CN-121981369-A - Rail area traffic safety supervision system and method based on AI large model

CN121981369ACN 121981369 ACN121981369 ACN 121981369ACN-121981369-A

Abstract

The invention relates to the field of rail transit supervision, in particular to a rail transit safety supervision system and method based on an AI large model, which adopts a lightweight micro-service architecture to establish a rail transit safety supervision system, comprises a multi-source data fusion engine, a multi-agent collaboration module, an intelligent decision support module and a multi-mode interaction unit, and integrates multi-source data such as police, ticket cards, electronic accounts and the like, the intelligent analysis and decision-making are realized by means of the technologies such as an AI large model, a knowledge graph and the like, the intelligent analysis and decision-making system has the functions of intelligent data fusion, active early warning, automatic office assistance and visual analysis, can optimize the patrol strategy, improve the emergency response efficiency, obviously reduce the manual intervention cost, and ensure the safety and the high-efficiency supervision of the rail transit operation, thereby solving the problems of the existing rail transit supervision data such as rupture, high manual dependence, slow response and lack of intelligent decision-making support.

Inventors

HUANG LUYAO
JIANG RENJIE
CAI JIONG
WANG XINGMING
XU MINGYU
Hou Xiaoshun

Assignees

中电智安科技有限公司

Dates

Publication Date: 20260505
Application Date: 20251219

Claims (10)

1. The rail area traffic safety supervision system based on the AI large model is characterized by comprising a multi-source data fusion engine, a multi-agent cooperation module, an intelligent decision support module and a multi-mode interaction unit: the multi-source data fusion engine adopts a distributed architecture, is used for data input and preprocessing, integrates multi-source heterogeneous data and constructs a data analysis view; the multi-agent cooperative module comprises a plurality of functional agents, realizes data interaction and cooperative work through a standardized interface, enables each agent to respectively conduct targeted analysis on specific types of data, and outputs intermediate analysis results of subdivision scenes; The intelligent decision support module builds an intelligent decision engine based on the AI large model, integrates unified data provided by the multi-source data fusion engine and intermediate analysis results output by the multi-agent cooperation module by combining the knowledge graph, performs cross-scene comprehensive research and judgment, finally outputs a standardized decision result, and presents the standardized decision result to a user through an interface of the multi-mode interaction unit so as to realize intelligent analysis and decision of rail area traffic safety supervision.
2. The rail transit security supervision system according to claim 1, wherein the multi-source data fusion engine comprises a data acquisition layer, a data processing layer, and a data analysis layer: The data acquisition layer acquires police condition data through a special line protocol to the police command platform, receives ticket traffic data through an FTP service, and acquires ticket identity information through transfer of a data domain server; the data processing layer realizes multi-source data real-time stream cleaning, format conversion, loading and warehousing based on APACHE KAFKA; the data analysis layer utilizes Hive data warehouse technology to complete data association analysis and output standardized unified data views.
3. The rail transit security supervision system according to claim 1, wherein the multi-agent collaboration module comprises an alert analysis agent, a document assistant agent, a ticket analysis agent, and an electronic ledger analysis agent: the alarm analysis intelligent body realizes alarm analysis, risk prediction and patrol optimization based on the BERT variant deep learning model and ARIMA/LSTM time sequence analysis; the document assistant agent provides legal searching, document generation and knowledge base management functions by means of natural language processing technology; the ticket card analysis agent realizes passenger flow anomaly detection and personnel track association analysis through K-means clustering and DBSCAN algorithm; the electronic ledger analysis agent identifies equipment hidden trouble and generates correction advice based on ALBERT fine tuning models and rule engines.
4. The rail regional traffic safety supervision system according to claim 1, wherein the intelligent decision support module comprises an AI large model unit, a knowledge-graph unit, and a multi-modal interaction unit: The AI large model unit calls DeepSeek-R1-Qwen-32B model int4 quantized version, supports natural language processing and multi-modal interaction; the knowledge graph unit builds a quaternary relation graph of equipment-hidden danger-police condition-site; The multi-mode interaction unit supports text and picture interaction and provides visual BI view and decision suggestion display.
5. A method for rail regional traffic safety supervision using the system of claim 1, comprising the steps of: 1) Deploying public security network special hardware equipment, and building Ubuntu22.04 operating system, mySQL database and DeepSeek large model operating environment; 2) Collecting police, ticket and standing account data through a customized access protocol, and constructing a unified data analysis view after cleaning and conversion; 3) Each functional intelligent agent respectively completes police situation research and judgment, official document generation, passenger flow early warning and hidden danger identification, and outputs a special analysis result; 4) Generating decision suggestions based on the AI large model and the knowledge graph, and displaying the decision suggestions through a multi-modal interface; 5) And triggering an event bus to complete work order circulation and effect feedback of the abnormal event, and realizing full-flow supervision closed loop.
6. The method of claim 5, wherein in step 3), the processing flow of the alert analysis agent comprises: 3-a-1) accessing 110 police condition data in real time, and extracting characteristics of patrol areas, time and police condition categories through NER technology; 3-a-2) identifying repeated alarms based on SimHash algorithm, and judging important alarm conditions through XGBoost model; 3-a-3) pushing and manually rechecking the irregular warning situation according to the priority, and finally importing the irregular warning situation into a structured warning situation data pool; 3-a-4) predicting warning trend by using a Prophet model, and generating a patrol scheme by combining an integer programming algorithm.
7. The method according to claim 5, wherein in step 3), the document assistant agent includes a knowledge base management layer, a writing process layer and a business closed loop layer, and is capable of calling periodic statistics report data of the police analysis agent, hidden danger data of the standing book analysis agent, and passenger flow early warning data of the ticket analysis agent, so as to generate a standardized document adapted to a traffic safety supervision scene of a track area, and the specific workflow is as follows: 3-b-1) the knowledge base management layer realizes knowledge asset storage and retrieval through a relational database and a vector database, firstly, a user executes knowledge addition, modification, deletion and query operations in a knowledge base management background, then text is converted into a high-dimensional semantic vector based on Embedding capacity of a BERT pre-training model, and knowledge semantic level accurate retrieval is realized through a cosine similarity algorithm, so that a knowledge base supporting document creation is formed; 3-b-2) the writing flow layer receives the document creation basic information input by a user, generates a writing outline and supports closed loop confirmation of the user after retrieving related content from a knowledge base, then invokes an AI large model based on a Transformer framework to generate an article first draft, and then carries out scene splitting on the first draft and respectively processes four scenes of editing writing, light writing, overview writing and tool writing, wherein the overview writing adopts a TextRank text abstract algorithm to extract key information of the first draft, the tool writing completes variable replacement through a template filling algorithm, and finally carries out unified color rendering and batch derivation on multi-scene content; 3-b-3) the business closed loop layer judges the effectiveness of the output document results through a cosine similarity algorithm or a random forest classification model, and transmits the effective results back to a knowledge base to finish knowledge precipitation, so that a forward business closed loop of 'knowledge maintenance-creation-precipitation' is formed.
8. The method of claim 5, wherein in step 3), the ticket analysis agent passenger flow pre-warning method comprises: 3-c-1) adopting an isolated forest algorithm to count real-time passenger flow data at fixed time intervals, identifying abnormal conditions deviating from a normal passenger flow mode, and providing support for supervision, early warning and emergency treatment; 3-c-2) identifying the tracks of the staffs of the same person through DBSCAN clustering; 3-c-3) dynamically adjusting an early warning threshold according to the formula early warning score = current passenger flow/history peak value x hidden danger risk coefficient, and realizing risk superposition early warning.
9. The method according to claim 5, wherein in step 3), the hidden danger identification process of the electronic ledger analysis agent is: 3-d-1) adopting an ETL tool to integrate 15 types of ledgers into 6 business big tables, and establishing a retrieval index; 3-d-2) extracting equipment and hidden danger entities through ALBERT models; 3-d-3) constructing a hidden danger-warning situation association map based on GRAPHSAGE algorithm, and realizing intelligent SQL query and treatment closed-loop tracking.
10. The method of claim 5, wherein in step 4), the multi-agent linkage logic of the intelligent decision support module is: 4-1) calling ticket passenger flow data and standing account hidden danger data by the police analysis agent to finish multidimensional risk study and judgment; 4-2) automatically generating a standardized document by combining the analysis results of other intelligent agents by the document assistant intelligent agent. 4-3) All the process documents reflow the knowledge base to form a knowledge closed loop of analysis-decision-precipitation.

Description

Rail area traffic safety supervision system and method based on AI large model Technical Field The invention relates to the field of rail transit supervision, in particular to a rail regional traffic safety supervision system and method based on an AI large model. Background Along with the continuous expansion of the scale of urban rail transit, rail transit has become an important traffic mode for daily travel of urban residents. In order to ensure normal operation of rail transit and passenger safety, effective supervision of rail transit is required. When facing complicated changeable operation scene, because traditional rail transit supervision mainly relies on the manual work to go on, mainly there is cost of labor height, inefficiency, response speed and accuracy subalternation problem: 1. And the data island and the response are lagged, the risk early warning is triggered manually, and the equipment failure and the passenger flow mutation are difficult to respond in a linkage way. The centralized architecture results in independent storage of multi-source data, and static batch modes cannot analyze dynamic risk in real time. In other words, the data architecture of the existing rail transit supervision system presents a serious island state, the cooperation among all service subsystems is difficult, the data transfer is carried out manually, the overall operation efficiency is low, and the interconnection, intercommunication and sharing utilization of the data are difficult to realize; 2. Unstructured text processing bottleneck, namely manual verification is needed for alert description, complex decisions still depend on experience, and misclassification risks are high. Traditional NLP techniques have insufficient resolution capability for unstructured text (e.g., alert descriptions) and lack of automated decision support tools, still driven empirically. In other words, when the traditional rule engine scheme processes complex and changeable operation scenes, the response speed and accuracy are insufficient, obvious shortboards exist in the system load capacity, and the requirements of real-time analysis and decision making are difficult to meet; 3. And the manual dependence and operation redundancy are that the functional module is cracked, and an intelligent interactive interface is not needed, so that a worker needs to manually fill in an analysis report. In other words, the existing system is highly dependent on manual intervention, has excessively high operation specialization requirement, excessively long emergency response period and high labor cost, is easily influenced by human factors, and reduces the overall operation efficiency of the system; 4. The passive response and simulation are absent, the risk early warning depends on manual triggering, the event handling takes more than 2 hours on average, and prospective prevention and control are absent. In other words, the existing system lacks an effective intelligent decision support tool, cannot automatically complete complex analysis and decision tasks, and has the defects of insufficient accuracy and timeliness of analysis results, so that intelligent management of rail transit is greatly influenced. The nature of these problems is derived from long-term studies by those skilled in the art from structural contradictions between static technical architecture and dynamic business requirements, and serious inadequacies in the degree of algorithmic intelligence. Therefore, various intelligent rail transit monitoring systems are developed in combination with the AI technology, and attempt to thoroughly solve the data splitting problem (such as difficulty in cooperation among various service subsystems, low overall operation efficiency caused by manual data transfer), the problem of insufficient adaptive capacity to complex and variable operation scenes, obvious short board problem of system load capacity (because of slow response of the adopted traditional rule engine scheme), and the outstanding contradiction of high dependence on manual intervention, excessive operation specialization requirements, overlong emergency response period and the like exposed in practical application. For example, CN118644054a discloses a rail transit passenger flow early warning method and system based on an edge flow processing engine, and the system implements passenger flow data processing and analysis with huge data volume and high real-time requirements at a position close to a data source by a real-time data processing module, a large-scale data processing module, a multi-source data fusion module and an intelligent passenger flow early warning module, without transmitting passenger flow data to a cloud or a remote data center first and then processing the passenger flow data. However, when the system calculates passenger flow data at the edge side, the problems of low algorithm efficiency and long calculation time still exist. CN117459559a discloses a syst