CN-121979857-A - Traffic field AI Agent based on blood margin analysis and fine adjustment large model

CN121979857ACN 121979857 ACN121979857 ACN 121979857ACN-121979857-A

Abstract

The invention discloses an AI Agent in the traffic field based on a blood-margin analysis and fine adjustment large model, belongs to the technical field of crossing of artificial intelligence and traffic data processing, and is suitable for traffic core business scenes such as network traffic reduction operation and the like. The implementation route comprises the steps of collecting traffic core business original construction table data, implementing three-level standardized operation, constructing a blood-margin knowledge map containing traffic field-core index-traffic scene association, carrying out vectorization coding on traffic semantics enhancement on the map, constructing a vector knowledge base, constructing a pre-training corpus by means of Qwen and a Transformer fusion framework, taking the vector knowledge base as a support, fusing traffic industry standards and the like, adapting traffic proprietary tasks by using LoRA light fine-tuning strategies after pre-training, constructing an interaction model Agent, carrying out blood-margin analysis on traffic business data by means of Agent, and generating a structured diagnosis report containing business insight and decision advice. The invention solves the problems of insufficient accuracy and the like of the existing AI Agent when processing the traffic data in the traffic field, improves the traffic data quality and decision reliability, and provides technical support for abnormal positioning and operation optimization of the traffic data.

Inventors

XU HAICHAO
CHEN JUNYAN
WAN JIAN

Assignees

南京道智明交通科技有限公司

Dates

Publication Date: 20260505
Application Date: 20251229

Claims (8)

1. The traffic field AI Agent intelligent Agent based on the blood margin analysis and fine adjustment large model is characterized by comprising the following steps: S1, directionally acquiring original table construction data corresponding to traffic field target service, and carrying out three-level standardization processing for traffic service on the original table construction data to construct a blood-margin knowledge graph containing traffic link mapping relation, wherein the three-level standardization processing comprises the steps of respectively storing the original table construction data into a traffic service index table and a traffic service data table after inquiring and cleaning the original table construction data, and describing inter-table traffic service logic between the traffic service index table and the traffic service data table; S2, carrying out vectorization coding of traffic semantic enhancement on the blood margin knowledge graph, and constructing a vector knowledge base adapting to traffic business real-time query requirements; S3, expanding traffic industry indexes, designing traffic business flow specifications, building a pre-training corpus, relying on Qwen and a Transformer fusion framework, taking a vector knowledge base as a core support, pre-training a Qwen model by relying on the pre-training corpus to obtain a pre-training model, performing iterative tuning on the pre-training model by adopting a LoRA light fine tuning strategy to obtain a large language model adapting to the traffic field, and building an AI Agent in the traffic field based on the large language model; S4, developing blood margin analysis aiming at real-time traffic service query demands by means of an AI Agent in the traffic field through a stepped thinking chain, and outputting a structural diagnosis report, wherein the stepped thinking chain is a traffic service logic hierarchical reasoning process based on query demand disassembly, blood margin data matching, service logic deduction and conclusion verification flow.
2. The traffic field AI Agent based on the blood margin analysis and fine adjustment large model of claim 1, wherein in S1, the original SQL data is collected and three-level standardization processing is implemented, and further a blood margin knowledge graph is constructed, comprising the following steps: S11, directionally collecting original table building data comprising exclusive fields of traffic services, data types and inter-table traffic service logic from a traffic service association database according to a general data storage rule; S12, firstly querying original table building data by using an aggregate SQL sentence, splitting the aggregate SQL sentence into single sentences, and secondarily querying a query result by using the single sentences; for the secondary query result, firstly screening out non-traffic service data including test data and temporary debugging data, then removing redundant traffic service data through operations including deleting comments, replacing line changing symbols and tab making and compressing redundant blank spaces, and finally extracting traffic service indexes to store in a traffic service index table, and extracting traffic service data to store in a traffic service data table; s13, performing de-duplication processing on the traffic service index table and the traffic service data table, describing traffic service logic between the traffic service index and the traffic service data, and performing structural storage on the structure according to the key value of { "blood margin": traffic service index table, "sub-table structure": traffic service data table, "service association": traffic service logic }, thereby forming a blood margin knowledge graph containing traffic link mapping relation.
3. The traffic field AI Agent based on the blood edge analysis and fine adjustment large model of claim 1, wherein in S2, the blood edge knowledge graph is vectorized and encoded, a vector knowledge base is built, firstly, the SQL Converter tool is used for vectorizing the blood edge knowledge graph to implement traffic service semantic enhancement, traffic service logic contained in structured SQL data is embedded into a vector space to complete feature mapping, then exclusive semantic features of traffic service are integrated into the vector space, an index optimization mechanism adapting to a traffic service real-time query scene is built, structured storage of the blood edge knowledge graph is realized based on the vectorization coding and the index optimization mechanism, a vector knowledge base adapting to traffic service real-time query requirements is formed, and the vector knowledge base can achieve a millisecond-level call response effect.
4. The traffic domain AI Agent based on the blood margin analysis and fine adjustment large model of claim 1, wherein in S3, the traffic domain AI Agent is constructed by relying on Qwen and a Transformer fusion architecture, and comprises a pre-training stage and a fine adjustment stage; In the pre-training stage, extracting complete blood edge link description in a vector knowledge base, expanding traffic industry indexes and designing traffic flow specifications, constructing a traffic exclusive pre-training corpus with total scale not less than N, wherein the complete blood edge link description comprises specified traffic exclusive fields, corresponding traffic data tables, associated traffic index tables and a real-time traffic query scene to which the traffic exclusive fields belong; In the fine tuning stage, carrying out association labeling of a traffic service real-time query scene based on a blood edge knowledge graph, converting traffic service real-time query requirements into an instruction fine tuning format, constructing at least M test case sample sets, freezing core network parameters except for low-rank adapter parameters in a pre-training model by adopting a LoRA light fine tuning strategy, only updating the low-rank adapter parameters, constructing an evaluation system by taking blood edge link identification accuracy, abnormal traffic service positioning accuracy and response time as core evaluation indexes, carrying out iterative tuning on the pre-training model based on the test case sample sets, integrating a vector knowledge base into the pre-training model through the iterative tuning, and constructing an AI Agent in the traffic field based on the pre-training model, wherein the scale of the pre-training corpus meets N > M.
5. The traffic domain AI Agent based on the large blood-margin analysis and fine-tuning model of claim 1, wherein in S4, the blood-margin analysis for the traffic service real-time query requirement is performed by means of the traffic domain AI Agent, firstly, query intention and traffic service core entity corresponding to the traffic service real-time query requirement are disassembled by means of an entity identification unit embedded in the traffic domain AI Agent, then, traffic service logic is deduced hierarchically by invoking traffic service blood-margin data and historical interaction data associated in a vector knowledge base based on the query intention and the traffic service core entity, and finally, a structural diagnosis report is constructed based on the deduction result, wherein the structural diagnosis report comprises a query instruction, a blood-margin relation visual analysis chart, traffic service logic matched with the query intention and a targeted solution.
6. The traffic field AI Agent based on the blood margin analysis and fine adjustment large model is characterized by comprising a data acquisition and standardization module, a vector knowledge base construction module, an Agent construction module and a blood margin analysis and diagnosis module; The data acquisition and standardization module is used for directionally acquiring original form construction data corresponding to traffic field target business, carrying out three-level standardization processing for traffic business on the original form construction data, and constructing a blood margin knowledge graph containing a traffic link mapping relation; The vector knowledge base construction module is used for carrying out vectorization coding of traffic semantic enhancement on the blood margin knowledge map and constructing a vector knowledge base adapting to traffic business real-time query requirements; The Agent constructing module is used for expanding traffic industry indexes and designing traffic business flow specifications to construct a pre-training corpus, relying on Qwen and a Transformer fusion framework, taking a vector knowledge base as a core support, pre-training a Qwen model by relying on the pre-training corpus to obtain a pre-training model, adopting a LoRA light fine tuning strategy to implement iterative tuning on the pre-training model, integrating the vector knowledge base into the pre-training model to obtain a large language model adapting to the traffic field, and constructing an AI Agent in the traffic field based on the large language model; the blood margin analysis and diagnosis module is used for carrying out blood margin analysis aiming at traffic service real-time query requirements through a stepped thinking chain by means of an AI Agent in the traffic field and outputting a structural diagnosis report of the traffic service.
7. The traffic field AI Agent based on the blood margin analysis and fine adjustment large model of claim 6, wherein the data acquisition and standardization module is in butt joint with a traffic service association database, the traffic service association database comprises a road condition monitoring database, a vehicle scheduling database, a network bus order database and a bus passenger flow database, and the data acquisition and standardization module has an extraction function of writing traffic service data into a class target sentence, supports configuration of an original construction table data acquisition rule according to traffic service types, and has the technical capability of screening non-traffic service data and eliminating redundant traffic service data.
8. The traffic field AI Agent based on the blood margin analysis and fine adjustment large model of claim 6, wherein the Agent construction module comprises an embedded pre-training corpus construction unit, a corpus expansion unit, a word segmentation training unit and a fine adjustment optimizing unit, The pre-training corpus construction unit extracts complete blood-edge link description in the vector knowledge base, expands traffic industry indexes and designs traffic business flow specifications at the same time, and builds a pre-training corpus; the corpus expansion unit is used for supplementing a traffic field exclusive word list to the pre-training corpus; the word segmentation training unit is used for integrating a Byte-PairEncoding word segmentation device of a Transformer and a Qwen traffic field exclusive word list expansion mechanism, adopting an MLM task framework to combine a multi-granularity optimization scheme, and implementing special mask training on a traffic service core entity and associated traffic service logic based on a pre-training corpus to obtain a pre-training model with basic blood-edge analysis capability and traffic service understanding capability; The fine tuning optimization unit is used for freezing core network parameters except for low-rank adapter parameters in the pre-training model by adopting LoRA light fine tuning strategies, only updating the low-rank adapter parameters, constructing an evaluation system by taking blood-margin link identification accuracy, abnormal traffic service positioning accuracy and response time as core evaluation indexes, and performing iterative tuning on the pre-training model based on a test case sample set to obtain a large language model adapting to the traffic field.

Description

Traffic field AI Agent based on blood margin analysis and fine adjustment large model Technical Field The invention relates to an AI Agent in the traffic field of blood margin analysis and fine adjustment large models, belonging to the crossing technology of artificial intelligence and traffic data processing. Background The AI Agent belongs to an intelligent system capable of running autonomously in a specific environment, has core capabilities of sensing, reasoning, deciding, executing and the like, and can respond autonomously according to the dynamic change of the environment so as to achieve a task target. The core characteristics of the system are autonomy and intelligence, and complex tasks can be executed without manual intervention. In the traffic field, the technology can be deeply adapted to core service scenes such as road condition analysis, vehicle scheduling, order optimization, passenger flow adjustment and the like, and provides key support for urban traffic digital operation. However, three major core service pain points exist in the practical application of the current AI Agent in the traffic field, and it is difficult to meet the high reliability requirement of traffic operation: Firstly, the data processing link has high dependency on a general large language model, lacks the semantic adaptation capability of traffic service, and when the special traffic demands such as analysis of the reasons of traffic gaps of a regional network in the early peak, abrupt tracing of the public traffic and abnormal positioning of road condition data are faced, the system is easy to induce the problems of data compression, redundant integration, dirty data output and the like due to insufficient service understanding, for example, when the public traffic scheduling data are processed, test data in a non-operating period are erroneously included into statistics, so that departure interval optimization decision is deviated. Secondly, the transparency of the data generation process is insufficient and is limited by the data illusion and the black box problem, the generated data is difficult to support traffic service report preparation and supervision audit work, for example, the road condition congestion index generated by the prior art lacks data source link description and cannot be used as a reliable basis of a traffic dispersion scheme. Thirdly, the data traceability has obvious defects that when the problems of order statistics deviation, passenger flow data errors, road condition information distortion and the like occur, the full-link circulation track from collection (such as monitoring points and vehicle-mounted terminals) to summary (such as business reports) of the data cannot be traced, so that abnormal data disposal is delayed, and the traffic operation stability is affected. In addition, in the prior art, the accuracy of general Agent data output aiming at traffic scenes is only 80%, and when the services such as order slip attribution and the like are processed, core traffic factors such as the online rate of drivers, regional transport capacity distribution and the like cannot be associated, and the full-link traceability of data is difficult to realize, so that the method is generally difficult to meet the fine operation requirements of the traffic field. Disclosure of Invention Aiming at the special pain point in the traffic field, the invention provides an AI Agent intelligent Agent in the traffic field based on a blood margin analysis and fine adjustment large model, which effectively solves the problems of low accuracy, insufficient transparency, poor traceability and the like in traffic service data processing in the prior art by means of the core design thought of 'Qwen and a transform fusion framework+ LoRA light fine adjustment (adaptive traffic service dynamic change) +traffic special data blood margin analysis mechanism', thereby enhancing the traffic data quality and decision reliability and adapting the operation requirements of multiple scenes such as traffic service, public transportation, road condition management and the like. The technical scheme adopted by the invention is as follows: The traffic field AI Agent based on the blood margin analysis and fine adjustment large model can adapt to traffic core services such as data abnormity attribution, power scheduling traceability, order fluctuation analysis and the like, and comprises the following steps: S1, directionally acquiring original construction table data (covering traffic scenes such as network traffic orders, bus scheduling, urban road condition monitoring and the like) corresponding to traffic field target services, and implementing three-level standardization processing for traffic services on the original construction table data to construct a blood-source knowledge graph containing traffic link mapping relations, wherein the three-level standardization processing comprises the steps of respectively storing the