sources; is the subscript index; For the preset The weight coefficient of each data source; For the Data source The feature vector of dimension Data source Features is a tensor product operator, which indicates the fusion operation of different data sources; is the time embedding matrix; spatial embedding matrix; is the modulus of the tensor Product operation; For the The characteristic dimension of a data source.

[0010] Furthermore, the time embedding matrix ;in, represents all time steps in a time series; Represents the time step The time embedding vector of The temporal embedding vector Calculated by the following formula:

[0011] ;

[0012] in, For time The embedding vector of is the current time step; is the dimension of time embedding, which determines the length of the time embedding vector; is the subscript index.

[0013] Furthermore, the high-order tensor is decomposed by the following formula to obtain the tensor decomposition result: :

[0014] ;

[0015] in, It is Factor matrix on the module; define a nonlinear mapping function , decompose the tensor into Mapping to high-dimensional feature space : ; Use kernel function To calculate the similarity between tensor samples, The result of tensor decomposition Middle elements, The result of tensor decomposition Middle elements; is the modulus of the tensor Product operation; is the dimension of the high-order tensor; let the kernel matrix Middle Row, No. Elements of a column ;in, is the bandwidth parameter of the kernel function, which determines the nonlinearity of the kernel function. The kernel matrix is ​​given by the following formula: Centralization:

[0016] ;

[0017] in, is the centralized kernel matrix; Is the size of The identity matrix of is the dimension of the kernel matrix; for the centralized kernel matrix Perform eigenvalue decomposition and obtain the eigenvalue and the corresponding eigenvector :

[0018] ;

[0019] Before selection The eigenvector corresponding to the maximum eigenvalue is used as the key feature for dimensionality reduction .

[0020] Furthermore, the process of constructing a traffic knowledge graph based on a hypergraph by the traffic knowledge graph construction unit specifically includes: using the dimensionality reduction key features Constructing node feature matrix ; Define the nodes and edges of the hypergraph and construct the association matrix and the hyperedge weight matrix ; Calculate the node degree matrix and the hyperedge degree matrix ; Construct the hypergraph Laplacian matrix ; Fusion of node features and hypergraph structure to obtain updated node feature matrix .

[0021] Furthermore, using the key features of dimensionality reduction , construct node feature matrix ;in, Key features for dimensionality reduction No. elements, indicating the The feature vector of nodes; in the constructed traffic knowledge graph, the total number of nodes is ; Node feature matrix , is the number of nodes, is the feature dimension after dimensionality reduction; let Indicates The feature vector of each node; the association matrix Represents the association relationship between nodes and hyperedges:

[0022] ;

[0023] in, is the number of hyperedges; let the eigenvector As the nodes; For the hyperedges; hyperedge weight matrix Represents the weight of each hyperedge, reflecting the strength of the relationship between the nodes connected by the hyperedge. ; Indicates The weight of the hyperedge; the node degree matrix is ; Hyperedge degree matrix ;in, , ; Representation Node The degree of represents the sum of the weights of the hyperedges associated with it; Represents a hyperedge The degree indicates the number of nodes connected by the hyperedge.

[0024] Furthermore, the hypergraph Laplacian matrix Use the following formula to express it:

[0025] ;

[0026] in, ; , is the identity matrix; represents the inverse matrix of the hyperedge degree matrix; represents the inverse square root of the node degree matrix; is the transpose of the hypergraph association matrix; the updated node feature matrix Calculated using the following formula:

[0027] .

[0028] Furthermore, the specific execution process of the business impact and processing unit includes: using the pre-trained large language model to convert the user request Encoded into high-dimensional semantic vector ; Calculated using cosine similarity With each node feature vector Similarity ; for The Elements:

[0029] ;

[0030] in, For user requests and Similarity of nodes; Based on similarity , select the top nodes, forming a candidate node set ; For the candidate node set , build a probabilistic graph model, and reason about highly relevant nodes, including: extracting candidate nodes and their directly related neighbor nodes and edges from the traffic knowledge graph to form a subgraph ; Use Markov random fields or conditional random fields to Modeling; Calculate the posterior probability of each node in the subgraph through the message passing algorithm .

[0031] Furthermore, the posterior probability of each node in the subgraph is calculated using the message passing algorithm using the following formula:

[0032] ;

[0033] in, For Node The potential function reflects and The compatibility is taken as similarity ; For edge The potential function represents the node and The strength of the association between is the normalization factor; For a given user request, the node The posterior probability of .

[0034] By adopting the above technical scheme, the present invention produces the following beneficial effects: the present invention demonstrates excellent ability in the fusion processing of multi-source heterogeneous data. The data in the traditional traffic system often comes from a variety of sensors and devices, including real-time traffic flow, vehicle GPS trajectory, road infrastructure data, weather information, etc. These data formats and spatiotemporal characteristics are different, and the existing system is difficult to process efficiently. By introducing tensor decomposition technology and the construction of high-order tensors, the present invention can fuse these traffic data from different sources into a unified high-dimensional representation, retaining the spatiotemporal characteristics and multi-dimensional associations in the original data. This not only improves the accuracy of data fusion, but also ensures that important detail information will not be lost during the processing process, thereby providing a more comprehensive and accurate basis for subsequent analysis and decision-making. The present invention has a high degree of flexibility and real-time response capability when dealing with emergencies in the traffic system. By constructing a hypergraph Laplacian matrix and dynamically updating the node feature matrix, the present invention can effectively capture the global dependencies and dynamic changes between nodes in the traffic network. When an emergency occurs (such as a traffic accident, bad weather, etc.), the system can quickly identify the affected key nodes and their associated neighbor nodes, thereby providing a timely response plan for traffic managers. This dynamic adjustment mechanism based on node degree, hyperedge degree and correlation matrix ensures that the system can flexibly adjust according to real-time data when facing complex traffic conditions, providing accurate traffic scheduling and accident warning capabilities. Compared with traditional emergency response methods based on rules or simple statistical models, the present invention has faster response speed and more accurate prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the system structure of a transportation business rapid response processing system based on large language model technology provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0037] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other alternative features that are equivalent or have similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0038] Example 1: Reference Figure 1 A rapid response processing system for transportation business based on large language model technology, the system includes: a multi-source transportation data fusion unit, which is used to integrate multi-source transportation data from different data sources, use tensor fusion technology to represent the fusion of multi-source transportation data as a high-order tensor, extract features from the fused high-order tensor, and use a method combining tensor decomposition and kernel principal component analysis to extract key features and reduce dimensions to obtain reduced-dimensional key features; a transportation knowledge graph construction unit, which is used to construct a hypergraph-based transportation knowledge graph based on reduced-dimensional key features; a business impact and processing unit, which is used to receive user requests, use large language model technology to encode user requests into high-dimensional semantic vectors, calculate the cross-modal semantic similarity between user requests and nodes and relationships in the transportation knowledge graph, find the most relevant nodes, build a probabilistic graph model, reason about the most relevant nodes, obtain the posterior probability of each node in the transportation knowledge graph under a given user request, and send the data corresponding to the node for the user request based on the posterior probability.

[0039] Specifically, the multi-source traffic data fusion unit is a core component of the transportation business rapid response processing system based on the large language model technology, which is responsible for effectively fusing traffic data from different sources to provide efficient traffic response capabilities. In the field of transportation, different data sources, such as GPS positioning data, camera monitoring video, vehicle sensor data, road monitoring data, etc., usually record different types of traffic information respectively. These information are highly heterogeneous and dispersed in time, space and type, which makes it difficult to directly integrate and uniformly analyze the data. The multi-source traffic data fusion unit in the present invention realizes the fusion of these heterogeneous data through an innovative method, thereby constructing a unified data representation, laying the foundation for the subsequent construction of traffic knowledge graphs and user request processing. The unit first standardizes the data from different sources through a preprocessing step to eliminate the differences between data dimensions, sampling rates and formats. For example, GPS data is a sequence of spatiotemporal position information, while camera video data is a continuous sequence of image frames, and sensor data may be real-time speed or temperature information. In traditional systems, these data are usually processed independently, while in the present invention, the multi-source traffic data fusion unit adopts tensor fusion technology to uniformly represent these heterogeneous data sources as a high-order tensor. A tensor is a mathematical structure that can contain multiple dimensions and is suitable for processing complex data with multi-dimensional attributes. By representing various types of traffic data as high-order tensors, this system can integrate all data sources into a data structure that is commonly represented in a multi-dimensional space, achieving information unification in multiple dimensions.

[0040] In this process, the system uses an advanced feature extraction technology to ensure that the key information of the data is retained. Multi-source traffic data may contain redundant information or noise, so simple fusion cannot ensure the extraction of the most effective feature information. To solve this problem, the present invention adopts a feature extraction and dimensionality reduction method that combines tensor decomposition and kernel principal component analysis (KPCA). Tensor decomposition technology allows the system to decompose high-order tensors into multiple low-order tensors, helping to identify the potential correlation structure between different traffic data sources. For example, the change in traffic flow in camera video data may be directly related to the vehicle movement path in GPS location data, and tensor decomposition can effectively reveal this complex relationship. KPCA further enhances this process by mapping nonlinear data features into high-dimensional space and performing dimensionality reduction processing to extract key traffic data features. This feature extraction process not only reduces the dimensionality of the data, but also ensures that important information in the fused data is retained. In addition, data fusion is not just a superposition at the physical level, but also involves semantic associations. In the transportation business rapid response system, different data sources may logically represent different levels of traffic information. For example, sensor data may represent the real-time status of a vehicle, while surveillance camera data may be used to identify traffic accidents or road congestion. These data have different semantics, and the information transmission rate and update frequency are also different. Therefore, during the fusion process, the system also uses a weighting strategy to assign weights according to the importance and timeliness of the data source. This ensures that the system can prioritize and use the most valuable data sources under different traffic conditions. For example, when a traffic accident occurs, video data may have a higher priority than GPS data because video can directly provide real-time information at the scene of the accident. Finally, the fused multi-source traffic data is represented as a high-dimensional tensor structure, and key low-dimensional features are obtained through dimensionality reduction processing. These features not only represent the core information of various types of traffic data, but also serve as the basis for the subsequent construction of traffic knowledge graphs. In the rapid response processing of transportation business, the system needs to quickly build a traffic knowledge graph based on these key features and respond to user requests. Therefore, the role of the multi-source traffic data fusion unit is to provide a stable and efficient multi-source data input terminal for the entire system, and to ensure the accuracy and timeliness of the data through a series of complex fusion and feature extraction processes. This technology not only effectively solves the integration difficulties caused by the heterogeneity of traffic data, but also greatly improves the accuracy and speed of system response, providing a solid technical foundation for the intelligent development of the transportation industry.

[0041] The core principle of the traffic knowledge graph construction unit is to use the structure of a hypergraph to represent multiple entities and their relationships. Unlike traditional graph structures, the edges of a hypergraph can connect multiple nodes, and this flexibility gives it a greater advantage in expressing multiple relationships. Specifically, the nodes in the traffic knowledge graph represent various entities in the traffic system, such as roads, vehicles, traffic signals, accidents, etc., and the hyperedges are used to represent the multi-dimensional associations between these entities. For example, a hyperedge may connect three nodes, namely, vehicles, roads, and traffic signals, to represent the complex interactive relationships between these entities. Through this hypergraph structure, the present invention can better reflect the diverse and multi-dimensional relationships in the traffic system, thereby providing a solid foundation for subsequent semantic analysis and reasoning. In the process of constructing the traffic knowledge graph, the attributes and associations of various entities in the traffic field are first determined based on the dimensionality reduction key features extracted from multi-source data. The system uses large language model technology to extract semantic features from the original traffic data and map these features to nodes and hyperedges in the knowledge graph. For example, traffic congestion information captured by surveillance cameras is associated with road nodes, vehicle nodes, and accident nodes. Through semantic analysis of a large language model, the system can identify the potential associations between these nodes and express them through a hypergraph structure. Such semantic extraction can not only identify explicit relationships between entities, but also mine implicit relationships that are difficult to identify through traditional means, making the traffic knowledge graph more comprehensive and accurate.

[0042] The construction of the traffic knowledge graph is not just a static structural representation, it also has adaptive capabilities in a dynamic environment. The traffic system is a complex system that changes in real time. The flow of vehicles, changes in traffic signals, changes in weather, etc. will affect the state of the entire system. In order to solve these problems, the traffic knowledge graph construction unit has designed a mechanism based on dynamic updates. The system will update the knowledge graph regularly or based on event triggers. When new traffic data enters, the system will re-analyze the data through a large language model and adjust the existing knowledge graph. For example, when a traffic accident occurs, the system will automatically identify the accident node through the combination of multi-source information such as camera data and GPS data, and update the status of the road and vehicle nodes related to the accident. This dynamic update mechanism ensures that the knowledge graph can always keep pace with the real-time traffic conditions, providing the latest and most accurate information basis for rapid business response. In addition, the traffic knowledge graph also has reasoning capabilities, which can perform semantic reasoning based on the existing graph structure, helping the system to quickly find the most relevant nodes and relationships in user requests. When a user makes a request, the system will first use a large language model to convert the request into a high-dimensional semantic vector, and perform cross-modal semantic similarity calculations with the nodes in the knowledge graph. By comparing the semantic vector of the user's request with the semantic vectors of each node in the knowledge graph, the system can quickly find the traffic entity most relevant to the request. For example, a user may ask about the current traffic conditions of a certain road. The system will find the road node through semantic matching, and trace the related vehicles, traffic signals and other information along the hyperedges related to it to comprehensively judge the current traffic status. This mechanism based on cross-modal semantic reasoning enables the system to not only be limited to direct data matching, but also to perform more complex logical inferences through semantic relationships, greatly improving the level of intelligence of the response.

[0043] The business impact and processing unit first receives natural language requests from users. These requests may involve a variety of traffic business scenarios, such as querying the current road congestion, requesting real-time traffic information for a certain route, obtaining the rescue progress of the accident site, etc. In traditional systems, the processing of such requests usually relies on keyword matching or simple predefined rules, which makes it difficult to understand the complex and ambiguous expressions of users. The present invention gives the system stronger natural language understanding capabilities through the large language model technology. Specifically, the user's natural language request will be semantically encoded through the large language model and converted into a high-dimensional semantic vector. This semantic vector can not only capture the explicit meaning in the user's request, but also extract the implicit semantic hierarchy and context information, laying the foundation for subsequent traffic data matching. Next, the system will perform cross-modal semantic similarity calculation on the semantic vector of the user's request and the nodes and relationships in the traffic knowledge graph. The nodes in the traffic knowledge graph represent traffic entities, such as roads, vehicles, accidents, etc., and the edges represent the relationships between these entities. Cross-modal semantic similarity calculation can help the system find the nodes in the knowledge graph that are most relevant to the user's request. For example, when a user asks about the traffic conditions of a certain road, the system will not only find nodes related to the road, but also identify related nodes such as traffic signals, congestion conditions, accident information, etc. related to the road through semantic similarity matching. This cross-modal semantic matching enables the system to find the most relevant content from different types of traffic data and provide accurate and comprehensive responses.

[0044] After completing the semantic similarity calculation, the business impact and processing unit will further build a probabilistic graph model for reasoning. In this step, the system not only relies on static semantic matching results, but also predicts the possible status of each node under the user request through the reasoning mechanism. For example, the user may ask about the traffic conditions at a certain time in the future. The system will predict the traffic conditions of the future road through the probabilistic graph model based on the existing traffic data. This reasoning process is based on the deep semantic understanding of the large language model, combined with the historical data, relational structure and current traffic status in the traffic knowledge graph. By calculating the posterior probability of the node, the system can dynamically adjust the response result to ensure that the most likely answer is provided under uncertain circumstances. Another core function of the business impact and processing unit is response optimization. In most cases, user requests in the transportation industry need to be processed in real time, especially when it involves emergency situations such as accident rescue or traffic diversion. Therefore, when generating a response, the system will combine business impact analysis to evaluate the urgency and importance of the current request. For example, when a traffic accident occurs, the system will not only provide relevant information about the accident location, but also generate more targeted response plans through a comprehensive analysis of the scope of the accident, the progress of the rescue and the congestion of the surrounding roads. This business impact analysis relies on the real-time fusion of multi-source data and the dynamic update of the knowledge graph, which enables the system to adjust the response strategy according to the actual situation and provide processing results with high business relevance. In the final stage of user request processing, the business impact and processing unit will provide users with specific data responses based on the reasoning results and business impact analysis. For example, when a user requests real-time traffic information for a route, the system will not only return basic road information, but also generate a detailed and dynamic route status report based on the congestion, vehicle flow, traffic accidents and other data of the road section. If the user requests the progress of the accident handling, the system will generate a comprehensive report by analyzing the rescue progress data, on-site vehicle traffic data, etc., to help users understand the current progress and make decisions. This response process greatly improves the user experience and ensures that users can obtain the most relevant traffic data in a short time.

[0045] Example 2: The multi-source traffic data includes: real-time traffic flow, vehicle GPS trajectory, road infrastructure data and weather information; the road infrastructure data includes: road length, road width and road type; the weather information includes: temperature, precipitation, wind speed and wind direction; high-order tensor Calculated by the following formula:

[0046] ;

[0047] in, is the number of data sources; is the subscript index; For the preset The weight coefficient of each data source; For the Data source The feature vector of dimension Data source Features is a tensor product operator, which indicates the fusion operation of different data sources; is the time embedding matrix; spatial embedding matrix; is the modulus of the tensor Product operation; For the The characteristic dimension of a data source.

[0048] Specifically, the overall idea of ​​the formula is based on the concept of tensor fusion, which effectively integrates traffic information from multiple data sources. Traffic data includes multi-dimensional attributes, such as real-time traffic flow, vehicle GPS trajectory, road infrastructure, and weather information. The feature vector in each data source is Indicates the first dimensional features, such as the width of a road or the real-time vehicle speed. These different dimensional features are combined using the tensor product operator are integrated into a unified representation. This operation can effectively capture the complex associations between data sources, allowing information from different sources to be represented in a unified high-order tensor. This multi-dimensional fusion method can better reflect the interaction between data sources than traditional splicing or simple superposition, ensuring that the system has a more comprehensive and profound understanding of traffic conditions.

[0049] Weight coefficient The formula balances the impact of different data sources. Different data sources may have different importance in different scenarios. For example, in bad weather, weather information (such as precipitation and wind speed) has a greater impact on traffic. The system can increase the corresponding data source. In some cases, the real-time traffic flow may be more important, so the corresponding weight coefficient will be increased. Through this flexible weight adjustment mechanism, the system can dynamically adjust the priority of various data sources according to real-time needs, thereby providing users with the most timely and accurate traffic data response. and is another important component in the formula, which ensures the organic combination of multi-source traffic data in the temporal and spatial dimensions. The traffic system is highly dynamic, and information such as traffic flow and vehicle trajectory changes over time. Therefore, the time embedding matrix The introduction of can capture the time-varying characteristics of the data source. For example, the traffic flow data at a certain moment cannot represent the state of the entire system alone, but needs to be analyzed in time series with the past and future traffic conditions. Similarly, the spatial embedding matrix It is responsible for linking the data source with its geographic location, so that the system can understand the differences in traffic information in different geographical locations. For example, the road congestion in the city center may be completely different from that in the suburbs. Through the spatial embedding matrix, the system can distinguish the traffic patterns in different locations and perform reasonable spatial normalization on the traffic data according to the actual situation. Product Operation This operation further enhances the ability to express temporal and spatial information in tensor fusion. and By effectively combining with high-order tensors, the consistent representation of multi-source traffic data in the time and space dimensions can be achieved. This process enables the system to not only process traffic data at a single moment or location, but also conduct comprehensive analysis in a wider range of time periods and spaces, providing a more reliable foundation for the subsequent construction of traffic knowledge graphs. In addition, the characteristic dimensions of the data source It also plays a key role in the formula. Each data source has its own unique dimensions and attributes. For example, road infrastructure data may include road length and width, while GPS trajectory data may contain vehicle location and speed information. By unifying and fusing these different data dimensions, the formula ensures that all types of traffic data can be fully expressed in the same high-order tensor. This fusion of high-dimensional data not only enhances the system's ability to handle complex traffic situations, but also provides richer data information for rapid business response.

[0050] Example 3: Time Embedding Matrix ;in, represents all time steps in a time series; Represents the time step The time embedding vector of The temporal embedding vector Calculated by the following formula:

[0051] ;

[0052] in, For time The embedding vector of is the current time step; is the dimension of time embedding, which determines the length of the time embedding vector; is the subscript index.

[0053] Specifically, the time embedding matrix Each time step Corresponding to a time embedding vector, the vector captures the relative relationship between time steps, allowing the system to understand and process the time-varying characteristics of traffic data. Specifically, the formula uses sine and cosine functions to generate time embedding vectors. The advantage of this method is that it can express the progressive characteristics of time through periodic functions. For example, in the formula and Function Utilization This time step is taken as an input parameter and the embedding dimension is adjusted by to control the length of the vector. This design allows not only a linear relationship to be maintained between time steps, but also to capture more complex periodic changes, such as the daily repeated peak hours in traffic data, the impact of special events on traffic, etc. Data in transportation systems usually have obvious time dependence. For example, during peak hours in the morning and evening, traffic flow will show relatively fixed pattern changes, while traffic flow at night and during non-peak hours is relatively stable. Through the time embedding vector generated by this formula, the system can generate a unique vector representation for data at different time points. This representation not only reflects the absolute time of the current moment, but also reveals the intrinsic relationship between different time points. The periodic characteristics of the sine and cosine functions enable the system to perceive the periodic changes in time, which is of great significance for predicting traffic flow trends and identifying time-related abnormal events. Furthermore, the formula adjusts the constant To adjust the differences between different time steps. This part is designed to adapt to changes in different time spans and ensure that both short-term and long-term time information can be reasonably encoded. For example, in short-term traffic flow fluctuations, the system needs to respond quickly to the instantaneous state of the vehicle. At this time, the difference between time steps is small, while long-term trend prediction, such as traffic flow changes within a week, requires the system to capture a larger span of time step differences. Through this embedding strategy, temporal information can be accurately encoded at multiple levels, so the system can process traffic data at different time scales.

[0054] In addition, the dimension of the time embedding matrix Directly affects the length of the embedding vector. The choice of dimension determines the fineness of the system's capture of information in time series analysis. A higher embedding dimension can represent more time details, thereby improving the system's ability to recognize complex time patterns in traffic data. This is particularly important when dealing with the impact of weather changes on traffic. For example, rainfall may quickly affect the road's traffic capacity in a short period of time. The system can more accurately capture the time series characteristics of such emergencies through high-dimensional time embedding vectors and generate corresponding processing strategies in a timely manner. By introducing this time embedding mechanism, the transportation business rapid response processing system of the present invention can simultaneously perform efficient traffic data analysis in the time and space dimensions. The time embedding vector is not only closely related to the multi-source data fusion unit in the system, but also provides information support in the time dimension for the construction of the traffic knowledge graph. In the traffic knowledge graph, the dynamic changes of nodes and edges are often highly correlated with time. For example, the traffic conditions of a road not only depend on its physical properties, but are also strongly affected by the time of the day. Through time embedding, the system can more accurately predict changes in traffic conditions and provide users with the most timely and accurate traffic information. Combined with large language model technology, the system can also use the time embedding matrix to further improve the understanding and processing capabilities of natural language requests. User requests may involve traffic information for a specific time period, such as "traffic conditions this afternoon" or "congestion during tomorrow's morning rush hour." Time embedding enables the system to convert the time information in these natural languages ​​into specific vector representations, and then match and reason with the time nodes in the traffic knowledge graph. This combination of semantics and time enables the system to respond to user requests more intelligently, providing strong support for traffic management and decision-making.

[0055] Example 4: Perform tensor decomposition on the high-order tensor using the following formula to obtain the tensor decomposition result :

[0056] ;

[0057] in, It is Factor matrix on the module; define a nonlinear mapping function , decompose the tensor into Mapping to high-dimensional feature space : ; Use kernel function To calculate the similarity between tensor samples, The result of tensor decomposition Middle elements, The result of tensor decomposition Middle elements; is the modulus of the tensor Product operation; is the dimension of the high-order tensor; let the kernel matrix Middle Row, No. Elements of a column ;in, is the bandwidth parameter of the kernel function, which determines the nonlinearity of the kernel function. The kernel matrix is ​​given by the following formula: Centralization:

[0058] ;

[0059] in, is the centralised kernel matrix; Is the size of The identity matrix of is the dimension of the kernel matrix; for the centralized kernel matrix Perform eigenvalue decomposition and obtain the eigenvalue and the corresponding eigenvector :

[0060] ;

[0061] Before selection The eigenvector corresponding to the maximum eigenvalue is used as the key feature for dimensionality reduction .

[0062] Specifically, the data sources of the transportation industry are extensive, including real-time traffic flow, vehicle GPS trajectories, road infrastructure information, weather conditions, etc. Each data source has its own unique characteristic dimensions. In order to process these high-dimensional and complex traffic data, the system adopts the tensor decomposition method to convert the original high-order tensor Decomposition into core tensors and a series of factor matrices The core idea of ​​tensor decomposition is to extract the main patterns and features in the data by reducing the dimension of the data. This process can effectively reduce the complexity of the data while retaining the key information in the traffic data. In this way, the system can process large-scale traffic data more quickly and provide a basis for subsequent business responses. Next, the system introduces a nonlinear mapping function , which decomposes the tensor into the core tensor Mapped to a high-dimensional feature space. The role of nonlinear mapping is to capture the nonlinear characteristics in the data. In the transportation system, the change of data is often not a simple linear relationship. For example, the change of traffic flow is not only affected by road conditions and weather, but also by emergencies or crowd behavior, which often manifests as complex nonlinear patterns. By mapping the tensor decomposition results to a high-dimensional feature space, the system can reveal these complex patterns, thereby better understanding and predicting traffic conditions.

[0063] The introduction of kernel function further enhances the system's ability to measure data similarity. It is used to calculate the similarity between tensor samples. The core idea is to measure the similarity between them by calculating the inner product of data samples in a high-dimensional feature space. Specifically, the kernel function takes the form of a Gaussian kernel, which measures the Euclidean distance between samples and introduces a bandwidth parameter , to control the nonlinearity of the kernel function. The value makes the kernel function more sensitive and can capture small differences between data; larger This makes the kernel function smoother and more suitable for processing the overall trend in the data. In the process of processing traffic data, this flexibility of the kernel function enables the system to adjust its similarity measure according to different traffic scenarios. For example, when analyzing traffic flow during peak hours in the city center, the system can adjust the kernel function to prioritize capturing those small but important traffic fluctuations, thereby making more accurate predictions. The centering of the kernel matrix is ​​to eliminate the mean deviation in the data set and ensure that the similarity measure can truly reflect the relationship between samples. The centering process subtracts the mean value of the sample so that each element in the kernel matrix can be calculated based on uniformly distributed samples, avoiding errors introduced by uneven sample distribution. This step is particularly important because the distribution of traffic data in different time periods or locations often varies greatly. Through centralization, the system can better eliminate this difference, thereby ensuring the stability and accuracy of the similarity calculation. After completing the centering of the kernel matrix, the system performs eigenvalue decomposition on it to extract the most representative features in the data. The result of the eigenvalue decomposition is a set of eigenvalues ​​and their corresponding eigenvectors, which reflect the degree of change of the data in different directions. The system selects the front The eigenvector corresponding to the maximum eigenvalue is used as the final key feature for dimensionality reduction. These feature vectors represent the most important dimensional information in the data. By mapping the original high-dimensional data to these key feature spaces after dimensionality reduction, the system can significantly reduce the dimensionality and complexity of the data without losing important information. This dimensionality reduction process provides a more efficient computing basis for subsequent traffic prediction, analysis and decision-making.

[0064] Embodiment 5: The process of constructing a traffic knowledge graph based on a hypergraph by a traffic knowledge graph construction unit specifically includes: using the dimension reduction key features Constructing node feature matrix ; Define the nodes and edges of the hypergraph and construct the association matrix and the hyperedge weight matrix ; Calculate the node degree matrix and the hyperedge degree matrix ; Construct the hypergraph Laplacian matrix ; Fusion of node features and hypergraph structure to obtain updated node feature matrix .

[0065] Specifically, the node feature matrix The construction is based on the key features of dimensionality reduction , which are the core features extracted from the previous stage through high-order tensor decomposition and dimensionality reduction. In traffic scenarios, these features may represent important information of different traffic entities (such as roads, vehicles, accidents, etc.). By integrating these dimensionality reduction features into the node feature matrix In the hypergraph, the system is able to assign a specific vector representation to each traffic entity, which not only reflects the attributes of the entity, but also captures their mutual relationships with other entities. Next, the system defines the nodes and hyperedges in the hypergraph. Unlike traditional graph structures, the hyperedges of a hypergraph can connect multiple nodes, which means that it can more flexibly represent complex relationships between multiple entities. For example, in a traffic system, a hyperedge can simultaneously connect multiple traffic entities, such as vehicles, roads, and traffic signals, representing the dynamic interactive relationships between these entities. In this way, the system can more accurately reflect the association structure between different entities in the traffic system, especially in scenarios such as changes in traffic flow and the impact of accidents, where these complex relationships are crucial.

[0066] After defining the nodes and hyperedges, the system constructs the incidence matrix and the hyperedge weight matrix . Incidence Matrix It reflects the relationship between each node and its corresponding hyperedge. Specifically, it describes which nodes are connected to each other through certain hyperedges. Each hyperedge is assigned a weight to quantify its importance in the knowledge graph. Different entity relationships in the traffic system have different impacts on the overall traffic conditions. The relationships between some entities may have a significant impact on events such as traffic congestion and accidents, while other entity relationships are less important. By assigning different weights to hyperedges, the system can give priority to those relationships that have a greater impact on the system when analyzing and predicting traffic conditions, thereby improving the accuracy of the system's response. Next, the system calculates the node degree matrix and the hyperedge degree matrix . Node degree matrix Describes the degree of association of each node, that is, the strength of the connection between a traffic entity and other entities. It is used to measure the activity or influence of each hyperedge. By calculating these degree matrices, the system can further understand the structural characteristics in the traffic knowledge graph. Specifically, the calculation of node degree and hyperedge degree helps to identify key nodes and key relationships in the traffic system. For example, some roads may be key nodes in the entire traffic network, and their traffic conditions will have a significant impact on the overall traffic; similarly, some hyperedges may connect multiple important nodes, and their weights will play a key role in the overall prediction and analysis of the system. With these degree matrices, the system constructs the hypergraph Laplace matrix . The hypergraph Laplacian matrix is ​​a core mathematical representation of the hypergraph, which is used to capture the global structural characteristics between nodes in the graph. It reflects the propagation characteristics and information flow patterns of nodes in the hypergraph structure by combining the node degree matrix and the hyperedge degree matrix. In the context of the traffic knowledge graph, the construction of the hypergraph Laplacian matrix means that the system can understand the complex global interaction patterns between traffic entities. For example, through this matrix, the system can identify which nodes have the most frequent information flow, thereby predicting the possibility of traffic congestion or accidents on certain roads. Finally, the system fuses the node features with the hypergraph structure to generate an updated node feature matrix This process is a key step in the construction of the entire knowledge graph. The node feature matrix The initial features in are combined with the global structure information of the hypergraph Laplacian matrix to obtain a more accurate feature representation. The nodes in the traffic knowledge graph are not just isolated traffic entities, they influence each other through hyperedges and global structures. Therefore, the node feature matrix It reflects the relative position and influence of these entities in the overall traffic system. Through this fusion process, the system can process and analyze traffic data more intelligently. For example, the system can , predict traffic flow changes in a specific area, identify potential accident risk points, and provide corresponding solutions for traffic managers.

[0067] Example 6: Using dimensionality reduction key features , construct node feature matrix ;in, Key features for dimensionality reduction No. elements, indicating the The feature vector of nodes; in the constructed traffic knowledge graph, the total number of nodes is ; Node feature matrix , is the number of nodes, is the feature dimension after dimensionality reduction; let Indicates The feature vector of each node; the association matrix Represents the association relationship between nodes and hyperedges:

[0068] ;

[0069] in, is the number of hyperedges; let the eigenvector As the nodes; For the hyperedges; hyperedge weight matrix Represents the weight of each hyperedge, reflecting the strength of the relationship between the nodes connected by the hyperedge. ; Indicates The weight of the hyperedge; the node degree matrix is ; Hyperedge degree matrix ;in, , ; Representation Node The degree of represents the sum of the weights of the hyperedges associated with it; Represents a hyperedge The degree indicates the number of nodes connected by the hyperedge.

[0070] Specifically, dimension reduction features are obtained through the aforementioned tensor decomposition and kernel function processing, which represent the main characteristic information of each node in the traffic system. Here, the total number of nodes is , which is the number of entities in the traffic system, and the characteristic dimension of each node is This means that the matrix The size is , which contains the feature representation of all traffic entities. The feature vector of each node It reflects the traffic characteristics of the node, such as the width and traffic volume of a road, or the driving path and speed of a vehicle. In this way, the system can map different information in multi-source data into a unified feature space, which is convenient for subsequent analysis and calculation. Secondly, the system defines the association matrix , which is used to describe the association relationship between nodes and hyperedges. The introduction of hyperedges allows the system to more flexibly represent complex relationships in the traffic system, especially when there are associations between multiple nodes at the same time. A hyperedge is a structure that can connect multiple nodes, which can well describe complex interactions such as roads, vehicles, and traffic lights. In a traffic system, some nodes (such as important intersections) may be connected to multiple roads, traffic lights, and other traffic entities through the same hyperedge, indicating their association at a certain moment or in a certain scenario. Through the association matrix , the system can map these associations into binary relationships in the matrix: when the node Super edge hour, , otherwise it is 0. This representation is concise and clear, and can effectively capture the structural relationship between traffic entities.

[0071] In addition, the hyperedge weight matrix It is used to reflect the weight of the hyperedge, that is, the strength of the relationship between the nodes connected by each hyperedge. Different hyperedges in the transportation system are not equally important. Some hyperedges may connect key nodes in the transportation network and their weights should be higher, while other hyperedges may only reflect secondary relationships. By introducing the weight matrix , the system can weight the influence of hyperedges, making the analysis process more flexible. Corresponding to the influence degree of each hyperedge, for example, on a main road with heavy traffic, a hyperedge connecting multiple nodes may have a higher weight, thus affecting the mobility and congestion risk of the entire traffic network. Next, the node degree matrix and the hyperedge degree matrix This further improves our understanding of the structure of the traffic knowledge graph. By recording the degree of each node, that is, the node The sum of the weights of the associated hyperedges reflects the importance of each traffic entity in the system. A node with a larger degree usually means that it has more connections or a more important role in the entire network. For example, a node with a larger degree may represent an important transportation hub, and the traffic conditions of the hub will have a significant impact on the flow distribution of the entire system. Similarly, the hyperedge degree matrix The degree of each hyperedge is recorded, which reflects the number of nodes connected by the hyperedge. A hyperedge with a higher degree usually means that it connects more nodes and thus has a greater impact on the overall condition of the transportation system.

[0072] For example, suppose a node represents a transportation hub, and the node is connected to several main roads and secondary roads through a hyperedge. will be higher because it connects other important nodes through multiple hyperedges. It reflects the number and importance of nodes associated with this hyperedge. Through the node degree matrix and hyperedge degree matrix, the system can dynamically identify key nodes and key hyperedges in the traffic system, thereby providing more targeted management suggestions for traffic managers. For example, the system can use this information to predict which traffic hubs may be congested during peak hours and which traffic changes between nodes will have a significant impact on the overall traffic situation. Finally, this hypergraph-based traffic knowledge graph structure provides a highly flexible and intelligent traffic data processing framework by combining node features, hyperedge weights, node degrees and hyperedge degrees. The system can not only accurately describe the relationship between traffic entities through the association matrix and weight matrix, but also identify the most important nodes and hyperedges through the degree matrix, thereby improving the processing efficiency of the system in large-scale traffic scenarios. A significant advantage of this method is that it can structure the complex relationships in multi-source traffic data into a matrix representation that is easy to operate and analyze. This structured representation enables the system to handle a variety of complex traffic scenarios, from traffic flow prediction, congestion warning to accident analysis, and can use these matrices for rapid calculation and response. In addition, the hypergraph-based structure provides great scalability for the system. In the future, more nodes and hyperedges can be introduced according to different traffic needs to further enhance the intelligence level and response capabilities of the system.

[0073] Example 7: Hypergraph Laplacian Matrix Use the following formula to express it:

[0074] ;

[0075] in, ; , is the identity matrix; represents the inverse matrix of the hyperedge degree matrix; represents the inverse square root of the node degree matrix; is the transpose of the hypergraph association matrix; the updated node feature matrix Calculated using the following formula:

[0076] .

[0077] Specifically, the hypergraph Laplacian matrix It is a tool used by this system to capture the global interactions between nodes in the transportation network. In the transportation system, various entities (such as roads, vehicles, traffic signals, etc.) influence each other through a complex network of relationships. The relationship between these entities is not limited to simple bilateral connections, but often a complex many-to-many interaction. This many-to-many association can be described by a hyperedge. Hyperedge is a flexible structure that can connect multiple nodes at the same time, which makes it particularly suitable for dealing with complex scenarios such as transportation hubs and intersections. In this scenario, multiple roads, traffic lights, vehicles and other entities may be connected together through hyperedges at the same time, thereby jointly affecting the distribution and dynamic changes of traffic flow. The hypergraph Laplacian matrix globally models the transportation network by combining these complex multivariate relationships. In the formula, the node degree matrix and the hyperedge degree matrix Represents the degree of each node and the degree of the hyperedge respectively. The node degree reflects the importance or influence of each traffic entity in the system. For example, a node with a higher degree may be an important traffic hub, which is connected to multiple roads and traffic lights, and therefore occupies a key position in the entire traffic network. The hyperedge degree is used to describe the number of nodes connected by a hyperedge and the strength of the association between them. Through the node degree matrix and the hyperedge degree matrix, the system can identify the most important nodes and hyperedges in the traffic network and give them higher weights in subsequent analysis. This is crucial for dealing with large-scale traffic systems, especially when responding to emergencies such as traffic accidents and peak congestion. The system needs to be able to quickly identify which nodes and relationships are the most critical, so as to respond efficiently.

[0078] In the process of constructing the hypergraph Laplacian matrix, the incidence matrix The role of cannot be ignored. It represents the relationship between nodes and hyperedges and is the basis of the entire hypergraph structure. In the traffic knowledge graph, different nodes may be connected to each other through the same hyperedge. This relationship may reflect the interdependence of traffic flow or a common traffic management mechanism. For example, at an important intersection, multiple roads control traffic flow through the same traffic light, and the state of this traffic light will affect the traffic conditions of all these roads at the same time. In this case, The relationship between nodes and hyperedges can be represented in a binary way to help the system clearly understand which nodes are connected to each other through hyperedges. and They are the inverse operations of the node degree matrix and the hyperedge degree matrix, respectively. They are introduced to balance the influence of different nodes and hyperedges. In a complex transportation system, different nodes and hyperedges have unequal influence on the overall network. Some nodes (such as major roads or transportation hubs) have greater influence because of their higher connectivity, while other nodes may be on the edge and have less influence. Similarly, some hyperedges are more effective because they connect more key nodes. In order to ensure that the entire system can treat different nodes and hyperedges fairly during calculations, and Through the standardization operation, these differences are reasonably adjusted so that the final Laplace matrix can accurately reflect the global structural characteristics of the transportation network. The core function of is to describe the propagation pattern of information in the entire network. For the transportation system, the propagation of information can be understood as the dynamic changes of traffic flow. For example, a traffic accident on a road may affect the traffic conditions of multiple surrounding roads. The system can simulate the propagation path of this information and identify the interactive relationship between certain key nodes. This is of great significance for rapid response and early warning. For example, after an accident occurs on a main road, the system can predict the impact of the accident on surrounding roads through the hypergraph Laplace matrix and provide traffic managers with actionable intervention measures, such as adjusting traffic lights or directing vehicles to detour, to reduce the impact of the accident on the overall traffic flow.

[0079] Next, the node feature matrix The updating process makes full use of the structural information of the Laplacian matrix. The initial node feature matrix It reflects the initial characteristics of each node, such as the width of the road, the speed of the vehicle, the status of the traffic light, etc. Although this information helps the system understand the independent attributes of each node, it does not capture the dynamic interaction between nodes in the entire network. , the system combines the Laplace matrix Update the node features to ensure that the features of each node not only reflect its own attributes, but also its relative position and role in the entire transportation network. Specifically, the Laplace matrix describes the global interdependencies between nodes, so in During the update process, the features of each node are adjusted according to the features of other nodes associated with it. For example, if a node is connected to multiple nodes with higher degrees, its features will become more important. The system can generate a more accurate node representation by updating the feature, especially in a dynamic traffic environment. This feature update can help the system capture changes in the traffic network in real time. For example, the traffic status of a road may be affected by other roads or traffic lights connected to it. The system can dynamically adjust the characteristics of the road to reflect its changes in the current traffic network. This process is crucial for the rapid response of the transportation industry, especially when responding to emergencies or traffic rush hours. The system needs to quickly adjust node characteristics according to actual conditions in order to provide the most timely and accurate decision support for traffic managers.

[0080] Example 8: The specific execution process of the business impact and processing unit includes: using a pre-trained large language model to convert the user request Encoded into high-dimensional semantic vector ; Calculated using cosine similarity With each node feature vector Similarity ; for The Elements:

[0081] ;

[0082] in, For user requests and Similarity of nodes; Based on similarity , select the top nodes, forming a candidate node set ; For the candidate node set , build a probabilistic graph model, and reason about highly relevant nodes, including: extracting candidate nodes and their directly related neighbor nodes and edges from the traffic knowledge graph to form a subgraph ; Use Markov random fields or conditional random fields to Modeling; Calculate the posterior probability of each node in the subgraph through the message passing algorithm .

[0083] Specifically, the first step of the process is to send the user request Encoded into high-dimensional semantic vector At this stage, the system uses the pre-trained large language model to perform semantic analysis and representation of the user request. The large language model has powerful natural language processing capabilities and can extract deep semantic information from the text of the user request. Through this semantic encoding process, the user's request is converted into a vector , this vector not only contains the literal meaning of the user's request, but also captures deeper context and semantic relationships. For example, if a user asks "What is the current traffic condition of a certain road?", the model can not only understand that this question is related to road traffic, but also capture information such as time and specific location, thereby providing a semantically rich vector representation for subsequent processing. Next, the system calculates the semantic vector of the user's request and the feature vector of each node in the traffic knowledge graph The similarity between Here, cosine similarity is used as a measurement tool because it can measure the directional similarity of two vectors in a high-dimensional space and is particularly suitable for comparing the similarity between semantic vectors. Formula Indicates the user request and The similarity of nodes In this process, each node feature vector Represents the characteristics of a specific traffic entity in the traffic knowledge graph, such as the traffic flow of a road, the status of a traffic light, etc. and The system can determine which nodes are most relevant to the user request by measuring the similarity between them. The value range of cosine similarity is arrive , the closer the similarity is , which means the higher the semantic relevance between the user request and the node. , the system then selects the top nodes, forming a candidate node set This candidate node set is the basis for subsequent reasoning of the system, and it contains those traffic entities that are most likely to be related to the user request. By limiting the number of candidate nodes , the system can focus resources on the most relevant nodes and avoid processing a large number of irrelevant nodes in the entire knowledge graph, thereby improving efficiency. For example, if a user asks about the traffic conditions of a specific road, the system can add nodes related to the road (such as the road itself, adjacent traffic lights, and other surrounding roads) to the candidate node set through semantic matching, without having to process the entire city's traffic network.

[0084] After getting the candidate node set After that, the system starts to build a probabilistic graph model to reason about these highly correlated nodes. The key to this step is to extract neighbor nodes and edges directly associated with the candidate nodes from the traffic knowledge graph to form a subgraph. This subgraph includes not only the candidate nodes, but also the neighboring nodes related to them and the relationships between these nodes. For example, a user may ask about the traffic conditions of a certain road. The system will not only find the nodes of the road, but also include the entities such as the roads, traffic lights, and vehicles adjacent to it into the subgraph to ensure the comprehensiveness and accuracy of the analysis. Next, the system uses Markov random field MRF or conditional random field CRF to analyze the subgraph. Modeling. MRF and CRF are commonly used probabilistic graph models. They establish a global probability distribution by defining the dependencies between nodes. In the traffic system, the interactions and dependencies between different nodes are often non-independent. For example, the vehicle flow on the road depends not only on the road itself, but also on factors such as the congestion of adjacent roads and changes in traffic signals. MRF and CRF capture these complex dependencies between nodes by defining conditional probabilities, and can more accurately reflect the dynamic changes in the traffic system. Finally, the system calculates the posterior probability of each node in the subgraph through the message passing algorithm. The core idea of ​​the message passing algorithm is to gradually update the state of each node through information exchange between neighboring nodes until the system converges to a stable probability distribution. In the context of traffic knowledge graphs, the posterior probability Indicates that given a user request In the case of The probability of a state or event occurring. For example, if a user asks about the traffic flow of a certain road, the system calculates the posterior probabilities of all nodes related to the road through a message passing algorithm, thereby inferring which road is most likely to be congested or which traffic light may need to be adjusted. This reasoning method based on probabilistic graphical models enables the system to make more accurate predictions and responses when faced with complex traffic systems.

[0085] Example 9: Use the following formula to calculate the posterior probability of each node in the subgraph through the message passing algorithm:

[0086] ;

[0087] in, For Node The potential function reflects and The compatibility is taken as similarity ; For edge The potential function represents the node and The strength of the association between is the normalization factor; For a given user request, the node The posterior probability of .

[0088] Specifically, the structure of the formula is based on the message passing algorithm of the probabilistic graph model, which is used to calculate the posterior probability of each node in the subgraph Specifically, the posterior probability Indicates that given a user request In the case of The state probability of a node in the traffic knowledge graph represents a specific traffic entity, such as a road, a traffic signal, or a vehicle, and the user requests is a traffic-related question raised by the user, such as "Is this road congested?" Through this formula, the system can infer the status of related nodes based on the user's request, such as the traffic flow of a certain road, the changing trend of a certain traffic light, etc. The potential function in the formula It is one of the core components for calculating the posterior probability. The role of the potential function is to reflect the node With user request compatibility, which measures the similarity between user requests and node features. Each has its own eigenvector , these feature vectors may include road width, traffic flow, vehicle speed, etc. In the previous step, the cosine similarity was calculated to obtain It is used to measure user requests With node features The degree of match between them. Potential function The value of is the similarity , which represents the node Given a user request, the degree of match with the user's query content. A higher degree of match means that the node is more likely to be related to the traffic conditions that the user is concerned about, so its posterior probability will also be greater.

[0089] Potential Function is another key element in the formula, which is used to describe the node Its neighbor nodes The strength of the association between nodes. In a traffic system, nodes do not exist in isolation, but form a complex network of relationships through roads, traffic signals, etc. For example, the traffic flow of a road is not only affected by its own situation, but also by the flow of other roads connected to it and changes in traffic lights. Reflecting the interdependence between nodes, it measures the With Node The greater the strength of the association, the more likely the states of the two nodes are to affect each other. For example, a main road and its connected branches may have a strong correlation, so their potential function values ​​are high, indicating that the state change of one node is likely to affect the state of another node. By combining the potential function and the potential function, the system can consider both the individual characteristics of the nodes and the relationship between the nodes. Specifically, in the formula The part is the product of the potential function, which shows that the node The state of a node depends not only on its own characteristics, but also on the states of all neighboring nodes connected to it. In this way, the system can capture the complex global dependency structure in the traffic knowledge graph. For example, if a user asks about the traffic conditions of a certain road, the system will not only consider the road's own characteristics (such as road width, traffic volume, etc.), but also analyze the traffic conditions of other roads connected to it in order to make a more comprehensive inference. This relationship model between nodes and neighboring nodes is particularly suitable for dealing with complex traffic systems, because in actual traffic networks, the mutual influence between nodes is often nonlinear and full of dynamic changes. The normalization factor in the formula The sum of the posterior probabilities of all nodes is ensured to be 1, thus forming a valid probability distribution. The purpose of normalization is to prevent the probability value from being too large or too small and to maintain the stability and rationality of the results. In the reasoning process of the traffic knowledge graph, the values ​​of the potential function and potential function of multiple nodes may be significantly different. With the introduction of , the system can standardize the posterior probabilities of all nodes so that these probabilities can be compared within the same range, thereby determining which node states best meet the user's query requirements. The message passing algorithm is the main reasoning tool behind this formula. The core idea of ​​the algorithm is to gradually update the posterior probability of each node through information exchange between nodes until it converges to a stable probability distribution. For traffic systems, the message passing algorithm can effectively capture the mutual influence between different nodes. For example, when a user inquires about the congestion of a main road, the system receives information from adjacent roads, traffic lights and other nodes through the message passing algorithm, and gradually updates the posterior probability of the main road, thereby inferring the most likely traffic conditions. Through this iterative reasoning process, the system is able to make intelligent decisions in complex traffic networks. Finally, the system calculates the posterior probability of each node. , and uses these probabilities to respond to user requests. For example, if a user asks about the traffic conditions at a certain intersection, the system generates a comprehensive traffic condition report by calculating the posterior probabilities of all nodes related to the intersection (such as connected roads, traffic lights, etc.). This probability-based reasoning method can handle complex, multi-dimensional traffic information, ensuring that the system can make more accurate and intelligent responses when dealing with emergencies or complex traffic situations.

[0090] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A rapid response processing system for transportation business based on large language model technology, characterized in that: The system includes: a multi-source traffic data fusion unit, which is used to integrate multi-source traffic data from different data sources, adopt tensor fusion technology to express the fusion of multi-source traffic data as a high-order tensor, extract features from the fused high-order tensor, and adopt a method combining tensor decomposition and kernel principal component analysis to extract key features and reduce the dimension to obtain reduced-dimensional key features; a traffic knowledge graph construction unit, which is used to construct a traffic knowledge graph based on a hypergraph based on the reduced-dimensional key features; a business impact and processing unit, which is used to receive user requests, encode user requests into high-dimensional semantic vectors using large language model technology, and calculate the user request. The method uses the cross-modal semantic similarity between user requests and nodes and relationships in the traffic knowledge graph, finds the most relevant nodes, builds a probabilistic graph model, and infers the most relevant nodes to obtain the posterior probability of each node in the traffic knowledge graph under a given user request, and sends the data corresponding to the node for the user request based on the posterior probability; the multi-source traffic data includes: real-time traffic flow, vehicle GPS trajectory, road infrastructure data and weather information; the road infrastructure data includes: road length, road width, and road type; the weather information includes: temperature, precipitation, wind speed and wind direction; high-order tensors Calculated by the following formula: ； in, is the number of data sources; is the subscript index; For the preset The weight coefficient of each data source; For the Data source The feature vector of dimension Data source Features is a tensor product operator, which indicates the fusion operation of different data sources; is the time embedding matrix; spatial embedding matrix; is the modulus of the tensor Product operation; For the feature dimensions of data sources; time embedding matrix ;in, represents all time steps in a time series; Represents the time step The time embedding vector of The temporal embedding vector Calculated by the following formula: ； in, For time The embedding vector of is the current time step; is the dimension of time embedding, which determines the length of the time embedding vector; is the subscript index; the tensor decomposition of the high-order tensor is performed through the following formula to obtain the tensor decomposition result : ； in, It is Factor matrix on the module; define a nonlinear mapping function , decompose the tensor into Mapping to high-dimensional feature space : ; Use kernel function To calculate the similarity between tensor samples, The result of tensor decomposition Middle elements, The result of tensor decomposition Middle elements; is the modulus of the tensor Product operation; is the dimension of the high-order tensor; let the kernel matrix Middle Row, No. Elements of a column ;in, is the bandwidth parameter of the kernel function, which determines the nonlinearity of the kernel function. The kernel matrix is ​​given by the following formula: Centralization: ； in, is the centralized kernel matrix; Is the size of The identity matrix of is the dimension of the kernel matrix; for the centralized kernel matrix Perform eigenvalue decomposition and obtain the eigenvalue and the corresponding eigenvector : ； Before selection The eigenvector corresponding to the maximum eigenvalue is used as the key feature for dimensionality reduction .

2. The transportation business rapid response processing system based on large language model technology as claimed in claim 1 is characterized in that: The process of constructing a traffic knowledge graph based on a hypergraph by the traffic knowledge graph construction unit specifically includes: using the key features of dimensionality reduction Constructing node feature matrix ; Define the nodes and edges of the hypergraph and construct the association matrix and the hyperedge weight matrix ; Calculate the node degree matrix and the hyperedge degree matrix ; Construct the hypergraph Laplacian matrix ; Fusion of node features and hypergraph structure to obtain updated node feature matrix .

3. The transportation business rapid response processing system based on large language model technology as claimed in claim 2 is characterized in that: Using key features of dimensionality reduction , construct node feature matrix ;in, Key features for dimensionality reduction No. elements, indicating the The feature vector of nodes; in the constructed traffic knowledge graph, the total number of nodes is ; Node feature matrix , is the number of nodes, is the feature dimension after dimensionality reduction; let Indicates The feature vector of each node; the association matrix Represents the association relationship between nodes and hyperedges: ； in, is the number of hyperedges; let the eigenvector As the nodes; For the hyperedges; hyperedge weight matrix Represents the weight of each hyperedge, reflecting the strength of the relationship between the nodes connected by the hyperedge. ; Indicates The weight of the hyperedge; the node degree matrix is ; Hyperedge degree matrix ;in, , ; Representation Node The degree of represents the sum of the weights of the hyperedges associated with it; Represents a hyperedge The degree indicates the number of nodes connected by the hyperedge.

4. The transportation business rapid response processing system based on large language model technology as claimed in claim 3 is characterized in that: Hypergraph Laplacian Matrix Use the following formula to express it: ； in, ; , is the identity matrix; represents the inverse matrix of the hyperedge degree matrix; represents the inverse square root of the node degree matrix; is the transpose of the hypergraph association matrix; the updated node feature matrix Calculated using the following formula: 。 5. The transportation business rapid response processing system based on large language model technology as claimed in claim 4 is characterized in that: The specific execution process of the business impact and processing unit includes: using the pre-trained large language model to convert user requests Encoded into high-dimensional semantic vector ; Calculated using cosine similarity With each node feature vector Similarity ; for The Elements: ； in, For user requests and Similarity of nodes; Based on similarity , select the top nodes, forming a candidate node set ; For the candidate node set , build a probabilistic graph model, and reason about highly relevant nodes, including: extracting candidate nodes and their directly related neighbor nodes and edges from the traffic knowledge graph to form a subgraph ; Use Markov random fields or conditional random fields to Modeling; Calculate the posterior probability of each node in the subgraph through the message passing algorithm .

6. The transportation business rapid response processing system based on large language model technology as claimed in claim 5 is characterized in that: Use the following formula to calculate the posterior probability of each node in the subgraph through the message passing algorithm: ； in, For Node The potential function reflects and The compatibility is taken as similarity ; For edge The potential function represents the node and The strength of the association between is the normalization factor; For a given user request, the node The posterior probability of .

Images