Transportation hub large passenger flow emergency decision method and system, readable storage medium

Abstract

The application provides a traffic hub large passenger flow emergency decision-making method and system based on swarm intelligence emergence and a large model, and a readable storage medium, the system comprises: a distributed sensing module responsible for edge intelligence processing at the data source; a swarm intelligence emergence module responsible for forming a global control strategy through local interaction between intelligent agents; a large model meta-intelligent agent evaluation module responsible for macro monitoring and evolution guidance of the intelligent agent group; and an emergency decision-making linkage execution module. The application has the beneficial effects that: the mode change from traditional single center decision-making to distributed collaborative autonomous decision-making is realized, the robustness, adaptability and response real-time performance of the system in a dynamic uncertain environment are significantly enhanced, and finally a closed-loop autonomous emergency response system integrating sensing, decision-making, execution and optimization is formed, and the intelligent level and operation safety guarantee capability of high-speed rail station large passenger flow management are comprehensively improved.

Description

TECHNICAL FIELD

[0001] The application relates to a passenger flow emergency decision method, in particular to a traffic hub large passenger flow emergency decision method and system based on swarm intelligence emergence and a large model and a readable storage medium. BACKGROUND

[0002] A subway station large passenger flow emergency command linkage method, equipment and storage medium (CN202411876448.X): the patent provides a subway station large passenger flow emergency command linkage method. Its core lies in that machine learning algorithm (such as LSTM) is used to fuse historical and real-time data (including video image, AFC card data, train weighing data) to predict short-term passenger flow (future 15 minutes) and the number of people getting off at the station, and multi-level warning (red, orange, yellow) is carried out based on the preset threshold. After confirming the alarm, the system can automatically or semi-automatically start the emergency plan, guide the passengers through PIS, broadcast, BIM heat map visualization and other ways, and can generate train adjustment strategy (such as adding trains). The technology emphasizes real-time prediction, multi-level alarm and linkage response.

[0003] The core defect of the existing system (a subway station large passenger flow emergency command linkage method, equipment and storage medium CN202411876448.X) of the patent lies in the rigidity and centralization of the decision mechanism. Although it uses machine learning algorithm to realize relatively accurate passenger flow prediction, the subsequent emergency response completely depends on "preset threshold" and "pre-prepared emergency plan". This is a typical "if-else" rule system, and its intelligent level is only reflected in the front-end prediction, while the decision and execution of the middle and back-end are seriously dependent on artificial preset logic, which cannot cope with complex and variable sudden scenes (for example, a composite event of multiple regional simultaneous different types of passenger flow congestion). The whole system is centrally commanded by a central brain, which has the risk of single point failure, and all decisions need to go through the "confirmation alarm" link, so the response chain is long and the execution is slow in the actual emergency scene.

[0004] Railway passenger flow prediction method and system based on knowledge graph and multi-modal large model (CN202510613433.2): This patent proposes a railway passenger flow prediction method and system based on knowledge graph and multi-modal large model. Its innovation lies in the explicit introduction of knowledge graph technology to extract entities and relationships from multi-source heterogeneous data (historical passenger flow, weather, economic indicators, social public opinion), and support dynamic updating through small sample learning. At the same time, the system uses a multi-modal large model (supports text, image, time series data) to associate and reason (based on retrieval-screening-fusion strategy) real-time data and knowledge graph, output passenger flow prediction results, and provide scheduling optimization suggestions (such as capacity allocation, ticket price adjustment). This technology significantly improves the multi-dimensionality and interpretability of the prediction, and is the closest existing technology to your patent design idea, but its focus is still on the prediction and suggestion stage, and it does not extend to the closed-loop linkage and real-time decision automation level of emergency command execution.

[0005] Existing system (railway passenger flow prediction method and system based on knowledge graph and multi-modal large model CN202510613433.2): The core function of this patent is still limited to the "prediction" and "suggestion" level, and it has not formed a closed loop of decision-making and execution. It perfectly solves the problem of multi-source data fusion and interpretable prediction, but the "scheduling optimization suggestions" output still need to be judged and operated by human experts. The large model is only a reference decision-making role, not a self-control terminal that can directly command action. The entire system is an "open loop" system, which cannot automatically convert intelligent decisions into control instructions and issue them to the execution terminal (such as broadcast, gate, indication system), so it cannot realize real "emergency linkage" automation.

[0006] A subway large passenger flow scenario emergency command system (CN202410020231.2): This patent relates to a subway large passenger flow scenario emergency command system, which consists of four modules: intelligent perception, analysis and judgment, early warning and prediction, and contact disposal. The system integrates multi-source data (AFC, video monitoring, BIM, GIS, operation data), and through passenger flow prediction engine, dense crowd monitoring engine, risk assessment model, etc. for comprehensive analysis, realizes passenger flow prediction and early warning. Its characteristics are the use of digital twin and three-dimensional visualization technology (based on BIM+GIS) to intuitively display the line network and station heat map on the command screen, and has a preplan recommendation function (such as flow limiting, adjusting the interval between trains). This system embodies the closed-loop idea of multi-source perception, visualized analysis and preplan recommendation, but it does not introduce knowledge graph construction to build a knowledge system, nor does it use large models for cross-modal semantic reasoning and intelligent decision-making generation. The preplan recommendation is still based on rule base and traditional model, and the degree of intelligence is limited.

[0007] The existing system (a subway large passenger flow scene emergency command system CN202410020231.2) constructs a digital twin visualization command platform with rich functions, but its intelligent kernel is relatively traditional. The "analysis and judgment" and "early warning and prediction" modules are based on conventional risk assessment models and prediction engines, and the "preplan recommendation" also depends on a traditional rule matching library. The whole system lacks an intelligent center capable of deep semantic understanding, complex reasoning and autonomous decision-making. Digital twin mainly plays a visualization role here, improving the situation awareness capability, but it does not fundamentally improve the intelligent level of decision-making. SUMMARY

[0008] In order to solve the problems in the prior art, the application provides a traffic hub large passenger flow emergency decision method and system based on swarm intelligence emergence and large model, and a storage medium, aiming at solving the problems of centralized decision architecture, weak system self-adaptability, slow real-time response, and disconnection between prediction and execution in the existing emergency command technology.

[0009] The application provides a traffic hub large passenger flow emergency decision system based on swarm intelligence emergence and large model, comprising:

[0010] A distributed sensing module is responsible for edge intelligence processing at the data source. The distributed sensing module completes the collection, cleaning, feature extraction and state judgment of local data in parallel through the intelligent agents deployed on each physical facility of the traffic hub, and generates a concise local state summary to provide a sensing basis for the swarm intelligence emergence module;

[0011] A swarm intelligence emergence module is responsible for forming a global control strategy through local interaction between intelligent agents. The swarm intelligence emergence module receives the local state summary of each intelligent agent, and based on the pre-defined interaction rules and reward functions, performs high-speed information exchange and negotiation in the intelligent agent network, and finally emerges a set of consistent distributed collaborative decision instructions;

[0012] A large model meta-intelligent agent evaluation module is responsible for macro monitoring and evolution guidance of the intelligent agent group. The large model meta-intelligent agent evaluation module evaluates the overall efficiency of group decision-making from a global perspective beyond local interaction, and when the group behavior is inefficient, optimizes the underlying interaction rules through simulation and deduction, realizing continuous self-improvement of the system;

[0013] An emergency decision linkage execution module is responsible for converting the instructions confirmed by the swarm intelligence emergence module or the large model meta-intelligent agent evaluation module into actual control actions. The emergency decision linkage execution module provides a unified, reliable and safe interface, parses and issues abstract decision instructions to various heterogeneous execution terminals, and ensures the atomicity and traceability of the instructions.

[0014] As a further improvement of the application, the distributed perception module is implemented by deploying lightweight agent containers on each edge computing node; a dedicated agent software instance is defined for each type of facility within the distributed perception module; each agent is built-in with a data acquisition adapter that reads raw data from local sensors, controllers or system logs in real time; after the raw data enters the agent, it first passes through a local data cleaning submodule, which uses a lightweight rule engine or a miniature anomaly detection model to eliminate transient outliers; then it enters a feature extraction and state judgment submodule, which is preloaded with state judgment logic related to the function of the facility; finally, through a state summary generation submodule, the judgment result is packaged into a standardized JSON message body, which contains the agent ID, timestamp, device type, own coordinates, current state code and confidence level; each agent publishes the local state summary to the message topic through a lightweight communication protocol for the swarm intelligence emergence module to receive and process.

[0015] As a further improvement of the application, the feature extraction and state judgment submodule preloads state judgment logic related to the function of the facility, including:

[0016] a gate agent, whose built-in algorithm judges its state as "smooth", "busy" or "congested" based on the passage rate and the number of people in line;

[0017] a camera agent that analyzes video streams in real time through a built-in lightweight CNN model to output the crowd density level and movement speed in the current area.

[0018] As a further improvement of the application, the swarm intelligence emergence module is implemented by building an interaction bus on a high-performance message middleware and running a dynamic rule engine; the swarm intelligence emergence module first has an interaction network management submodule that is responsible for maintaining the network topology relationship of all active agents, defining the "neighbor" range of each agent, and each agent subscribing to the state topics of its "neighbors" and related agents; when an agent receives a state update of itself or other agents, it triggers a distributed decision submodule; the distributed decision submodule is built-in with a decision logic based on a reward function, and the agent calculates the expected reward value of a series of potential actions based on its own and its neighbors' current state; then, the negotiation and consensus submodule lets the agents reach a local consensus on the final action through a specific interaction protocol; finally, through an instruction generation submodule, the consensus-reached action is converted into a control instruction that can be directly issued and pushed to the instruction queue, and the entire decision-making process does not require central scheduling, but emerges global intelligence completely through parallel and local interaction.

[0019] As a further improvement of the present application, the large model meta-agent evaluation module is implemented by calling the API interface of the cloud large-scale pre-trained language model and combining the reinforcement learning simulation environment; a global situation monitoring and evaluation submodule is first provided in the large model meta-agent evaluation module, the global situation monitoring and evaluation submodule subscribes to the state summary and final decision instruction of all agents in real time, and constructs a global situation view; the decision effectiveness of the swarm intelligence is quantitatively evaluated by calculating a series of key performance indicators KPI, if the decision effectiveness is effective, the instruction is automatically issued to the execution terminal of the traffic hub for execution, if the decision effectiveness is inefficient or invalid, high-level intervention is triggered.

[0020] As a further improvement of the present application, the global situation monitoring and evaluation submodule maintains a global state diagram, and periodically calculates the global average evacuation rate GAER, the congestion point duration ratio CDR, the instruction conflict rate CCR, and the resource utilization variance RUV, assigns different weights to the above-mentioned key performance indicators KPI, and inputs a comprehensive evaluation function, the comprehensive evaluation function is shown in formula (1):

[0021] (1)

[0022] Wherein,

[0023] is a comprehensive health score;

[0024] is a weight coefficient, the value of which is determined by a hybrid decision-making process, the process includes: system initialization based on analytic hierarchy process, offline optimization based on reinforcement learning and digital twinning, and online adaptive fine-tuning based on monitoring the long-term correlation of each KPI indicator and overall effectiveness;

[0025] denotes the normalized value;

[0026] When the comprehensive health score is lower than the preset threshold, it is determined that the decision effectiveness of the swarm intelligence is inefficient or invalid, and high-level intervention is triggered.

[0027] The high-level intervention includes:

[0028] When it is monitored that the decision effectiveness of the swarm intelligence is inefficient, the global situation monitoring and evaluation submodule starts a simulation optimization environment based on digital twinning and reinforcement learning, the system automatically captures the current global state snapshot and the rule set R_old and reward function F_old currently used by all agents, the global state snapshot includes the state and position of all agents, on this basis, a high-fidelity simulation environment is quickly constructed, and the large model meta-agent evaluation module receives the following Prompt:

[0029] The current group agent's comprehensive health score is [Health_Score],

[0030] The main problem is [main KPI problem],

[0031] The current rule is [R_old],

[0032] The reward function is [F_old],

[0033] Please analyze the root cause of the problem and generate new reward function candidates that may optimize the group behavior, which encourages [including: load balancing, rapid evacuation],

[0034] Among them,

[0035] The reward function introduces the "long-term return" and "coordination penalty" mechanism, and the reward function is shown in formula 2.

[0036] (2)

[0037] Among them,

[0038] represents the expected reward value,

[0039] Q_i(t) represents the local queue length of the agent at time t,

[0040] Q_i(t+1) represents the change in the agent's own traffic in unit time,

[0041] represents the increase in traffic, i.e. deterioration, and gives a negative reward,

[0042] represents the decrease in traffic, i.e. improvement, and gives a positive reward,

[0043] Guide the agent to focus on trends rather than just the present situation;

[0044] represents the variance of the queue length of the neighbor agent within the communication range of the agent, the larger the variance, the worse the coordination effect, and the greater the penalty,

[0045] is the weight coefficient, which is determined by the large model meta-agent evaluation module in simulation deduction.

[0046] The emergency decision linkage execution module is realized by deploying an enterprise service bus or an API gateway, and encapsulates all differences of execution systems internally; the emergency decision linkage execution module is internally provided with a first instruction analysis and routing submodule, which receives an instruction queue from the swarm intelligence emergence module or the large model meta-intelligent agent evaluation module, the instruction contains a target device ID, an action type and parameters, and an analyzer converts the standardized instruction into a specific protocol and control instruction recognizable by a downstream execution subsystem according to the device ID and the action type through a preset device driver adapter submodule; subsequently, a safe execution and transaction management submodule is responsible for reliable delivery of the instruction, adopts an asynchronous confirmation mechanism, ensures that the success or failure state of the instruction execution is accurately recorded, and the emergency decision linkage execution module is provided with an execution feedback collection submodule, which actively pulls or receives state feedback from each execution terminal, and packages and publishes the execution effect data to a specified feedback topic to form a closed loop and provide data basis for system optimization.

[0047] The application further provides a traffic hub large passenger flow emergency decision method based on swarm intelligence emergence and a large model.

[0048] S1, the intelligent agents deployed on each physical facility of the traffic hub independently complete data acquisition, cleaning and state summary generation based on local rules;

[0049] S2, each intelligent agent performs rapid local negotiation with only adjacent nodes within a communication range based on simple rules and a reward function;

[0050] S3, through the local interaction of step S2, global collaborative decision instructions are emerged from bottom to top to form a distributed collaborative decision instruction set; at the same time, the large model meta-intelligent agent evaluation module monitors the global situation in the cloud and evaluates the effectiveness of the distributed collaborative decision instruction set; if the distributed collaborative decision instruction set is effective, the instructions are automatically issued to each execution terminal of the traffic hub for execution, and the execution effect is fed back to the large model meta-intelligent agent evaluation module; if the distributed collaborative decision instruction set is inefficient or invalid, the large model meta-intelligent agent evaluation module starts a high-order intervention mechanism, optimizes the interaction rules and the reward function of the intelligent agents through simulation and deduction, and dynamically issues updates, and the updated rules guide the intelligent agent group to generate a more optimal distributed collaborative decision instruction set, thereby forming a perception-decision-execution-optimization closed loop autonomous response system.

[0051] The application further provides a readable storage medium, wherein the readable storage medium stores execution instructions, and the execution instructions are executed by a processor to implement the method.

[0052] The application has the beneficial effects that the application provides a traffic hub large passenger flow emergency decision method and system based on swarm intelligence emergence and a large model, a decentralized distributed intelligent decision architecture is constructed, the agents deployed in each key physical facility interact with each other according to local information, and a global emergency control strategy emerges from bottom to top; and a large model is introduced as a meta-agent, the online monitoring and rule optimization of the group behavior mode are realized, the overall decision effectiveness is improved and high-order guidance is realized; the mode is changed from traditional single center decision to distributed collaborative autonomous decision, the robustness, adaptability and response real-time of the system in a dynamic uncertain environment are significantly enhanced, and finally a closed-loop autonomous emergency response system integrating perception, decision, execution and optimization is formed, and the intelligent level and operation safety guarantee capability of high-speed rail station large passenger flow management are comprehensively improved. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other solutions can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0054] Figure 1 is a flow chart of a traffic hub large passenger flow emergency decision method based on swarm intelligence emergence and a large model of the application;

[0055] Figure 2 is a whole architecture diagram of a traffic hub large passenger flow emergency decision system based on swarm intelligence emergence and a large model of the application;

[0056] Figure 3 is a running flow chart of a swarm intelligence emergence module;

[0057] Figure 4 is a running flow chart of a large model meta-agent evaluation module;

[0058] Figure 5 is a running flow chart of an emergency decision linkage execution module. DETAILED DESCRIPTION

[0059] It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0060] In the description of the present application, it needs to be understood that the terms "center", "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the scope of protection of the present application. In addition, the terms "first", "second" and the like are only for the purpose of description and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined with "first", "second" and the like can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0061] In the description of the present application, it needs to be explained that, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting" should be understood in a broad sense, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be directly connected, or indirectly connected through an intermediate medium; it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood through specific circumstances.

[0062] The present application will be further described below in conjunction with the description and specific embodiments of the drawings.

[0063] As shown in the drawings, a high-speed rail station large passenger flow emergency decision method based on swarm intelligence emergence and large model includes the following processes: Figure 1

[0064] The agent (also known as facility agent) deployed on each key facility of the high-speed rail station independently completes data collection, cleaning and state summary generation based on local rules;

[0065] Subsequently, each agent performs rapid local negotiation with only the adjacent nodes within the communication range based on simple rules and reward functions;

[0066] ​Through such local interaction, global collaborative decision-making instructions emerge from the bottom up, forming a distributed collaborative decision-making instruction set; At the same time, the large model meta-intelligent agent evaluation module monitors the global situation in the cloud and evaluates the effectiveness of group decision-making; If the group decision-making is effective, the instructions are automatically issued to each execution terminal (such as the passenger guidance system, operation scheduling system, and station equipment system) for execution, and the execution effect is fed back to the system; If the group decision-making is inefficient or ineffective, the meta-intelligent agent starts the high-level intervention mechanism, optimizes the interaction rules and reward functions of the intelligent agent through simulation and deduction, and dynamically issues updates; The updated rules guide the intelligent agent group to produce better decisions, thus forming a closed-loop operation system that continuously optimizes and adapts.

[0067] As shown in Figure 2 A high-speed rail station large passenger flow emergency decision-making system based on swarm intelligence emergence and large model is used to implement the above-mentioned high-speed rail station large passenger flow emergency decision-making method based on swarm intelligence emergence and large model. The high-speed rail station large passenger flow emergency decision-making system based on swarm intelligence emergence and large model is built by constructing a technical architecture that integrates distributed sensing, collaborative interaction, intelligent emergence, and meta-cognition optimization, aiming to solve the problems of centralized decision-making bottleneck, weak self-adaptability, and real-time response lag of existing emergency systems. The system mainly includes a distributed sensing module, a swarm intelligence emergence module, a large model meta-intelligent agent evaluation module, and an emergency decision-making linkage execution module. The system first performs local data collection and edge computing through intelligent agents deployed in various physical facilities to form distributed situational awareness; Then each intelligent agent interacts and negotiates with neighboring nodes based on a pre-defined reward function, and emerges a global diversion strategy through self-organization; The large model meta-intelligent agent continuously monitors the effectiveness of group behavior and starts a high-level intervention mechanism when the strategy is inefficient, optimizes the interaction rules of the intelligent agent through simulation and deduction, and dynamically issues updates; Finally, the effective emerging strategy is automatically issued to each execution terminal through an API interface, forming a closed-loop autonomous response system of sensing-decision-execution-optimization, significantly improving the robustness and self-adaptability of the system in dynamic environments.

[0068] The distributed perception module is responsible for edge intelligence processing at the data source. Its core function is to complete the collection, cleaning, feature extraction and state judgment of local data in parallel through the agents deployed on key physical facilities in the high-speed rail station, and generate a concise local state summary to provide the perception basis for group collaboration. This module is implemented by deploying lightweight agent containers (such as Docker) on each edge computing node (such as intelligent cameras, gate controllers, edge servers). The module defines a dedicated agent software instance for each type of facility, such as gate agent, broadcast agent, etc. Each agent has a specific data collection adapter built-in to read raw data from local sensors, controllers or system logs in real time (such as gate passage count, camera video stream, broadcast status signal). After the data enters the agent, it first passes through the local data cleaning submodule, which uses a lightweight rule engine (such as Drools Lite) or a micro anomaly detection model (such as a moving average filtering algorithm for sensor data) to remove transient outliers. Then it enters the feature extraction and state judgment submodule, which preloads state judgment logic related to the function of the facility (for example, for the gate agent, its built-in algorithm will judge its state as "smooth", "busy" or "congested" based on the passage rate and the number of people in line; for the camera agent, it analyzes the video stream in real time through a built-in lightweight CNN model to output the crowd density level and movement speed in the area). Finally, through the state summary generation submodule, the judgment result is packaged into a standardized JSON message body, which contains the agent ID, timestamp, device type, own coordinates, current state code and confidence. Each agent publishes the local state summary to a specific message topic (Topic) through a lightweight communication protocol (such as MQTT) for the group intelligence emergence module to receive and process.

[0069] As Figure 3As shown, the swarm intelligence emergence module is responsible for self-organizing a global control strategy through local interactions among agents. Its core function is to receive the state summaries of each agent, and based on predefined interaction rules and reward functions, conduct high-speed information exchange and negotiation in the agent network, and finally emerge a coordinated and consistent distributed execution instruction. This module is implemented by building an interaction bus on a high-performance message middleware (such as Redis Pub / Sub or RabbitMQ) and running a dynamic rule engine. Inside the module, there is first an interaction network management submodule, which is responsible for maintaining the network topology relationship of all active agents (such as defining the "neighbor" range of each agent). Each agent subscribes to the state topics of its "neighbors" and related agents. When an agent receives a state update of itself or other agents, it triggers the distributed decision-making submodule. This submodule has built-in decision logic based on the reward function, and the agent will calculate the expected reward value (Reward) of a series of potential actions based on its current state and the state of its neighbors. For example, a gate agent in a "congestion" state will calculate the reward value of actions such as "send a guidance request to display screen A", "send a guidance request to display screen B", or "maintain the current state", and the design goal of the reward function is to maximize global evacuation efficiency. Subsequently, the negotiation and consensus submodule lets the agents reach a local consensus on the final action through a specific interaction protocol (such as a market auction mechanism-based protocol or a gossip-based protocol). Finally, through the instruction generation submodule, the consensus action is converted into a control instruction that can be directly issued (such as {"target": "display_screen_01", "action": "show_arrow", "direction": "north"}), and pushed to the instruction queue. The entire decision-making process does not require central scheduling, and the global intelligence is completely emerged through parallel and local interactions.

[0070] As Figure 4As shown, the large model meta-agent module is responsible for macro monitoring and evolution guidance of swarm intelligence. Its core function is to evaluate the overall effectiveness of group decision-making from a global perspective beyond local interaction, and to optimize the underlying interaction rules when the group behavior is inefficient, thereby realizing the continuous self-improvement of the system. This module calls the API interface of the cloud large-scale pre-trained language model (such as GPT-4) and combines the reinforcement learning simulation environment (such as Ray framework) to achieve it. Inside the module, there is a global situation monitoring and evaluation submodule. This submodule subscribes to the state summary and final decision instructions of all agents in real time and builds a global situation view. It quantitatively evaluates the decision-making effectiveness of swarm intelligence by calculating a series of key performance indicators (KPIs). It maintains a global state graph and periodically calculates the global average evacuation rate (GAER), congestion duration ratio (CDR), command conflict rate (CCR), and resource utilization variance (RUV). The above KPIs are assigned different weights and input into a comprehensive evaluation function, which is shown in equation (1).

[0071] (1)

[0072] wherein is the weight coefficient, and its value is determined through a hybrid decision-making process, which includes: system initialization based on the analytic hierarchy process, offline optimization based on reinforcement learning and digital twinning, and online adaptive fine-tuning based on the long-term correlation between monitoring KPI indicators and overall effectiveness, represents the normalized value. When is below the preset threshold, it is determined that the group behavior is inefficient or ineffective, triggering high-level intervention.

[0073] When the group behavior is detected to be inefficient, this submodule initiates a simulation optimization environment based on digital twin and reinforcement learning (RL). The system automatically captures the current global state snapshot (including the state and position of all agents) and the current rule set R_old and reward function F_old that all agents are using. Based on this, a high-fidelity simulation environment is quickly built. The large model receives the following Prompt: the current group of agents has a health score of [Health_Score], the main problem is [main KPI problem, such as high CDR]. The current rules are [R_old], and the reward function is [F_old]. Please analyze the root cause of the problem and generate 3 new reward function candidates (F_candidate1, F_candidate2, F_candidate3) that may optimize the group behavior. The new reward function should encourage [e.g. load balancing, rapid evacuation]. Among them, the reward function introduces "long-term return" and "coordination penalty" mechanisms, and the reward function is shown in formula (2).

[0074] (2)

[0075] where, denotes the local queue length of the agent at time t. This term has a negative sign, indicating that the longer the queue (i.e. the higher the cost), the greater the penalty to the total reward, the change in the agent's own traffic volume per unit time. represents an increase in traffic (deterioration), giving a negative reward; represents a decrease in traffic (improvement), giving a positive reward. This guides the agent to focus on trends rather than just the present situation. denotes the variance of the queue length of neighboring agents within the communication range of the agent. The larger the variance, the worse the coordination effect, and the greater the penalty. is a weight coefficient determined by the large model in the simulation deduction.

[0076] For example, Figure 5As shown, the emergency decision linkage execution module is responsible for converting the instructions confirmed by the group emergence or meta-agent into actual control actions. Its core function is to provide a unified, reliable and safe interface to parse and issue abstract decision instructions to various heterogeneous execution terminals, and to ensure the atomicity and traceability of the instructions. This module is implemented by deploying an enterprise service bus (ESB) or an API gateway (such as Kong), which encapsulates the differences of all execution systems internally. Inside the module, there is first an instruction parsing and routing submodule, which receives the instruction queue from the group intelligence emergence module or the large model meta-agent module. The instruction contains the target device ID, action type and parameters. The parser converts the standardized instruction into a specific protocol and control instruction that can be recognized by the downstream execution subsystem (for example, converts {"action": "show_arrow"} into a TCP / IP control command sent to the specified PIS screen; converts the transport adjustment scheme into a Web Service API call request conforming to the specifications of the railway dispatching system). Subsequently, the safe execution and transaction management submodule is responsible for the reliable issuance of instructions, which adopts an asynchronous confirmation mechanism to ensure that the success or failure status of instruction execution is accurately recorded. At the same time, the module is equipped with an execution feedback collection submodule that actively pulls or receives state feedback from each execution terminal (such as the actual number of people passing through the gate, the current display content of the PIS screen), and packages and publishes this execution effect data to the specified feedback topic, forming a closed loop to provide data basis for system optimization.

[0077] The main embodiment of the present application is first to deploy a distributed facility agent network in the high-speed rail station digital twin platform, and configure an initial behavior strategy for each agent based on local perception and simple rules; then establish a management and control interface between the digital twin platform and the large model, and build a group interaction simulation environment for simulation training and rule optimization. On this basis, the agent interaction behavior is trained through historical passenger flow data and multi-modal real-time information, so that the large model can learn and optimize the group decision rules. After training, real-time sensor data, device status and passenger flow information in actual operation are connected to the agent network to promote the emergence of decision-making through local interaction in the real environment, while the large model is used to monitor and dynamically optimize the overall operation efficiency of the system. Finally, the scheduling instructions generated by the group intelligence agent are mapped to the digital twin visualization interface in real time, and are simultaneously issued to the execution terminals such as station broadcasting, guidance, gate, etc., forming a closed-loop linkage system of perception-decision-execution-optimization.

[0078] Compared with the static decision mode relying on centralized and preset rules in the prior art, the traffic hub large passenger flow emergency decision method and system based on swarm intelligence emergence and large model, and the storage medium provided by the application fundamentally overcome inherent defects such as decision rigidity, slow response and weak self-adaptive ability of the traditional system. The application sinks decision intelligence to each edge facility unit, enables the system to generate a globally optimized strategy through local interaction and self-organization, realizes the mode change from 'central command' to 'group cooperation', and greatly improves the real-time performance and robustness of emergency response. At the same time, the large model meta-agent as a high-order optimizer, through macroscopic monitoring and dynamic rule optimization of group behavior, gives the system the ability of continuous self-evolution, so that it can always maintain efficient decision-making in a complex dynamic environment. Finally, the application not only significantly improves the efficiency and safety of large passenger flow dredging, but also realizes the leap of the emergency management system from 'passive response' to 'autonomous intelligence', fully embodies the adaptability and evolvability that the intelligent system should have.

[0079] The above is a further detailed description of the application in combination with specific preferred embodiments, and the specific implementation of the application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the application belongs, without departing from the concept of the application, a number of simple deductions or substitutions can be made, which should be regarded as falling within the protection scope of the application.

Claims

1. A traffic hub large passenger flow emergency decision system, characterized in that, Comprise: A distributed sensing module responsible for edge intelligence processing at the data source; the distributed sensing module completes the collection, cleaning, feature extraction and state judgment of local data in parallel through intelligent agents deployed on various physical facilities in the transportation hub, and generates a concise local state summary to provide a sensing basis for the swarm intelligence emergence module; A swarm intelligence emergence module responsible for forming a global control strategy through local interaction between intelligent agents; The swarm intelligence emergence module receives local state summaries from each intelligent agent, and based on predefined interaction rules and reward functions, exchanges and negotiates information at high speed in the intelligent agent network, and finally emerges a set of consistent distributed collaborative decision instructions; A large model meta-agent evaluation module responsible for macro monitoring and evolution guidance of the intelligent agent group; the large model meta-agent evaluation module evaluates the overall effectiveness of group decision-making from a global perspective, and when group behavior is inefficient, optimizes the underlying interaction rules through simulation and deduction, achieving continuous self-improvement of the system; An emergency decision linkage execution module responsible for converting instructions confirmed by the swarm intelligence emergence module or the large model meta-agent evaluation module into actual control actions; the emergency decision linkage execution module provides a unified, reliable and safe interface, parses and issues abstract decision instructions to various heterogeneous execution terminals, and ensures the atomicity and traceability of the instructions; The large model meta-agent evaluation module is implemented by calling the API interface of the cloud large-scale pre-trained language model and combining the reinforcement learning simulation environment; the large model meta-agent evaluation module first has a global situation monitoring and evaluation submodule, which subscribes to the state summary and final decision instructions of all intelligent agents in real time and builds a global situation view; the decision effectiveness of the swarm intelligence is quantitatively evaluated by calculating a series of key performance indicators KPI, if the decision effectiveness is effective, the instructions are automatically issued to the execution terminals of the transportation hub for execution, if the decision effectiveness is inefficient or invalid, high-level intervention is triggered; The global situation monitoring and evaluation submodule maintains a global state graph and periodically calculates the global average evacuation rate GAER, the congestion point duration ratio CDR, the instruction conflict rate CCR, and the resource utilization variance RUV, assigns different weights to the above key performance indicators KPI, and inputs a comprehensive evaluation function, the comprehensive evaluation function is shown in formula (1): (1) Wherein, Comprehensive health score; are weight coefficients, whose values are determined by a hybrid decision-making process, which includes: system initialization based on analytic hierarchy process, offline optimization based on reinforcement learning and digital twinning, and online adaptive fine-tuning based on monitoring the long-term correlation of each KPI indicator and overall performance; represents the normalized value; When the comprehensive health score When the comprehensive health score is lower than the preset threshold, it is determined that the decision effectiveness of the swarm intelligence is inefficient or ineffective, triggering high-level intervention.

2. The mass passenger flow emergency decision system for traffic hub according to claim 1, characterized in that: The distributed perception module is implemented by deploying lightweight agent containers on each edge computing node; the distributed perception module defines a dedicated agent software instance for each type of facility; each agent has a data acquisition adapter built-in, which reads raw data from local sensors, controllers or system logs in real time; after the raw data enters the agent, it first passes through the local data cleaning submodule, which uses a lightweight rule engine or a miniature anomaly detection model to eliminate transient outliers; then it enters the feature extraction and state judgment submodule, which has pre-installed state judgment logic related to the function of the facility; finally, through the state summary generation submodule, the judgment result is packaged into a standardized JSON message body, which contains the agent ID, timestamp, device type, own coordinates, current state code and confidence; each agent publishes the local state summary to the message topic through a lightweight communication protocol for the swarm intelligence emergence module to receive and process.

3. The mass passenger flow emergency decision system for traffic hub according to claim 2, characterized in that: The feature extraction and state judgment submodule pre-installed state judgment logic includes: Gate agent, whose built-in algorithm judges its state as "smooth", "busy" or "congested" based on the passage rate and the number of people in line; Camera agent, which analyzes video streams in real time through a built-in lightweight CNN model to output the crowd density level and movement speed in the current area.

4. The mass passenger flow emergency decision system for traffic hub according to claim 1, characterized in that: The swarm intelligence emergence module is implemented by building an interaction bus on a high-performance message middleware and running a dynamic rule engine; the swarm intelligence emergence module first has an interaction network management submodule, which is responsible for maintaining the network topology relationship of all active agents, defining the "neighbor" range of each agent, and each agent subscribing to the state topics of its "neighbors" and related agents; when an agent receives a state update of itself or other agents, it triggers the distributed decision submodule; the distributed decision submodule has built-in decision logic based on a reward function, and the agent calculates the expected reward value of a series of potential actions based on its current state and the state of its neighbors; then, the negotiation and consensus submodule lets the agents reach a local consensus on the final action through a specific interaction protocol; finally, through the instruction generation submodule, the consensus reached action is converted into a control instruction that can be directly issued and pushed to the instruction queue; the whole decision-making process does not require central scheduling, and the global intelligence emerges completely through parallel and local interaction.

5. The mass passenger flow emergency decision system for traffic hub according to claim 1, characterized in that: High-level intervention includes: When the decision-making effectiveness of the swarm intelligence is monitored to be inefficient, the global situation monitoring and evaluation submodule starts a simulation optimization environment based on digital twinning and reinforcement learning; the system automatically captures the current global state snapshot and the current rule set R_old and reward function F_old used by all agents; the global state snapshot includes the state and position of all agents; based on this, a high-fidelity simulation environment is quickly built; the large model meta-agent evaluation module receives the following Prompt: The comprehensive health score of the current swarm intelligence agent is [Health_Score], The main problem is [main KPI problem], The current rule is [R_old], The reward function is [F_old], Please analyze the root cause of the problem and generate new reward function candidates that may optimize group behavior, new reward functions encourage [include: load balancing, rapid evacuation], Among them, The reward function introduces "long-term return" and "synergy penalty" mechanisms, and the reward function is shown in formula (2); (2) Among them, Rt= represents the immediate reward value for the action taken by the agent at time step t, Let L(t) denote the local queue length of the agent at time t, The change in the agent's own flow over a unit of time. represents an increase in traffic, i.e. deterioration, giving a negative reward, representing a decrease in flow, i.e. an improvement, giving a positive reward, Guide the agent to pay attention to the trend rather than the present situation; denotes the variance of the queue length of the neighboring agents within the communication range of the agent, the larger the variance, the worse the coordination effect, the larger the punishment, are determined by the large model meta-agent evaluation module in the simulation deduction.

6. The mass passenger flow emergency decision system for traffic hub according to claim 1, characterized in that: The emergency decision linkage execution module is realized by deploying an enterprise service bus or an API gateway, encapsulating all the differences of the execution system internally; The instruction analysis and routing submodule is first provided in the emergency decision linkage execution module, which receives the instruction queue from the swarm intelligence emergence module or the large model meta-agent evaluation module, and the instruction contains the target device ID, action type and parameter. The parser converts the standardized instruction into a specific protocol and control instruction that can be recognized by the downstream execution subsystem through the preinstalled device driver adapter submodule according to the device ID and action type; Subsequently, the safe execution and transaction management submodule is responsible for the reliable delivery of instructions, which adopts an asynchronous confirmation mechanism to ensure that the success or failure status of instruction execution is accurately recorded. At the same time, the emergency decision linkage execution module is provided with an execution feedback collection submodule, which actively pulls or receives state feedback from each execution terminal and packages and publishes the execution effect data to the designated feedback topic, forming a closed loop to provide data basis for system optimization.

7. A traffic hub mass flow emergency decision method, characterized in that, The traffic hub large-flow emergency decision system according to any one of claims 1-6 performs the following steps: S1, the agent deployed on each physical facility in the traffic hub independently completes data collection, cleaning and state summary generation based on local rules; S2, each agent based on simple rules and reward functions only conducts rapid local negotiation with adjacent nodes within the communication range; S3, through the local interaction of step S2, a global collaborative decision instruction set is emerged from bottom to top, and at the same time, the large model meta-agent evaluation module monitors the global situation in the cloud and evaluates the effectiveness of the distributed collaborative decision instruction set; if the distributed collaborative decision instruction set is effective, the instruction is automatically issued to each execution terminal in the traffic hub for execution, and the execution effect is fed back to the large model meta-agent evaluation module; if the distributed collaborative decision instruction set is inefficient or invalid, the large model meta-agent evaluation module starts a high-order intervention mechanism, optimizes the interaction rules and reward functions of the agent through simulation and deduction, and dynamically issues updates. The updated rules guide the agent group to generate a more optimal distributed collaborative decision instruction set, thereby forming a closed-loop autonomous response system of perception-decision-execution-optimization.

8. A readable storage medium characterized by: The readable storage medium stores execution instructions, and the execution instructions are executed by the processor to implement the method of claim 7. The readable storage medium stores execution instructions, and the execution instructions are executed by the processor to implement the method of claim 7.

Images