Intelligent Transportation Q&A Platform Based on Knowledge Graph and Large Language Model

Abstract

The present invention belongs to the field of traffic Q&A, and particularly relates to an intelligent transportation Q&A platform based on a knowledge graph and a large language model, which includes an interaction interface, a deep search module, and an execution module. The interaction interface is responsible for obtaining the user input demand vector and displaying the Q&A results. The deep search module, based on the input demand vector, uses a distributed deep search strategy network constructed by a multi-level distributed storage sub-library and a variable real-time chain to obtain a first Q&A feedback vector, where the length of the variable real-time chain can reflect the running state of the corresponding road section in real time. After the first Q&A feedback vector is input into the execution module, a first Q&A implementation evaluation probability is obtained. After setting an implementation evaluation threshold, it is fed back to the deep search module together with the evaluation probability, and then a second effective feedback Q&A is obtained, realizing efficient, fast, and comprehensive answers to transportation-related questions.

Description

Technical Field

[0001] The present invention belongs to the field of traffic Q&A, and particularly relates to an intelligent transportation Q&A platform based on a knowledge graph and a large language model. Background Art

[0002] In the field of transportation, with the continuous expansion of the industry scale and the gradual improvement of the informatization level, a large amount of complex data of various types has been generated, covering aspects such as highway topological networks, industry entities, and management affairs. However, there are many deficiencies in the existing information processing and query methods. Traditional Q&A systems are difficult to effectively integrate and deeply analyze these multi-source data, cannot reflect the highway transportation status in real time and accurately, have poor accuracy and comprehensiveness in answering users' diverse transportation-related questions, and also lack an efficient and reasonable data storage and management mechanism. In particular, the feedback on relevant search information is not timely enough, and the display status cannot be updated in a timely manner. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, the present invention proposes an intelligent transportation Q&A platform based on a knowledge graph and a large language model, which includes an interaction interface, a deep search module, and an execution module; the interaction interface is responsible for obtaining the user input demand vector and displaying the Q&A result. The deep search module, based on the input demand vector, obtains the first Q&A feedback vector by means of a distributed deep search strategy network constructed by a multi-level distributed storage sub-library and a variable real-time chain, where the length of the variable real-time chain can reflect the running status of the corresponding road section in real time. After the first Q&A feedback vector is input into the execution module, the first Q&A implementation evaluation probability is obtained. After setting the implementation evaluation threshold, it is fed back to the deep search module together with the evaluation probability, and then the second effective feedback Q&A is obtained, realizing efficient, fast, and comprehensive answers to transportation-related questions.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] An intelligent transportation Q&A platform based on a knowledge graph and a large language model, comprising: at least one interaction interface, a deep search module, and an execution module;

[0006] The interaction interface is used to obtain the input demand vector and the feedback display of the Q&A result;

[0007] The deep search module, based on the input demand vector, obtains the first Q&A feedback vector through the configured distributed deep search strategy network;

[0008] The distributed deep search strategy network is constructed by a hierarchical distributed subnet and a variable real-time chain, and the transportation status of the corresponding road section is represented in real time based on the length of the variable real-time chain between nodes in the hierarchical distributed subnet;

[0009] Input the first Q&A feedback vector into the execution module to obtain the first Q&A implementation evaluation probability;

[0010] Set the implementation evaluation threshold, and feedback the implementation evaluation threshold and the first Q&A implementation evaluation probability to the deep search module to obtain the second effective feedback Q&A.

[0011] Specifically, the distributed deep search policy network includes a first-level policy search subnet, a second-level theme retrieval subnet, and a third-level distributed storage subnet; the second-level theme retrieval subnet includes basic theme information and management theme information; the basic theme information includes: sub-class information of operating enterprises, sub-class information of employees, sub-class information of operating vehicles, and sub-class information of highways; the management theme information includes sub-class information of credit management, sub-class information of administrative licenses, sub-class information of administrative law enforcement, sub-class information of public services, sub-class information of industry supervision, and n a remaining number of extended sub-class information;

[0012] The third-level distributed storage subnet includes n +9 double-layer knowledge graph node sub-libraries; the n +9 double-layer knowledge graph node sub-libraries correspond one-to-one with each sub-class information under the second-level theme retrieval subnet, and are used to correspondingly store the real-time information and historical sub-class information corresponding to each sub-class information.

[0013] Specifically, the construction steps of the first-level policy search subnet include:

[0014] Obtain the highway topology network information, and set the entrances and exits corresponding to each section of the highway in the highway topology network as node pairs of the corresponding sections;

[0015] According to the obtained highway topology network information and the set node pairs, obtain the highway topology subnet through graph algorithms;

[0016] Based on the real-time operation information in different time periods, weather simulation information, and the real-time sub-class information set included in each node pair within the highway topology network stored in the highway topology subnet, construct a variable real-time chain between each node pair;

[0017] Build the variable real-time chain between each node pair into the corresponding node pairs within the highway topology subnet to obtain the first-level policy search subnet.

[0018] Specifically, the construction steps of the first-level policy search subnet further include:

[0019] Connect the variable real-time chain between each node pair with each sub-class information under the basic theme information and management theme information to obtain a theme information search chain between the first-level policy search subnet and the second-level theme retrieval subnet;

[0020] Construct a topic search verification model based on the Bayesian algorithm, combined with an intention decomposition sub-model and an information type selection key;

[0021] Based on the topic search verification model, use the real-time operation information by time period, weather simulation information, input demand vector, and all subclass information stored in the three-level distributed storage subnet for sub-topic selection verification training to obtain a trained topic search verification model;

[0022] Configure the trained topic search verification model into all topic information search chains to obtain a topic search selection verification chain between the primary policy search subnet and the secondary topic retrieval subnet.

[0023] Specifically, for the variable real-time chain, obtain the transportation state adjustment factor between corresponding node pairs through a two-way weight adjustment model to adjust the length of the variable real-time chain between corresponding node pairs in real time;

[0024] The construction process of the two-way weight adjustment model includes:

[0025] Take the length of the variable real-time chain corresponding to each node pair as the main dependent variable, and take the parameter set corresponding to each subclass information stored in the three-level distributed storage subnet as the collaborative independent variable;

[0026] Input the main dependent variable and collaborative independent variable under each node pair into the factor analysis algorithm to obtain the contribution factor matrix of the collaborative independent variable set corresponding to each main dependent variable and the collinearity matrix between all collaborative independent variables under the corresponding main dependent variable.

[0027] Specifically, the construction process of the two-way weight adjustment model also includes:

[0028] Configure a collinearity threshold, and based on the contribution factor matrix of the collaborative independent variable set corresponding to each main dependent variable, the collinearity matrix between all collaborative independent variables under the corresponding main dependent variable, and the collinearity threshold, obtain the effective collaborative independent variable set corresponding to each main dependent variable;

[0029] Based on each main dependent variable and the corresponding effective collaborative independent variable set, obtain the adjustment factor function corresponding to each node pair through a multiple logistic regression function;

[0030] Configure the benchmark length of the variable real-time chain with the normal passing time between the corresponding node pairs on the road, build the adjustment factor function corresponding to each node pair into the corresponding variable real-time chain, and adjust the real-time length of the variable real-time chain between the corresponding node pairs by combining the real-time obtained adjustment factor with the benchmark length of the variable real-time chain.

[0031] Specifically, the construction process of the information type selection key includes:

[0032] Input the input demand vector into the intention decomposition sub-model to obtain a spectrum of the proportion of subclass information of the demand;

[0033] Based on the spectrum of the proportion of subclass information of the demand, select and activate the variable real-time chain between corresponding node pairs and the theme search selection verification chain between the corresponding subclass information;

[0034] Combine the proportion weights of the corresponding subclass information in the spectrum of the proportion of subclass information of the demand with multi-threading technology, configure search resources for each activated theme search selection verification chain, and perform directional multi-level search on the double-layer knowledge graph node sub-library to obtain the feedback demand information of each subclass corresponding to the input demand vector.

[0035] Specifically, the construction process of the information type selection key further includes:

[0036] Feed back the obtained feedback demand information of each subclass to the corresponding theme search selection verification chain, and perform verification by time period and weather condition through the built-in theme search verification model;

[0037] Feed back the information that passes the verification to the variable real-time chain corresponding to the search of the input demand vector, adjust the length of the variable real-time chain corresponding to the current moment through the adjustment factor function corresponding to the variable real-time chain, and at the same time display the searched information in real time;

[0038] When the verification of a certain subclass information fails, feed back the corresponding information verification result to the corresponding theme search selection verification chain for secondary search until all the searched subclass information passes, and update the length of the variable real-time chain and the feedback display information in real time through the subclass information that passes the verification;

[0039] When there is no search demand between the corresponding node pairs at the current moment, search and update the real-time information in the real-time knowledge sub-node graph in the double-layer knowledge graph node sub-library through the subclass information type corresponding to the variable real-time chain corresponding to the current node pair.

[0040] Specifically, the construction steps of the double-layer knowledge graph node sub-library include:

[0041] Configure a corresponding tree-shaped data sub-library based on each subclass information type under the secondary theme retrieval subnet, and store the real-time information corresponding to each subclass information in the root node of the corresponding tree-shaped data sub-library;

[0042] Store the historical subclass information of the corresponding type in the corresponding tree-shaped data sub-library in the order from near to far in time, and establish a backward storage mechanism between the root node and the first sub-node under the same tree-shaped data sub-library.

[0043] Specifically, the construction steps of the double-layer knowledge graph node sub-library further include:

[0044] Construct a real-time knowledge sub-node graph through a graph algorithm based on the root nodes corresponding to all tree-like data sub-libraries;

[0045] The root nodes corresponding to the tree-like data sub-libraries correspond one-to-one with the sub-nodes in the real-time knowledge sub-node graph;

[0046] Based on the subclass information corresponding to all root nodes and the frequency of subclass information types in historical joint searches, construct the connection relationships between all nodes in the real-time knowledge sub-node graph through a correlation algorithm, and configure the connection relationships between the corresponding nodes between the corresponding sub-nodes;

[0047] Based on the establishment process of the real-time knowledge sub-node graph, construct historical knowledge sub-node graphs of corresponding historical subclass information in different tree-like data sub-libraries at corresponding time points in different time dimensions from near to far in time.

[0048] Specifically, the establishment process of the backward storage mechanism includes:

[0049] Assume that the tree-like data sub-library corresponding to a certain subclass information includes a root node and m sub-nodes mounted hierarchically, and set a real-time information time length threshold and a historical information storage time length threshold;

[0050] Configure the real-time information time length threshold between the root node and the first-level sub-node connected to the root node. When the real-time information timestamp stored in the root node is greater than the real-time information time length threshold, automatically generate a new first-level sub-node and save the subclass information corresponding to the time greater than the real-time information time length threshold to the new first-level sub-node;

[0051] Lower the storage level of the original first-level sub-node by one level and mount it to the new first-level sub-node, and establish an index based on the relationship between the subclass information stored in the new first-level sub-node and the original first-level sub-node. At the same time, lower the level of each level of sub-node mounted under the original first-level sub-node by one level according to the same process, and establish an information index relationship between the historical subclass information stored in the lowered-level sub-nodes and the subclass information stored in the new first-level sub-node for index connection.

[0052] Specifically, the establishment process of the backward storage mechanism further includes:

[0053] Based on the historical information storage time length threshold, obtain the sub-nodes corresponding to the subclass information stored for the same time length and set them as critical sub-nodes;

[0054] Save the historical information storage time length threshold to the critical child nodes corresponding to each tree-shaped data sub-library. When the storage time length of the subclass information corresponding to the corresponding child node is greater than the historical information storage time length threshold, delete the corresponding child node from the corresponding tree-shaped data sub-library.

[0055] Specifically, the double-layer knowledge graph node sub-library is also configured with a search restriction mechanism, specifically:

[0056] When the input demand vector is received at the current moment, perform directional search on the real-time knowledge sub-node graph and all historical knowledge sub-node graphs according to the established index relationship and the input demand vector;

[0057] When the input demand vector is not received at the current moment, only search for the effective collaborative independent variable set information included in the variable real-time chain between corresponding nodes in the real-time knowledge sub-node graph, and update the length of the corresponding variable real-time chain in real time.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] In view of the deficiencies of the prior art, the present invention constructs a distributed depth search strategy network composed of a hierarchical distributed subnet and a variable real-time chain, and uses a bidirectional weight adjustment model to adjust the length of the variable real-time chain in real time to accurately reflect the highway transportation state, so as to realize the deep analysis of multi-source data. When answering users' diverse transportation problems, with the help of the intention decomposition sub-model, information type selection key, and theme search verification model, perform directional multi-level search and sub-theme selection verification training based on the input demand vector, significantly improving the accuracy and comprehensiveness of the answer. In terms of data storage and management, by constructing a double-layer knowledge graph node sub-library including a backward storage mechanism and a search restriction mechanism, realize efficient storage and reasonable retrieval, and can update the length of the variable real-time chain and feedback display information in real time according to the search information, timely meet the user's needs, and greatly improve the efficiency and quality of information processing and query in the transportation field. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is the architecture diagram of the intelligent transportation Q&A platform in Embodiment 1 of the present invention;

[0061] Figure 2 It is the core module diagram of the intelligent transportation Q&A platform based on the knowledge graph and large language model in Embodiment 1 of the present invention;

[0062] Figure 3 It is the display diagram of the backward storage mechanism in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0063] Embodiment 1

[0064] Please refer toFigure 1 , which is the architecture diagram of the intelligent transportation Q&A platform in this embodiment. It includes cloud + network, data layer, energy layer, service layer and application layer. Each layer has complex functions. In particular, the data layer includes two large databases, the basic library and the theme library. Each database contains different sub-class information, such as operating enterprises, practitioners, highway information, etc. And under each sub-class information, there are numerous data classifications. These complex data types lead to problems such as incomplete information search, untimely feedback of road condition updates, and poor accuracy of information search during data storage and retrieval due to data islands and lack of corresponding storage and retrieval mechanisms. For this reason, an embodiment provided by the present invention: an intelligent transportation Q&A platform based on knowledge graph and large language model. Please refer to Figure 2 , including at least one interaction interface, a deep search module, an execution module and a search adjustment module;

[0065] The interaction interface is used to obtain the input demand vector and the feedback display of the Q&A result. Please refer to Figure 2 , where N represents the Nth interaction interface;

[0066] The deep search module, based on the input demand vector, obtains the first Q&A feedback vector through the configured distributed deep search strategy network;

[0067] The distributed deep search strategy network is constructed by a hierarchical distributed subnet and a variable real-time chain, and the transportation state of the corresponding road section is represented in real time based on the length of the variable real-time chain between nodes in the hierarchical distributed subnet;

[0068] Furthermore, in this embodiment, the distributed deep search strategy network includes a primary strategy search subnet, a secondary theme retrieval subnet and a tertiary distributed storage subnet;

[0069] As Figure 1 shown, the secondary theme retrieval subnet includes basic theme information and management theme information; the basic theme information includes: sub-class information of operating enterprises, sub-class information of practitioners, sub-class information of operating vehicles and sub-class information of highways; the management theme information includes sub-class information of credit management, sub-class information of administrative license, sub-class information of administrative law enforcement, sub-class information of public service, sub-class information of industry supervision and n n remaining extended sub-class information;

[0070] Furthermore, the n remaining extended sub-class information in this embodiment is a sub-class information type reserved for newly emerging sub-class information, which is specifically filled by those skilled in the art according to the actual data processing process.

[0071] The tertiary distributed storage subnet includes n +9 double-layer knowledge graph node sub-libraries; then + Nine double-layer knowledge graph node sub-libraries correspond one-to-one with each subclass information under the secondary topic retrieval subnet, and are used to correspondingly store the real-time information and historical subclass information corresponding to each subclass information.

[0072] Furthermore, the construction steps of the primary policy search subnet in this embodiment include:

[0073] Obtain highway topology network information, and set the entrances and exits corresponding to each section of the highway in the highway topology network as node pairs of the corresponding sections;

[0074] According to the obtained highway topology network information and the set node pairs, obtain a highway topology subnet through a graph algorithm;

[0075] Based on the real-time operation information by time period, weather simulation information between each node pair stored in the highway topology subnet, and the real-time subclass information set included in each node pair, construct a variable real-time chain between each node pair;

[0076] Furthermore, the real-time subclass information set included here is at least two types of subclass information among the basic topic information and the management topic information.

[0077] Furthermore, the weather simulation information in this embodiment is obtained by simulating according to the historical weather data of the corresponding highway section through a weather simulation algorithm, and is used to simulate the operating state corresponding to the corresponding highway section under different weather conditions to adjust the length of the variable real-time chain.

[0078] Furthermore, the corresponding real-time operation information by time period in this embodiment includes holiday and non-holiday nodes. For each day, with a two-hour small time segment, the real-time operation state data of the corresponding highway section is divided and collected.

[0079] Build the variable real-time chain between each node pair into the corresponding node pairs in the highway topology subnet to obtain a primary policy search subnet;

[0080] Connect the variable real-time chain between each node pair with each subclass information under the basic topic information and the management topic information to obtain a topic information search chain between the primary policy search subnet and the secondary topic retrieval subnet;

[0081] Based on the Bayesian algorithm, combine the intention decomposition sub-model and the information type selection key to construct a topic search verification model;

[0082] Furthermore, the intention decomposition sub-model in this embodiment is constructed by a Chinese pre-trained Bert model and an attention network, which is used to obtain the input demand vector corresponding to the user input and decompose the proportion weight of the corresponding search subclass information for the input demand vector. For example, if the user mainly wants to know whether there is a congestion state of the vehicles running on the corresponding highway section, then the corresponding highway subclass information is the main subclass information at this time, while other information such as whether there is administrative law enforcement subclass information on the corresponding highway section, such as vehicle inspections causing slow vehicle operation, these are information with unimportant proportion, and can be searched and fed back accordingly, but the corresponding search proportion weight will be very small.

[0083] Based on the theme search verification model, use the real-time operation information by time period, weather simulation information, input demand vector and all subclass information stored in the three-level distributed storage subnet to perform sub-theme selection verification training, and obtain the trained theme search verification model;

[0084] Configure the trained theme search verification model into all theme information search chains to obtain the theme search selection verification chain between the first-level policy search subnet and the second-level theme retrieval subnet.

[0085] This process constructs a highway topology subnet based on the highway topology network information, and combines the real-time operation information by time period, weather simulation information and real-time subclass information to construct a variable real-time chain, which can accurately and real-time reflect the highway transportation state. Because the data division by time period and weather simulation enable the variable real-time chain to dynamically adapt to the road condition changes in different time periods and weather, and the length change intuitively reflects the transportation state. Embed the variable real-time chain into the highway topology subnet, so that the first-level policy search subnet has the ability to perceive and present the real-time situation of highway transportation. Connect the subclass information of the variable real-time chain and the second-level theme retrieval subnet to form a theme information search chain, laying a foundation for multi-source data associated query. Based on the Bayesian algorithm, combine the intention decomposition sub-model with the information type selection key to construct a theme search verification model, and use multi-source information to perform sub-theme selection verification training, which can more accurately screen and verify relevant information according to the user input demand vector, improve the accuracy and relevance of information retrieval. The finally obtained theme search selection verification chain ensures that in the face of complex and diverse transportation problems, the system can quickly and accurately obtain and verify relevant information, so as to provide users with comprehensive and reliable question-and-answer feedback, effectively improving the efficiency and quality of information processing and query in the transportation field.

[0086] Furthermore, the variable real-time chain in this embodiment obtains the transportation state adjustment factor between the corresponding node pairs through a two-way weight adjustment model to adjust the length of the variable real-time chain between the corresponding node pairs in real time;

[0087] Furthermore, the construction process of the two-way weight adjustment model in this embodiment includes:

[0088] Taking the length of the variable real-time chain corresponding to each node pair as the main dependent variable, and taking the parameter sets corresponding to each subclass information stored in the three-level distributed storage subnet as the collaborative independent variables;

[0089] Inputting the main dependent variable and the collaborative independent variables under each node pair into a factor analysis algorithm to obtain the contribution factor matrix of the collaborative independent variable set corresponding to each main dependent variable and the collinearity matrix between all collaborative independent variables under the corresponding main dependent variable;

[0090] Configuring a collinearity threshold, and based on the contribution factor matrix of the collaborative independent variable set corresponding to each main dependent variable, the collinearity matrix between all collaborative independent variables under the corresponding main dependent variable, and the collinearity threshold, obtaining the effective collaborative independent variable set corresponding to each main dependent variable;

[0091] Based on each main dependent variable and the corresponding effective collaborative independent variable set, obtaining the adjustment factor function corresponding to each node pair through a multiple logistic regression function;

[0092] Configuring the reference length of the variable real-time chain with the normal passing time between the roads corresponding to the corresponding node pairs, embedding the adjustment factor function corresponding to each node pair into the corresponding variable real-time chain, and adjusting the real-time length of the variable real-time chain between the corresponding node pairs by combining the real-time obtained adjustment factors with the reference length of the variable real-time chain.

[0093] Furthermore, the construction process of the information type selection key in this embodiment includes:

[0094] Inputting the input demand vector into the intention decomposition sub-model to obtain the proportion spectrum of demand subclass information;

[0095] Based on the proportion spectrum of demand subclass information, selecting and activating the theme search selection verification chain between the variable real-time chain corresponding to the corresponding node pair and the corresponding subclass information;

[0096] Combining the proportion weights of the corresponding subclass information in the proportion spectrum of demand subclass information with multi-threading technology, configuring search resources for each activated theme search selection verification chain to perform directional multi-level search on the two-layer knowledge graph node sub-library, and obtaining the feedback demand information of each subclass corresponding to the input demand vector.

[0097] Feeding back the obtained feedback demand information of each subclass to the corresponding theme search selection verification chain, and performing verification by time period and weather status through the built-in theme search verification model;

[0098] Feeding back the information that passes the verification to the variable real-time chain corresponding to the search of the input demand vector, adjusting the length of the variable real-time chain corresponding to the current moment through the adjustment factor function corresponding to the variable real-time chain, and simultaneously displaying the searched information in real time;

[0099] When the verification of a certain subclass of information fails, the corresponding information verification result is fed back to the corresponding topic search selection verification chain for secondary search until all the searched subclass information passes, and the length of the variable real-time chain and the feedback display information are updated in real time through the verified subclass information.

[0100] When there is no search requirement between the corresponding nodes at the current moment, the real-time information in the real-time knowledge sub-node graph of the double-layer knowledge graph node sub-library is searched and updated according to the subclass information type corresponding to the variable real-time chain corresponding to the current node pair.

[0101] Furthermore, the construction steps of the double-layer knowledge graph node sub-library in this embodiment include:

[0102] Based on each subclass information type under the secondary topic retrieval subnet, a corresponding tree-like data sub-library is configured, and the real-time information corresponding to each subclass information is stored in the root node of the corresponding tree-like data sub-library.

[0103] The historical subclass information of the corresponding type is stored hierarchically in the corresponding tree-like data sub-library in the order from the most recent to the oldest, and a backward storage mechanism is established between the root node and the first sub-node under the same tree-like data sub-library.

[0104] Based on the root nodes corresponding to all the tree-like data sub-libraries, a real-time knowledge sub-node graph is constructed through a graph algorithm.

[0105] The root nodes corresponding to the tree-like data sub-libraries correspond one-to-one with the sub-nodes in the real-time knowledge sub-node graph.

[0106] Based on the subclass information corresponding to all the root nodes and the frequency of the subclass information types in the historical joint search, the connection relationships between all the nodes in the real-time knowledge sub-node graph are constructed through a correlation algorithm, and the corresponding connection relationships between the nodes are configured between the corresponding sub-nodes.

[0107] Based on the establishment process of the real-time knowledge sub-node graph, historical knowledge sub-node graphs of the corresponding historical subclass information in different tree-like data sub-libraries at different time points in different time dimensions are constructed in the order from the most recent to the oldest.

[0108] Furthermore, each historical knowledge sub-node graph in this embodiment has the same structure as the real-time knowledge sub-node graph, and an index connection between the real-time knowledge sub-node graph and the historical knowledge sub-node graph and between all the historical knowledge sub-node graphs in the same tree-like data sub-library is established through a time dimension variable. The historical subclass information stored in the same historical knowledge sub-node graph is different subclass information in the same time dimension.

[0109] Further, the establishment process of the backward storage mechanism in this embodiment includes:

[0110] Assume that the tree - shaped data sub - library corresponding to a certain subclass information includes a root node and m several hierarchically - mounted sub - nodes, and set the real - time information time - length threshold and the historical information storage time - length threshold;

[0111] Configure the real - time information time - length threshold between the root node and the first - level sub - nodes connected to the root node. When the real - time information timestamp saved in the root node is greater than the real - time information time - length threshold, automatically generate a new first - level sub - node and save the subclass information corresponding to the timestamp greater than the real - time information time - length threshold to the new first - level sub - node;

[0112] Lower the storage level of the original first - level sub - nodes by one level and mount them to the new first - level sub - node, and establish an index according to the relationship between the subclass information stored in the new first - level sub - node and the original first - level sub - nodes. At the same time, for each level of sub - nodes mounted under the original first - level sub - nodes, lower their levels by one level according to the same process, and establish an information index relationship between the historical subclass information stored in the lowered - level sub - nodes and the subclass information stored in the new first - level sub - node for index connection.

[0113] Please refer to Figure 3 , assume that A is the root node of a certain tree - shaped data sub - library, and A1, A2, A3 are the hierarchical sub - nodes under the root node, where A1 is the first - level sub - node connected to the root node. When the real - time information timestamp saved in the root node is greater than the real - time information time - length threshold, automatically generate A1' to save the outdated information corresponding to the previous real - time information timestamp in the root node, and the original A1 is lowered by one level, and A2 and A3 are also automatically lowered by one level for storage, so that the information with an earlier timestamp is stored in the top - layer space of the database in real - time, reducing the difficulty of retrieval.

[0114] Based on the historical information storage time - length threshold, obtain the sub - nodes corresponding to the subclass information saved for the same time length and set them as critical sub - nodes;

[0115] Save the historical information storage time - length threshold to each critical sub - node corresponding to the tree - shaped data sub - library. When the time length of the subclass information saved in the corresponding sub - node is greater than the historical information storage time - length threshold, delete the corresponding sub - node from the corresponding tree - shaped data sub - library.

[0116] Further, the double - layer knowledge graph node sub - library in this embodiment is also configured with a search restriction mechanism, specifically:

[0117] When the input demand vector is received at the current moment, a directional search is performed on the real-time knowledge sub-node graph and all historical knowledge sub-node graphs according to the established index relationship and the input demand vector;

[0118] When the input demand vector is not received at the current moment, only the real-time knowledge sub-node graph is searched for the effective collaborative independent variable set information included in the variable real-time chain between corresponding node pairs, and the length of the corresponding variable real-time chain is updated in real time.

[0119] Furthermore, in the construction of the information type selection key, the input demand vector is input into the intention decomposition sub-model to obtain the proportion graph of demand sub-class information, which can accurately analyze the user's needs. Based on this, the corresponding theme search selection verification chain is activated, and combined with multi-thread technology for directional multi-level search, the feedback demand information of each sub-class can be efficiently obtained from the double-layer knowledge graph node sub-library. After verification by time period and weather conditions, the accuracy and applicability of the information are ensured. The information that passes the verification is fed back to the variable real-time chain, which can not only adjust the chain length in real time to intuitively reflect the transportation status, but also display the information to the user in real time, enhancing the user experience. For the information that fails to pass the verification, a secondary search is performed to ensure the integrity and reliability of the information. When there is no search demand, the real-time knowledge sub-node graph is updated to maintain the timeliness of the data. The construction of the double-layer knowledge graph node sub-library greatly improves the efficiency of data storage and management. The tree-like data sub-library is configured according to the sub-class information type, the real-time information is stored in the root node, the historical information is stored hierarchically by time, and a backward storage mechanism is established, making the data storage structure clear and orderly, facilitating data update and query. The construction of the real-time knowledge sub-node graph and the historical knowledge sub-node graph, as well as the determination of the connection relationship between nodes, reflect the association between different sub-class information, providing convenience for the comprehensive query and analysis of information. The backward storage mechanism realizes the dynamic management of data by setting a time threshold, ensuring the freshness of the data and the storage efficiency of the system. The search restriction mechanism flexibly adjusts the search scope according to whether the input demand vector is received, which not only meets the comprehensive needs of users during queries, but also focuses on real-time updating of the variable real-time chain when there is no query, improving the utilization efficiency of system resources.

[0120] Input the first Q&A feedback vector into the execution module to obtain the first Q&A implementation evaluation probability;

[0121] Furthermore, the first Q&A implementation evaluation probability in this embodiment is obtained through the theme search verification model;

[0122] Set the implementation evaluation threshold, and feedback the implementation evaluation threshold and the first Q&A implementation evaluation probability to the depth search module to obtain the second effective feedback Q&A.

[0123] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in the field may also change, modify, replace and modify the above-mentioned embodiments without departing from the scope of protection of the purpose of the present invention and the claims, and all of these are within the protection of the present invention.

[0124] If the disclosed technical solution involves personal information, the product using the disclosed technical solution has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the disclosed technical solution involves sensitive personal information, the product using the disclosed technical solution has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, when the personal information processing rules are notified by obvious signs / information, the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. A transportation intelligent Q&A platform based on a knowledge graph and a large language model, characterized in that It includes at least one interaction interface, a deep search module, and an execution module; The interaction interface is used to obtain the input requirement vector and the feedback display of the Q&A result; The deep search module, based on the input requirement vector, obtains the first Q&A feedback vector through the configured distributed deep search strategy network; The distributed deep search strategy network is constructed by a hierarchical distributed subnet and a variable real-time chain, and based on the length of the variable real-time chain between nodes in the hierarchical distributed subnet, it represents the transportation status of the corresponding road section in real time; Input the first Q&A feedback vector into the execution module to obtain the first Q&A implementation evaluation probability; Set the implementation evaluation threshold, and feedback the implementation evaluation threshold and the first Q&A implementation evaluation probability to the deep search module to obtain the second effective feedback Q&A; The distributed deep search strategy network includes a primary strategy search subnet, a secondary theme retrieval subnet, and a tertiary distributed storage subnet; the secondary theme retrieval subnet includes basic theme information and management theme information; The said basic theme information includes: subclass information of enterprises engaged in the industry, subclass information of employees, subclass information of operating vehicles, and subclass information of highways; the said management theme information includes credit management subclass information, administrative licensing subclass information, administrative law enforcement subclass information, public service subclass information, industry supervision subclass information, and n information of n remaining extended subclasses; The three-level distributed storage subnet includes n + 9 double-layer knowledge graph node sub-libraries; the n + 9 double-layer knowledge graph node sub-libraries correspond one-to-one with each subclass information under the secondary topic retrieval subnet, and are used to correspondingly store the real-time information and historical subclass information corresponding to each subclass information; The construction steps of the primary strategy search subnet include: Obtain the highway topology network information, and set the entrances and exits corresponding to each section of the highway in the highway topology network as the node pairs of the corresponding road section; According to the obtained highway topology network information and the set node pairs, obtain the highway topology subnet through a graph algorithm; Based on the time-segmented real-time operation information, weather simulation information, and the real-time subclass information set included in each node pair in the highway topology network stored in the highway topology subnet, construct a variable real-time chain between each node pair; Build the variable real-time chain between each node pair into the corresponding node pairs in the highway topology subnet to obtain the primary strategy search subnet.

2. The intelligent transportation Q&A platform based on the knowledge graph and the large language model according to claim 1, characterized in that, The construction steps of the primary strategy search subnet further include: Connect the variable real-time chain between each node pair with each subclass information under the basic theme information and management theme information to obtain the theme information search chain between the primary strategy search subnet and the secondary theme retrieval subnet; Based on the Bayesian algorithm, combine the intention decomposition sub-model and the information type selection key to construct a theme search verification model; Based on the theme search verification model, use the time-segmented real-time operation information, weather simulation information, input requirement vector, and all subclass information stored in the tertiary distributed storage subnet for sub-theme selection verification training to obtain the trained theme search verification model; Configure the trained theme search verification model to all the theme information search chains to obtain the theme search selection verification chain between the primary strategy search subnet and the secondary theme retrieval subnet.

3. The transportation intelligent Q&A platform based on the knowledge graph and the large language model according to claim 2, characterized in that, The variable real-time chain obtains the transportation status adjustment factor between the corresponding node pairs through a two-way weight adjustment model to adjust the length of the variable real-time chain between the corresponding node pairs in real time; The construction process of the two-way weight adjustment model includes: Take the length of the variable real-time chain corresponding to each node pair as the main independent variable, and take the parameter set corresponding to each subclass information stored in the tertiary distributed storage subnet as the collaborative independent variable; Input the main dependent variable and co - independent variables under each node pair into the factor analysis algorithm to obtain the contribution factor matrix of the co - independent variable set corresponding to each main dependent variable and the collinearity matrix among all co - independent variables under the corresponding main dependent variable.

4. The intelligent transportation Q&A platform based on the knowledge graph and the large language model according to claim 3, characterized in that, The construction process of the two - way weight adjustment model further includes: Configure a collinearity threshold. Based on the contribution factor matrix of the co - independent variable set corresponding to each main dependent variable, the collinearity matrix among all co - independent variables under the corresponding main dependent variable, and the collinearity threshold, obtain the effective co - independent variable set corresponding to each main dependent variable; Based on each main dependent variable and the corresponding effective co - independent variable set, obtain the adjustment factor function corresponding to each node pair through the multiple logistic regression function; Configure the benchmark length of the variable real - time chain with the normal passing time between the roads corresponding to the corresponding node pair. Embed the adjustment factor function corresponding to each node pair into the corresponding variable real - time chain, and adjust the real - time length of the variable real - time chain between the corresponding node pairs by combining the real - time obtained adjustment factor with the benchmark length of the variable real - time chain.

5. The intelligent transportation Q&A platform based on the knowledge graph and the large language model according to claim 4, wherein The construction process of the information type selection key includes: Input the input demand vector into the intention decomposition sub - model to obtain the proportion map of demand subclass information; Based on the proportion map of demand subclass information, select and activate the theme search selection verification chain between the variable real - time chain corresponding to the corresponding node pair and the corresponding subclass information; Combine the proportion weight of the corresponding subclass information in the proportion map of demand subclass information with multi - thread technology, configure search resources for each activated theme search selection verification chain to conduct directional multi - level search on the double - layer knowledge graph node sub - library, and obtain the feedback demand information of each subclass corresponding to the input demand vector.

6. The intelligent transportation Q&A platform based on the knowledge graph and large language model according to claim 5, wherein, The construction process of the information type selection key further includes: Feed the obtained feedback demand information of each subclass back to the corresponding theme search selection verification chain, and conduct verification by time period and weather status through the built - in theme search verification model; Feed the verified information back to the variable real - time chain corresponding to the search of the input demand vector, adjust the length of the variable real - time chain corresponding to the current moment through the adjustment factor function corresponding to the variable real - time chain, and at the same time, display the searched information in real - time; When a certain subclass information fails the verification, feed the corresponding information verification result back to the corresponding theme search selection verification chain for secondary search until all searched subclass information passes, and update the length of the variable real - time chain and the feedback display information in real - time with the verified subclass information; When there is no search demand between the corresponding node pairs at the current moment, search and update the real - time information in the real - time knowledge sub - node map in the double - layer knowledge graph node sub - library according to the subclass information type corresponding to the variable real - time chain corresponding to the current node pair.

7. The intelligent transportation Q&A platform based on the knowledge graph and the large language model according to claim 6, characterized in that, The construction steps of the double - layer knowledge graph node sub - library include: Configure the corresponding tree - like data sub - library based on each subclass information type under the secondary theme retrieval subnet, and store the real - time information corresponding to each subclass information in the root node of the corresponding tree - like data sub - library; The historical subclass information of the corresponding type is hierarchically stored in the corresponding tree - shaped data sub - library in the order from the most recent to the oldest in time, and a backward storage mechanism is established between the root node and the first - level sub - node under the same tree - shaped data sub - library.

8. The intelligent transportation Q&A platform based on the knowledge graph and the large language model according to claim 7, characterized in that, The steps for constructing the double - layer knowledge graph node sub - library further include: Constructing a real - time knowledge sub - node graph based on the root nodes corresponding to all tree - shaped data sub - libraries through a graph algorithm; The root nodes corresponding to the tree - shaped data sub - libraries correspond one - to - one with the sub - nodes in the real - time knowledge sub - node graph; Based on the subclass information corresponding to all root nodes and the frequency of the subclass information types in historical joint searches, through a correlation algorithm, construct the connection relationships between all nodes in the real - time knowledge sub - node graph, and configure the connection relationships between the corresponding nodes to the corresponding sub - nodes; Based on the establishment process of the real - time knowledge sub - node graph, construct historical knowledge sub - node graphs of the corresponding historical subclass information in different tree - shaped data sub - libraries at corresponding time points in different time dimensions in the order from the most recent to the oldest in time.

9. The intelligent transportation Q&A platform based on the knowledge graph and the large language model according to claim 8, characterized in that, The establishment process of the backward storage mechanism includes: Suppose that the tree - shaped data sub - library corresponding to a certain subclass information includes a root node and m sub - nodes mounted hierarchically, and set the real - time information time - length threshold and the historical information storage time - length threshold; Configure the real - time information time - length threshold between the root node and the first - level sub - node connected to the root node. When the real - time information timestamp saved in the root node is greater than the real - time information time - length threshold, automatically generate a new first - level sub - node and save the subclass information corresponding to the value greater than the real - time information time - length threshold to the new first - level sub - node; Lower the storage level of the original first - level sub - node by one level and mount it to the new first - level sub - node, and establish an index based on the relationship between the subclass information stored in the new first - level sub - node and the original first - level sub - node. At the same time, for each level of sub - nodes mounted under the original first - level sub - node, lower their levels by one level according to the same process, and establish an information index relationship and index connection between the historical subclass information stored in the lowered - level sub - nodes and the subclass information stored in the new first - level sub - node.

10. The intelligent transportation Q&A platform based on a knowledge graph and a large language model according to claim 9, characterized in that, The establishment process of the backward storage mechanism further includes: Based on the historical information storage time - length threshold, obtain the sub - nodes corresponding to the subclass information saved for the same time length and set them as critical sub - nodes; Save the historical information storage time - length threshold to the critical sub - nodes corresponding to each tree - shaped data sub - library. When the time length of the subclass information saved in the corresponding sub - node is greater than the historical information storage time - length threshold, delete the corresponding sub - node from the corresponding tree - shaped data sub - library.

11. The intelligent transportation Q&A platform based on the knowledge graph and the large language model according to claim 10, characterized in that, The double - layer knowledge graph node sub - library is also configured with a search restriction mechanism, specifically: When the input demand vector is received at the current moment, then conduct a directed search on the real - time knowledge sub - node graph and all historical knowledge sub - node graphs according to the established index relationship and the input demand vector; When the input demand vector is not received at the current moment, then only conduct a search on the real - time knowledge sub - node graph for the effective collaborative independent variable set information included in the variable - type real - time chain between corresponding node pairs, and real - time update the length of the corresponding variable - type real - time chain.

Images