Methods and systems for implementing intelligent agents at the station level in rail transit

Abstract

This application relates to the field of data communication technology and discloses a method and system for implementing a station-level intelligent agent in rail transit. The method includes: constructing a specialized agent base using an ARM processor, designing a LiteIP secure communication bus based on an FPGA, building a multi-core CPU cluster to construct the intelligent agent's central hub, establishing a real-time and entity database to form a data foundation, integrating federated learning and neural network-based decision generation, and executing edge computing to achieve station-level autonomous control. This application achieves millisecond-level data exchange and intelligent collaborative control within the station-level system, significantly improving the system's response speed and processing efficiency in emergency situations.

Description

Technical Field

[0001] This application relates to the field of data communication technology, and in particular to a method and system for implementing a station-level intelligent agent in rail transit. Background Technology

[0002] With the increasing scale and complexity of urban rail transit systems, traditional station-level control systems face numerous challenges. Existing rail transit station-level systems typically employ a distributed architecture, deploying various specialized subsystems within the station (such as Building Automation System (BAS), Fire Alarm System (FAS), Passenger Information System (PIS), Public Address System (PA), and Platform Screen Door System (PSD)). Limited data exchange between these systems is achieved through standard protocols such as Modbus and BACnet, or proprietary protocols. These systems are generally built on traditional programmable logic controllers (PLCs) or industrial control computers, relying on pre-defined logic rules for control, and typically require coordination and monitoring through a central system. Regarding data communication, existing systems mostly use the traditional TCP / IP protocol stack, transmitting data via Ethernet or industrial buses. The data exchange process involves multiple connection establishment, data transmission, and disconnection processes, resulting in low communication efficiency.

[0003] However, this traditional architecture has significant shortcomings: First, the various specialized subsystems form "information silos," hindering efficient data sharing and resulting in poor inter-system collaboration; second, communication latency is high (typically on the order of hundreds of milliseconds), making it difficult to meet the demands of rapid response in emergencies; third, the lack of a unified data processing platform makes it difficult to achieve intelligent analysis and decision-making across systems; fourth, system scalability is limited, making it difficult to add new functions or integrate new equipment; fifth, the security mechanisms are simplistic, making it difficult to cope with increasingly complex cybersecurity threats; and finally, the system's adaptive capabilities are weak, making it unable to dynamically adjust control strategies based on actual operating conditions. These shortcomings lead to slow system response in emergencies such as fires, large passenger flows, and equipment failures, hindering efficient and intelligent collaborative handling and impacting the safety and service quality of rail transit operations. Summary of the Invention

[0004] One objective of this application is to provide a method and system for implementing intelligent agents at the station level in rail transit, which enables millisecond-level data exchange and intelligent collaborative control within the station-level system, significantly improving the system's response speed and processing efficiency in emergency situations.

[0005] Firstly, this application provides a method for implementing a station-level intelligent agent in rail transit, comprising: constructing a general professional agent base, configuring a real-time embedded operating system, and forming a modular professional agent architecture; designing a secure communication bus based on the professional agent architecture, performing differentiated data transmission processing, and establishing a high-speed communication network; combining multiple ARM processors into a station-level multi-core CPU cluster based on the high-speed communication network, dynamically allocating workload, and constructing a station-level intelligent agent central system; establishing a unified real-time database URTDB and a unified entity database UDB based on the station-level intelligent agent central system, and performing hierarchical management of system data to form the intelligent agent data foundation; performing multi-dimensional analysis and prediction of real-time station-level operation data to generate intelligent decision-making schemes; and executing edge computing and distributed linkage control based on the intelligent decision-making schemes to coordinate and schedule various professional subsystems, thereby realizing autonomous control of the station-level intelligent agent.

[0006] Secondly, this application provides a rail transit station-level intelligent agent implementation system, including:

[0007] Modules are used to build a general professional body base, configure a real-time embedded operating system, and form a modular professional body architecture.

[0008] The transmission module is used to design a secure communication bus according to the professional architecture, perform differentiated transmission processing on data, and establish a high-speed communication network.

[0009] The allocation module is used to combine multiple ARM processors into a station-level multi-core CPU cluster based on a high-speed communication network, dynamically allocate workloads, and build a station-level intelligent agent central system.

[0010] The hierarchical module is used to establish a unified real-time database URTDB and a unified entity database UDB based on the station-level intelligent agent central system, and to perform hierarchical management of system data to form the intelligent agent data foundation;

[0011] The prediction module is used to perform multi-dimensional analysis and prediction of real-time station-level operational data and generate intelligent decision-making solutions.

[0012] The control module is used to perform edge computing and distributed linkage control based on intelligent decision-making schemes, coordinate and schedule various professional subsystems, and realize the autonomous control of station-level intelligent agents.

[0013] Thirdly, a computer device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the computer device to execute the above-described rail transit station-level intelligent agent implementation method.

[0014] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described method for implementing a rail transit station-level intelligent agent.

[0015] The technical solution provided in this application constructs a general-purpose professional body base using a high-performance 64-bit ARM processor, configures a real-time embedded operating system, and forms a modular professional body architecture. This effectively solves the problem of independent deployment and difficulty in collaborative work of various professional subsystems in traditional rail transit station-level systems, achieving efficient integration and flexible configuration of hardware resources. Based on the professional body architecture, an FPGA is used to implement a LiteIP secure communication bus. A dynamic key encryption mechanism is used for differentiated data transmission processing, reducing communication latency from hundreds of milliseconds in traditional TCP / IP to sub-milliseconds. Furthermore, by differentiating between global shared data streams, group shared data streams, and point-to-point data streams, data transmission efficiency and security are significantly improved. Multiple ARM processors are combined into a station-level multi-core CPU cluster. A task priority allocation algorithm dynamically allocates workload, and the constructed station-level intelligent body central system fully utilizes computing resources, reducing the response time of critical tasks to less than 20ms, far superior to the 300ms response time of traditional systems. A unified real-time database URTDB and a unified entity database UDB are established, and a multi-level caching mechanism and DSI data service interface are used to manage the system. Data is managed in a hierarchical manner, completely solving the "data silo" problem in traditional systems. This allows various professional entities to easily access global data, providing a complete data foundation for intelligent analysis. Especially in the application of artificial intelligence, this solution integrates federated learning and deep neural network technologies, enabling various professional entities to share model knowledge while protecting local data privacy. This significantly improves the overall performance of AI models. For example, the context-aware engine, through the combination of convolutional neural networks and recurrent neural networks, successfully integrates and analyzes multimodal data (visual, temporal, sensor, etc.), accurately identifies complex scene states, predicts trend changes, and provides a reliable basis for intelligent decision-making. Finally, through edge computing and distributed linkage control, it achieves comprehensive autonomous control capabilities, from millisecond-level emergency control to long-term optimization. Especially in emergencies such as fires and large passenger flows, the entire processing flow from data acquisition, analysis and processing to instruction execution takes only 80ms, nearly four times faster than traditional systems. This significantly improves the safety level and operational efficiency of rail transit stations. At the same time, through adaptive parameter adjustment and continuous updates to the professional entity control strategy library, the system can continuously evolve and optimize to adapt to changing operating environments and needs. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0017] Figure 1 This is a schematic diagram of an embodiment of the rail transit station-level intelligent agent implementation method in this application.

[0018] Figure 2 This is a schematic diagram of a station-level intelligent agent in an embodiment of this application;

[0019] Figure 3 This is a schematic diagram of the database system in this application;

[0020] Figure 4 This is a schematic diagram of distributed linkage in this application;

[0021] Figure 5 This is a schematic diagram of one embodiment of the rail transit station-level intelligent agent implementation system in this application.

[0022] Figure 6 This is a schematic block diagram of the structure of the computer device in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] Please see Figure 1 One embodiment of the rail transit station-level intelligent agent implementation method in this application includes:

[0025] Step S101: Construct a general professional body base, configure a real-time embedded operating system, and form a modular professional body architecture;

[0026] Step S102: Design a secure communication bus based on the professional architecture, perform differentiated data transmission processing, and establish a high-speed communication network.

[0027] Step S103: Based on the high-speed communication network, combine multiple ARM processors into a station-level multi-core CPU cluster, dynamically allocate the workload, and build a station-level intelligent agent central system.

[0028] Step S104: Establish a unified real-time database URTDB and a unified entity database UDB based on the station-level intelligent agent central system, and perform hierarchical management of system data to form the intelligent agent data foundation;

[0029] Step S105: Perform multi-dimensional analysis and prediction on real-time station-level operation data to generate intelligent decision-making solutions;

[0030] Step S106: Based on the intelligent decision-making scheme, perform edge computing and distributed linkage control to coordinate and schedule various professional subsystems, and realize the autonomous control of the station-level intelligent agent.

[0031] It is understood that the implementing entity of this application can be a rail transit station-level intelligent agent implementation system, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0032] Specifically, a general-purpose professional system base is constructed using a high-performance 64-bit ARM processor, configured with a real-time embedded operating system, forming a modular professional system architecture. The ARM processor was chosen for its high performance and low power consumption, achieving a processing power of 1.5GHz and supporting parallel task processing with its multi-core structure. The real-time embedded operating system is configured with a time resolution of 1ms to ensure accurate system response. The modular professional system architecture includes seven professional systems: BAS, FAS, PIS, PA, PSD, ATS, and SCADA. Each professional system is customized for specific functions but shares a unified base architecture. For example, the FAS professional system is specifically responsible for fire alarm signal processing. Its base is equipped with a 4-core ARM processor, 512MB of RAM, and 1GB of flash memory. A protocol adaptation and conversion module converts different communication protocols such as Modbus and BACnet into a unified standard data format, enabling seamless communication with fire detectors, manual alarm buttons, and other devices.

[0033] A station-level intelligent agent is an intelligent cabinet-style structure containing a core processing cluster, specialized subsystems, and a secure communication bus. The core processing cluster is a multi-level, multi-tasking cluster composed of 64-bit ARM processors based on domestically produced CPUs. It enables unified and efficient storage of real-time data (urtdb) and historical data (udb). Each specialized agent connects to the intelligent agent cluster via efficient private FPGA communication, such as... Figure 2 The diagram shown is a schematic of a station-level intelligent agent in an embodiment of this application.

[0034] Based on a professional system architecture, an FPGA-based LiteIP secure communication bus is designed. A dynamic key encryption mechanism is used for differentiated data transmission processing to establish a high-speed communication network. The FPGA employs hardware circuitry with 100,000 logic gates to achieve efficient processing of the LiteIP protocol stack. The LiteIP communication bus divides data streams into three types: globally shared data streams, group shared data streams, and point-to-point data streams, assigning different transmission priorities to each. The dynamic key encryption mechanism, based on elliptic curve cryptography, automatically updates the key pair every 30 seconds to encrypt transmitted data. Data packets are marked with hardware timestamps and transmission frame sequence numbers, combined with dynamic bandwidth allocation technology, achieving sub-millisecond communication latency, improving efficiency by 10 times compared to traditional TCP / IP. Multiple ARM processors are combined into a station-level multi-core CPU cluster based on the high-speed communication network. A task priority allocation algorithm dynamically distributes workloads, constructing a station-level intelligent agent central system. Multiple ARM processors are networked according to a master-slave architecture, forming a processor array, and are divided into three roles: master control processor, computing processor, and communication processor. The task priority allocation algorithm assigns weight coefficients to different tasks based on their urgency, resource requirements, and the current system load, allocating urgent tasks to low-load processors. The station-level intelligent agent central system maintains real-time updates to the resource state graph through inter-processor communication channels, ensuring balanced workload distribution and keeping system response time within 20ms.

[0035] A unified real-time database (URTDB) and a unified entity database (UDB) are established based on the station-level intelligent agent central system. A multi-level caching mechanism and a DSI data service interface are used to manage system data hierarchically, forming the foundation of the intelligent agent data. URTDB allocates high-speed memory space to store real-time control data, while UDB manages parameter configurations and historical data through data persistence mapping technology. The multi-level caching mechanism stores high-frequency access data in the L1 cache, medium-frequency access data in the L2 cache, and low-frequency access data in the main storage area, accelerating data access speed. The DSI data service interface divides the data structure into a global concept layer, a local concept layer, a memory physical layer, an object layer, and an application physical layer, providing a unified access standard for various professional agents and solving the data silo problem. Based on the intelligent agent data foundation, federated learning and deep neural network technologies are integrated to perform multi-dimensional analysis and prediction of real-time station-level operational data, generating intelligent decision-making schemes. The data acquisition and preprocessing unit extracts multi-source heterogeneous data from URTDB and UDB, and forms a standardized dataset through denoising, normalization, and feature extraction. The hierarchical federated learning architecture enables each professional entity to perform local computations and transmit only model gradient information, protecting data privacy while improving learning efficiency. The context-aware engine processes visual data through convolutional neural networks, analyzes time-series data using recurrent neural networks, integrates multimodal information to identify station-level operational status, generates a multivariate time-series prediction matrix, and outputs intelligent decision-making schemes using reinforcement learning algorithms. Based on the intelligent decision-making schemes, edge computing and distributed linkage control are implemented to coordinate and schedule various professional subsystems, achieving autonomous control of the station-level intelligent entity. The intelligent decision mapper decomposes the decision scheme into a set of execution instructions for each professional entity, and the edge computing task distribution mechanism divides the instruction set into real-time control instructions and non-real-time control instructions, constructing a dual-channel scheduling framework. The scheduling priority evaluator sorts the instruction execution order, forming a linkage control timing table, and issues control commands to each professional subsystem through a distributed transaction mechanism. The state feedback collector acquires execution status data, generates control execution reports, and performs adaptive parameter adjustments based on non-real-time control instructions, completing the closed loop of autonomous control for the station-level intelligent entity.

[0036] Taking an emergency fire scenario as an example, when the FAS (Fire Safety System) detects a smoke signal, it transmits the data to the station-level multi-core CPU cluster via the FPGA-accelerated LiteIP security communication bus, and the data is updated in real time in the URTDB (Ultra-Record Database). The agent uses a deep neural network to analyze the smoke detection data and compare it with historical data to confirm the fire situation and determine its spread trend, generating an emergency decision plan. The intelligent decision mapper translates the plan into a set of execution instructions for the agent. These instructions trigger coordinated responses through a dual-channel scheduling framework, including the BAS (Balanced Air System) controlling the ventilation system, the PA (Publication Automation System) issuing evacuation broadcasts, the PSD (Platform Shielding Device) controlling the platform screen doors, and the ATS (Automatic Train Management System) adjusting train operation plans. The entire process, from data acquisition to instruction execution, takes only 80ms, far less than the 300ms response time of traditional systems, significantly improving the efficiency of emergency response and demonstrating the significant advantages of high-speed communication for station-level agents.

[0037] It should be noted that the station-level intelligent agent is the central hub of a station-level intelligent system, capable of unified management and control of all relevant sensors and controllers within the station, and even data, models, and contingency plans related to intelligence, including personnel and objects. For each specialized subsystem, it can be directly connected to a PLC, or processed and executed by the corresponding specialized agent. Each specialized agent is responsible for data acquisition and control within its own specialty. The intelligent agents communicate with each other through a special FPGA-based network adapter, significantly improving the transmission efficiency between the specialized agent's base and the intelligent agent itself.

[0038] The station-level intelligent agent architecture introduces a multi-CPU cluster framework into the construction of station intelligent agents. Each specialized subsystem is treated as a CPU or an organ in the human body, efficiently connected to the central nervous system via vascular and neural links, thus becoming a human entity. The station-level intelligent agent is structured as an intelligent cabinet containing a core processing cluster, specialized subsystems, and a secure communication bus. The core processing cluster is a multi-level, multi-tasking cluster composed of 64-bit ARM processors based on domestically produced CPUs. It enables unified and efficient storage of real-time data (URTDB) and historical data (UDB). Each specialized subsystem connects to the intelligent agent cluster via efficient private FPGA communication.

[0039] The database system is centered around a high-efficiency in-memory database: such as Figure 3 The diagram shown is a schematic representation of the database system in this application.

[0040] Within the professional unit, it transforms into a computer similar to a multi-CPU core. The real-time operating system has a time resolution of 1ms, and the core cluster is the most powerful brain, working closely with the professional unit. Each professional unit is fully capable of handling control response within 5-10ms. Combined with the 5-10ms actual calculation response of the core cluster, it can achieve an autonomous driving level response efficiency of 20ms.

[0041] Working mechanism:

[0042] UDB is a unified entity database that manages parameters and historical data within intelligent entities. URTDB is a unified real-time database, an in-memory database designed for real-time control. The core cluster includes UDB, URTDB, specialized entity images, and a cluster monitoring system. It performs high-speed computation and processing through UDB and URTDB to generate real-time responses and executes related operations through the data bus with relevant specialized entities.

[0043] The station-level intelligent agent consists of a general-purpose specialized agent base, similar to human gene tissue—a unified, domestically produced hardware configuration, equipped with a real-time embedded operating system, an independent real-time database, and real-time message and communication middleware. Each specialized agent directly connects to UDB and URTDB via a bus to form a specialized system, achieving specialized operation and control, and interacting with the core cluster of intelligent agents. It serves as both a specialized control brain and a key element of the entire station-level intelligent agent, closely united around the core cluster, responding to control commands at the fastest possible speed.

[0044] Each professional unit can achieve flexible configuration for single-machine or multi-machine operation at the station level, and the device itself is plug-and-play. It has rich communication protocols with its respective PLC, and these protocols are plug-in / plug-and-play for communication. Internal private cloud deployment allows for automatic configuration and installation of hardware for each professional unit as needed.

[0045] Traditional TCP-based data standards often require multiple processes of establishing connections, data exchange, and disconnection, resulting in mediocre efficiency. To accommodate high-volume, high-efficiency transmission, this intelligent agent design employs a secure LiteIP communication protocol based on FPGA with private dynamic key management between various specialized subsystems. This effectively differentiates and treats globally shared data, group-shared data, and point-to-point interconnected data, forming an optimal communication strategy.

[0046] The security layer uses simple dynamic code encryption, and data is automatically encrypted during the serialization process during transmission.

[0047] A custom transport layer enables efficient transmission of proprietary protocols using LiteIP technology between professional entities. As shown above, even disregarding the efficiency improvements at the transport layer, this approach significantly enhances the timeliness and reliability of site-level intelligence.

[0048] The AI-powered intelligent decision-making and self-learning subsystem incorporates federated learning, computer vision, and deep learning to achieve real-time emergency decision-making, predictive maintenance, and customized information services at the station-level system level. Through vision and deep learning, combined with collected data, and multi-dimensional data analysis, it confirms data reliability and the status of equipment and personnel. It can make accurate and rapid judgments without relying on the computing power of a central system. For example, when a platform screen door alarm occurs or there are smoke, fire, or water-related alarms in a tunnel, visual and infrared image analysis can accurately confirm whether an alarm has actually been triggered, reducing false alarms. This enhances operation and maintenance capabilities.

[0049] At the station level, local data processing and intelligent scheduling are implemented, reducing the burden on the central system and improving overall response time. As a station-level intelligent agent, its greatest value lies in its ability to achieve station-level edge computing and distributed linkage, maximizing the safety level and response speed of the subway. A typical scenario is a fire alarm in the FAS (Fire Alarm System). This immediately triggers an emergency plan or a temporary plan based on AI decision-making, coordinating with relevant BAS (Balanced Air System) components for fire-related emergency response, such as ventilation fan exhaust, elevator rerouting from two-way to one-way, and triggering other escape devices. Simultaneously, announcements are made within the station to deploy personnel, and passengers outside the station are advised to avoid entering. Furthermore, the emergency locking mechanism of the ATS (Automatic Train Protection System) is triggered, quickly removing departing trains and locking incoming trains to prevent them from entering. This system engineering transforms decisions previously made at the central level into independent actions by the station-level intelligent agent, significantly improving immediacy and reliability. Figure 4 The diagram shown is a schematic of distributed linkage in this application;

[0050] In this embodiment, a general-purpose professional body base is constructed using a high-performance 64-bit ARM processor, configured with a real-time embedded operating system, forming a modular professional body architecture. This effectively solves the problem of independent deployment and difficulty in collaborative work among various professional subsystems in traditional rail transit station-level systems, achieving efficient integration and flexible configuration of hardware resources. An FPGA based on the professional body architecture implements a LiteIP secure communication bus, utilizing a dynamic key encryption mechanism for differentiated data transmission processing. This not only reduces communication latency from the hundreds of milliseconds of traditional TCP / IP to sub-milliseconds, but also significantly improves data transmission efficiency and security through differentiated processing of global shared data streams, group shared data streams, and point-to-point data streams. Multiple ARM processors are combined into a station-level multi-core CPU cluster, and workload is dynamically allocated through a task priority allocation algorithm. The constructed station-level intelligent body central system fully utilizes computing resources, reducing the response time of critical tasks to less than 20ms, far superior to the 300ms response time of traditional systems. A unified real-time database URTDB and a unified entity database UDB are established, and system data is processed through a multi-level caching mechanism and a DSI data service interface. By implementing hierarchical management, the "data silo" problem in traditional systems is completely solved, enabling various professional entities to easily access global data and providing a complete data foundation for intelligent analysis. Especially in the application of artificial intelligence, this solution integrates federated learning and deep neural network technologies, allowing various professional entities to share model knowledge while protecting local data privacy. This significantly improves the overall performance of AI models. For example, the context-aware engine, through the combination of convolutional neural networks and recurrent neural networks, successfully integrates and analyzes multimodal data (visual, temporal, sensor, etc.), accurately identifies complex scene states, predicts trend changes, and provides a reliable basis for intelligent decision-making. Finally, through edge computing and distributed linkage control, it achieves comprehensive autonomous control capabilities from millisecond-level emergency control to long-term optimization. Especially in emergency situations such as fires and large passenger flows, the entire processing flow from data acquisition, analysis and processing to instruction execution takes only 80ms, nearly four times faster than traditional systems. This significantly improves the safety level and operational efficiency of rail transit stations. At the same time, through adaptive parameter adjustment and continuous updates to the professional entity control strategy library, the system can continuously evolve and optimize to adapt to changing operating environments and needs.

[0051] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0052] By setting up a multi-channel timing control unit to precisely control the clock signal of the ARM processor, a real-time processing capability with a time resolution of 1ms can be obtained.

[0053] The real-time embedded operating system is divided into a real-time task layer and an application layer, with the real-time task layer managing hardware resource allocation.

[0054] A protocol adaptation and conversion module is built into the general professional body base to convert different communication protocols into a unified standard data format;

[0055] Based on their functions, professional fonts are divided into seven types: BAS professional font, FAS professional font, PIS professional font, PA professional font, PSD professional font, ATS professional font, and SCADA professional font, forming a professional font family.

[0056] By using a private cloud resource allocation algorithm, the computing resources and storage space of the professional body base are dynamically configured to form an elastic computing architecture;

[0057] Select the appropriate professional entity from the professional entity family and establish a communication connection with the corresponding PLC device to form a complete modular professional entity architecture.

[0058] Specifically, the multi-channel timing control unit is a hardware module specifically designed for high-precision clock signal management. It divides the ARM processor's base clock signal into multiple independent signals and precisely controls them, with each signal maintained stable by an independent crystal oscillator and phase-locked loop. The multi-channel timing control unit performs frequency division processing on the clock signal, converting the main frequency into multiple levels of clock frequencies to ensure that tasks of different priorities receive corresponding time accuracy, with a maximum accuracy of 1ms. This control unit also incorporates a time compensation algorithm, which monitors clock drift in real time and automatically adjusts parameters to maintain clock synchronization, thereby achieving real-time processing capability with a 1ms time resolution. The real-time embedded operating system adopts a layered architecture, divided into a real-time task layer and an application layer. The real-time task layer is located at the bottom layer and directly controls hardware resources, responsible for core functions such as interrupt handling, task scheduling, and memory management. This layer maintains a task priority queue, allocates CPU time slices according to priority, and implements priority inversion protection for critical resources. The real-time task layer abstracts hardware resources into resource objects. Each resource object contains three attributes: status identifier, occupancy information, and access permissions. A resource allocation table records the current resource usage. Application tasks obtain the required resources through a resource request interface, and the real-time task layer makes allocation decisions based on priority and the current resource status. The application layer, located above, is responsible for implementing specific business logic and interacts with the real-time task layer through standard APIs.

[0059] The general-purpose professional base station incorporates a built-in protocol adaptation and conversion module. This module adopts a plug-in architecture and supports multiple industrial communication protocols such as Modbus, BACnet, CIP, and Profibus. The protocol adaptation and conversion module comprises three core parts: a protocol parsing engine, a data mapper, and a standard data encapsulator. The protocol parsing engine is responsible for identifying and parsing data frames of different protocol formats and extracting valid data content; the data mapper maps the parsed data to a preset data model to establish a unified data structure; and the standard data encapsulator encapsulates the mapped data into a unified standard data format for easy subsequent processing. Data transmitted by various devices, after being converted by this module, is transformed into standard data packets with the same structure, regardless of the differences in the original protocols. Based on the different functional requirements of rail transit stations, the professional base station is divided into seven types. The BAS (Balanced Assurance System) is responsible for station environmental control, managing equipment such as air conditioning, lighting, and elevators; the FAS (Fire Alarm System) is responsible for fire alarms, processing data from smoke detectors, temperature sensors, and other equipment; the PIS (Personal Information System) is responsible for passenger information display, managing information screens, LED displays, and other equipment; the PA (Public Address System) is responsible for public address systems, controlling speakers, amplifiers, and other equipment; the PSD (Platform Speaker Control System) is responsible for platform screen door control; the ATS (Automatic Train Management System) is responsible for automatic train dispatching; and the SCADA (Supervisory Control and Data Acquisition) system is responsible for monitoring and data acquisition. These seven systems together form a family of systems, with each system optimizing its hardware configuration and software structure for its specific functional area.

[0060] The private cloud resource allocation algorithm dynamically configures the computing resources and storage space of the professional entity's infrastructure. Based on the resource pool management principle, this algorithm integrates all available hardware resources into a unified resource pool, including CPU cores, memory space, and storage capacity. The algorithm monitors the resource utilization and task queue length of each professional entity using a load prediction model, calculates resource demand trends, and then automatically adjusts the resource allocation scheme according to priority strategies and resource availability. During resource allocation, resources required for critical tasks are reserved first, and remaining resources are dynamically allocated as needed. When resource contention occurs, it is processed according to a preset priority order. The resource allocation algorithm also supports elastic scaling strategies, allowing automatic expansion or contraction of resource configurations to form a resilient computing architecture when the load changes. When selecting a professional entity from the professional entity group to establish a communication connection with the corresponding PLC device, device scanning and identification are first performed to obtain a list of PLC devices and their communication parameters in the connected network. Based on a preset function association table, different PLC devices are matched with corresponding professional entities to establish a device-professional entity mapping relationship. Then, through the aforementioned protocol adaptation and conversion module, the communication connection is initialized, handshake authentication is completed, and a data channel is established. Each specialized entity acquires control point information from PLC devices, builds a control point database, and forms an equipment abstraction layer. Multiple specialized entities work collaboratively, each responsible for equipment management and data processing within its own domain, collectively forming a complete modular specialized entity architecture.

[0061] Taking emergency fire response at a rail transit station as an example, when a fire occurs within the station, smoke detectors detect smoke signals and transmit the data to the PLC controller via the Modbus protocol. The FAS (Fixed Air System) module uses a protocol adaptation and conversion module to parse the Modbus data frame into a standard data format, extracting information such as smoke concentration and temperature. A multi-channel timing control unit ensures that the FAS module processes real-time data with 1ms time precision, and the real-time task layer immediately allocates high-priority resources to process fire alarm information. The FAS module transmits the fire alarm information to other modules via a high-speed communication network, triggering a coordinated response: the BAS (Balanced Air System) module controls the air conditioning system to switch to emergency smoke extraction mode, the PA (Public Air Utility) module issues evacuation broadcasts, the PSD (Platform Shielding Device) module controls the platform screen doors to open, and the ATS (Automatic Train Management System) module adjusts train operation plans. The entire process is dynamically adjusted by a private cloud resource allocation algorithm, allocating more resources to modules handling emergency tasks to ensure efficient execution of the emergency response. From smoke detection to the completion of a station-wide coordinated response, the entire process takes less than 100ms, demonstrating the efficient communication and collaboration capabilities of the modular module architecture in emergency response.

[0062] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0063] The LiteIP protocol stack is customized using FPGA hardware circuitry to build a high-performance protocol processing unit.

[0064] The communication data stream is divided into three types: global shared data stream, group shared data stream, and point-to-point data stream, and each is assigned a different transmission priority.

[0065] Dynamic key pairs are generated based on the principles of elliptic curve cryptography, and encryption strategies of different strengths are applied to different types of data streams.

[0066] A reliable data packet transmission confirmation mechanism is formed by marking data packets with hardware timestamps and transmission frame sequence numbers;

[0067] The system automatically calculates the optimal transmission path based on topology information and performs intelligent routing and forwarding of data packets.

[0068] By utilizing dynamic bandwidth allocation technology, data transmission between multiple professional entities can be coordinated and managed to establish a high-speed communication network.

[0069] Specifically, the FPGA hardware circuitry implements the LiteIP protocol stack in a customized manner. The process of building a high-performance protocol processing unit first involves transforming the software processing logic of traditional network protocol stacks into hardware circuit design. The LiteIP protocol stack is a lightweight communication protocol that, compared to traditional TCP / IP, eliminates redundant handshake confirmation mechanisms and directly implements data transmission functions at the data link layer. The FPGA hardware circuitry uses dedicated logic units to implement functions such as data fragmentation, packet header encapsulation, checksum calculation, and data encryption. Each functional module employs a pipelined architecture for parallel processing. Inside the FPGA, a multi-channel processing structure is established, with each channel containing independent transmit / receive buffers and state control logic. Multiple channels share an encryption engine and checksum unit. When a data packet enters the FPGA, it first passes through a message classifier to identify its data type, then is allocated to the corresponding processing channel. After pipelined processing, it is output as a standard LiteIP data packet. The entire processing is completed at the hardware level without CPU intervention. The communication data stream is divided into three types: globally shared data stream, group shared data stream, and point-to-point data stream, representing a functional classification of data communication modes. Globally shared data streams refer to public information that needs to be accessed by all professional entities, such as system clocks and global alarm information. A broadcast flag is set in the data packet header to ensure that the data is received by all nodes. Group-shared data streams refer to data shared only within a specific functional group, such as temperature and humidity information within the environmental control group. The data packet header contains a group identifier, and only professional entities belonging to that group process this data. Point-to-point data streams are private communication data between two specific professional entities, such as control command transmissions. The data packet header contains the source and destination addresses. Each of the three data streams is assigned a different transmission priority: point-to-point data streams have the highest priority (value 1), group-shared data streams have a middle priority (value 2), and globally shared data streams have the lowest priority (value 3). Priority values ​​are passed through the priority field in the data packet header. During data exchange processing, the LiteIP protocol stack prioritizes higher-priority data packets to ensure that critical control commands are delivered in a timely manner.

[0070] The core security mechanism of the LiteIP secure communication bus is the generation of dynamic key pairs based on elliptic curve cryptography (ECC) principles, applying different encryption strengths to different types of data streams. ECC is a public-key cryptosystem based on elliptic curve mathematics. Compared to the traditional RSA algorithm, it provides equivalent security strength with a shorter key length, making it suitable for resource-constrained environments. The dynamic key generator generates a basic key pair based on preset elliptic curve parameters, then adds a timestamp and a random number factor to calculate the dynamic key. During generation, an elliptic curve and its base points are selected, a private key integer is randomly generated, and the public key point is calculated using elliptic curve dot multiplication. Different encryption strengths are applied to different types of data streams: point-to-point data streams are fully encrypted with a 256-bit ECC key, including both the header and payload; group-shared data streams are encrypted with a 192-bit ECC key, encrypting only the data payload; and globally shared data streams are simply encrypted with a 128-bit ECC key, primarily protecting data integrity. The key update mechanism automatically updates the key pairs at preset time intervals (usually 30 seconds) and distributes the new public key to all communicating parties through a secure channel.

[0071] Hardware timestamps and transmission frame sequence numbers mark data packets, forming a reliable data packet transmission acknowledgment mechanism that solves the data transmission reliability problem. Each FPGA module has a built-in high-precision clock source, providing microsecond-level timestamps. When a data packet enters the transmission process, the FPGA automatically adds the current timestamp and a counter-based frame sequence number to the packet header. After receiving the data packet, the receiver extracts the timestamp and sequence number information, compares it with the local clock to calculate the transmission delay, and detects packet loss using the sequence number. If packet loss is detected, different measures are taken depending on the severity of the packet loss: a retransmission request is sent when a single packet is lost; the sender is notified to reduce the transmission rate when multiple packets are lost consecutively; and a link quality assessment is triggered when intermittent packet loss exceeds a threshold. Data packet reception acknowledgment uses a cumulative acknowledgment mechanism, with the receiver periodically sending acknowledgment frames containing the highest sequence number that has been correctly received, effectively reducing acknowledgment traffic.

[0072] The process of automatically calculating the optimal transmission path based on topology information and intelligently routing and forwarding data packets involves dynamic route calculation and forwarding decisions. The topology discovery module collects network connection status information by periodically broadcasting probe packets, constructing a global topology map that includes inter-node connectivity, link delay, and bandwidth information. The path calculator calculates the shortest path from the source node to the destination node based on Dijkstra's algorithm, using link delay as the primary metric while also considering bandwidth and load factors. The path calculation results form a routing table containing the destination address, next-hop address, and path delay information. Data packet forwarding decisions are made based on the routing table. When a data packet arrives at a forwarding node, the destination address is extracted, the routing table is consulted to obtain the next-hop information, and if multiple paths exist, the optimal path is selected based on the current load. The routing table is updated periodically to adapt to changes in network topology, ensuring that data packets are always transmitted along the optimal path.

[0073] Dynamic bandwidth allocation technology coordinates and manages data transmission between multiple professional entities, achieving rational utilization of network resources during the establishment of a high-speed communication network. The bandwidth monitoring module continuously collects traffic data from each link, calculating current bandwidth utilization and remaining available bandwidth. The QoS (Quality of Service) manager allocates network resources based on data flow priority and bandwidth demand, establishing a multi-level queue structure where high-priority data receives bandwidth allocation first. The traffic shaper controls the data flows sent by each professional entity, limiting the transmission rate according to preset strategies to prevent any single professional entity from consuming excessive bandwidth resources. The congestion control module monitors network congestion status; when excessive link load is detected, it triggers a congestion notification, requiring relevant professional entities to reduce their data transmission rate. The bandwidth reservation mechanism allows critical services to pre-apply for bandwidth resources, ensuring sufficient transmission resources are available even during peak network load periods.

[0074] In station-level applications of rail transit, when an emergency occurs such as a platform screen door malfunction, the PSD (Power Segment Detection) agent first detects the abnormal status data of the platform screen door through sensors, encapsulates this data into a point-to-point data stream (priority 1), and processes it through the LiteIP protocol stack implemented on the FPGA. The data packet is encrypted with a 256-bit ECC key and appended with the current timestamp and frame sequence number. The path calculator determines the optimal path from the PSD agent to the station-level multi-core CPU cluster based on the current network topology, and the data packet is forwarded via the shortest latency path. The bandwidth manager identifies high-priority data streams and immediately allocates sufficient bandwidth resources to them, ensuring minimal transmission latency. After receiving the data, the station-level multi-core CPU cluster verifies data integrity by decryption, extracts the timestamp to calculate the end-to-end latency (typically less than 1ms), and returns an acknowledgment frame to the PSD agent. Simultaneously, the station-level intelligent agent generates linkage instructions based on the platform screen door malfunction information, creating a set of point-to-point data streams and sending them to the PA (Publication Allocation) agent (broadcasting evacuation information), the ATS (Action Tracking System) agent (adjusting train operation), and other relevant agents. The entire information transmission process, from the detection of the platform screen door malfunction to the coordinated response of multiple professional entities, was completed with a total delay of less than 5ms, demonstrating the high-speed communication capability of the FPGA-based LiteIP secure communication bus in the intelligent entities of rail transit stations.

[0075] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0076] Multiple ARM processors are physically networked together in a master-slave architecture to form a processor array;

[0077] The processors in the processor array are defined by their roles and classified into three types: main control processor, computing processor and communication processor.

[0078] Establish inter-processor communication channels based on high-speed communication networks and create processor resource pools;

[0079] The processor load in the processor resource pool is monitored in real time through a status monitoring mechanism to generate a resource status diagram.

[0080] Based on the resource status diagram, identify task types and urgency levels, and assign priority weight coefficients to different tasks;

[0081] The workload is dynamically allocated based on priority weight coefficients, and tasks are distributed to the corresponding processors for execution, thus building a station-level intelligent agent central system.

[0082] Specifically, the process of physically networking multiple ARM processors in a primary / standby architecture to form a processor array first involves physical hardware deployment. The ARM processors employ a 64-bit architecture, each with a multi-core design and a clock speed of 1.5-2.0 GHz, providing strong parallel computing capabilities. The physical network uses a dual-backbone redundant topology, with each processor connected to two independent high-speed communication backbone networks via dual network cards, ensuring uninterrupted communication in the event of a single network failure. In the primary / standby architecture, processors are deployed in pairs, with one processor in primary mode and the other in standby mode, maintaining real-time data synchronization. The primary processor handles actual business processing, while the standby processor receives processing status information in real time. When the primary processor fails, the standby processor can take over the business within milliseconds, ensuring continuous system operation. The processor array maintains state synchronization through a heartbeat mechanism. The primary processor sends a heartbeat packet every 100 ms, and the standby processor monitors the heartbeat interval. If no heartbeat packet is received three times consecutively, a primary / standby switchover process is triggered. The processors in the processor array are defined by role, categorized into three types: main control processors, computing processors, and communication processors, establishing a clear functional division. The main control processor, as the coordination center of the entire system, is responsible for management functions such as task distribution, resource scheduling, and system monitoring. It is typically configured with a high-performance ARM processor with a large built-in cache to optimize the execution efficiency of management instructions. In a station-level intelligent agent central system, the main control processor is configured with dual-machine redundancy, with only one active and the other in standby mode. Computing processors focus on data-intensive computing tasks, such as data analysis, pattern recognition, and predictive computing. They are configured with multi-core ARM processors with optimized floating-point units, suitable for parallel data processing. Communication processors are dedicated to external data interaction, handling network communication, protocol conversion, and data routing. They are equipped with high-speed network interfaces and dedicated buffers to optimize data throughput performance. The number of each type of processor is configured based on the station size and business needs. Generally, a medium-sized station is configured with 2 main control processors, 4-8 computing processors, and 2-4 communication processors.

[0083] The process of establishing inter-processor communication channels and creating a processor resource pool based on a high-speed communication network enables unified management and flexible resource allocation. The high-speed communication network is implemented using an FPGA-based LiteIP protocol stack, providing sub-millisecond communication latency. The inter-processor communication channels include two types: control channels and data channels. The control channel transmits control data such as scheduling instructions and status information, using a point-to-point communication mode. The data channel transmits business data, supporting both point-to-point and broadcast modes. When each processor joins the resource pool, it first registers with the master processor, submitting a resource description file containing information such as processor type, number of cores, memory capacity, and current load. The master processor collects all resource description files, constructs a global resource pool view, and records all available computing resources and their status. The processor resource pool adopts a dynamic membership mechanism, allowing processors to join or leave the resource pool at any time, adapting to system expansion and maintenance needs. A global resource view is established through real-time monitoring of processor load in the processor resource pool and the generation of a resource status graph. The status monitoring mechanism deploys a monitoring agent on each processor to periodically collect processor load data, including metrics such as CPU utilization, memory usage, network traffic, and the number of active tasks. This data is sent to the main control processor via the control channel and aggregated and analyzed by the state aggregator. The resource state diagram is a multi-dimensional data structure that records the current resource usage of each processor and its trend over time. The resource state diagram uses color coding to visually display resource load: red indicates high load (above 85%), yellow indicates medium load (50%-85%), and green indicates low load (below 50%). The resource state diagram also includes historical load data for load trend analysis and prediction. The main control processor calculates load change trends based on historical data to predict resource demand in the near future.

[0084] The process of identifying task types and urgency based on resource status diagrams and assigning priority weight coefficients to different tasks enables differentiated task processing. Task type identification is based on task characteristic analysis, including factors such as source, resource requirements, and execution duration, classifying tasks into three main categories: control, computation, and communication. Urgency assessment is determined based on task time sensitivity and business importance, categorized into four levels: urgent (requiring immediate handling, such as equipment alarms), high (requiring priority handling, such as personnel safety related), medium (routine business processing), and low (backend analysis processing). Priority weight coefficient calculation comprehensively considers task type and urgency, using a weighted method to determine the final priority. Control tasks have higher weights than computation and communication tasks, and urgent tasks have higher weights than non-urgent tasks. Tasks with higher weight coefficients receive more system resource allocation and are scheduled for execution first. Dynamically allocating workloads according to priority weight coefficients and distributing tasks to appropriate processors for execution, the process of building a station-level intelligent agent central system achieves efficient task scheduling. The task scheduler first sorts the queue of tasks to be processed in descending order of priority weight coefficients, ensuring that high-priority tasks are processed first. Then, based on the task type, the appropriate processor category is selected: control tasks are assigned to the master processor, computation tasks to computation processors, and communication tasks to communication processors. After determining the processor category, the task scheduler queries the resource status graph for processors with lower current loads and assigns tasks to those processors for execution. During task distribution, the scheduler generates a task description packet containing information such as task ID, priority, and resource requirements, and sends it to the target processor through the control channel. After receiving the task, the target processor arranges the execution order according to the task priority. After the task is completed, it returns the execution result and resource usage information to the master processor.

[0085] For example, in a rail transit station-level intelligent agent, when a large passenger flow occurs at the station, passenger flow monitoring sensor data is received and initially processed by the communication processor to generate passenger flow density distribution data. The main control processor obtains this data from the communication processor, identifies that the passenger flow density in multiple areas of the platform exceeds the safety threshold, and determines it as a high-urgency control task, assigning it a priority weight coefficient of 0.9 (out of 1). The main control processor queries the resource status diagram and finds that the current load of the main control processor MCU-02 is only 30%, far lower than the 75% load of MCU-01, so it assigns the passenger flow control task to MCU-02 for processing. After receiving the task, MCU-02 analyzes the passenger flow data and historical patterns, calculates the optimal evacuation route, and generates a linkage control scheme. This scheme includes several measures such as adjusting station broadcasts, display screen information, and platform screen door opening and closing times, which need to be sent to the PA, PIS, and PSD professional agents for execution, respectively. Each control task is packaged and assigned a corresponding priority, and then distributed to the corresponding communication processor for execution through the task scheduler. The entire process, from receiving passenger flow data to issuing control commands, takes only 65ms, enabling a rapid response to sudden surges in passenger flow and demonstrating the efficient collaborative processing capabilities of the station-level intelligent multi-core CPU cluster in rail transit scenarios.

[0086] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0087] In the station-level intelligent agent central system, a memory region is divided, and high-speed access memory space is allocated for the unified real-time database URTDB.

[0088] A unified entity database (UDB) is built using data persistence mapping technology to store and manage parameter configuration information and historical data.

[0089] System data is classified according to access frequency. High-frequency access data is stored in L1 cache, medium-frequency access data is stored in L2 cache, and low-frequency access data is stored in main storage.

[0090] Establish data association between URTDB and UDB using a data mapping table, and build an automatic switching mechanism for hot and cold data;

[0091] Create DSI data service interfaces for each professional entity, and divide the data structure into global concept layer, local concept layer, memory physical layer, object layer and application physical layer;

[0092] Data read / write operations and transaction processing are performed based on the DSI data service interface, forming the data foundation for intelligent agents.

[0093] Specifically, the process of allocating high-speed access memory space for the unified real-time database URTDB within the station-level intelligent agent central system involves the precise planning and allocation of memory resources. URTDB is a memory-based database structure specifically designed for storing and processing real-time control data requiring millisecond-level response times. The memory region partitioning employs a segmented paging memory management method, dividing the physical memory space into multiple functional areas, with the URTDB dedicated area located in the high-speed access area closest to the CPU. Specifically, the physical memory space is first mapped to a logical address space using memory page tables. Then, a dedicated segment is allocated for URTDB within the logical address space, typically occupying 30%-40% of the total memory. Internally, URTDB is further divided into three functional areas: a data area, an index area, and a buffer. The data area stores the actual data records, the index area stores fast lookup indexes, and the buffer is used for temporary data exchange. The high-speed access mechanism uses memory locking technology to prevent URTDB data from being swapped to disk by the operating system, ensuring that data always resides in physical memory and achieving sub-millisecond data access performance. The Unified Entity Database (UDB) built using data persistence mapping technology addresses the need for persistent data storage by managing parameter configuration information and historical data. UDB employs a disk-based relational database architecture with a multi-tiered storage structure. Data persistence mapping technology is a mechanism that establishes a mapping relationship between in-memory data objects and disk storage space. Through this technology, UDB can efficiently manage large amounts of non-real-time data. Specifically, UDB divides data files into fixed-size data pages. Each data page contains a header and a body. The header records page attributes and index information, while the body stores the actual data record. When accessing a data record, the data page location is first located using the index, and then the entire data page is loaded into memory, completing the data mapping between memory and disk. UDB specifically stores parameter configuration information and historical data. Parameter configuration information includes static data such as device parameters, system configurations, and business rules; historical data includes operation records, device status changes, event logs, and other data that needs to be stored long-term. UDB uses an incremental storage strategy, periodically writing new data in batches to disk, reducing I / O operation frequency and improving storage efficiency.

[0094] System data is categorized according to access frequency. High-frequency access data is stored in the L1 cache, medium-frequency access data in the L2 cache, and low-frequency access data in the main memory, achieving multi-level acceleration of data access. Access frequency classification is based on statistical analysis of data access, calculating an access popularity value by recording the access frequency and time distribution of each data entry. The access popularity value calculation comprehensively considers recent access frequency and access time distribution, assigning higher weight to accesses within the most recent period. Data is divided into three categories based on the popularity value: data with a popularity value above a preset threshold T1 is high-frequency access data; data with a popularity value between T1 and T2 is medium-frequency access data; and data with a popularity value below T2 is low-frequency access data. The L1 cache is located inside the CPU, with a small capacity but extremely fast speed, typically a few hundred KB; the L2 cache is located between the CPU and main memory, with a moderate capacity and moderate access speed, typically a few MB; the main memory has a large capacity but relatively slow access speed, typically a few GB. The data cache manager is responsible for maintaining the scheduling and updating of data across different cache levels. It uses an improved LRU (Least Recently Used) algorithm for cache replacement. When the cache space is insufficient, it prioritizes the elimination of data with lower access frequency to ensure that high-frequency access data always maintains the best access performance.

[0095] By establishing a data relationship between URTDB and UDB using a data mapping table, and constructing an automatic hot / cold data switching mechanism, intelligent data flow between different storage media is achieved. The data mapping table is a centralized metadata management structure that records the location and attributes of data items in URTDB and UDB. Each record in the mapping table contains fields such as data ID, URTDB address, UDB address, data type, timestamp, and access counter. When data changes, it is first updated in URTDB, and the data item is marked as "dirty data," indicating that it is out of sync with the version in UDB. The data synchronization thread periodically scans the "dirty data" in URTDB and writes it to UDB in batches, completing data persistence. The automatic hot / cold data switching mechanism dynamically adjusts the distribution of data between URTDB and UDB based on changes in data access frequency. When the access frequency of low-frequency data increases, the data loader reads the data from UDB and loads it into URTDB; conversely, when the access frequency of high-frequency data decreases, the data aging device removes the data from URTDB, retaining it only in UDB, freeing up valuable memory resources. During the switchover process, the data mapping table is updated in real time to record the current location of the data and ensure that the data access path is correct.

[0096] A unified data access standard was established by creating DSI (Data Service Interface) data service interfaces for various professional entities, dividing the data structure into a global concept layer, a local concept layer, a memory physical layer, an object layer, and an application physical layer. DSI is a standardized set of APIs that provides a unified data access method for different professional entities. The global concept layer, at the highest level of abstraction, defines a site-wide common data model and concepts, such as core concepts like devices, events, and alarms. The local concept layer defines specialized data models for specific professional fields, such as the smoke detector and temperature sensor models for the FAS professional entity. The memory physical layer defines the physical storage structure of data in the URTDB, including memory address mapping and index structure. The object layer abstracts data into object-oriented data entities, defining the attributes and methods of data objects. The application physical layer defines the physical storage structure of data in the UDB, including table structure and field types. When professional entities access data through the DSI interface, they do not need to worry about the physical storage location and storage method of the data; they only need to operate through the unified interface. The DSI service layer is responsible for converting logical operations into corresponding physical operations, simplifying data access complexity and solving the data silo problem.

[0097] The process of establishing the agent's data foundation through data read / write operations and transaction processing based on the DSI data service interface achieves standardization and consistency in data operations. The DSI interface provides four basic operations: Read, Write, Query, and Subscribe. Read operations retrieve the current value of one or more data items, supporting both synchronous and asynchronous modes; write operations update data values, including single-item and batch writes; query operations retrieve data based on conditions, supporting complex combinations of conditions and aggregate functions; and subscription operations allow agents to register for data change notifications, automatically receiving notifications when data changes. The transaction processing mechanism ensures the atomicity, consistency, isolation, and durability of complex operations, implementing distributed transactions through a two-phase commit protocol, thus solving the data consistency problem in multi-agent collaborative operations. Data security control uses a role-based access control model, assigning different data access permissions to different agents to prevent unauthorized access. Through this series of mechanisms, a complete, unified, and efficient agent data foundation is established, providing support for data processing and intelligent decision-making at the station-level agent level.

[0098] Taking the platform screen door system monitoring in a rail transit station as an example, when the PSD professional entity needs to monitor the status of all platform screen doors in real time, it first initiates a data subscription request to the URTDB through the DSI interface to subscribe to the status data of all platform screen doors. The DSI interface converts this request into an internal data subscription operation, looking up the location of all platform screen door status data in the data mapping table. Since platform screen door status is frequently accessed data, this data has been cached in the L1 cache. The DSI interface directly retrieves the latest data from the L1 cache and returns it to the PSD professional entity, while simultaneously establishing a data change notification mechanism. When the status of a platform screen door changes, the controller sends the status change data to the station-level intelligent agent central system through the communication processor. The data update processor first writes the new status to the data area of ​​the URTDB, updates the relevant indexes in the index area, and marks the data as "dirty data". The DSI interface detects the data change and immediately sends a change notification to the PSD professional entity that has subscribed to the data. At the same time, the data synchronization thread periodically writes the "dirty data" in the URTDB to the UDB, completing data persistence. If a PSD professional needs to query the historical fault records of a certain platform gate, it initiates a query request through the DSI interface. The DSI interface recognizes this as a historical data query, directly retrieves the relevant records from the UDB, and returns them. Throughout the process, the PSD professional only needs to call the unified DSI interface, without needing to worry about whether the data is stored in URTDB or UDB, nor handling complex issues such as data synchronization and cache management. This greatly simplifies the data access process and improves data processing efficiency.

[0099] In one specific embodiment, the process of executing step S105 may specifically include the following steps:

[0100] The data acquisition and preprocessing unit extracts multi-source heterogeneous data from the unified real-time database URTDB and the unified entity database UDB. The multi-source heterogeneous data is then denoised, normalized, and feature extracted to form a standardized dataset.

[0101] The standardized dataset is evaluated and filtered using a feature selection algorithm to construct a multidimensional feature vector and generate a training sample library.

[0102] Based on the training sample library, a hierarchical federated learning architecture is designed. After local computation processing of the data of each professional body, only the model gradient information is transmitted to complete the aggregation of model parameters and build a global deep neural network model.

[0103] A context-aware engine is created based on a global deep neural network model. Visual data is processed through convolutional neural networks, time-series data is analyzed using recurrent neural networks, multimodal information is fused, and the current station-level operating status is identified.

[0104] A multivariate time series prediction matrix is ​​generated by the context-aware engine to perform forward prediction of key station-level indicators and to build a dynamic adjustment mechanism for early warning thresholds.

[0105] Decision tree analysis is performed based on a dynamic adjustment mechanism for early warning thresholds and a preset rule base to generate a multi-level set of alternative solutions. The multi-level set of alternative solutions is then optimized and evaluated using a reinforcement learning algorithm to output an intelligent decision solution.

[0106] Specifically, the process of extracting multi-source heterogeneous data from the unified real-time database URTDB and the unified entity database UDB by the data acquisition and preprocessing unit first involves data source identification and access. Multi-source heterogeneous data refers to data sets with different sources and formats, including various types such as real-time sensor data, video surveillance data, equipment status data, and historical operation records in the rail transit station-level environment. The data acquisition and preprocessing unit initiates data requests to URTDB and UDB through the DSI data service interface, specifying query conditions such as data type, time range, and data source. For real-time data, it is mainly obtained from URTDB; for historical data and parameter configurations, it is extracted from UDB. After data acquisition, data quality detection is first performed to mark missing values, outliers, and duplicate values. Then, noise reduction processing is performed, using algorithms such as median filtering and wavelet transform to remove noise interference from the data. Data normalization processing unifies data with different dimensions to the same scale range. Commonly used normalization methods include min-max normalization and Z-score normalization. Feature extraction is the process of extracting representative features from raw data, including statistical feature extraction (mean, variance, kurtosis, etc.), time-domain feature extraction (trend, periodicity, etc.), and frequency-domain feature extraction (spectral analysis, etc.). After this series of processing steps, a standardized dataset with consistent structure and reliable quality is formed, laying the foundation for subsequent analysis. Feature selection algorithms evaluate and filter the standardized dataset, construct multi-dimensional feature vectors, and generate a training sample library, achieving data dimensionality reduction and quality improvement. Feature selection algorithms evaluate the importance of each feature and select the most representative feature subset, reducing data redundancy and noise impact. In rail transit station-level intelligent agents, commonly used feature selection methods include filtering, wrapping, and embedding. Filtering ranks features based on statistical indicators (such as information gain, Pearson correlation coefficient, etc.) and selects the top-ranked features; wrapping uses the performance of the target model as an evaluation criterion, using strategies such as forward selection or backward elimination to find the optimal feature subset; embedding integrates feature selection into the model training process, such as the LASSO algorithm based on L1 regularization. The feature evaluation process calculates an importance score for each feature, sorts them in descending order of score, and selects the top N features to form a feature subset. Multidimensional feature vectors combine the selected features into vector form, with each vector representing a data feature at a specific time point or for a single sample. The training sample library consists of multidimensional feature vectors and corresponding label data. The label data indicates the category to which the sample belongs (e.g., device status category) or the target value (e.g., predicted passenger flow). The training sample library is divided into three parts: a training set, a validation set, and a test set, used for subsequent model training and evaluation.

[0107] Based on the training sample library, a hierarchical federated learning architecture is designed. After local computation processing of data from each professional entity, only model gradient information is transmitted to complete model parameter aggregation. The process of constructing a global deep neural network model solves the problems of data privacy protection and collaborative learning. Federated learning is a distributed machine learning method that allows multiple parties to jointly train a model without sharing the original data. In the rail transit station-level intelligent agent, the hierarchical federated learning architecture is divided into two layers: intra-station federated learning and inter-station federated learning. In intra-station federated learning, each professional entity (such as BAS, FAS, PIS, etc.) acts as a client, and the station-level intelligent agent central system acts as the server; in inter-station federated learning, the intelligent agents at each station act as clients, and the network center server acts as the aggregation server. The federated learning process includes five steps: model initialization, local training, gradient upload, parameter aggregation, and model update. First, the server initializes global model parameters and distributes them to each client. Then, each client trains the model using local data and calculates parameter gradients. Next, the client uploads the gradient information (not the original data) to the server. The server aggregates the gradients uploaded by all clients and updates the global model parameters. Finally, the server distributes the updated global model parameters to each client to begin a new round of training. Through multiple iterations, a global deep neural network model integrating knowledge from multiple sources is obtained, which has the ability to handle complex pattern recognition and predictive analysis. Based on the global deep neural network model, a context-aware engine is created. This engine processes visual data using convolutional neural networks, analyzes time-series data using recurrent neural networks, and integrates multimodal information to identify the current station-level operating status, thus achieving intelligent perception of complex scenes. The context-aware engine is an intelligent system capable of understanding and interpreting the current operating environment of a rail transit station. Convolutional neural networks (CNNs) primarily process visual data from surveillance cameras, extracting spatial features through multi-layer convolution and pooling to identify visual information such as passenger flow density, abnormal behavior, and equipment status. Recurrent Neural Networks (RNNs), especially variants of Long Short-Term Memory (LSTM), primarily process time-series data, such as time-dependent data like passenger flow changes and equipment operating parameters, capturing long-term temporal dependencies through memory units. Multimodal information fusion comprehensively analyzes different types of data (visual, auditory, sensor, etc.), employing an attention mechanism to assign different weights to different modalities, resulting in a more comprehensive understanding of the scene. The output of the context-aware engine is a comprehensive description of the current station-level operational status, including multiple dimensions such as passenger flow, equipment operating status, and safety risk assessment, providing a basis for subsequent decision-making.

[0108] The context-aware engine generates a multivariate time series prediction matrix to perform forward prediction of key station-level indicators. The process of constructing a dynamic adjustment mechanism for early warning thresholds enables the extrapolation from the current state to future trends. The multivariate time series prediction matrix is ​​a three-dimensional data structure containing predicted values ​​of multiple key indicators at multiple future time points. Forward prediction, using recurrent neural networks or temporal convolutional networks, predicts the changing trends of each key indicator over a future period based on historical data and the current state. The prediction process considers the mutual influence between indicators, modeling the relationships between indicators through attention mechanisms or graph neural networks. The dynamic adjustment mechanism for early warning thresholds adaptively adjusts the early warning thresholds for each indicator based on the prediction results and historical data distribution. Traditional fixed threshold methods cannot adapt to changes in different scenarios, while dynamic thresholds update the early warning thresholds in real time by calculating the statistical distribution and trend changes of historical data, combined with expert rules. The dynamic threshold calculation considers multi-dimensional influencing factors such as time factors (weekdays / holidays), environmental factors (weather, temperature), and operational factors (passenger flow, train frequency), forming a more accurate early warning judgment standard.

[0109] Based on a dynamic adjustment mechanism for early warning thresholds and a pre-defined rule base, decision tree analysis is performed to generate a multi-level set of alternative solutions. Reinforcement learning algorithms are then used to optimize and evaluate this set, ultimately outputting an intelligent decision solution. This process achieves the intelligent transformation from early warning to decision-making. The pre-defined rule base is a knowledge base defined by domain experts, containing processing rules and procedures for various scenarios. Decision tree analysis, based on early warning information and the rule base, constructs decision paths, forming a hierarchical decision structure. The decision-making process first identifies the event type and severity, then matches applicable handling solutions according to rules, considering various constraints and resource limitations to generate multiple feasible alternative solutions. The multi-level set of alternative solutions includes three levels: emergency response, short-term adjustment, and long-term optimization, corresponding to decision-making needs at different time scales. The reinforcement learning algorithm learns the optimal decision-making strategy through interaction with the environment. In the rail transit station-level intelligent agent, reinforcement learning mainly employs algorithms such as Deep Q-Network (DQN) or policy gradient, guiding the model to learn the optimal decision path through a reward function. The reward function design comprehensively considers multiple factors such as safety, efficiency, resource consumption, and passenger experience, obtaining a comprehensive evaluation score through weighted summation. The reinforcement learning model gradually optimizes the decision-making strategy through multiple simulations and training, and finally selects the optimal solution from the set of alternative solutions as the intelligent decision-making solution output.

[0110] Taking emergency response to a fire at a rail transit station as an example, when a smoke detector in the station triggers an alarm, the data acquisition and preprocessing unit first obtains multi-source data from the URTDB, including real-time smoke detector data, temperature sensor data, and wind direction and speed data. After denoising (eliminating signal jitter) and normalization (unifying data from different units to the 0-1 range), key features such as the rate of change in smoke concentration and the rate of temperature increase are extracted to form a standardized dataset. A feature selection algorithm evaluates the importance of each feature and selects the three most representative features—smoke concentration, temperature change rate, and wind speed—to construct a multi-dimensional feature vector. In the hierarchical federated learning architecture, the FAS (Focused Automation System), BAS (Balanced Automation System), and SCADA (Supervisory Control and Controller Administration) agents each train their initial models based on local data, only uploading the model gradients to the station-level intelligent agent central system. The central system aggregates the gradient information to update the global model. The context-aware engine processes surveillance camera footage using a convolutional neural network to identify the fire location as the northwest corner of the station hall. It analyzes time-series smoke concentration data using a recurrent neural network to determine the fire's development trend. After fusing multimodal information, it confirms the current state as "initial fire, not yet affecting main passageways." A multivariate time-series prediction matrix predicts the smoke spread path and speed within the next 30 minutes, dynamically adjusting safety risk thresholds for each area. Based on early warning information and a pre-set rule base, decision tree analysis generates a set of alternative solutions including "evacuation route planning," "ventilation system adjustment," and "fire equipment activation." A reinforcement learning algorithm evaluates the comprehensive effectiveness of each solution, ultimately outputting the optimal intelligent decision: activate fire equipment in Zone B, adjust fans 1 and 3 to smoke extraction mode, shut down escalators, evacuate passengers through exits 2 and 4, and simultaneously broadcast a customized evacuation announcement via the PA system. This solution is sent to various professional entities for execution via the station-level intelligent agent. The entire process, from data acquisition to solution execution, takes only 80ms, demonstrating the efficiency and intelligence of the rail transit station-level intelligent agent implementation method in emergency handling.

[0111] In one specific embodiment, the process of executing step S106 may specifically include the following steps:

[0112] The intelligent decision-making scheme is decomposed into instructions by the intelligent decision mapper to generate a set of professional body execution instructions.

[0113] By utilizing the edge computing task distribution mechanism, the professional body's execution instruction set is divided into real-time control instructions and non-real-time control instructions according to functional attributes, and a dual-channel scheduling framework is constructed.

[0114] The emergency response process is triggered by the real-time control commands in the dual-channel scheduling framework. The scheduling priority evaluator sorts the execution order of the commands to form a linkage control timing table.

[0115] Based on the linkage control sequence table, control commands are issued to each professional subsystem through a distributed transaction mechanism, and the execution programs in the foundation of each professional entity are started synchronously to generate a collaborative control effect.

[0116] The status feedback collector obtains execution status data from various professional subsystems, performs correlation analysis on the execution status data, and generates a control execution report;

[0117] Adaptive parameter adjustments are made based on control execution reports and non-real-time control commands to update the professional agent control strategy library and complete the station-level intelligent agent autonomous control closed loop.

[0118] Specifically, the intelligent decision mapper's process of decomposing intelligent decision schemes into instruction sets for professional entities first involves the structuring of the decision content. The intelligent decision mapper is a functional module specifically designed to convert high-level decisions into concrete operational instructions. Its core function is to transform abstract decision intentions into specific control commands. The instruction decomposition process employs a top-down recursive strategy. First, it identifies the main action objectives in the decision scheme, then subdivides each objective into specific control actions, and finally matches the control actions to the corresponding professional entity functions. During instruction decomposition, the mapper queries an instruction mapping table, which maintains the correspondence between decision actions and professional entity functions, ensuring that each decomposed instruction can find an execution entity. The professional entity execution instruction set is a collection of instructions organized according to professional entity type. Each instruction includes attributes such as operation type, operation object, parameter value, execution time, and priority. The instruction set is encapsulated in XML format for easy cross-system transmission and parsing. Each instruction is accompanied by a unique identifier for subsequent execution tracking and status feedback. The edge computing task distribution mechanism divides the professional entity execution instruction set into real-time control instructions and non-real-time control instructions according to functional attributes. The process of constructing a dual-channel scheduling framework enables differentiated task processing. Edge computing task distribution is a technical architecture that decentralizes computing tasks to near the data source. In rail transit station-level intelligent agents, edge nodes are primarily the foundations of various professional entities. During task distribution, the functional attributes of each instruction in the instruction set are first analyzed. Based on time sensitivity, instructions are divided into two categories: real-time control instructions, which have extremely high time response requirements (typically requiring millisecond-level response), such as emergency stops and fire alarm responses; and non-real-time control instructions, which have relatively relaxed time requirements (second-level or minute-level response), such as environmental parameter adjustments and equipment maintenance instructions. The dual-channel scheduling framework is a parallel processing architecture comprising two independent processing pipelines: a real-time channel and a non-real-time channel. The real-time channel uses an interrupt-driven mode with the highest processing priority, dedicated to handling real-time control instructions; the non-real-time channel uses a polling mode, processing non-real-time control instructions according to a preset scheduling strategy. This architecture design ensures that time-critical tasks receive immediate responses without blocking the execution of other ordinary tasks.

[0119] The emergency response process, triggered by real-time control commands within the dual-channel scheduling framework, establishes precise task execution timing control by using a scheduling priority evaluator to sort the command execution order and form a linkage control timing table. The emergency response process is a dedicated processing mechanism for real-time control commands, comprising four main stages: command parsing, resource preparation, execution scheduling, and status monitoring. The scheduling priority evaluator is the functional module responsible for determining the command execution order in the emergency response process; its core algorithm is based on a multi-factor weighted scoring mechanism. During the evaluation process, firstly, the internal priority markers, time sensitivity, and resource dependencies of each command are extracted; then, a comprehensive priority score is calculated based on preset weights; finally, commands are arranged in descending order of priority scores to form an execution queue. For commands with the same priority, they are sorted according to their arrival time; for commands with dependencies, the preceding command is executed before the following command. The linkage control timing table is a detailed data structure recording the command execution plan, including the execution time point, execution subject, estimated execution duration, preconditions, and subsequent operations for each command, providing precise execution references for various professional entities and ensuring the timing coordination of complex linkage operations.

[0120] Based on the linkage control timing table, control commands are issued to various professional subsystems through a distributed transaction mechanism, synchronously starting the execution programs within each professional entity's base, thus generating a collaborative control effect and achieving cross-system collaborative operation. The distributed transaction mechanism is a technical means to ensure that operations across multiple systems either all succeed or all are rolled back, suitable for scenarios requiring multi-system collaboration. In the rail transit station-level intelligent agent, the distributed transaction adopts a two-phase commit protocol, including a preparation phase and a commit phase. In the preparation phase, the transaction coordinator sends pre-execution requests to each professional subsystem. Each subsystem verifies the validity of the instructions and locks the required resources, then returns ready or rejects the response. In the commit phase, if all subsystems return ready, the coordinator sends the formal execution command; otherwise, it sends a cancel execution command. The control command issuance uses an FPGA-based LiteIP secure communication bus to ensure the real-time performance and reliability of command transmission. After receiving the control command, each professional entity's base calls the corresponding execution program according to the command content. Before program execution, parameter verification and resource checks are performed to ensure that the execution conditions are met. Synergistic control effect is the result of multiple professional subsystems performing their respective tasks in a predetermined sequence to jointly achieve complex control objectives. For example, the linkage control after a fire alarm involves the coordinated work of multiple systems such as fire protection, ventilation, broadcasting, and access control.

[0121] The status feedback collector acquires execution status data from various professional subsystems, performs correlation analysis on this data, and generates a control execution report, establishing an execution monitoring and evaluation mechanism. The status feedback collector is a distributed data acquisition component responsible for collecting execution status information from various professional subsystems in real time. The collection process employs a push-and-pull model: for critical status changes, professional subsystems proactively push status data; for routine status monitoring, the collector periodically pulls status data. Execution status data includes multiple dimensions such as instruction execution progress, execution results, resource consumption, and anomaly information, transmitted in a structured data format. Correlation analysis is the process of integrating and correlating status data from different professional subsystems. It combines status data within the same time period through timestamp matching to construct a multi-dimensional status view. During the analysis, causal reasoning identifies the triggering relationships of status changes, pattern matching identifies typical status sequence patterns, and anomaly detection identifies deviations and problems in the execution process. The control execution report is the output of the correlation analysis, systematically recording the entire instruction execution process, including overall execution status, execution status at each stage, abnormal events and handling methods, resource usage, and execution effect evaluation.

[0122] Adaptive parameter adjustment based on control execution reports and non-real-time control commands updates the specialized agent control strategy library, completing the station-level intelligent agent autonomous control closed loop and achieving continuous system optimization. Adaptive parameter adjustment is a technique for dynamically adjusting control parameters based on execution feedback, primarily applied in non-real-time control scenarios in rail transit station-level intelligent agents. The adjustment process first extracts key performance indicators (KPIs) from the control execution report, such as response time, resource utilization, and control accuracy; then, it compares these with expected targets to calculate the performance gap; next, based on the performance gap and preset adjustment rules, it determines the direction and magnitude of parameter adjustment; finally, it applies the adjusted parameters to relevant control algorithms. The specialized agent control strategy library is a knowledge base storing various control strategies and parameter configurations, organized in a hierarchical structure, including a basic strategy layer, a scenario strategy layer, and an optimization strategy layer. The strategy update mechanism, based on the evaluation results of the control execution report, marks inefficient or problematic strategies, introduces optimized new strategies, and updates the strategy scores. The station-level intelligent agent autonomous control closed loop refers to a complete control cycle from perception, decision-making, execution, feedback, and optimization. Through continuous data acquisition, analysis, execution, and feedback, an adaptive control system is formed, continuously improving control performance and intelligence.

[0123] Taking the emergency evacuation of large passenger flows at rail transit stations as an example, the intelligent decision mapper first receives the intelligent decision plan "Activate large passenger flow evacuation plan A". It then decomposes this plan into multiple specific instructions, including "PIS - Display evacuation guidance information", "PA - Play evacuation broadcast", "BAS - Adjust air conditioning airflow", and "PSD - Adjust door opening time", forming a professional execution instruction set. The edge computing task distribution mechanism analyzes the instruction attributes, identifying "PIS - Display evacuation guidance information" and "PA - Play evacuation broadcast" as real-time control instructions and assigning them to the real-time channel, while identifying "BAS - Adjust air conditioning airflow" as a non-real-time control instruction and assigning it to the non-real-time channel. The scheduling priority evaluator calculates the priority of each real-time instruction. "PA - Play evacuation broadcast" receives the highest priority (90 points) due to its direct impact on personnel safety, while "PIS - Display evacuation guidance information" receives 85 points. Based on this, a linkage control timing table is formed: t=0 seconds, PA plays the evacuation broadcast; t=0.5 seconds, PIS displays the evacuation guidance information; t=2 seconds, PSD adjusts the door opening time. The distributed transaction mechanism initiates a two-phase commit: first, a preparation request is sent to PA, PIS, and PSD. After confirming resource readiness, the three entities return a ready response. Then, the formal execution command is sent, and the three entities synchronously start the execution program. PA activates the emergency broadcast, PIS switches the display screen content, and PSD modifies the control parameters, jointly completing the evacuation guidance. The status feedback collector collects the execution status in real time: PA returns "Broadcast activation successful, volume 85 decibels," PIS returns "Display content switching complete, all 15 displays updated," and PSD returns "Parameter modification successful, door opening time adjusted to 12 seconds." Correlation analysis integrates this data to generate a control execution report, recording the execution status, time points, and effect data of the entire evacuation process. Based on the execution report, it was found that the PSD door opening time was slightly insufficient. Combining passenger flow data, an adaptive parameter adjustment mechanism adjusted the door opening time from 12 seconds to 15 seconds and updated relevant parameters in the professional body control strategy library to provide a better configuration for similar scenarios in the future, thereby completing the station-level intelligent body autonomous control closed loop. The entire process demonstrates the efficient collaborative control capability of the rail transit station-level intelligent body implementation method in complex scenarios.

[0124] The above describes the implementation method of the rail transit station-level intelligent agent in the embodiments of this application. The following describes the implementation system of the rail transit station-level intelligent agent in the embodiments of this application. Please refer to [link / reference]. Figure 5 One embodiment of the rail transit station-level intelligent agent implementation system in this application includes:

[0125] Modules are used to build a general professional body base, configure a real-time embedded operating system, and form a modular professional body architecture.

[0126] The transmission module is used to design a secure communication bus according to the professional architecture, perform differentiated transmission processing on data, and establish a high-speed communication network.

[0127] The allocation module is used to combine multiple ARM processors into a station-level multi-core CPU cluster based on a high-speed communication network, dynamically allocate workloads, and build a station-level intelligent agent central system.

[0128] The hierarchical module is used to establish a unified real-time database URTDB and a unified entity database UDB based on the station-level intelligent agent central system, and to perform hierarchical management of system data to form the intelligent agent data foundation;

[0129] The prediction module is used to perform multi-dimensional analysis and prediction of real-time station-level operational data and generate intelligent decision-making solutions.

[0130] The control module is used to perform edge computing and distributed linkage control based on intelligent decision-making schemes, coordinate and schedule various professional subsystems, and realize the autonomous control of station-level intelligent agents.

[0131] Through the collaborative efforts of the aforementioned components, a general-purpose professional system base is constructed using a high-performance 64-bit ARM processor, configured with a real-time embedded operating system, forming a modular professional system architecture. This effectively solves the problem of independent deployment and difficulty in collaborative work among various professional subsystems in traditional rail transit station-level systems, achieving efficient integration and flexible configuration of hardware resources. Based on the professional system architecture, an FPGA is used to implement a LiteIP secure communication bus, employing a dynamic key encryption mechanism for differentiated data transmission processing. This not only reduces communication latency from the hundreds of milliseconds of traditional TCP / IP to sub-milliseconds, but also significantly improves data transmission efficiency and security through differentiated processing of global shared data streams, group shared data streams, and point-to-point data streams. Multiple ARM processors are combined into a station-level multi-core CPU cluster, and workload is dynamically allocated through a task priority allocation algorithm. The constructed station-level intelligent system fully utilizes computing resources, reducing the response time of critical tasks to less than 20ms, far superior to the 300ms response time of traditional systems. A unified real-time database URTDB and a unified entity database UDB are established, utilizing a multi-level caching mechanism and a DSI data service interface. The system employs hierarchical data management, completely resolving the "data silo" problem in traditional systems. This allows various professional entities to easily access global data, providing a complete data foundation for intelligent analysis. Particularly in artificial intelligence applications, this solution integrates federated learning and deep neural network technologies, enabling different professional entities to share model knowledge while protecting local data privacy. This significantly improves the overall performance of AI models. For example, the context-aware engine, through the combination of convolutional and recurrent neural networks, successfully integrates and analyzes multimodal data (visual, temporal, sensor, etc.), accurately identifying complex scene states and predicting trend changes, providing a reliable basis for intelligent decision-making. Ultimately, through edge computing and distributed linkage control, it achieves comprehensive autonomous control capabilities, from millisecond-level emergency control to long-term optimization. Especially in emergencies such as fires and large passenger flows, the entire processing flow from data acquisition, analysis, and instruction execution takes only 80ms, nearly four times faster than traditional systems. This significantly improves the safety and operational efficiency of rail transit stations. Furthermore, through adaptive parameter adjustments and continuous updates to the professional entity control strategy library, the system can continuously evolve and optimize to adapt to changing operating environments and needs.

[0132] Reference Figure 6 This invention also provides a computer device, which can be a server, and its internal structure can be as follows: Figure 6As shown, the computer device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0133] Those skilled in the art will understand that Figure 6 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0134] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0135] Those skilled in the art will understand that implementing all or part of the processes in the methods of the above embodiments can be accomplished by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the present invention and embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0136] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0137] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0138] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A rail transit station level agent implementation method, characterized in that, The application relates to a station-level intelligent agent system and a method for constructing the station-level intelligent agent system. The application comprises the following steps: A general professional body base is constructed, a real-time embedded operating system is configured, and a modular professional body architecture is formed; A safe communication bus is designed according to the professional body architecture, differential transmission processing is carried out on data, and a high-speed communication network is established; Based on the high-speed communication network, a plurality of ARM processors are combined into a station-level multi-core CPU cluster, work loads are dynamically allocated, and a station-level intelligent agent central system is constructed; A unified real-time database URTDB and a unified entity database UDB are established relying on the station-level intelligent agent central system, system data is managed in layers, and an intelligent agent data foundation is formed; Multi-dimensional analysis and prediction are carried out on real-time station-level operation data, and an intelligent decision scheme is generated, which comprises the following steps: multi-source heterogeneous data is extracted from the unified real-time database URTDB and the unified entity database UDB through a data acquisition preprocessing unit, the multi-source heterogeneous data is subjected to denoising, normalization and feature extraction processing, and a standardized data set is formed; a multi-dimensional feature vector is constructed by evaluating and screening the standardized data set by using a feature selection algorithm, and a training sample library is generated; a hierarchical federated learning architecture is designed according to the training sample library, model gradient information is transmitted after local calculation processing of each professional body data, model parameter aggregation is completed, a global deep neural network model is constructed, each professional body can share model knowledge under the premise of protecting local data privacy; a situation awareness engine is created according to the global deep neural network model, visual data is processed through a convolutional neural network, time series data is analyzed by using a recurrent neural network, multi-modal information is fused, and the current station-level operation state is identified; a multivariate time series prediction matrix is generated by the situation awareness engine, forward prediction is carried out on key indicators of the station level, and a warning threshold dynamic adjustment mechanism is constructed; decision tree analysis is carried out based on the warning threshold dynamic adjustment mechanism and a preset rule library, a multi-level alternative scheme set is generated, the multi-level alternative scheme set is optimized and evaluated by using a reinforcement learning algorithm, and an intelligent decision scheme is output; In the hierarchical federated learning architecture, station-level intelligent agents are divided into two levels of intra-station federated learning and inter-station federated learning; in the intra-station federated learning, each professional body acts as a client, and the station-level intelligent agent central system acts as a server; in the inter-station federated learning, the intelligent agents of each station act as clients, and a line network center server acts as an aggregation server; 2.The rail transit station-level agent implementation method according to claim 1, characterized in that, Based on the intelligent decision scheme, edge computing and distributed linkage control are carried out, each professional subsystem is cooperatively scheduled, and station-level intelligent agent autonomous control is realized. The general professional body base is constructed, the real-time embedded operating system is configured, and the modular professional body architecture is formed, which comprises the following steps: A plurality of time sequence control units are arranged to accurately control clock signals of the ARM processors, and real-time processing capability with 1ms time resolution is obtained; The real-time embedded operating system is divided into a real-time task layer and an application program layer, and the real-time task layer manages hardware resource allocation; A protocol adaptation conversion module is arranged in the general professional body base, and different communication protocols are converted into a unified standard data format; According to the function is divided into BAS professional body, FAS professional body, PIS professional body, PA professional body, PSD professional body, ATS professional body, SCADA professional body seven types of professional body, constitute professional body population; Through the private cloud resource allocation algorithm to the professional body base computing resources and storage space for dynamic configuration, form elastic computing architecture; From the professional body population, select the corresponding professional body and the corresponding PLC equipment to establish communication connection, constitute the complete modular professional body architecture. 3.The rail transit station-level agent implementation method according to claim 1, characterized in that, The safe communication bus is designed according to the professional body architecture, and the data is differentially transmitted and processed to establish a high-speed communication network, including: Customize the LiteIP protocol stack through FPGA hardware circuit to build a high-performance protocol processing unit; The communication data stream is divided into three types: global shared data stream, group shared data stream and point-to-point data stream, which are respectively assigned different transmission priorities; Based on the principle of elliptic curve cryptography, a dynamic key pair is generated, and different encryption strategies are applied to different types of data streams; Through the hardware timestamp mechanism and the transmission frame sequence number, the data packet is marked to form a reliable data packet transmission confirmation mechanism; According to the topological structure information, the optimal transmission path is automatically calculated, and the data packet is intelligently routed and forwarded; Utilize the communication bandwidth dynamic allocation technology to coordinate and manage the data transmission between multiple professional bodies, and establish a high-speed communication network. 4.The rail transit station-level agent implementation method according to claim 1, characterized in that, The multiple ARM processors are combined into a station-level multi-core CPU cluster based on the high-speed communication network, and the work load is dynamically allocated to build a station-level agent hub system, including: The multiple ARM processors are connected in a physical network according to the master-slave architecture to form a processor array; The processors in the processor array are defined as roles, divided into three types: master processor, computing processor and communication processor; Based on the high-speed communication network, a communication channel between processors is established, and a processor resource pool is created; Through the state monitoring mechanism, the processor load in the processor resource pool is monitored in real time, and a resource state diagram is generated; According to the resource state diagram, identify the task type and urgency, and assign priority weight coefficients to different tasks; According to the priority weight coefficient, the work load is dynamically allocated, and the task is distributed to the corresponding processor for execution to build a station-level agent hub system.

5. The rail transit station level agent implementation method according to claim 1, characterized in that, Based on the station-level agent hub system, a unified real-time database URTDB and a unified entity database UDB are established, and the system data is managed in layers to form an agent data foundation, including: In the station-level agent hub system, the memory area is divided, and the high-speed access memory space is allocated for the unified real-time database URTDB; Through data persistence mapping technology, a unified entity database UDB is built to store and manage parameter configuration information and historical data; According to the access frequency, the system data is classified, the high-frequency access data is stored in L1 cache, the medium-frequency access data is stored in L2 cache, and the low-frequency access data is stored in main storage area; Use the data mapping table to establish the data association relationship between URTDB and UDB, and build a hot data and cold data automatic switching mechanism; The DSI data service interface is created for each professional body, and the data structure is divided into a global concept layer, a local concept layer, a memory physical layer, an object layer, and an application physical layer; Data reading and writing operations and transaction processing are performed based on the DSI data service interface to form an intelligent agent data foundation. 6.The rail transit station-level agent implementation method according to claim 1, characterized in that, Edge computing and distributed linkage control are performed based on the intelligent decision scheme to cooperatively schedule each professional subsystem and realize station-level intelligent agent autonomous control, including: An intelligent decision scheme is decomposed into a professional body execution instruction set by an intelligent decision mapper; The professional body execution instruction set is divided into real-time control instructions and non-real-time control instructions according to functional attributes by using an edge computing task distribution mechanism to construct a dual-channel scheduling framework; An emergency response processing flow is triggered by the real-time control instructions in the dual-channel scheduling framework, and the instruction execution order is sorted by a scheduling priority evaluator to form a linkage control time sequence table; Control commands are issued to each professional subsystem according to the linkage control time sequence table by a distributed transaction mechanism to synchronously start an execution program in each professional body base to produce a cooperative control effect; Execution state data is obtained from each professional subsystem by a state feedback collector, the execution state data is associated and analyzed to generate a control execution report; Adaptive parameter adjustment is performed based on the control execution report and the non-real-time control instructions to update a professional body control strategy library to complete a station-level intelligent agent autonomous control closed loop.

7. A rail transit station level agent implementation system, configured to implement the rail transit station level agent implementation method according to any one of claims 1-6, wherein, It includes: A construction module is configured to construct a general professional body base, configure a real-time embedded operating system, and form a modular professional body architecture; A transmission module is configured to design a secure communication bus according to the professional body architecture, perform differential transmission processing on data, and establish a high-speed communication network; An allocation module is configured to combine multiple ARM processors into a station-level multi-core CPU cluster based on the high-speed communication network, dynamically allocate workloads, and construct a station-level intelligent agent hub system; A hierarchical module is configured to establish a unified real-time database (URTDB) and a unified entity database (UDB) based on the station-level intelligent agent hub system, perform hierarchical management on system data, and form an intelligent agent data foundation. The prediction module is used for multi-dimensional analysis and prediction of real-time station-level operation data, and generates an intelligent decision scheme, including: extracting multi-source heterogeneous data from the unified real-time database URTDB and the unified entity database UDB through a data acquisition preprocessing unit, performing denoising, normalization and feature extraction processing on the multi-source heterogeneous data to form a standardized data set; using a feature selection algorithm to evaluate and screen the standardized data set, constructing a multi-dimensional feature vector, and generating a training sample library; designing a hierarchical federated learning architecture according to the training sample library, performing local calculation and processing on each professional body data, transmitting only model gradient information, completing model parameter aggregation, and constructing a global deep neural network model; creating a context-aware engine according to the global deep neural network model, processing visual data through a convolutional neural network, analyzing time series data using a recurrent neural network, fusing multi-modal information, and identifying the current station-level operation state; generating a multivariate time series prediction matrix from the context-aware engine, forward predicting key station-level indicators, and constructing a warning threshold dynamic adjustment mechanism; performing decision tree analysis based on the warning threshold dynamic adjustment mechanism and a preset rule base, generating a multi-level alternative scheme set, optimizing and evaluating the multi-level alternative scheme set through a reinforcement learning algorithm, and outputting an intelligent decision scheme; The control module is used for performing edge computing and distributed linkage control based on the intelligent decision scheme, cooperatively scheduling each professional subsystem, and realizing autonomous control of the station-level intelligent agent.

8. A computer device, comprising: The computer readable storage medium has a computer program stored thereon, and the computer program enables the processor to execute the rail transit station-level intelligent agent implementation method in any one of claims 1 to 6 when the computer program is run on the processor.

9. A computer readable storage medium having a computer program stored thereon, the computer program enabling the processor to execute the rail transit station-level intelligent agent implementation method in any one of claims 1 to 6 when the computer program is run on the processor.

Images