Decision support method for time-space dynamic data of Internet of Vehicles based on knowledge retrieval enhancement

Abstract

The invention discloses a decision support method for time-space dynamic data of Internet of Vehicles based on knowledge retrieval enhancement, which comprises the following steps: collecting perception data inside and outside a vehicle, and separating general data and private data; preprocessing the data, and compressing and encrypting the private data; performing data processing by adopting a data knowledge dual-drive mode and a large language model, and converting the data into structured data; constructing a super-relation knowledge graph, and carrying out vectorization processing based on an embedded vector model; performing node importance evaluation based on the super-relation knowledge graph, and performing data updating and increment; storing the private data locally, uploading the general data to a public database of the cloud, and defining a dynamic access range of the public database; and extracting a user questioning entity by using the local entity recognition model, matching data existing in the local and the cloud, obtaining background knowledge contents, and generating an enhanced retrieval result. According to the method, the problems of acquisition of decision support information and use safety of data in an Internet of Vehicles environment are solved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement. Background Art

[0002] In the digital age, Internet of Vehicles technology connects vehicles to the Internet, enabling information exchange between vehicles and between vehicles and infrastructure. This not only improves driving safety and efficiency, but also brings challenges to data security and privacy protection. In Internet of Vehicles technology, data security issues are particularly prominent. During driving, vehicles generate and exchange a large amount of real-time data, including vehicle status information, environmental perception data, and user private data. The privacy of these data requires us to take effective protection measures to ensure that users' privacy and data security are fully protected while enjoying intelligent decision support.

[0003] In the graph-based retrieval enhancement system, it is also a challenge to achieve real-time update of vehicle information. Most of the existing retrieval enhancement systems do not support the incremental generation of knowledge based on time series, which means that vehicle data cannot be effectively processed and updated. In order to meet this challenge, so that hyper-relational knowledge can be continuously and dynamically updated to ensure the accuracy and effectiveness of vehicle information in the database, it is necessary to propose a decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement. Summary of the invention

[0004] In view of this, the purpose of the present invention is to propose a decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement.

[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is: The present invention provides a decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement, comprising the following steps: Step 1: Collect the vehicle's internal and external sensing data, and separate the general data and privacy data from the vehicle's internal and external sensing data; Step 2: pre-process the general data and private data, compress and encrypt the private data, and mark and classify the general data; Step 3: The processed general data and privacy data are processed based on rule documents using a data knowledge dual-driven model and a large language model to convert the vehicle's internal and external perception data into structured data; Step 4: construct a hyper-relational knowledge graph based on structured data, and vectorize the index of information in the hyper-relational knowledge graph based on the embedded vector model; Step 5: Evaluate the importance of nodes based on the hyper-relational knowledge graph, and update and incrementally expand the data of the hyper-relational knowledge graph; Step 6: Save the private data locally and upload the general data to the public database in the cloud, and define the dynamic access scope of the public database based on the hyper-relational knowledge graph; Step 7: Use the local entity recognition model to extract the user question entities, match the local and cloud data, obtain background knowledge content, and generate enhanced retrieval results.

[0006] Furthermore, the step 1 specifically includes: Step 11: Collecting vehicle internal and external perception data, wherein the vehicle internal and external perception data includes vehicle environment information, vehicle status information, traffic flow information, Internet of Vehicles service platform information, vehicle location information, vehicle location information, Internet of Vehicles user community group information, parking lot layout information, user personal information, vehicle driving trajectory information, and vehicle audio interaction information; Step 12: converting the vehicle internal and external perception data into text data for recognition by a pre-trained language model, and saving the data in a structured format; Step 13: Use a pre-trained language model based on deep learning for pre-training. For the privacy data types in the Internet of Vehicles, freeze some parameters of the pre-trained language model by freezing the large model and repairing the small model. Only fine-tune the key layers to achieve the function of data separation and desensitization of the pre-trained language model. Step 14: Use the fine-tuned pre-trained language model to identify the vehicle's internal and external perception data, and separate general data and privacy data. The general data includes the vehicle's environmental information, vehicle status information, traffic flow information, Internet of Vehicles service platform information, vehicle location information, vehicle location information, Internet of Vehicles user community group information, and parking lot layout information; the privacy data includes user personal information, vehicle driving trajectory information, and in-vehicle audio interaction information.

[0007] Furthermore, the step 2 specifically includes: Step 21: Clean the general data and private data, and process missing values ​​and outliers; Step 22: construct a semantic knowledge base based on the privacy data, analyze the semantic meaning of the privacy data in the semantic knowledge base, and construct a Huffman tree based on the semantic features of different users and scenarios and the frequency of occurrence of characters, assign a code to each character for compression, and save the compressed privacy data and the code table; Step 23: Add noise to the private data through differential privacy technology to cover up the true value of the original data; Step 24: Use the AES encryption algorithm to encrypt the private data, and retain the encryption mark of the original text category at the encryption location; Step 25: Mark and classify common data to match local and cloud data interactions in different scenarios.

[0008] Furthermore, the step 3 specifically includes: Step 31, in the data-driven part, the vehicle's internal and external perception data are used to train and optimize the large language model, so that the large language model can automatically identify and extract entities and relationships in the vehicle's internal and external perception data; in the knowledge-driven part, the expert's knowledge is introduced to set up rule documents, including making rule documents, restricting the output structure of the large language model, providing entity and relationship extraction cases, relationship importance rating evaluation cases, and category attribute discrimination evaluation cases to guide the data processing and relationship evaluation of the large language model, and then combining data-driven and knowledge-driven to form a data-knowledge dual-driven model; Step 32: Select a large language model, deploy the large language model in the cloud, and the local vehicle uses an API to call the large language model from the cloud; or deploy the large language model in the Internet of Vehicles block, and use a private intranet to call the large language model through port forwarding; Step 33: Use the data-knowledge dual-drive model and the large language model to evaluate the vehicle's internal and external perception data. Learn patterns and rules from the actual vehicle's internal and external perception data, provide rule documents to limit the output content and format of the model interaction, formulate a document segmentation strategy, and use prior knowledge to guide the output of the large language model. Step 34, determining the block size according to the amount of information of the vehicle interior and exterior sensing data, using a small block for the vehicle interior and exterior sensing data whose amount of information is greater than a preset value, and using a large block for the vehicle interior and exterior sensing data whose amount of information is not greater than the preset value; Step 35: Obtain entity information by interacting with the large language model based on rule documents through block data, extract the entity information required to build a hyper-relational knowledge graph, and convert it into structured data.

[0009] Furthermore, it is characterized in that the step 35 specifically comprises: Step 351: When using the large language model for the first interaction, record the data block location of the information source; Step 352: extract knowledge according to the knowledge density and block size of the data, and complete the first large language model interaction; Step 353: After the first interaction with the large language model, the extracted data is transmitted to the large language model for secondary judgment, the relationship between the nodes and edges in the hyper-relational knowledge graph is evaluated, and the rating result is outputted in terms of relationship strength rating; Step 354: Make a third judgment on the data whose edge description relationship rating in the relationship strength rating has changed, extract the data block location of the information source recorded during the first interaction, interact with the large language model again, obtain entity information, and convert it into structured data, wherein the structured data includes entities, relationships and their attribute information.

[0010] Furthermore, the step 4 specifically includes: Step 41: construct a hyper-relational knowledge graph based on structured data, and represent entities and relationships in the structured data as nodes and edges in the hyper-relational knowledge graph, respectively. Each node contains entity information, and the edge represents the relationship between entities. The hyper-relation consists of a main triple and several auxiliary attribute value descriptions. The main triple is represented as (subject, relationship, object), and the auxiliary attribute value description is represented in the form of key-value pairs. The hyper-relational knowledge graph is represented as:

[0011] in, is a set of nodes, each node Contains attributes , Represents the node key value category, which stores multiple pairs of key values ​​in list form, where: Include Key and value ; is a set of edges, defined as , that is, the edge is a set of nodes An ordered pair of two nodes in Connecting Nodes and nodes ,side Contains attributes and properties , It describes the relationship between edges. Described are relationship strength ratings; Step 42: Add description information to each node and edge, and use key-value pairs to save the supplementary information of the node; Step 43: For nodes with multiple relationships, support multi-edge description, mark and distinguish the information of the same node and edge, evaluate the importance of different information, and mark the importance of the information; Step 44: Capture context information through the hyper-relational knowledge graph, enhance the context understanding ability of the hyper-relational knowledge graph, identify and integrate the multi-dimensional relationships between entities, and the attribute information of entities and relationships, and finally obtain a hyper-relational knowledge graph with rich semantics; Step 45: Use the embedded vector model to perform embedded vector encoding on the keys, nodes and relationships in the hyper-relational knowledge graph, and write the vector information into the lists of each key, node and relationship to complete the vectorization of the index.

[0012] Furthermore, the step 5 specifically includes: Step 51: Use EasyGraph technology to analyze the hyper-relational knowledge graph; Step 52: Use a temporal network node importance evaluation method based on inter-layer neighborhood information entropy to determine the importance of each node and edge, and write it into In the process, the user's weight tendency is combined to evaluate the importance of the final information matching selection; Step 53: incrementally expand the constructed hyper-relational knowledge graph, and perform another round of data processing based on rule documents to form the original hyper-relational knowledge graph and the new hyper-relational knowledge graph, with the original hyper-relational knowledge graph serving as the ontology graph and the new hyper-relational knowledge graph serving as the incremental graph; Step 54: define the scope of entity alignment according to the temporal network node importance evaluation method, set the block size according to the scale of the ontology graph, and divide the ontology graph into blocks according to the scope of entity alignment and the block size to form multiple node blocks; Step 55: Taking the important nodes identified in as the center, match the node blocks of the ontology graph with the nodes of the incremental graph to obtain the node blocks with the highest similarity; Step 56: In the structured data, use the large language model to , and Intelligent fusion of node information, and Perform knowledge fusion of the same entity, expand the parts that do not match the incremental graph and the ontology graph as new structured data to the ontology graph, and complete the incremental generation of the extended content in the ontology graph.

[0013] Furthermore, the timing network node importance evaluation method of step 52 specifically includes: Step 521: According to the neighbor topology overlap coefficient Compute the inter-layer similarity of the edges that appear consistently between a node and its neighbors:

[0014] in, Indicates nodes, Indicates nodes, Indicates time slice, Indicated in The adjacency matrix of time slices Midpoint and nodes The edge between Step 522: Construct a temporal super-adjacency matrix based on the impact between different time slices :

[0015] in, Shidi The adjacency matrix of the time slices is is a positive integer, s represents the sth time slice, s is a positive integer and s <t, represents the sth time slice and the The influence matrix between time slices, yes The temporal super-adjacency matrix is ​​used to describe the relationship or influence of each node between the sth time slice and the tth time slice; Step 523: Based on the constructed temporal super-adjacency matrix The eigenvector corresponding to the maximum eigenvalue, that is, the eigenvector centrality of each node in each time slice, is calculated to determine the importance of the node in the graph network.

[0016] Furthermore, the step 6 specifically includes: Step 61: Save the private data as a local hyper-relationship graph file locally, upload the common data to a public database in the cloud, and support multi-vehicle joint access; Step 62: By marking each semantic record with a permission scope based on a rule document, different scopes of semantic library sets are opened for different permissions: for neighbor nodes in a similar geographical range, traffic geographic information on the local map in the shared semantics is opened to the public; for neighbor nodes in the same unit group relationship, the geographical tags, exclusive parking spaces and historical trajectories of neighbor vehicles in the shared unit campus are opened to support multi-vehicle shared access; Step 63: By using the constructed hyper-relational knowledge graph, the clustering relationship between vehicles is identified based on various key-value pairs of the node key-value category, and vehicles with similar characteristics or behaviors are classified into the same group, so that vehicles can access each other without authorization within the scope of privacy security, and the data accessed are all valid data that is strongly related to each other in time and space; Step 64 : The administrator loads and updates the cloud data in conjunction with the infrastructure within the user group to achieve sharing and retrieval of common data.

[0017] Furthermore, the step 7 specifically includes: Step 71: extracting entities and expressions of the question based on the local entity recognition model; Step 72: vectorize the entity and expression of the problem through an embedding vector model, and convert the text information into vector information; Step 73: Identify the clustering relationship between vehicles through the clustering information in the hyper-relational knowledge graph, and dynamically determine the access scope; Step 74: perform similarity matching between the vector information and the local hyper-relationship graph file and the index vector in the public database in the cloud within the authority to obtain background knowledge content; Step 75: Combine the matched background knowledge content with the user's question content to form prompt content and interact with the large language model; Step 76: Return the response after the interaction of the large language model to the local machine, and reversely decrypt it based on the encryption process in step 2 to obtain the final enhanced retrieval result.

[0018] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: The present invention uses a pre-trained language model based on deep learning for pre-training to capture the complexity, diversity and spatiotemporal correlation of data in the Internet of Vehicles environment. For privacy data types, the model is fine-tuned using annotated data sets to improve the recognition and extraction performance of privacy data. The privacy data and general data processed initially are protected, including missing value processing and noise data processing, to ensure the security and accuracy of the data, build a semantic knowledge base, and achieve data compression and privacy protection. A large language model that combines data processing rules and data knowledge dual-drive mode is used to convert vehicle data into structured data and build a hyper-relational knowledge graph, which contains attribute information of nodes and edges. A BERT-based embedded vector model is constructed to encode vehicle knowledge and general knowledge, and the spatiotemporal correlation between neighbor nodes in the Internet of Vehicles is mined using the hyper-relational knowledge graph. The temporal network node importance evaluation method is introduced to determine the relationship strength between nodes, and knowledge increment generation is completed. The general data is uploaded to a public database in the cloud to achieve multi-vehicle common access and data sharing. Through this series of technical steps, an intelligent system that can protect vehicle and user privacy and efficiently process and utilize Internet of Vehicles data is constructed.

[0019] In summary, the decision support method for spatiotemporal dynamic data of the Internet of Vehicles based on knowledge retrieval enhancement proposed in the present invention combines the dual-driven and incremental retrieval enhancement generation of data and knowledge with the knowledge graph, semantic knowledge base, hyper-relationship, node importance and dynamic attention mechanism of data with the needs of the Internet of Vehicles. It not only effectively solves the problems of real-time and efficient data processing and security in the Internet of Vehicles, but also provides new ideas and technical means for users to make intelligent decisions, which is of great significance for promoting the intelligence and security of the Internet of Vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 It is an execution flow chart of a decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement provided by an embodiment of the present invention.

[0022] Figure 2 It is a schematic diagram of the knowledge graph provided by an embodiment of the present invention.

[0023] Figure 3 It is a structural framework provided by an embodiment of the present invention.

[0024] Figure 4 It is a schematic diagram of enhanced retrieval provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The present invention will be further described in detail below in conjunction with the accompanying drawings and examples. It is particularly noted that the following examples are only used to illustrate the present invention, but are not intended to limit the scope of the present invention. Similarly, the following examples are only partial embodiments of the present invention rather than all embodiments, and all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0026] See also Figure 1-Figure 4 , a decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement of the present invention comprises the following steps: Step 1: Collect the vehicle's internal and external sensing data, and separate the general data and privacy data from the vehicle's internal and external sensing data; In this embodiment, the step 1 specifically includes: Step 11: Collecting vehicle internal and external perception data, wherein the vehicle internal and external perception data includes vehicle environment information, vehicle status information, traffic flow information, Internet of Vehicles service platform information, vehicle location information, vehicle location information, Internet of Vehicles user community group information, parking lot layout information, user personal information, vehicle driving trajectory information, and vehicle audio interaction information; Step 12: converting the vehicle internal and external perception data into text data for recognition by a pre-trained language model, and saving the data in a structured format; Step 13: Use a pre-trained language model based on deep learning for pre-training to learn the complexity, diversity and spatiotemporal correlation of data in the spatiotemporal dynamic complex environment of the Internet of Vehicles. For the type of privacy data in the Internet of Vehicles, freeze some parameters of the pre-trained language model by freezing the large model and repairing the small model, and only fine-tune the key layers to reduce the amount of calculation while maintaining the model performance, so that the pre-trained language model has the function of data separation and desensitization; use the labeled data set to fine-tune the model to improve the model's recognition and extraction performance in privacy data; Step 14: Use the fine-tuned pre-trained language model to identify the vehicle's internal and external perception data, and separate general data and privacy data. The general data includes the vehicle's environmental information, vehicle status information, traffic flow information, Internet of Vehicles service platform information, vehicle location information, vehicle location information, Internet of Vehicles user community group information, and parking lot layout information; the privacy data includes user personal information, vehicle driving trajectory information, and in-vehicle audio interaction information.

[0027] Step 2: pre-process the general data and private data, compress and encrypt the private data, and mark and classify the general data; In this embodiment, step 2 specifically includes: Step 21: Clean the general data and private data, process missing values ​​and outliers to improve data quality and ensure data security and accuracy; Data cleaning: Remove stop words, spell correction, correct spelling errors, ensure the accuracy of the text, remove special characters and numbers, clean up non-text elements and unnecessary numbers in the text, stemming and lemmatization.

[0028] The missing value processing method is majority interpolation. For categorical labels or phrases, if there are missing values, majority interpolation can be used. For example, if the labels in a data set are "A", "B", and "C", and a label is missing, use the label with the highest frequency to fill it.

[0029] Outlier detection method use Law. ,in, is the first quartile, is the third quartile. If the data point is below or higher , it is considered an outlier.

[0030] Step 22: Build a semantic knowledge base based on the private data, analyze the semantic meaning of the private data in the semantic knowledge base, and build a Huffman tree based on the semantic features of different users and scenarios and the frequency of occurrence of characters, assign a code to each character for compression, and save the compressed private data and the code table for use in decompression; Step 23: Add noise to the private data through differential privacy technology to cover up the true value of the original data; in this way, even if the data is leaked, the original data cannot be directly identified, and the original nature of the data is protected from being directly disclosed, ensuring the security of data transmission and storage, and realizing a personalized privacy protection strategy.

[0031] Step 24: Use the AES encryption algorithm to encrypt the private data and retain the encryption mark of the original text category at the encryption location to ensure the security and privacy of the data during sharing and use; Encryption algorithm (AES): in, is ciphertext, Is to use the key The encryption function of It is plain text. Encrypt and clean the data for subsequent local and external encrypted indexing to prevent the leakage of private data.

[0032] Step 25: Mark and classify common data to match local and cloud data interactions in different scenarios.

[0033] Step 3: The processed general data and privacy data are processed based on rule documents using a data knowledge dual-driven model and a large language model to convert the vehicle's internal and external perception data into structured data; In this embodiment, step 3 specifically includes: Step 31, in the data-driven part, the vehicle's internal and external perception data are used to train and optimize the large language model, so that the large language model can automatically identify and extract entities and relationships in the vehicle's internal and external perception data; in the knowledge-driven part, the expert's knowledge is introduced to set up rule documents, including making rule documents, restricting the output structure of the large language model, providing entity and relationship extraction cases, relationship importance rating evaluation cases, and category attribute discrimination evaluation cases to guide the data processing and relationship evaluation of the large language model, and then combining data-driven and knowledge-driven to form a data-knowledge dual-driven model; Step 32: Select a large language model, deploy the large language model in the cloud, and the local vehicle uses an API to call the large language model from the cloud; or deploy the large language model in the Internet of Vehicles block, and use a private intranet to call the large language model through port forwarding; Step 33: Use the data-knowledge dual-drive model and the large language model to evaluate the vehicle's internal and external perception data. Learn patterns and rules from the actual vehicle's internal and external perception data, provide rule documents to limit the output content and format of the model interaction, formulate a document segmentation strategy, and use prior knowledge to guide the output of the large language model. Step 34: Determine the block size according to the amount of information of the vehicle's internal and external perception data. If the amount of information of the vehicle's internal and external perception data is greater than a preset value, a small block is used. If the amount of information of the vehicle's internal and external perception data is not greater than the preset value, a large block is used, so as to improve the information extraction efficiency and credibility of the large language model. Step 35: Obtain entity information by interacting with the large language model based on rule documents through block data, extract the entity information required to build a hyper-relational knowledge graph, and convert it into structured data.

[0034] In this embodiment, step 35 specifically includes: Step 351: When using the large language model for the first interaction, record the data block location of the information source; Step 352: extract knowledge according to the knowledge density and block size of the data, and complete the first large language model interaction; Step 353: After the first interaction with the large language model, the extracted data is transmitted to the large language model for secondary judgment, the relationship between the nodes and edges in the hyper-relational knowledge graph is evaluated, and the rating result is outputted in terms of relationship strength rating; Step 354: Make a third judgment on the data whose edge description relationship rating in the relationship strength rating has changed, extract the data block location of the information source recorded during the first interaction, interact with the large language model again, obtain entity information, and convert it into structured data, wherein the structured data includes entities, relationships and their attribute information.

[0035] Step 4: construct a hyper-relational knowledge graph based on structured data, use the hyper-relational knowledge graph to realize hyper-relational mining of spatiotemporal correlations between data of neighboring nodes in the Internet of Vehicles, and vectorize the index of information in the hyper-relational knowledge graph based on the embedded vector model; In this embodiment, step 4 specifically includes: Step 41: construct a hyper-relational knowledge graph based on the structured data, and represent the entities and relationships in the structured data as nodes and edges in the hyper-relational knowledge graph, respectively. Each node contains entity information, and the edge represents the relationship between entities. The hyper-relational knowledge graph is as follows: Figure 2 As shown, Figure 2 The circles in the figure represent nodes, and the lines represent edges; the hyper-relational knowledge graph is represented as follows:

[0036] in, is a set of nodes, each node Contains attributes , Represents the node key value category, which stores multiple pairs of key values ​​in list form, where: Include Key and value ; is a set of edges, defined as , that is, the edge is a set of nodes An ordered pair of two nodes in Connecting Nodes and nodes ,side Contains attributes and properties , It describes the relationship between edges. Described are relationship strength ratings; Among them, the super relation is composed of a main triple and several auxiliary attribute value descriptions. The main triple is expressed as (subject, relation, object), which are respectively represented by step 4 , while the auxiliary attribute value description is By introducing the concept of super-relationship, the relationship between entities is extended to a complex relationship that can include multiple entities, which involves identifying and integrating multi-dimensional relationships between entities and the attribute information of entities and relationships.

[0037] Among them, the format of the super relationship is as follows: If there is a piece of knowledge: "Zhang San attended the meeting at the Beijing International Convention Center on November 11, 2024. The reason for attending this meeting was the annual summary." The content of the super relationship is obtained as follows:

[0038] Step 42: Add description information to each node and edge, and use key-value pairs to save the supplementary information of the node; Step 43: For nodes with multiple relationships, support multi-edge description, mark and distinguish the information of the same node and edge, evaluate the importance of different information, and mark the importance of the information; Step 44: Capture context information such as source documents, time qualifiers, and confidence scores through the hyper-relational knowledge graph to provide the system with a more detailed knowledge representation. This information will be used to enhance the context understanding capability of the hyper-relational knowledge graph, identify and integrate multi-dimensional relationships between entities, and attribute information between entities and relationships, and ultimately obtain a hyper-relational knowledge graph with rich semantics; realize hyper-relational mining of spatiotemporal correlations between data of neighboring nodes in the Internet of Vehicles; Step 45: Use the embedded vector model (BERT-based embedded vector model) to encode the keys, nodes, and relationships in the hyper-relational knowledge graph into embedded vectors, and write the vector information into the lists of each key, node, and relationship to complete the vectorization of the index. , a more comprehensive vector representation is constructed using a BERT-based embedding vector model, which not only contains the semantic information of the text, but also incorporates the relationship between entities and key-value attribute information, providing richer structured data for building indexes and node description lists of knowledge graphs.

[0039] Step 5: Evaluate the importance of nodes based on the hyper-relational knowledge graph, and update and incrementally expand the data of the hyper-relational knowledge graph; In this embodiment, step 5 specifically includes: Step 51: Use EasyGraph technology to analyze the hyper-relational knowledge graph, and use its hybrid programming characteristics to accelerate graph processing and graph analysis, effectively manage and query the real-time status, driving path, environmental interaction and other information of the vehicle in the graph, and combine user decision-making needs to improve the response speed and decision-making efficiency of decision support; Step 52: Use a temporal network node importance evaluation method based on inter-layer neighborhood information entropy to determine the importance of each node and edge, and write it into In the process of evaluating the importance of information matching, the user's weight tendency is combined to accurately measure the influence of nodes in the network. Step 53: incrementally expand the constructed hyper-relational knowledge graph, and perform another round of data processing based on rule documents to form the original hyper-relational knowledge graph and the new hyper-relational knowledge graph, with the original hyper-relational knowledge graph serving as the ontology graph and the new hyper-relational knowledge graph serving as the incremental graph; Step 54: define the scope of entity alignment according to the temporal network node importance evaluation method, set the block size according to the scale of the ontology graph, and divide the ontology graph into blocks according to the scope of entity alignment and the block size to form multiple node blocks; Step 55: Taking the important nodes identified in as the center, match the node blocks of the ontology graph with the nodes of the incremental graph to obtain the node blocks with the highest similarity; Step 56: In the structured data, use the large language model to , and Intelligent fusion of node information, and Perform knowledge fusion of the same entity, expand the parts that do not match the incremental graph and the ontology graph as new structured data to the ontology graph, and complete the incremental generation of the extended content in the ontology graph.

[0040] In this embodiment, the timing network node importance evaluation method of step 52 specifically includes: Step 521: According to the neighbor topology overlap coefficient Compute the inter-layer similarity of the edges that appear consistently between a node and its neighbors:

[0041] in, Indicates nodes, Indicates nodes, Indicates time slice, Indicated in The adjacency matrix of time slices Midpoint and nodes The edge between Step 522: Construct a temporal super-adjacency matrix based on the impact between different time slices :

[0042] in, Shidi The adjacency matrix of the time slices is is a positive integer, s represents the sth time slice, s is a positive integer and s <t, represents the sth time slice and the The influence matrix between time slices, yes The temporal super-adjacency matrix is ​​used to describe the relationship or influence of each node between the sth time slice and the tth time slice; Step 523: Based on the constructed temporal super-adjacency matrix The eigenvector corresponding to the maximum eigenvalue, that is, the eigenvector centrality of each node in each time slice, is calculated to determine the importance of the node in the graph network.

[0043] Step 6: Save the private data locally and upload the general data to the public database in the cloud, and define the dynamic access scope of the public database based on the hyper-relational knowledge graph; In this embodiment, step 6 specifically includes: Step 61: Save the private data as a local hyper-relationship graph file locally, upload the common data to a public database in the cloud, and support multi-vehicle joint access; Step 62: By marking each semantic record with a permission scope based on a rule document, different scopes of semantic library sets are opened for different permissions: for neighbor nodes in a similar geographical range, traffic geographic information on the local map in the shared semantics is opened to the public; for neighbor nodes in the same unit group relationship, the geographical tags, exclusive parking spaces and historical trajectories of neighbor vehicles in the shared unit campus are opened to support multi-vehicle shared access; Step 63: By using the constructed hyper-relational knowledge graph, the clustering relationship between vehicles is identified based on various key-value pairs of the node key-value category, and vehicles with similar characteristics or behaviors are classified into the same group, so that vehicles can access each other without authorization within the scope of privacy and security, and the data accessed are all valid data that is strongly related to each other's time and space, so as to more accurately control the scope of data sharing and more efficiently realize the rapid retrieval of useful data; Step 64: The administrator loads and updates the cloud data in combination with the infrastructure within the user group to achieve the sharing and retrieval of common data. This not only protects the privacy of the vehicle, but also achieves the sharing and retrieval of common knowledge, and greatly improves the secondary knowledge construction rate of the system.

[0044] Step 7: Use the local entity recognition model to extract the user question entities, match the local and cloud data, obtain background knowledge content, and generate enhanced retrieval results.

[0045] In this embodiment, step 7 specifically includes: Step 71: extracting entities and expressions of the question based on the local entity recognition model; Step 72: vectorize the entity and expression of the problem through an embedding vector model, and convert the text information into vector information; Step 73: Identify the clustering relationship between vehicles through the clustering information in the hyper-relational knowledge graph, and dynamically determine the access scope; Step 74: perform similarity matching between the vector information and the index vectors in the local hyper-relational graph file and the public database in the cloud within the authority to obtain the background knowledge content; this process takes into account the importance and contextual information of the node to retrieve the most relevant background knowledge. Improve the accuracy of decision-making background support; Step 75: Combine the matched background knowledge content with the user's question content to form prompt content and interact with the large language model; Step 76: Return the response after the interaction of the large language model to the local computer, and reversely decrypt it based on the encryption process in step 2 to obtain the final enhanced search result. This completes the data usage closed loop and meets the user's data usage security and intelligent acquisition needs for decision support information in the Internet of Vehicles environment.

[0046] The above descriptions are only some embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any equivalent device or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement, characterized in that: The steps include: Step 1: Collect the vehicle's internal and external sensing data, and separate the general data and privacy data from the vehicle's internal and external sensing data; Step 2: pre-process the general data and private data, compress and encrypt the private data, and mark and classify the general data; Step 3: The processed general data and privacy data are processed based on rule documents using a data knowledge dual-driven model and a large language model to convert the vehicle's internal and external perception data into structured data; Step 4: construct a hyper-relational knowledge graph based on structured data, and vectorize the index of information in the hyper-relational knowledge graph based on the embedded vector model; Step 5: Evaluate the importance of nodes based on the hyper-relational knowledge graph, and update and incrementally expand the data of the hyper-relational knowledge graph; Step 6: Save the private data locally and upload the general data to the public database in the cloud, and define the dynamic access scope of the public database based on the hyper-relational knowledge graph; Step 7: Use the local entity recognition model to extract the user question entities, match the local and cloud data, obtain background knowledge content, and generate enhanced retrieval results.

2. The decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement as claimed in claim 1, characterized in that: The step 1 specifically includes: Step 11: Collecting vehicle internal and external perception data, wherein the vehicle internal and external perception data includes vehicle environment information, vehicle status information, traffic flow information, Internet of Vehicles service platform information, vehicle location information, vehicle location information, Internet of Vehicles user community group information, parking lot layout information, user personal information, vehicle driving trajectory information, and vehicle audio interaction information; Step 12: converting the vehicle internal and external perception data into text data for recognition by a pre-trained language model, and saving the data in a structured format; Step 13: Use a pre-trained language model based on deep learning for pre-training. For the privacy data types in the Internet of Vehicles, freeze some parameters of the pre-trained language model by freezing the large model and repairing the small model. Only fine-tune the key layers to achieve the function of data separation and desensitization of the pre-trained language model. Step 14: Use the fine-tuned pre-trained language model to identify the vehicle's internal and external perception data, and separate general data and privacy data. The general data includes the vehicle's environmental information, vehicle status information, traffic flow information, Internet of Vehicles service platform information, vehicle location information, vehicle location information, Internet of Vehicles user community group information, and parking lot layout information; the privacy data includes user personal information, vehicle driving trajectory information, and in-vehicle audio interaction information.

3. The decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement as claimed in claim 1, characterized in that: The step 2 specifically includes: Step 21: Clean the general data and private data, and process missing values ​​and outliers; Step 22: construct a semantic knowledge base based on the privacy data, analyze the semantic meaning of the privacy data in the semantic knowledge base, and construct a Huffman tree based on the semantic features of different users and scenarios and the frequency of occurrence of characters, assign a code to each character for compression, and save the compressed privacy data and the code table; Step 23: Add noise to the private data through differential privacy technology to cover up the true value of the original data; Step 24: Use the AES encryption algorithm to encrypt the private data, and retain the encryption mark of the original text category at the encryption location; Step 25: Mark and classify common data to match local and cloud data interactions in different scenarios.

4. The decision support method based on knowledge retrieval enhanced spatiotemporal dynamic data of Internet of Vehicles as claimed in claim 1, characterized in that: The step 3 specifically includes: Step 31, in the data-driven part, the vehicle's internal and external perception data are used to train and optimize the large language model, so that the large language model can automatically identify and extract entities and relationships in the vehicle's internal and external perception data; in the knowledge-driven part, the expert's knowledge is introduced to set up rule documents, including making rule documents, restricting the output structure of the large language model, providing entity and relationship extraction cases, relationship importance rating evaluation cases, and category attribute discrimination evaluation cases to guide the data processing and relationship evaluation of the large language model, and then combining data-driven and knowledge-driven to form a data-knowledge dual-driven model; Step 32: Select a large language model, deploy the large language model in the cloud, and the local vehicle uses an API to call the large language model from the cloud; or deploy the large language model in the Internet of Vehicles block, and use a private intranet to call the large language model through port forwarding; Step 33: Use the data-knowledge dual-drive model and the large language model to evaluate the vehicle's internal and external perception data. Learn patterns and rules from the actual vehicle's internal and external perception data, provide rule documents to limit the output content and format of the model interaction, formulate a document segmentation strategy, and use prior knowledge to guide the output of the large language model. Step 34, determining the block size according to the amount of information of the vehicle interior and exterior sensing data, using a small block for the vehicle interior and exterior sensing data whose amount of information is greater than a preset value, and using a large block for the vehicle interior and exterior sensing data whose amount of information is not greater than the preset value; Step 35: Obtain entity information by interacting with the large language model based on rule documents through block data, extract the entity information required to build a hyper-relational knowledge graph, and convert it into structured data.

5. The decision support method based on knowledge retrieval enhanced spatiotemporal dynamic data of Internet of Vehicles as claimed in claim 4, characterized in that: The step 35 specifically includes: Step 351: When using the large language model for the first interaction, record the data block location of the information source; Step 352: extract knowledge according to the knowledge density and block size of the data, and complete the first large language model interaction; Step 353: After the first interaction with the large language model, the extracted data is transmitted to the large language model for secondary judgment, the relationship between the nodes and edges in the hyper-relational knowledge graph is evaluated, and the rating result is outputted in terms of relationship strength rating; Step 354: Make a third judgment on the data whose edge description relationship rating in the relationship strength rating has changed, extract the data block location of the information source recorded during the first interaction, interact with the large language model again, obtain entity information, and convert it into structured data, wherein the structured data includes entities, relationships and their attribute information.

6. The decision support method based on knowledge retrieval enhanced spatiotemporal dynamic data of Internet of Vehicles as claimed in claim 1, characterized in that: The step 4 specifically includes: Step 41: construct a hyper-relational knowledge graph based on structured data, and represent entities and relationships in the structured data as nodes and edges in the hyper-relational knowledge graph, respectively. Each node contains entity information, and the edge represents the relationship between entities. The hyper-relation consists of a main triple and several auxiliary attribute value descriptions. The main triple is represented as (subject, relationship, object), and the auxiliary attribute value description is represented in the form of key-value pairs. The hyper-relational knowledge graph is represented as: in, is a set of nodes, each node Contains attributes , Represents the node key value category, which stores multiple pairs of key values ​​in list form, where: Include Key and value ; is a set of edges, defined as , that is, the edge is a set of nodes An ordered pair of two nodes in Connecting Nodes and nodes ,side Contains attributes and properties , It describes the relationship between edges. Described are relationship strength ratings; Step 42: Add description information to each node and edge, and use key-value pairs to save the supplementary information of the node; Step 43: For nodes with multiple relationships, support multi-edge description, mark and distinguish the information of the same node and edge, evaluate the importance of different information, and mark the importance of the information; Step 44: Capture context information through the hyper-relational knowledge graph, enhance the context understanding ability of the hyper-relational knowledge graph, identify and integrate the multi-dimensional relationships between entities, and the attribute information of entities and relationships, and finally obtain a hyper-relational knowledge graph with rich semantics; Step 45: Use the embedded vector model to perform embedded vector encoding on the keys, nodes and relationships in the hyper-relational knowledge graph, and write the vector information into the lists of each key, node and relationship to complete the vectorization of the index.

7. The decision support method for spatiotemporal dynamic data of Internet of Vehicles based on knowledge retrieval enhancement as claimed in claim 6, characterized in that: The step 5 specifically includes: Step 51: Use EasyGraph technology to analyze the hyper-relational knowledge graph; Step 52: Use a temporal network node importance evaluation method based on inter-layer neighborhood information entropy to determine the importance of each node and edge, and write it into In the process, the user's weight tendency is combined to evaluate the importance of the final information matching selection; Step 53: incrementally expand the constructed hyper-relational knowledge graph, and perform another round of data processing based on rule documents to form the original hyper-relational knowledge graph and the new hyper-relational knowledge graph, with the original hyper-relational knowledge graph serving as the ontology graph and the new hyper-relational knowledge graph serving as the incremental graph; Step 54: define the scope of entity alignment according to the temporal network node importance evaluation method, set the block size according to the scale of the ontology graph, and divide the ontology graph into blocks according to the scope of entity alignment and the block size to form multiple node blocks; Step 55: Taking the important nodes identified in as the center, match the node blocks of the ontology graph with the nodes of the incremental graph to obtain the node blocks with the highest similarity; Step 56: In the structured data, use the large language model to , and Intelligent fusion of node information, and Perform knowledge fusion of the same entity, expand the parts that do not match the incremental graph and the ontology graph as new structured data to the ontology graph, and complete the incremental generation of the extended content in the ontology graph.

8. The decision support method based on knowledge retrieval enhanced spatiotemporal dynamic data of Internet of Vehicles as claimed in claim 7, characterized in that: The timing network node importance evaluation method of step 52 specifically includes: Step 521: According to the neighbor topology overlap coefficient Compute the inter-layer similarity of the edges that appear consistently between a node and its neighbors: in, Indicates nodes, Indicates nodes, Indicates time slice, Indicated in The adjacency matrix of time slices Midpoint and nodes The edge between Step 522: Construct a temporal super-adjacency matrix based on the impact between different time slices : in, Indicates The adjacency matrix of the time slices is is a positive integer, s represents the sth time slice, s is a positive integer and s < t, represents the sth time slice and the The influence matrix between time slices, yes The temporal super-adjacency matrix is ​​used to describe the relationship or influence of each node between the sth time slice and the tth time slice; Step 523: Based on the constructed temporal super-adjacency matrix The eigenvector corresponding to the maximum eigenvalue, that is, the eigenvector centrality of each node in each time slice, is calculated to determine the importance of the node in the graph network.

9. The decision support method based on knowledge retrieval enhanced spatiotemporal dynamic data of Internet of Vehicles as claimed in claim 1, characterized in that: The step 6 specifically includes: Step 61: Save the private data as a local hyper-relationship graph file locally, upload the common data to a public database in the cloud, and support multi-vehicle joint access; Step 62: By marking each semantic record with a permission scope based on a rule document, different scopes of semantic library sets are opened for different permissions: for neighbor nodes in a similar geographical range, traffic geographic information on the local map in the shared semantics is opened to the public; for neighbor nodes in the same unit group relationship, the geographical tags, exclusive parking spaces and historical trajectories of neighbor vehicles in the shared unit campus are opened to support multi-vehicle shared access; Step 63: By using the constructed hyper-relational knowledge graph, the clustering relationship between vehicles is identified based on various key-value pairs of the node key-value category, and vehicles with similar characteristics or behaviors are classified into the same group, so that vehicles can access each other without authorization within the scope of privacy security, and the data accessed are all valid data that is strongly related to each other in time and space; Step 64 : The administrator loads and updates the cloud data in conjunction with the infrastructure within the user group to achieve sharing and retrieval of common data.

10. The decision support method based on knowledge retrieval enhanced spatiotemporal dynamic data of Internet of Vehicles as claimed in claim 1, characterized in that: The step 7 specifically includes: Step 71: extracting entities and expressions of the question based on the local entity recognition model; Step 72: vectorize the entity and expression of the problem through an embedding vector model, and convert the text information into vector information; Step 73: Identify the clustering relationship between vehicles through the clustering information in the hyper-relational knowledge graph, and dynamically determine the access scope; Step 74: perform similarity matching between the vector information and the local hyper-relationship graph file and the index vector in the public database in the cloud within the authority to obtain background knowledge content; Step 75: Combine the matched background knowledge content with the user's question content to form prompt content and interact with the large language model; Step 76: Return the response after the interaction of the large language model to the local machine, and reversely decrypt it based on the encryption process in step 2 to obtain the final enhanced retrieval result.

Images