Multi-source data fusion type passenger station intelligent management and control method and system based on AI drive

Abstract

The invention relates to a multi-source data fusion type passenger station intelligent management and control method and system based on AI drive, and belongs to the technical field of infrastructure intelligent management, and the intelligent management and control method comprises the steps: obtaining video stream data through an intelligent camera, obtaining environment monitoring data through an Internet of Things sensor, and obtaining ticket checking gate log data through a business system interface. And performing edge calculation processing on the video stream data to generate structured metadata. And according to the space grid ID alignment data, aggregating and generating a fusion feature matrix according to a time window. And calling a passenger flow prediction model to process the fused feature matrix, and outputting a passenger flow prediction value and a confidence score. The predicted value is compared with a dynamic threshold, and a standardized decision event is generated when the threshold is exceeded. And a linkage task package is generated through rule engine matching, the linkage task package is distributed to each device through message middleware to execute cooperative operation, and model parameters and threshold library data are updated according to a device response message. The management and control requirements of large-scale passenger flow in the passenger station can be effectively met.

Description

Technical Field

[0001] This application relates to the field of intelligent infrastructure management technology, and in particular to an AI-driven, multi-source data fusion-based intelligent control method and system for passenger stations. Background Technology

[0002] In modern transportation systems, highway passenger stations and passenger terminals within integrated transportation hubs of Class I and above serve as key nodes, and their operational management level directly impacts passenger travel efficiency, public safety experience, and the overall effectiveness of the transportation network. Currently, these terminals generally employ a certain level of information management system to digitize core operations such as ticketing, security monitoring, information dissemination, and public address guidance. These typical existing solutions often adopt a "subsystem integration" architecture, where dedicated software systems are independently built or procured around different business functions. Examples include independent ticketing management systems, independent video surveillance systems, and independent information dissemination and broadcasting systems. These subsystems, within their respective functional domains, enable the digitization of basic business operations; for instance, the ticketing system handles ticket purchase and inspection processes, while the video surveillance system records and stores images.

[0003] However, this architecture has gradually revealed its inherent limitations in practical applications. Because the various subsystems are often independent in their construction phases, technical standards, and supplier selections, significant differences exist in their physical deployments, data formats, communication protocols, and even data management platforms, creating isolated "information silos." For example, real-time ticket sales and check-in data generated by the ticketing system is difficult for the video surveillance system to acquire in real time and use to assist in analyzing the causes of passenger flow in specific areas; abnormal images captured by the video surveillance system, such as excessive crowding in local areas, cannot automatically and immediately trigger the broadcast system for voice guidance or information release system updates and prompts. System collaboration heavily relies on operators manually switching, observing, judging, and operating multiple independent interfaces. This sequential manual intervention mode is not only slow to respond but also prone to problems such as distorted information transmission, incomplete decision-making basis, and asynchronous response actions due to human factors when facing rapidly changing passenger flow dynamics or sudden emergencies. Furthermore, the system's functions are largely limited to recording and post-event queries of events that have already occurred, failing to provide strong support for forward-looking decision-making such as passenger flow prediction, precise allocation of transportation capacity, and diagnosis of service bottlenecks, thus making it difficult to cope with the needs of managing large-scale passenger flows. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides an AI-driven, multi-source data fusion-based intelligent management and control method and system for passenger stations.

[0005] Firstly, this application provides an AI-driven, multi-source data fusion-based intelligent management and control method for passenger stations, employing the following technical solution: A multi-source data fusion-based intelligent management and control method for passenger stations driven by AI, the intelligent management and control method comprising: Video stream data is obtained through smart cameras deployed in the passenger station, environmental monitoring data is obtained through IoT sensors, and ticket gate log data is obtained through business system interfaces. Edge computing processing is performed on the video stream data to generate structured metadata containing passenger flow density values ​​and abnormal behavior identifiers, while a unified timestamp and spatial grid ID label are attached to all acquired data. Structured metadata with spatiotemporal tags, environmental monitoring data, and ticket check log data are input into the streaming computing engine. Data from the same region are aligned according to the spatial grid ID, aggregated and calculated according to a preset time window, and a fusion feature matrix is ​​generated. The pre-configured passenger flow prediction model is invoked to process the fused feature matrix, and the passenger flow prediction value and confidence score for future time periods are output. The predicted passenger flow value is compared with the preset passenger flow threshold in the corresponding spatiotemporal dimension in the dynamic threshold library. When the predicted passenger flow value exceeds the preset passenger flow threshold, a standardized decision event is generated. The standardized decision events are input into a pre-configured rule engine for rule entry matching to generate a linkage task package containing a sequence of device control instructions; The linkage task package is distributed to the target subsystem through the message middleware, driving the broadcasting equipment, information display screen and access control controller to perform coordinated operations, and receiving the device response messages in real time; Based on the execution status code in the device response message, update the passenger flow prediction model parameters and dynamic threshold library data.

[0006] By adopting the above-mentioned technical solutions, the traditional, decentralized, manual, and slow-responding operation and management model of passenger stations is transformed into a data-driven, intelligent decision-making, automated collaborative, and continuously optimized smart management and control system. This technical solution achieves a fundamental shift in operation and management from passive response to proactive prediction and intervention, significantly improving the speed of emergency response, daily operational efficiency, scientific resource allocation, and passenger service quality of passenger stations. Simultaneously, the system's self-learning characteristics ensure its long-term applicability and advanced nature, providing a complete, efficient, and evolutionary intelligent solution for addressing the complex management challenges of large passenger transport hubs.

[0007] Secondly, this application provides an AI-driven multi-source data fusion-based intelligent management and control system for passenger stations, employing the following technical solution: An AI-driven, multi-source data fusion-based intelligent management and control system for passenger stations, the intelligent management and control system comprising: The multi-source sensing module is used to acquire video stream data through smart cameras deployed in the passenger station, acquire environmental monitoring data through IoT sensors, and acquire ticket gate log data through business system interfaces. The data processing module is used to perform edge computing processing on the video stream data, generate structured metadata containing passenger flow density values ​​and abnormal behavior identifiers, and attach a unified timestamp and spatial grid ID tag to all acquired data. The data feature fusion module is used to input structured metadata with spatiotemporal labels, environmental monitoring data, and ticket check log data into the streaming computing engine, align data in the same area according to the spatial grid ID, aggregate and calculate according to a preset time window, and generate a fusion feature matrix. The predictive analysis module is used to call a pre-configured passenger flow prediction model to process the fused feature matrix and output the passenger flow prediction value and confidence score for future time periods. The decision module is used to compare the passenger flow forecast with the preset passenger flow threshold in the corresponding spatiotemporal dimension in the dynamic threshold library, and generate a standardized decision event when the passenger flow forecast exceeds the preset passenger flow threshold. The linkage control module is used to input the standardized decision events into a pre-configured rule engine for rule entry matching and generate a linkage task package containing a sequence of device control instructions. The execution control module is used to distribute the linkage task package to the target subsystem through the message middleware, drive the broadcasting equipment, information display screen and access control controller to perform coordinated operations, and receive the device response messages in real time; The self-learning optimization module is used to update the passenger flow prediction model parameters and dynamic threshold library data based on the execution status code in the device response message.

[0008] Thirdly, this application provides a computer device, which adopts the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to perform the steps of the method as described in the first aspect.

[0009] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect. Attached Figure Description

[0010] Figure 1This is a schematic diagram of the first process of an AI-driven multi-source data fusion intelligent management and control method for passenger stations, according to one embodiment of this application.

[0011] Figure 2 This is a schematic diagram of the second process of an AI-driven multi-source data fusion intelligent management and control method for passenger stations, according to one embodiment of this application.

[0012] Figure 3 This is a schematic diagram of the third process of an AI-driven multi-source data fusion-based intelligent management and control method for passenger stations, according to one embodiment of this application.

[0013] Figure 4 This is a schematic diagram of the fourth process of an AI-driven multi-source data fusion-based intelligent management and control method for passenger stations, according to one embodiment of this application.

[0014] Figure 5 This is a schematic diagram of the fifth process of an AI-driven multi-source data fusion-based intelligent management and control method for passenger stations, according to one embodiment of this application.

[0015] Figure 6 This is a schematic diagram of the sixth process of an AI-driven multi-source data fusion intelligent management and control method for passenger stations, according to one embodiment of this application.

[0016] Figure 7 This is a schematic diagram of the seventh process of an AI-driven multi-source data fusion intelligent management and control method for passenger stations, according to one embodiment of this application. Detailed Implementation

[0017] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-7 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0018] This application discloses an AI-driven, multi-source data fusion-based intelligent management and control method for passenger stations.

[0019] Reference Figure 1 A method for intelligent management and control of passenger stations based on AI-driven multi-source data fusion, specifically including: Step S101: Obtain video stream data through smart cameras deployed in the passenger station, obtain environmental monitoring data through IoT sensors, and obtain ticket gate log data through the business system interface; Traditional passenger station management systems often rely on single, isolated data sources, failing to provide a comprehensive understanding of operational status. This application addresses this by constructing a holistic perception system using three different types of data sources: Video stream data from smart cameras provides the most intuitive and richest visual information, forming the foundation for passenger flow analysis, behavior recognition, and security monitoring. Its advantage lies in its ability to capture large-scale, non-contact dynamic scene information.

[0020] Environmental monitoring data from IoT sensors (such as temperature, humidity, air quality, and noise): This data reflects the comfort and safety of the passenger station's physical environment. Environmental factors directly affect passenger gathering behavior (for example, passengers may be more inclined to gather in well-ventilated areas when the temperature is too high) and facility operation status, serving as an important basis for refined management and emergency decision-making.

[0021] Ticket gate log data from the business system interface: This is structured data from the core operations system, providing accurate time-series records of passenger flow, such as the number of passengers checking tickets, train schedule correlations, and real-time station passenger estimates. It represents the core logic of passenger station operations.

[0022] By combining these three types of data, the system achieves multi-dimensional information coverage from visual perception (video), physical environment perception (sensors) to business logic perception (logs), laying a solid and comprehensive data foundation for subsequent data fusion and intelligent analysis, and fundamentally solving the "information silo" problem.

[0023] Step S102: Perform edge computing processing on the video stream data to generate structured metadata containing passenger flow density values ​​and abnormal behavior identifiers, and attach a unified timestamp and spatial grid ID label to all acquired data. Edge computing processing specifically includes running lightweight algorithms on embedded AI chips at the data generation source (i.e., smart cameras or their near-end edge computing devices) to directly analyze video streams in real time. This distributed computing model transforms unstructured video streams into structured information (i.e., metadata) that machines can directly understand and process. For example, instead of transmitting complete video footage, the algorithm outputs structured results such as "In grid A01 area, 15 people have been identified, with a density of 0.75 people / square meter, indicating a suspected gathering event." This significantly reduces network transmission load, lowers the computational complexity of the central platform, and significantly improves the real-time performance of data processing.

[0024] Next, the additional spatiotemporal labels specifically include: a unified timestamp ensures that all data from different sources and collection frequencies can be accurately aligned on the timeline, which is a prerequisite for time-series correlation analysis. The spatial grid ID label discretizes continuous physical space into standardized grid units (e.g., dividing waiting halls and ticket halls into 1m×1m or 5m×5m grids), providing a unified spatial coordinate reference for all data. This allows information that was originally physically dispersed and heterogeneous in data format (such as camera data from a corner, temperature and humidity readings for that area, and ticket checking records from nearby turnstiles) to be linked together through the key attribute of "spatial grid ID," achieving data normalization in the spatiotemporal dimension and creating the necessary conditions for the next step of multi-source data fusion.

[0025] Step S103: Input the structured metadata with spatiotemporal tags, environmental monitoring data and ticket check log data into the streaming computing engine, align the data in the same area according to the spatial grid ID, aggregate and calculate according to the preset time window, and generate a fusion feature matrix. The integrated feature matrix includes regional passenger flow saturation, average environmental indicators, and the difference between planned train schedules and the number of passengers who have already checked in. Specifically, unlike traditional batch processing, streaming computing engines (such as Apache Flink and Spark Streaming) can perform low-latency, high-throughput real-time processing on continuously generated data streams. This aligns with the high real-time requirements of passenger station management, ensuring that analytical decisions are based on the freshest data. The system correlates data from different sources within the same grid cell based on an attached "spatial grid ID." For example, it correlates "passenger flow density value of grid B10," "temperature sensor reading of grid B10," and "number of passengers checked by ticket gates serving the area near grid B10 in the most recent time window." This alignment breaks down physical and logical barriers between data sources, forming a unified data view describing the overall state of the grid cell.

[0026] Next, the data stream is sliced ​​and aggregated according to a preset time window (e.g., every 5 minutes). Within each window, statistical calculations are performed on the aligned data to generate aggregated features such as "regional passenger flow saturation" (the average or peak passenger flow density within the window), "average environmental indicators" (the average temperature and humidity within the window), and "difference between planned departures and the number of passengers who have checked tickets" (reflecting capacity gaps). These features together constitute a fused feature matrix, where rows can represent different spatial grids and columns represent different aggregated features. This matrix is ​​no longer a one-dimensional, isolated data point, but a multi-dimensional snapshot of the state rich in contextual information, capable of comprehensively and quantitatively describing the overall operational status of the passenger station at a specific time and in a specific area, providing high-quality input for the AI ​​model.

[0027] Step S104: Call the pre-configured passenger flow prediction model to process the fusion feature matrix and output the passenger flow prediction value and confidence score for future time periods. Passenger flow prediction models are typically machine learning models trained on historical data, such as Long Short-Term Memory (LSTM) networks or Gated Recurrent Unit (GRU) deep learning models. These models are particularly adept at learning long-term dependencies and periodic patterns (such as daily peaks and weekly patterns) from time-series data. The model takes a fused feature matrix (containing historical and current passenger flow, environmental, and business data) as input and, through its complex internal nonlinear calculations, outputs a passenger flow prediction for a specific future time period (the next 15 minutes).

[0028] In this embodiment, the confidence score quantifies the reliability of the prediction result. The model may calculate the confidence score based on the prediction interval, the model's own uncertainty estimate of the input data (such as using Bayesian deep learning), or historical prediction errors. A high confidence score means that the model is quite confident in the prediction, while a low confidence score suggests that the prediction result may be unreliable, requiring human intervention or a more conservative strategy. This increases the robustness and reliability of the system's decision-making and avoids the risks that may arise from blindly trusting a single prediction value.

[0029] Step S105: Compare the passenger flow forecast with the preset passenger flow threshold in the corresponding spatiotemporal dimension in the dynamic threshold library. When the passenger flow forecast exceeds the preset passenger flow threshold, a standardized decision event is generated. The standardized decision events include event type codes, confidence scores, and associated grid IDs; Specifically, unlike fixed thresholds, the thresholds in the dynamic threshold library are dynamically adjusted based on historical data from the same period, real-time operational status (such as current shift density), and even environmental factors (such as weather). For example, the passenger flow threshold for waiting area A at 10:00 AM on a weekend will automatically be higher than at the same time on a weekday, because historical patterns show that passenger flow is greater on weekends. This dynamism makes the thresholds more context-adaptive and scientific, avoiding the problem of fixed thresholds being either too sensitive (generating a large number of false alarms) or too insensitive (missing real risks) in complex real-world scenarios.

[0030] In this embodiment, when the predicted value exceeds a threshold, the system does not act directly, but instead generates a structured "decision event." This event includes an "event type code" (such as "passenger flow overload warning"), a "confidence score," and an "associated grid ID" (locating the event's occurrence area). This standardized format allows the event to be uniformly and efficiently parsed and processed by the subsequent rule engine, serving as a key interface for achieving automated processes. It transforms a complex prediction result into a clear, program-understandable "decision trigger signal."

[0031] Step S106: Input the standardized decision events into the pre-configured rule engine for rule entry matching to generate a linkage task package containing a sequence of device control instructions; The rules engine is a software component that pre-defines multiple "event-condition-action" rules. When a standardized decision event is input, the engine matches the event type, grid ID, confidence level, and other attributes against the conditions in the rule base. Once a match is found, the rules engine triggers the corresponding action sequence. These actions are predefined, cross-system collaborative operation instructions.

[0032] For example, a rule might specify: "IF Event Type = 'Passenger Flow Exceeds Limit Warning' AND Grid ID = 'Waiting Hall A' AND Confidence > 0.8, THEN Execute: 1. Call the broadcast system API to play the evacuation guidance message for Area A; 2. Call the information publishing system API to switch the content of the displays near Area A to the evacuation guidance message; 3. Call the access control system API to unlock the backup passage for Area A." The engine encapsulates these instructions into a linked task package. This mechanism transforms complex emergency procedures requiring multi-person collaboration into a rule-driven, millisecond-level automated instruction sequence, achieving true "integrated management and control."

[0033] Step S107: Distribute linkage task packages to the target subsystem through message middleware to drive broadcasting equipment, information display screen and access control controller to perform coordinated operations, and receive device response messages in real time; Using message middleware (such as RabbitMQ or Kafka) to deliver task packages is an asynchronous communication model. The sender (rules engine) can continue processing other transactions after sending the task package to the message queue without waiting for an immediate response from the receiver (each subsystem). This improves system throughput and responsiveness, and avoids system-wide blocking due to the temporary unavailability of a subsystem.

[0034] In this embodiment, the message middleware decouples the generator of instructions (rule engine) from the executors (each subsystem). Neither party needs to know the other's technical details or network address; they only need to communicate according to the agreed-upon message format. This greatly improves the system's flexibility, scalability, and maintainability. When adding or replacing subsystems, only their message interfaces need to be adapted; there is no need to modify the core rule engine.

[0035] In addition, the system receives response messages from each subsystem after executing instructions, and these messages contain execution status codes. This is a crucial part of achieving closed-loop control, providing real-time feedback data for subsequent performance evaluation and system optimization.

[0036] Step S108: Update the passenger flow prediction model parameters and dynamic threshold library data based on the execution status code in the device response message.

[0037] Specifically, the system compares the predicted passenger flow with the actual passenger flow data observed after the command is executed (which can be obtained through subsequent data collection). If there is a persistent deviation, the system can use this deviation data (as a loss signal) to fine-tune the parameters of the passenger flow prediction model (such as the weights of the LSTM network) through methods such as online learning or periodic retraining, so that future predictions are more accurate.

[0038] Meanwhile, the system optimizes the dynamic threshold library based on the effectiveness of command execution (e.g., whether passenger flow is alleviated as expected after the diversion command is executed) and long-term operational data. If a certain threshold is frequently triggered but does not produce the expected effect or fails to provide effective warnings, the system can automatically adjust the setting logic or value of that threshold to better meet actual management needs.

[0039] The above embodiments transform the traditional, decentralized, manual, and slow-responding operation and management model of passenger stations into a data-driven, intelligent decision-making, automated collaborative, and continuously optimized smart management and control system. This application's technical solution achieves a fundamental shift in operation and management from passive response to proactive prediction and intervention, significantly improving the speed of emergency response, daily operational efficiency, scientific resource allocation, and passenger service quality of passenger stations. Simultaneously, the system's self-learning characteristics ensure its long-term applicability and advanced nature, providing a complete, efficient, and evolutionary intelligent solution for addressing the complex management challenges of large passenger transport hubs.

[0040] Based on the core logic of "unified perception, integrated analysis, intelligent decision-making, and collaborative execution", this application transforms the scattered and isolated subsystems into controlled terminals and data sources of the platform by constructing a centralized intelligent platform. By deploying an AI analysis engine and a rule engine on the platform, it can achieve real-time understanding of the overall operational status of the site and cross-system automated control.

[0041] In terms of system setup, intelligent network cameras with embedded AI chips are deployed at key locations in the passenger station, connected to the station's dedicated security network via PoE switches. A central platform server cluster is deployed in the server room, interconnected with the security network and business network via core switches. A network interface module is added to the traditional system controller to enable it to receive platform control commands. In the software platform deployment, integrated management and control platform software is installed, device access drivers are configured, and administrators configure "event-condition-action" rules.

[0042] In the workflow, taking the scenario of excessive passenger gathering in a waiting hall as an example, the intelligent camera's built-in AI algorithm analyzes the video footage in real time. When the passenger density exceeds a preset threshold, a structured alarm event message is generated and sent to the platform's real-time data lake service via the security network. After receiving the event, the AI ​​capability layer calls on data from multiple cameras to generate a passenger flow heatmap. The data intelligence analysis engine queries relevant multi-source data and uses a short-term passenger flow prediction model to conclude that passenger flow is continuously increasing. The rules engine encapsulates the original event and the prediction conclusion into a comprehensive early warning event, matches and triggers corresponding control commands in the rule base, and simultaneously sends them to each subsystem for execution through standard API interfaces. The execution status is tracked through a closed-loop feedback mechanism, and complete process data is recorded for subsequent optimization.

[0043] This application transforms data flow from a closed loop within each subsystem to a centralized platform for convergence and sharing; control flow from manual, sequential operation to AI-driven rule engines automatically generating and issuing standardized control commands in parallel; and the decision-making mechanism from relying on lagging judgments based on personal experience to real-time predictive and assisted decision-making based on multi-source data fusion and AI models. Compared to the original system, this application achieves fundamental improvements from information silos to data fusion, from manual judgment to intelligent analysis, from passive response to proactive intervention, from experience-based decision-making to data-driven decision-making, and from closed and rigid to open and scalable. It upgrades the passenger station into an organic whole commanded and dispatched by a unified intelligent brain, significantly improving safety, operational efficiency, and service levels.

[0044] Reference Figure 2 As one implementation of step S102, the step of performing edge computing processing on the video stream data to generate structured metadata containing passenger flow density values ​​and abnormal behavior identifiers includes: Step S201: Perform real-time decoding and frame buffering on the video stream data to generate a standardized video frame sequence; Due to considerations of transmission and storage efficiency, the original video stream is a highly compressed encoded stream that cannot be directly processed by image recognition algorithms. Therefore, "real-time decoding" is the process of using a decoder (such as the FFmpeg library) to restore the compressed bitstream into individual frames of bitmap format (such as RGB or YUV) images. This is a necessary decapsulation operation for computer vision processing.

[0045] Next, frame buffering is used to address the potential speed mismatch between data production rate and computational consumption rate. The system allocates a buffer in memory to temporarily store and manage decoded image frames, enabling jitter reduction, frame drop compensation, or frame extraction at a fixed frequency. This provides a stable and uniform supply of video frames to downstream analysis modules, preventing data loss or analysis delays caused by excessive instantaneous computational load.

[0046] Ultimately, "generating a standardized video frame sequence" means normalizing the size of all cached frames (e.g., scaling them to 640x480), converting the color space (e.g., to RGB), and performing any necessary normalization. Standardization eliminates input differences caused by variations in camera model, resolution, and shooting parameters, ensuring that subsequent object detection models can operate under a consistent data distribution. This is a crucial preprocessing step for guaranteeing the model's recognition accuracy and stability.

[0047] Step S202: Perform human target detection on the standardized video frame sequence and output target detection data including human bounding box coordinates and confidence scores; Human target detection typically employs detection models based on deep convolutional neural networks (CNNs), such as YOLO, SSD, or Faster R-CNN series. This model receives standardized video frames as input, and its network structure extracts deep features from the image through multiple layers of convolution and pooling operations. Finally, the output layer performs a gridded scan of the image, predicting whether a human target exists in each grid region and providing its precise location and probability of presence.

[0048] In this embodiment, the human body bounding box coordinates are typically represented in the form of (x_center, y_center, width, height), that is, in pixels, marking the position and size of the rectangular box enclosing each detected human body in the image. The confidence score is a probabilistic output of the model's certainty that a human body actually exists within the bounding box, typically ranging from 0 to 1. A high confidence score indicates that the model is very certain that a valid human target has been detected, while a low confidence score may indicate a false detection (such as misidentifying a human poster or a specially shaped object as a person) or uncertainty regarding severely occluded targets.

[0049] Step S203: Calculate the population distribution density within the preset grid area based on the human body bounding box coordinates, and generate a passenger flow density value; The generation of passenger flow density values ​​typically requires first establishing a model mapping image pixel coordinates to the actual physical space based on camera calibration parameters or preset mapping relationships. During calculation, for a given "preset grid area" corresponding to a specific area of ​​the actual ground within the passenger station, the system filters out all human bounding boxes falling within the image projection range of that area. Then, the algorithm considers the pixel area occupied by each bounding box and its position in the image (reflecting the distance to the target), estimating the approximate area occupied by the individual represented by that bounding box in the real physical grid through perspective transformation or area weighting methods.

[0050] Ultimately, by summing up the areas occupied by all individuals within the grid and comparing them with the total physical area of ​​the grid, or by employing a more complex kernel density estimation algorithm, a standardized "passenger flow density value" is generated. This value, for example, is between 0 and 1, representing the degree of population occupancy per unit area in that region. This value has more spatial significance than simply the number of people and can more accurately reflect the state of congestion, density, or sparseness.

[0051] Step S204: Extract human behavior feature vectors based on target detection data, match abnormal behavior patterns through a pre-set behavior classification model, and generate abnormal behavior identifiers when the confidence level exceeds the judgment threshold. The system uses a sequence of human bounding boxes as a foundation to extract the motion and posture features of each continuously tracked individual within a time window. For example, it can further detect skeletal keypoints and calculate the motion trajectory, velocity, and acceleration of these keypoints (such as the head, shoulders, elbows, and knees) across multiple frames. It can even perform Fourier transforms on these trajectories to extract their feature coefficients in the frequency domain. All these low-level features are combined into a high-dimensional "behavioral feature vector," which is a digital summary of the individual's action patterns during that time period.

[0052] Next, this feature vector is input into a "pre-defined behavior classification model." This model is typically a temporal classification network, such as LSTM, Transformer, or 3D CNN, trained on a large amount of labeled data (e.g., video clips of normal walking, running, gathering, and falling). The model analyzes the input feature vector and outputs the probability value of it belonging to each pre-defined abnormal behavior category (e.g., "gathering," "running," "falling"). "Matching abnormal behavior patterns" is the process by which the model compares and infers the current feature with the learned pattern library. Only when the model's output "confidence" (i.e., probability value) for a certain abnormal category exceeds the pre-defined "judgment threshold" for that category does the system "generate an abnormal behavior label." This label is a pre-defined, machine-readable code (e.g., "Crowd_Gathering," "Fall_Down") that indicates that an abnormal event with specific semantics has been successfully detected and confirmed from continuous video signals.

[0053] Step S205: Associate the passenger flow density value with the corresponding timestamp and spatial grid ID of the abnormal behavior identifier, and encapsulate them into structured metadata.

[0054] The timestamp records the precise moment when the density value or anomaly occurred, ensuring the alignment of all station data on the timeline; the spatial grid ID accurately locates the physical area corresponding to the information (such as "grid 12 in waiting hall A"), ensuring spatial fusion with data from other sources such as temperature and humidity sensors and turnstile data.

[0055] Finally, according to a predetermined data schema, such as JSON or Protocol Buffers, these elements are organized into a structured, self-describing, and lightweight data packet. This data packet is completely separated from the huge original video stream and only contains highly refined and semantically clear analysis results, so that it can be uploaded to the cloud or the central processing system in real time at an extremely low bandwidth cost, providing standardized and high-quality video dimension information input for "multi-source data fusion".

[0056] In the above embodiments, real-time and accurate in-depth understanding and purification of video information are achieved on the edge side. The unstructured, massive, and high-bandwidth-required video stream is transformed into a structured, lightweight, and semantic-rich "structured metadata packet". This transformation not only greatly reduces the pressure on network transmission and central storage, but more importantly, it maps image pixels into "passenger flow density" and "abnormal behavior" symbols that can be directly understood and processed by machines and business systems, providing high-quality, low-latency, and directly computable video semantic input for subsequent multi-source data fusion, passenger flow prediction, and intelligent linkage control.

[0057] Refer to Figure 3 , as an embodiment of step S103, the steps of inputting the structured metadata with spatio-temporal tags, environmental monitoring data, and ticket inspection log data into a streaming computing engine, aligning the data in the same area according to the spatial grid ID, and aggregating and calculating according to a preset time window to generate a fusion feature matrix include: Step S301, receiving the structured metadata with spatio-temporal tags, environmental monitoring data, and ticket inspection log data, where the spatio-temporal tags include a unified timestamp and a spatial grid ID; Among them, the structured metadata usually refers to highly condensed semantic information extracted from the video stream through edge computing, such as passenger flow density and abnormal behavior identification; the environmental monitoring data comes from Internet of Things sensors, such as temperature and humidity, air quality readings; the ticket inspection log data comes from the business system and records ticket verification and passenger passage records. Fundamental differences exist in the physical sources, collection frequencies, data formats, and semantic dimensions of these three types of data. The bond that enables them to be analyzed collaboratively is the "spatio-temporal tag" attached to each data point.

[0058] Step S302, spatially aligning the structured metadata, environmental monitoring data, and ticket inspection log data in the same physical area according to the spatial grid ID to generate a grouped data set grouped by grid; In traditional passenger stations, the various subsystems (such as security video, building automation, and ticketing systems) typically operate independently, resulting in fragmented spatial data descriptions: video data is based on pixel coordinates, environmental data on sensor physical locations, and business data on device logical IDs. This step performs a logical "spatial connection" operation using the "spatial grid ID" as a common key.

[0059] Specifically, the system maintains a spatial registry that maps the field of view of each smart camera, the installation location of each IoT sensor, and the physical location of each ticket gate to one or more "spatial grid IDs". When data flows in, the system parses the grid ID carried by each data entry and merges all video metadata (reflecting the crowd status in that area), environmental monitoring data (reflecting the physical environment of that area), and ticket check log data (reflecting the passenger flow associated with that area) that are marked as belonging to the same physical grid (e.g., "grid 12 in waiting hall A") into the same data group.

[0060] Understandably, this process achieves "spatial alignment," which means linking observational information from different dimensions describing the same physical space in the digital world to form a multi-dimensional "raw dataset" based on a grid. This dataset serves as the fundamental unit for subsequent spatiotemporal statistical analysis. It ensures that any feature calculated (such as the congestion level of a region) is indeed associated with environmental indicators (temperature) and business context (capacity) from the same physical space, avoiding erroneous associations and decisions caused by spatial mismatch.

[0061] Step S303: Based on a preset time window period, perform multi-dimensional aggregation calculation on the grouped dataset to generate a real-time feature vector containing regional passenger flow saturation, average environmental indicators, and the difference between planned trains and the number of passengers who have already checked in. To capture patterns, smooth noise, and match the pace required for decision analysis, the system introduces a "preset time window period" (e.g., 5 minutes). For each grid group, the streaming engine performs "multi-dimensional aggregation calculations" on all data within the window using a sliding or scrolling window approach. This calculation is a selective statistical summary oriented towards business objectives.

[0062] Specifically, for passenger flow indicators, the "maximum value" (reflecting peak congestion pressure) and "moving average value" (reflecting sustained congestion level) of passenger flow density within the window might be calculated. For environmental indicators, such as temperature, the average value might be taken to resist interference from instantaneous abnormal readings. For capacity indicators, it is necessary to correlate with business data and calculate the difference between the total number of seats for planned trains and the actual number of passengers who have checked tickets at the end of a window to obtain the "real-time capacity gap value".

[0063] Understandably, these carefully selected aggregation calculations extract core information from their respective dimensions from different perspectives, condensing a massive amount of fine-grained data points within a window into a fixed, low-dimensional "real-time feature vector." Each element of this vector represents an indicator with specific business implications, comprehensively reflecting a snapshot of the overall status of a grid area within a specific time period across three key dimensions: passenger flow, environment, and transportation capacity. This represents a crucial step in upgrading the information density of data from the "data level" to the "feature level."

[0064] Step S304: Combine the feature vectors of all grids within each time window into a matrix structure to obtain the spatiotemporal dimensional fusion feature matrix.

[0065] The system arranges and combines the feature vectors generated by all grids within the same time window according to the spatial logical relationship of the grids (e.g., according to the order of grid IDs). Typically, the feature vector of each grid can be considered as a row; therefore, stacking the vectors of all N grids forms an N-row matrix. Each row of this matrix represents the state of a spatial location (grid) within that time window, and all rows together constitute a "spatial state snapshot" of the entire passenger station at that moment. This matrix is ​​the embodiment of the "spatiotemporal dimensional fusion feature matrix" on a time slice.

[0066] In this embodiment, when the system processes multiple time windows consecutively, it generates a matrix sequence arranged in chronological order. This three-dimensional data structure (number of spatial grids × number of feature dimensions × number of time windows) is an ideal input for spatiotemporal data analysis. It uses rows as spatial units, columns as feature dimensions, and different time slices as depth to completely and systematically encode the joint evolution of passenger flow stations in various spatial regions and multi-dimensional indicators over a past period.

[0067] In the above implementation, data streams from different physical devices, protocols, and semantics are precisely aligned, correlated, and aggregated within a unified spatiotemporal grid coordinate system. This transforms the original low-value data streams into a high-quality, high-information-density "spatiotemporal fusion feature matrix." This technical solution fundamentally changes the application model of multi-source data, elevating it from scattered and isolated state monitoring to a unified, computable, and deeply reflective intelligent analysis foundation that reflects the complex relationships and evolution of the "human-environment-business" system. This provides a solid, reliable, and interpretable data core for the entire intelligent management and control system to achieve accurate prediction and collaborative decision-making.

[0068] Reference Figure 4 As one implementation of step S104, the step of calling a pre-configured passenger flow prediction model to process the fused feature matrix and outputting the passenger flow prediction value and confidence score for future time periods includes: Step S401: Obtain the fusion feature matrix, which includes the regional passenger flow saturation, average environmental indicators, and the difference between planned train schedules and the number of passengers who have already checked tickets for multiple spatial grids within a continuous time window; The fusion feature matrix, as the direct output of the preceding data fusion method, has the intrinsic value of systematically integrating multiple driving factors affecting passenger flow dynamics. Each row of the matrix represents a specific spatial grid, each column represents a specific past time window, and each cell stores a feature vector. This vector simultaneously includes the grid's "regional passenger flow saturation" (reflecting historical congestion levels), "average environmental indicators" (such as temperature and humidity, reflecting the potential impact of the physical environment on personnel gathering and movement), and "capacity gap value" (reflecting the real-time gap between transportation plans and actual demand, a key business signal guiding passenger flow direction).

[0069] Step S402: The spatiotemporal dimension of the fused feature matrix is ​​reorganized to generate a three-dimensional tensor data structure with spatial grids as rows and time windows as columns; Specifically, these three dimensions (space, time, and features) are explicitly separated and organized into a unified three-dimensional array, namely a "three-dimensional tensor". In particular, after reorganization, the first dimension (rows) of the tensor indexes different spatial grids, maintaining spatial independence; the second dimension (columns) indexes continuous time windows, forming a time series; and the third dimension (depth) corresponds to multiple feature values ​​(passenger flow saturation, environmental average, capacity gap, etc.) within each grid-time point unit.

[0070] Understandably, this data structure is similar to a multi-channel video: the spatial grid is like pixel locations, the temporal window is like the frame sequence, and multiple features are like different color channels (such as RGB), collectively depicting a dynamically evolving, multi-attribute station situation map. This tensor structure is crucial because it allows subsequent predictive models to use convolutional operations to capture the interactions between neighboring grids and passenger flow diffusion patterns in the spatial dimension (e.g., how congestion in the entrance area gradually spreads to the security checkpoint), while using recurrent connections to learn long-term historical dependencies and trend cycles in the temporal dimension, thereby achieving "spatiotemporal joint modeling" of passenger flow dynamics, rather than treating space and time separately.

[0071] Step S403: Input the three-dimensional tensor data structure into the pre-trained passenger flow prediction model and output the passenger flow prediction value sequence for the future preset time period. This involves using a complex machine learning model that has learned passenger flow evolution patterns from a large amount of historical data to perform forward inference on the reconstructed tensor data, thereby extrapolating future passenger flow patterns. The "pre-trained passenger flow prediction model" is typically a deep learning architecture designed specifically for spatiotemporal sequence data, such as a Convolutional Long Short-Term Memory (ConvLSTM) network. The core capability of this model lies in its simultaneous integration of the spatial feature extraction capabilities of a Convolutional Neural Network (CNN) and the temporal dependency modeling capabilities of a Long Short-Term Memory (LSTM) network.

[0072] In this embodiment of the application, when the three-dimensional tensor is input into the model, the convolutional layer inside the model will slide on each time slice to automatically learn and extract the pattern and spatial correlation of passenger flow distribution between different spatial grids (for example, to identify which areas are usually passenger flow hotspots and which areas are connected).

[0073] Meanwhile, the LSTM unit processes these continuous spatial feature maps along the time dimension. Its internal gating mechanism (forget gate, input gate, output gate) can selectively remember long-term historical states, forget irrelevant information, and integrate the current input to capture the trends, cycles (such as daily peaks), and abrupt change patterns of passenger flow. After the model "learns" from several past time windows represented by the entire input tensor, its output layer recursively or stepwise generates a "sequence of passenger flow predictions for future preset time periods" based on the learned spatiotemporal dynamics. This sequence predicts passenger flow indicators (such as number of people or density) for each spatial grid in multiple future time windows (such as three 5-minute intervals), thus providing a forward-looking, fine-grained outlook on passenger flow trends and providing a time window for advance decision-making.

[0074] Step S404: Generate a confidence score for the corresponding passenger flow prediction value sequence through the uncertainty quantification module embedded in the passenger flow prediction model; Because real-world passenger flow is affected by numerous unknown or unforeseen factors (such as temporary train delays and extreme weather), any model's predictions inherently contain uncertainty. The "uncertainty quantification module" is an algorithmic mechanism integrated within or in conjunction with the prediction model, used to estimate the model's confidence level in each prediction.

[0075] In this embodiment, the Monte Carlo Dropout method can be used. During model inference (prediction), this method does not disable the Dropout layer used to prevent overfitting, but keeps it in a randomly activated state, and performs forward propagation sampling multiple times (e.g., 100 times) on the same input tensor. Due to the randomness of Dropout, each sampling is equivalent to using a slightly different sub-model, resulting in a set of slightly different predicted values. The variance of this set of predicted values ​​intuitively reflects the degree of prediction fluctuation caused by data noise, model structure uncertainty, etc. Calculating the standard deviation of this set of predicted values ​​and mapping its reciprocal to the 0-1 interval generates a "confidence score". A high score indicates that the prediction results are highly consistent, and the model has high certainty; a low score indicates that the prediction results are scattered, and the model is "confused" or uncertain about the current input pattern. This score corresponds one-to-one with the predicted value sequence, providing a key trade-off for downstream decision-making logic (such as threshold comparison, rule triggering), enabling the system to actively intervene when the prediction confidence is high, and to adopt a more conservative strategy or trigger manual review when the confidence is low.

[0076] Step S405: Encapsulate the passenger flow prediction value sequence and confidence score into a standardized prediction output package.

[0077] This data packet contains not only a sequence of predicted passenger flow values ​​for a preset future time period and a corresponding array of confidence scores, but also crucial business and spatiotemporal identifiers, such as the associated spatial grid ID (indicating which area the prediction targets), the prediction duration (e.g., the next 15 minutes), and the starting point information of the time window. This encapsulation ensures that each data packet is self-contained and meaningful. It transforms the technical results of model predictions into standardized decision input events that downstream dynamic threshold library comparisons and rule engines can directly and accurately identify and process. For example, the rule engine can parse this JSON packet and extract the structured fact that "the predicted passenger flow value for grid B2 continuously exceeds the threshold in the next 15 minutes, and the average confidence score is higher than 0.8," thereby reliably triggering the corresponding linkage plan.

[0078] In the above implementation, a data representation method of "spatiotemporal three-dimensional tensor" and an advanced spatiotemporal prediction model integrating an uncertainty quantification mechanism are introduced to construct a precise, reliable, and self-aware passenger flow prediction capability. This technical solution transforms the complex spatiotemporal correlations and causal information contained in the multi-source fusion feature matrix into a quantitative inference of future passenger flow trends, and assigns a confidence score to each prediction point, improving the accuracy and spatiotemporal granularity of the prediction, enabling the control system to proactively identify congestion risk points; at the same time, the confidence score injects a crucial "uncertainty awareness" into the entire intelligent decision-making closed loop, enabling the system to distinguish between high-reliability predictions and low-reliability speculations, thereby making more robust and reasonable automated decisions.

[0079] Reference Figure 5 As one implementation of step S106, the step of inputting standardized decision events into a pre-configured rule engine for rule entry matching to generate a linkage task package containing a sequence of device control instructions includes: Step S501: Receive standardized decision events, including event type codes, confidence scores, and associated grid IDs; In this context, a standardized decision event is not a simple alarm signal, but a data object that encapsulates a multi-dimensional decision context.

[0080] Specifically, the event type code is a pre-defined enumeration value or code that precisely defines the business nature of the event, such as "Category A congestion warning," "Category B environmental anomaly," or "Category C emergency evacuation." This definition is the fundamental basis for all subsequent response logic selections. The confidence score is a floating-point number between 0 and 1, quantifying the certainty of the upstream AI model's (such as a passenger flow prediction model) judgment on the event. It provides a key reliability metric for rule execution, allowing response actions to be tiered or selectively triggered based on the degree of certainty in the prediction. The associated grid ID precisely anchors the physical spatial location of the event, such as "Waiting Hall A-Grid 12," which is the key geographical link mapping abstract decision logic to concrete physical execution equipment.

[0081] Step S502: Retrieve the preset rule base according to the event type code and match rule entries that meet the confidence threshold condition; This step is the core of the rules engine. Its logic involves performing a rapid, conditional retrieval and matching based on the policy knowledge base, mapping the received decision events to predefined, optimal business response logic. The pre-built rule base is a structured collection of policy knowledge, typically organized using an "event-condition-action" paradigm.

[0082] Specifically, when the engine receives a decision event, it first performs a quick search using the "event type code" as the primary key to filter out all potential rule entries designed to handle that type of event, while simultaneously setting a check for "meeting the confidence threshold condition." Each rule, when defined, can, in addition to associating with the event type, also preset a minimum confidence threshold required to execute the rule. For example, for high-cost or high-impact operations like "opening an emergency exit," the required confidence threshold might be set to 0.9; while for "playing a regular evacuation broadcast," the threshold might be only 0.7. The engine compares the "confidence score" carried by the event with the preset threshold for the rule; only rules with scores reaching or exceeding the threshold are ultimately "matched" and activated. This mechanism is crucial; it constructs a risk-based dynamic filter, effectively preventing erroneous or ambiguous predictions from upstream AI models under low confidence (i.e., high uncertainty) conditions from triggering inappropriate or even harmful device linkages, thereby significantly improving the robustness and safety of the entire system's automated decision-making.

[0083] Step S503: Parse the execution action template in the matched rule entry and map the associated mesh ID to the target device physical address; The principle behind this step is to transform the matched, abstract response strategy into an operational blueprint that can be executed on a specific physical device through two steps of parsing and mapping.

[0084] First, the engine "parses the execution action template in the matching rules." This template is not the final control command, but a logical framework that defines "what to do" and "to whom to do it," such as "send guidance voice text to the broadcast system of the target area" or "control the display screen of the target area to switch to the guide map template." "Target area" and "guidance voice text" in the template are variables to be populated.

[0085] Next, the engine executes the "mapping of associated grid IDs to target device physical addresses," a precise conversion from spatial logic to physical topology. The system maintains a "grid-device mapping table," which dynamically records the network identifiers or addresses of all controllable physical devices associated with each spatial grid ID within the passenger station. For example, grid "B2" might be associated with a "display with IP address 192.168.10.201" and a "broadcast terminal with MAC address 00:1A:3F:FE:DC:BA." By querying this table, the engine replaces the abstract "associated grid ID" in the decision event with a series of specific, routable "target device physical addresses." Thus, the action template is concretized from "broadcast system in grid B2" to "device with MAC address 00:1A:3F:FE:DC:BA." This step ensures that the response action accurately targets the actual physical location where the event occurs, forming the foundation for precise spatial control.

[0086] Step S504: Generate a standardized control instruction sequence based on the execution action template and the target device physical address; After obtaining the parsed action template and the specific device physical address, the engine needs to generate the final control instructions. The generated instruction sequence is not targeted at a specific brand or model of device protocol, but describes an abstract operational intent.

[0087] For example, for a broadcasting system, the instruction is "play the specified text content", rather than "call a function in a specific DLL file to play a specific MP3 file"; for a display screen, the instruction is "display the content under a certain template", rather than "send a set of binary pixel data through a specific private protocol".

[0088] Furthermore, the engine contains driver adapters for various types of devices, responsible for translating standardized operational intentions (such as PlayText) and parameters (such as voice text and template ID) into specific protocol instructions that the target device can recognize, combined with the target device's physical address. For example, it might generate an HTTP request conforming to the brand's broadcast system API specifications or a register write instruction conforming to the access controller's Modbus protocol. These device-specific instructions, while adhering to internal standards, are organized into an ordered "sequence" to ensure that the actions of multiple devices are executed in the order or in parallel as required by business logic.

[0089] Step S505: Encapsulate the standardized control instruction sequence into a linkage task package that can be transmitted across systems.

[0090] Among them, the encapsulation process does not simply stack instructions together. Instead, it creates a data packet containing the complete execution context according to a widely accepted standard protocol in the industry (such as OPC UA, MQTT with specific payload formats) or a structured data format agreed internally (such as XML, JSONSchema). In addition to containing the "control instruction sequence" itself, this data packet usually also includes the globally unique identifier (event ID) of this linkage, task priority, timestamp, and possible receipt requirements. Such encapsulation makes the linkage task packet an independent and atomic work unit, which can be safely sent to the message queue and asynchronously consumed and executed by the clients of each subsystem.

[0091] In the above implementation, a highly automated, precise, reliable, and decoupled intelligent response center from the underlying devices is constructed. By efficiently and safely translating and issuing the predictive decisions of the upstream AI model into a series of collaborative control instructions that can be precisely executed in the physical world, the confidence threshold mechanism endows the automated system with risk awareness, the grid-device mapping mechanism realizes the spatial precision of management and control, and finally, reliable execution in a complex heterogeneous environment is ensured through standardized task packets.

[0092] Refer to Figure 6 , as an implementation of step S108, the steps of updating the passenger flow prediction model parameters and the dynamic threshold library data based on the execution status code in the device response message include: Step S601, receive the device response message returned by the target subsystem, which includes the execution status code, the associated decision event identifier, and the timestamp; Among them, the device response message is not a simple operation confirmation, but a structured feedback digital twin containing multiple layers of semantics. The "execution status code" is the authoritative encoding of the execution result of the control instruction received by the underlying hardware or subsystem. It precisely distinguishes different states such as "successfully executed", "partially executed", "communication failure", "device failure", etc., providing a fundamental criterion for the operational effectiveness of subsequent analysis. The "associated decision event identifier" is a globally unique ID, like a transaction number, which strongly associates the specific physical device response this time with the abstract, standardized decision event (and earlier prediction events) generated by the rule engine that triggered it before, establishing traceability throughout the entire "perception - prediction - decision - execution" link. The "timestamp" records the precise moment when the instruction is responded by the device, ensuring that the feedback event can be accurately inserted into the unified time series for analysis.

[0093] Step S602, retrieve the corresponding standardized decision event according to the decision event identifier, and extract the original prediction value, confidence score, and associated grid ID; Upon receiving a feedback message, the engine uses the "decision event identifier" as the primary key to quickly retrieve and locate the "standardized decision event" that was generated and issued by the rules engine. From this historical event record, three key pieces of information are extracted: the "original predicted value sequence" is the AI ​​model's quantitative estimate of future passenger flow trends, representing the target state that this coordinated operation attempted to address or prevent; the "confidence score" is the model's self-assessment of the certainty of its judgment when making the prediction, reflecting the prior reliability of the prediction; and the "associated grid ID" precisely indicates the spatial location of interest in the prediction and decision-making process.

[0094] Step S603: Obtain the actual passenger flow monitoring data of the associated grid ID area within a preset time period after execution; The system locates the specific physical area whose effects need to be evaluated based on the associated grid IDs extracted from historical decision events. It then focuses on the "preset time window after execution," which corresponds to the initially predicted future time period (the next 15 minutes). The purpose is to examine whether the actual passenger flow evolution in the area after the intervention measures are implemented is as predicted, or whether the trajectory has changed due to the intervention. Obtaining "actual passenger flow monitoring data" typically involves querying the edge computing nodes of the smart cameras in the area again to obtain a new stream of "structured metadata" starting from the moment the instruction was completed, and extracting a timestamped "passenger flow density numerical sequence" from it.

[0095] Step S604: Compare the deviation between the original predicted value and the actual passenger flow monitoring data, and combine the execution status code to generate model correction parameters; The process involves comparing the extracted "original predicted value sequence" with the acquired "actual passenger flow data sequence" at specific time points or along the overall trend to calculate a quantified "deviation" (such as root mean square error or mean absolute percentage error). Simultaneously, execution status codes are used to intelligently distinguish error sources. If the "execution status code" indicates successful execution (e.g., successful broadcast, gate opening), any deviation in passenger flow is primarily attributed to an inaccurate prediction model. In this case, the calculated deviation is directly used to "generate model correction parameters." These parameters (such as gradient descent direction and learning rate adjustment coefficient) specify how model weights should be adjusted to reduce such prediction errors. Conversely, if the "execution status code" indicates execution failure (e.g., equipment offline, instruction rejected), even if the actual passenger flow differs from the prediction, the primary cause is not a model error but a failure in the execution process. In this situation, the system may generate different signals, such as triggering an alarm to notify maintenance, or generating an indication parameter that requires the model to "ignore" this abnormal sample, preventing the model from learning erroneous associations caused by execution failure. This mechanism ensures that model optimization is only responsible for its own prediction errors and will not be misled by external execution failures, greatly improving learning efficiency and model robustness.

[0096] Step S605: Update the preset passenger flow thresholds in the corresponding spatiotemporal dimension of the dynamic threshold library according to the distribution characteristics of the deviation degree. The dynamic threshold library is not statically configured; its "baseline value" needs to be iterated as the system's understanding of passenger flow patterns deepens. The system analyzes the distribution characteristics of deviation. For example, over a period of time, for a specific grid at a specific time (such as area A at 8 am on a weekday), if the deviation remains consistently small and the confidence level is high, it indicates that the model's prediction for that scenario is very accurate, and the system can have more "confidence." In this case, it is advisable to "proportionally reduce the passenger flow threshold baseline value" to make the warning line more sensitive, thereby triggering diversion measures earlier to pursue a better operational state (proactive optimization). Conversely, if the actual passenger flow data frequently exceeds the existing threshold without warning (i.e., "continuous over-limit of actual passenger flow"), or if the prediction deviation is large and unstable, it may indicate that the passenger flow pattern in the area has changed (such as the opening of a new commercial outlet), or that the original threshold setting is too lenient.

[0097] At this point, the system needs to increase the baseline value of the passenger flow threshold based on an exponentially weighted average, making the early warning mechanism more closely reflect the new reality and avoiding frequent false alarms or missed alarms (passive adaptation). Through this two-way adjustment, the threshold database can act like an experienced administrator, continuously fine-tuning its judgment criteria based on historical performance and real-time feedback, enabling the entire control system to find a dynamic balance between accuracy and sensitivity.

[0098] Step S606: Input the model correction parameters into the passenger flow prediction model to perform incremental parameter updates.

[0099] Incremental parameter updates are an efficient paradigm for continuous model learning. Unlike retraining the model with the full dataset, the system inputs generated "model correction parameters" (such as weight gradients calculated through backpropagation) into the passenger flow prediction model. The model uses these parameters to adjust the connection weights of each layer of its internal neural network with a very small "learning step".

[0100] Specifically, based on feedback from a single or batch of decision-making practices, the model subtly modifies its understanding of the complex mapping relationship between "input features (fusion matrix) and output (future passenger flow)." This update is incremental, meaning it builds upon the model's existing knowledge, performs only local optimizations, has low computational cost, and can be performed frequently. This allows the model to quickly adapt to gradual changes in passenger flow patterns (such as seasonal changes or the impact of new transportation routes) or correct systematic biases in certain specific scenarios. Ultimately, a model that has undergone continuous incremental updates will see its predictive accuracy, robustness, and generalization ability continuously improve with the system's runtime, driving the entire intelligent management and control system to autonomously evolve towards greater precision and intelligence.

[0101] The above implementation constructs an intelligent closed loop that feeds back the execution results to the core decision-making process. By precisely linking execution feedback with the original decision, a refined attribution analysis of system performance (covering prediction accuracy and execution efficiency) is achieved. This technical solution not only optimizes the internal parameters of the passenger flow prediction model in a targeted manner, continuously improving its accuracy in predicting the future, but also dynamically adjusts the dynamic threshold benchmark used to trigger decisions, ensuring that the early warning mechanism maintains optimal sensitivity as the operating environment changes. This fundamentally solves the pain points of poor adaptability and long-term performance degradation in traditional automated systems, ensuring that the entire intelligent management and control solution maintains a high level of adaptability and reliability throughout its entire lifecycle.

[0102] Reference Figure 7 As a further implementation of the multi-source data fusion-based intelligent management and control method for passenger stations, the intelligent management and control method also includes: Step S701: Monitor the crowd gathering characteristic value in the abnormal behavior identifier in real time. When the characteristic value of the crowd gathering exceeds the dynamic threshold for a continuous preset time window, generate a group event warning signal. Specifically, the crowd aggregation characteristic value is a multi-dimensional quantitative indicator that integrates key parameters such as crowd density per unit area, variance of group movement speed, and frequency of limb contact. It is calculated in real time using computer vision algorithms (such as YOLO-v5 object detection for human body localization and counting, optical flow for motion trajectory analysis, and the OpenPose model for pose estimation to identify abnormal limb interactions). Its core lies in transforming unstructured video footage into quantifiable, structured data that reflects the dynamics of the crowd.

[0103] The dynamic threshold is not a fixed value, but rather generated through offline analysis and online learning of historical group event databases using machine learning algorithms. It can adapt to the normal passenger flow fluctuations in different time periods (such as off-peak and peak hours) and different areas (such as waiting halls and ticket halls), thus possessing contextual awareness capabilities. For example, an alert can be triggered only if the feature value exceeds the dynamic threshold for three consecutive time windows (e.g., each window is 30 seconds). This is an event confirmation strategy based on a sliding window model. Its logic is similar to a filter in digital signal processing, aiming to effectively filter out occasional noise and false alarms caused by brief obstructions, changes in lighting, or abnormal individual behavior, ensuring that the alert signal has high statistical significance and reliability.

[0104] Step S702: Associate the group event warning signal with the ticket check log data of the corresponding spatial grid ID, and filter the planned departure times within the area in the future preset time period; The spatial grid ID carried in the mass incident early warning signal (usually obtained by discretizing the WGS-84 coordinate system into a grid using a geocoding system such as Geohash) serves as the core spatial positioning index. The system uses this ID as a key to query the ticket check-in log database in real time. The ticket check-in log data records passenger passage events at each check-in gate, including gate number, timestamp, and associated train / flight information. The purpose of the association operation is to establish a mapping relationship between "risk physical locations" and "affected operational resources."

[0105] Specifically, the system first parses the list of main gates located in or serving the area based on the spatial grid ID. Then, using the timestamp of the warning signal as a benchmark, it scrolls forward 15 minutes to query all scheduled departures from these gates. This departure information typically comes from the production scheduling system and includes key attributes such as departure number, destination, vehicle type, rated passenger capacity, and scheduled departure time. The essence of this step is to inject operational context into the sudden event, linking the abstract cluster warning with the specific, impending capacity supply situation. This allows the system to move from a perception level of "what happened" to a cognitive level of "what will be affected," providing indispensable input data for the next step of precise capacity gap analysis.

[0106] Step S703: Calculate the capacity gap ratio based on the remaining ticket data of departure schedules and passenger flow forecasts, and generate a set of instructions for tiered traffic management strategies. After obtaining information on planned departure times for the relevant area in the near future, the system needs to dynamically assess whether existing capacity can meet potential passenger demand.

[0107] Specifically, the calculation of the capacity gap ratio is a quantitative means of achieving this assessment. Its input includes two key data: one is the future passenger flow forecast value output by the passenger flow forecasting model (this model is usually based on historical data and real-time characteristics), and the other is the real-time remaining ticket data from the ticketing system (i.e. the remaining capacity after subtracting the number of passengers who have checked their tickets from the total number of seats on the train).

[0108] Next, based on the calculated gap ratio, the system invokes a pre-built tiered traffic management strategy engine. This engine embeds strategy mapping rules based on business knowledge and historical experience, mapping continuous gap ratios to discrete, different levels of response. For example, a smaller gap ratio may only trigger an automated equipment response (such as opening a backup channel for diversion broadcasting), while a larger gap ratio requires a higher level of response, such as coordinating additional temporary shifts, mobilizing on-site manpower, or even linking external emergency resources. The final generated tiered traffic management strategy instruction set not only includes the specific sequence of actions to be executed (such as "broadcast system plays guidance voice X", "access controller opens channel Y"), but also defines metadata such as the execution priority and resource dependencies of the operations.

[0109] Step S704: Logically overlay the hierarchical guidance strategy instruction set with standardized decision events to generate an enhanced linkage task package; The standardized decision events include basic event attributes such as type, location, and confidence level, representing the system's fundamental judgment of the current situation. The tiered guidance strategy instruction set provides specific response plans for this situation. Logical overlay is not simply information splicing, but rather utilizes semantic web frameworks such as JSON-LD to associate and bind action elements in the strategy instruction set with the contextual attributes of the standardized events, forming a more complete and semantically richer "enhanced event object." For example, the "passenger flow exceeding limit" event can be organically combined with the instruction to "open backup channel 1 and play guidance announcements."

[0110] More importantly, this process includes a conflict resolution mechanism. When the system processes multiple events simultaneously (for example, both passenger flow overrun and fire alarm are triggered in the same area), simple instruction superposition may lead to contradictions (such as conflict between evacuation broadcast and fire evacuation broadcast). The conflict resolution algorithm (usually based on rules such as plan priority, risk level, and resource exclusivity) will be activated to arbitrate and optimize the instruction sequence, and may choose to execute a higher-priority instruction set, or generate a coordinated composite instruction sequence (such as giving priority to fire evacuation but adjusting the evacuation path to avoid conflicts).

[0111] The finally generated enhanced linkage task package is a structured executable plan that includes a coordinated sequence of device operations, a resource scheduling list, and a timeout rollback mechanism designed to ensure reliability.

[0112] Step S705: Push the key warning fields in the enhanced linkage task package to the mobile police terminal through the message middleware.

[0113] Among them, the message middleware (such as RabbitMQ, Kafka), as the information center of the system, adopts an asynchronous communication mechanism to decouple the generation and transmission of instructions, improving the overall throughput and response ability of the system. The enhanced linkage task package may contain a large amount of information, but when facing the mobile police terminal, in order to improve the transmission efficiency and terminal response speed, information extraction is required to push its core content, that is, the key warning fields.

[0114] In the embodiments of this application, these fields usually include the most essential information such as event level, precise location (grid ID), capacity gap ratio, and core disposal suggestions. The message middleware usually adopts a flexible routing mechanism such as Topic Exchange. The system will generate a routing key based on information such as warning level and event occurrence area to ensure that the message can be accurately delivered to the police terminal or terminal group responsible for that area and corresponding level, realizing the directional distribution of information. After receiving the message, the terminal needs to return a confirmation signal, and the system will monitor this feedback. If the confirmation is not received after the timeout, the message retransmission or escalation notification mechanism (such as notifying the upper-level command unit) will be triggered, thus building a closed-loop communication guarantee system. This ensures that key warnings and instructions can be reliably delivered to on-site disposal personnel, providing timely and accurate information support for manual intervention or collaborative disposal. The principle of this step is to use the enterprise-level integration mode to ensure low-latency and highly reliable delivery of key instructions in a distributed heterogeneous environment and establish an effective command feedback loop.

[0115] In the above implementation, the system can complete the closed loop from risk identification to initial response decision in a short time, reducing the response time from several minutes in the traditional mode to seconds; through accurate calculation of capacity gaps and matching of hierarchical strategies, it avoids insufficient or wasted emergency resource investment and achieves matching of response intensity with event scale; finally, through information fusion of enhanced task packages and reliable information push to mobile terminals, it ensures that the automated control of on-site equipment and manual intervention by police officers can be efficiently coordinated based on unified situational awareness, thereby comprehensively improving the safety assurance capabilities and operational efficiency of passenger stations in high-passenger-flow scenarios.

[0116] This application also discloses an AI-driven multi-source data fusion intelligent management and control system for passenger stations.

[0117] An AI-driven, multi-source data fusion-based intelligent management and control system for passenger stations specifically includes: The multi-source sensing module is used to acquire video stream data through smart cameras deployed in the passenger station, acquire environmental monitoring data through IoT sensors, and acquire ticket gate log data through business system interfaces. The data processing module is used to perform edge computing processing on video stream data, generate structured metadata containing passenger flow density values ​​and abnormal behavior identifiers, and attach a unified timestamp and spatial grid ID label to all acquired data. The data feature fusion module is used to input structured metadata with spatiotemporal labels, environmental monitoring data, and ticket check log data into the streaming computing engine, align data in the same area according to the spatial grid ID, aggregate and calculate according to a preset time window, and generate a fusion feature matrix. The predictive analysis module is used to call the pre-configured passenger flow prediction model to process the fusion feature matrix and output the passenger flow prediction value and confidence score for future periods. The decision module is used to compare the passenger flow forecast with the preset passenger flow thresholds in the corresponding spatiotemporal dimensions in the dynamic threshold library. When the passenger flow forecast exceeds the preset passenger flow threshold, a standardized decision event is generated. The linkage control module is used to input standardized decision events into a pre-configured rule engine for rule entry matching, and generate a linkage task package containing a sequence of device control instructions; The execution control module is used to distribute linkage task packages to the target subsystem through the message middleware, drive the broadcasting equipment, information display screen and access control controller to perform coordinated operations, and receive the device response messages in real time; The self-learning optimization module is used to update the parameters of the passenger flow prediction model and the dynamic threshold library data based on the execution status code in the device response message.

[0118] The AI-driven multi-source data fusion intelligent management and control system for passenger stations according to this application embodiment can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.

[0119] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0120] The embodiments described above in this application elaborate on the optimal technical solution centered on a centralized AI capability layer and an integrated linkage engine. However, to achieve the same objective—namely, to build a smart passenger station system capable of multi-source data fusion and integrated intelligent management and control—those skilled in the art can reasonably replace or modify certain technical details or implementation paths without departing from the core concept of this application.

[0121] Regarding alternative deployment locations for AI capabilities, the main solution primarily concentrates AI analysis capabilities, especially complex computer vision analysis, on the central platform's AI capability layer. A feasible alternative is to adopt a collaborative "cloud-edge-device" AI computing architecture. This involves further decentralizing lightweight AI models with extremely high real-time requirements or involving raw video stream processing, such as simple passenger flow counting and area intrusion detection, to edge computing units in the intelligent perception layer, such as more powerful AI cameras and edge computing boxes. Complex models requiring multi-source data fusion, such as cross-camera tracking and macro-level passenger flow prediction, remain executed on the central cloud platform. The edge and cloud exchange model parameters and processing results through a collaborative mechanism. This solution alleviates network bandwidth pressure, reduces the instantaneous peak demand on the central platform's computing resources, and further improves the real-time response to local events. Although the system architecture is slightly more complex, it still achieves intelligent data processing and fusion at multiple levels, ultimately serving the goal of integrated management and control. It can be considered an optimized option for scenarios with limited network conditions or central computing power.

[0122] Regarding the alternative solution for the data fusion and processing flow, in the main solution, the data resource layer clearly distinguishes between the real-time data lake and the fusion theme database, and adopts the ETL method combining batch and stream for fusion. The alternative solution can adopt the "streaming fusion and unified data middle platform" architecture. By using an advanced streaming computing engine such as Apache Flink, real-time association, cleaning, and fusion calculations are performed immediately after the data is accessed, and the processed fusion results are directly written into a unified real-time data warehouse or data middle platform that supports fast querying, weakening or combining the batch processing link, and emphasizing the real-time fusion and serviceization of data. This solution further reduces the latency from the original data to the analyzable and decision-making data, and is particularly suitable for dynamic scheduling scenarios with extremely high requirements for real-time performance. It may slightly adjust the support for data consistency and complex historical association analysis, but the core data fusion and supply capabilities are stronger, and it also supports the upper-layer intelligent analysis and linkage.

[0123] Regarding the alternative solution for the definition of the integrated linkage engine rules, the rule engine in the main solution mainly relies on the pre-static configuration of the "event-condition-action" rules. An enhanced or alternative solution is to introduce the mechanism of "dynamically generating strategies based on machine learning". The rule engine not only includes a static rule library but also accesses a policy learning module. This module learns through methods such as reinforcement learning from the continuously accumulated historical handling case events, execution actions, and final effect feedback, dynamically generates or recommends better linkage strategies, and can automatically optimize and adjust the thresholds and action combinations of the existing rule library. This solution enables the system to have certain self-learning and adaptive capabilities, be able to handle complex new scenarios that are not predefined in advance, or automatically optimize the handling strategies during long-term operation to pursue better effects. It can be used as a supplement or advanced evolution form to the static rule library, evolving the intelligent decision-making ability of the system from "rule-based" to "rule and learning combined".

[0124] Regarding alternative solutions for system integration and module communication, the main solution is based on microservice architecture and standardized APIs for inter-module communication. Alternative solutions can adopt different integration modes depending on the specific technology stack and deployment environment. Specifically, they include: (1) Message bus mode: Introducing an enterprise service bus or message queue as the core communication infrastructure, all modules perceive data reporting, event publishing, and instruction issuance through asynchronous communication via publish / subscribe topics, achieving higher decoupling and scalability; (2) Service mesh mode: In more complex distributed deployments, service mesh technologies such as Istio can be used to manage inter-service communication, load balancing, circuit breaking, etc., further improving the reliability and observability of the system under large-scale, multi-instance deployments. These alternative solutions differ in the technical implementation of communication mechanisms, but all serve the same purpose: to achieve efficient, reliable, and decoupled data exchange and instruction transmission between various functional modules within the system. They can all support the core processes of "multi-source data fusion" and "integrated management and control", and are suitable for projects of different scales or with specific technical preferences.

[0125] The above alternative solutions offer feasible technical variations from different perspectives, including computing power deployment, data processing, decision-making logic, and system integration. While they may differ from the main solution in implementation complexity, performance focus, resource consumption, or level of intelligence, they all share the core inventive concept of "aggregating multi-source data through a unified platform and achieving cross-system collaborative control using intelligent means," and therefore fall within the scope of protection of this application. Any equivalent substitution or reasonable evolution of specific technical means based on the core idea of ​​this application should be considered an extension of this application.

[0126] This application also discloses a computer device.

[0127] Computer equipment, including memory, processor, and computer program stored on memory and executable on processor, wherein the processor executes computer program to implement the above-described AI-driven multi-source data fusion intelligent management and control method for passenger stations.

[0128] This application also discloses a computer-readable storage medium.

[0129] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the AI-driven multi-source data fusion intelligent management and control methods for passenger stations.

[0130] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0131] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. An AI-driven multi-source data fusion-based intelligent management and control method for a passenger station, characterized in that, The intelligent management and control method comprises: Obtaining video stream data through an intelligent camera deployed in the passenger station, obtaining environmental monitoring data through an Internet of Things sensor, and obtaining ticket gate log data through a business system interface; Performing edge computing processing on the video stream data to generate structured metadata containing passenger flow density values and abnormal behavior identifiers, while adding a unified timestamp and a space grid ID label to all obtained data; Inputting the structured metadata with space-time labels, environmental monitoring data, and ticket log data into a stream computing engine, aligning data of the same region according to the space grid ID, and performing aggregation calculation according to a preset time window to generate a fusion feature matrix; Calling a pre-configured passenger flow prediction model to process the fusion feature matrix, and outputting passenger flow prediction values and confidence scores of future time periods; Comparing the passenger flow prediction values with preset passenger flow threshold values of corresponding space-time dimensions in a dynamic threshold library, and generating a standardized decision event when the passenger flow prediction values exceed the preset passenger flow threshold values; Inputting the standardized decision event into a pre-configured rule engine for rule item matching to generate a linkage task package containing a device control instruction sequence; Distributing the linkage task package to a target subsystem through a message middleware to drive broadcast devices, information display screens, and access controllers to perform cooperative operations, and receiving device response messages in real time; Updating passenger flow prediction model parameters and dynamic threshold library data based on execution status codes in the device response messages.

2. The AI-driven multi-source data fusion-based intelligent management and control method for a passenger station according to claim 1, characterized in that, The step of performing edge computing processing on the video stream data to generate structured metadata containing passenger flow density values and abnormal behavior identifiers comprises: Performing real-time decoding and frame buffer processing on the video stream data to generate a standardized video frame sequence; Performing human target detection on the standardized video frame sequence to output target detection data containing human body bounding box coordinates and confidence scores; Calculating the personnel distribution density in a preset grid area according to the human body bounding box coordinates to generate passenger flow density values; Extracting a human behavior feature vector based on the target detection data, matching an abnormal behavior pattern through a pre-set behavior classification model, and generating an abnormal behavior identifier when the confidence score exceeds a judgment threshold; Associating the passenger flow density values and the abnormal behavior identifier with corresponding timestamps and space grid IDs, and encapsulating them as structured metadata.

3. The AI-driven multi-source data fusion-based intelligent management and control method for a passenger station according to claim 1, characterized in that, The step of inputting the structured metadata with space-time labels, environmental monitoring data, and ticket log data into a stream computing engine, aligning data of the same region according to the space grid ID, and performing aggregation calculation according to a preset time window to generate a fusion feature matrix comprises: Receiving structured metadata with space-time labels, environmental monitoring data, and ticket log data, wherein the space-time labels include a unified timestamp and a space grid ID; Spatially aligning the structured metadata, environmental monitoring data, and ticket log data of the same physical region according to the space grid ID to generate a grouped data set grouped by grids; Performing multi-dimensional aggregation calculation on the grouped data set based on a preset time window period to generate real-time feature vectors containing regional passenger flow saturation, environmental index mean values, and differences between planned trips and ticketed passenger numbers; The feature vectors of all grids in each time window are combined into a matrix structure to obtain a fusion feature matrix of space-time dimensions.

4. The AI-driven multi-source data fusion-based intelligent management and control method for a passenger station according to claim 3, characterized in that, The step of calling a pre-configured passenger flow prediction model to process the fusion feature matrix and output passenger flow prediction values and confidence scores of a future period includes: Obtaining the fusion feature matrix includes the regional passenger flow saturation, the average of the environmental indicators, the difference between the planned number of trips and the number of checked passengers of a plurality of spatial grids in a continuous time window. Reorganizing the fusion feature matrix in space-time dimensions to generate a three-dimensional tensor data structure with spatial grids as rows and time windows as columns. Inputting the three-dimensional tensor data structure into a pre-trained passenger flow prediction model to output a passenger flow prediction value sequence of a future preset period. Generating a confidence score corresponding to the passenger flow prediction value sequence through an uncertainty quantification module embedded in the passenger flow prediction model. Packaging the passenger flow prediction value sequence and the confidence score into a standardized prediction output package.

5. The AI-driven multi-source data fusion-based intelligent management and control method for a passenger station according to claim 4, characterized in that, The step of inputting the standardized decision event into a pre-configured rule engine to match rule items and generate a linkage task package containing a device control instruction sequence includes: Receiving a standardized decision event containing an event type code, a confidence score and a related grid ID. Retrieving a pre-set rule library according to the event type code and matching rule items that meet the confidence threshold condition. Analyzing the execution action template in the matched rule item and mapping the related grid ID to the target device physical address. Generating a standardized control instruction sequence based on the execution action template and the target device physical address. Packaging the standardized control instruction sequence into a linkage task package that can be transmitted across systems. 6.The AI-driven multi-source data fusion-based intelligent management and control method for a passenger station according to claim 1, characterized in that, The step of updating the passenger flow prediction model parameters and the dynamic threshold library data based on the execution status code in the device response message includes: Receiving a device response message returned by a target subsystem, containing an execution status code, an associated decision event identifier and a timestamp. Retrieving the corresponding standardized decision event according to the decision event identifier and extracting the original prediction value, the confidence score and the related grid ID. Obtaining actual passenger flow monitoring data of the related grid ID area within a preset period after execution. Comparing the deviation degree of the original prediction value and the actual passenger flow monitoring data, and generating a model correction parameter combined with the execution status code. Updating the preset passenger flow threshold of the corresponding space-time dimension in the dynamic threshold library according to the distribution characteristics of the deviation degree. Inputting the model correction parameter into the passenger flow prediction model to perform parameter incremental update.

7. The AI-driven multi-source data fusion-based intelligent management and control method for a passenger station according to any one of claims 1 to 6, characterized in that, The intelligent management and control method further includes: Monitoring the personnel gathering feature value in the abnormal behavior identifier in real time, and generating a group event warning signal when the feature value of a continuous preset time window exceeds a dynamic threshold; Associating the group event warning signal with the ticket checking log data of the corresponding spatial grid ID, and screening the planned departure trips within a future preset period in the area; Calculating the transport capacity gap ratio based on the departure trip remaining ticket data and the passenger flow prediction value to generate a hierarchical diversion strategy instruction set; Logically superimposing the hierarchical diversion strategy instruction set and the standardized decision event to generate an enhanced linkage task package; Pushing the key warning fields in the enhanced linkage task package to a mobile police terminal through a message middleware.

8. An AI-driven multi-source data fusion-based intelligent management and control system for a passenger station, characterized in that, The intelligent management and control system includes: A multi-source perception module is configured to acquire video stream data through intelligent cameras deployed in the passenger station, acquire environmental monitoring data through Internet of Things sensors, and acquire ticket gate log data through a business system interface; A data processing module is configured to perform edge computing processing on the video stream data, generate structured metadata containing passenger flow density values and abnormal behavior identifiers, and attach a uniform timestamp and a spatial grid ID label to all acquired data; A data feature fusion module is configured to input the structured metadata with spatial and temporal labels, environmental monitoring data, and ticket log data into a stream computing engine, align data in the same region according to the spatial grid ID, aggregate and calculate according to a preset time window, and generate a fusion feature matrix; A prediction analysis module is configured to call a pre-configured passenger flow prediction model to process the fusion feature matrix, and output passenger flow prediction values and confidence scores for future time periods; A decision module is configured to compare the passenger flow prediction values with preset passenger flow threshold values in a dynamic threshold library corresponding to the spatial and temporal dimensions, and generate standardized decision events when the passenger flow prediction values exceed the preset passenger flow threshold values; A linkage control module is configured to input the standardized decision events into a pre-configured rule engine for rule item matching, and generate a linkage task package containing a device control instruction sequence; An execution control module is configured to distribute the linkage task package to target subsystems through a message middleware, drive broadcast devices, information display screens, and access controllers to perform cooperative operations, and receive device response messages in real time; A self-learning optimization module is configured to update passenger flow prediction model parameters and dynamic threshold library data based on execution status codes in the device response messages.

9. A computer device, characterized by: A computer program is stored in a memory and executable on a processor, and the processor implements the method according to any one of claims 1 to 7 when executing the program.

10. A computer-readable storage medium, characterized in that: A computer program is stored in a memory and executable on a processor, and the processor implements the method according to any one of claims 1 to 7 when executing the program.

Images