A metro traffic shielding door congestion risk prediction method and system

Abstract

The application discloses a kind of urban rail transit shield door congestion risk prediction method and system, it is related to urban rail transit intelligent operation and safety technical field.The method comprises: when urban rail train reaches platform area, by platform edge information processing device, the monitoring video stream of each door area is acquired in real time and the operating state data reported by the shield door corresponding to each door;By platform edge information processing device, according to the monitoring video stream and operating state data obtained, the multidimensional feature vector of each door area is extracted;By the platform edge information processing device, according to the multidimensional feature vector of each door area, the pre-trained gradient boosting tree model built-in is called in parallel, and the shield door congestion risk value of each door area at the current time is calculated respectively.The application realizes the risk prediction of multiple doors in parallel on the edge side of station by fusing visual data and physical state data, and provides active prevention and control support for urban rail transit safety operation.

Description

Technical Field

[0001] This application relates to the field of urban rail transit safety monitoring and intelligent control technology, and in particular to a method for predicting congestion risk of urban rail transit platform screen doors, as well as a computing device. Background Technology

[0002] With the increasing density of urban rail transit networks, platform passenger flow pressure is enormous during peak hours. Platform screen doors, as a crucial safety barrier between the platform and the train, play a vital role in ensuring the safety of their opening and closing process. However, during morning and evening rush hours, passengers frequently rush to board and alight, or leave behind large luggage or people with mobility impairments (such as wheelchairs or strollers) in the door area, which can easily lead to repeated anti-pinch issues, delayed closing times, and even accidents involving people or objects being trapped.

[0003] Existing platform screen door control technology mainly relies on anti-pinch sensors built into the doors (such as current detection and light curtain detection). This technology is a passive response mechanism, triggering a rebound only when the door has already contacted an obstacle or when resistance increases abnormally. This approach has significant lag: on the one hand, repeated opening and closing of the doors severely reduces train operating efficiency, causing delays across the entire line; on the other hand, in extremely crowded situations, passive anti-pinch mechanisms may not be effective in recognizing certain flexible obstacles or complex human postures, posing safety hazards.

[0004] Furthermore, while some existing technologies attempt to utilize video surveillance for passenger flow analysis, they often deploy video analysis servers in the station's central control room or in the cloud. Due to the massive volume of video data, limited network bandwidth, and latency, the time from data acquisition to analysis results is excessively long, failing to meet the millisecond-level real-time control requirements of platform screen doors. Simultaneously, existing passenger flow analyses primarily focus on macroscopic platform density statistics, lacking in-depth fusion analysis of microscopic characteristics of individual train door areas (such as fast-moving targets in specific locations or objects in special scenarios) and the physical operating status of platform screen doors (such as resistance curves and time-series parameters), resulting in low accuracy in risk prediction.

[0005] Therefore, there is an urgent need for a method and system that can predict the risk of congestion in single-door areas in real time and accurately, and can proactively intervene and control strategies. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] In view of the above-mentioned shortcomings and deficiencies of the existing technology, this application provides a method and system for predicting congestion risk of urban rail transit platform screen doors. It overcomes the defects of existing methods that can only respond passively and lack refined perception and active intervention capabilities. By deploying an information processing device at the edge of the platform, it realizes local real-time fusion analysis of video stream and operation status data, extracts multi-dimensional key features, uses a pre-trained gradient boosting tree model to accurately calculate the congestion risk value, and generates differentiated active control commands accordingly. This enables early prediction and accurate warning of congestion risk in the platform screen door area, providing active prevention and control support for the safe operation of urban rail transit.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the main technical solutions adopted in this application include:

[0010] In a first aspect, embodiments of this application provide a method for predicting congestion risks associated with platform screen doors in urban rail transit, the specific steps of which include:

[0011] S1. When the urban rail train arrives at the platform area, the platform edge information processing device acquires in real time the monitoring video stream of each door area, the operating status data reported by each door and the corresponding platform screen door; wherein, the monitoring video stream and the operating status data both have door identification or platform screen door identification and have the same timestamp.

[0012] S2. The platform edge information processing device extracts a multi-dimensional feature vector for each car door based on the acquired monitoring video stream and operating status data of each car door;

[0013] S3. The platform edge information processing device calculates the congestion risk value of the platform screen doors for each door area at the current moment by calling its built-in pre-trained gradient boosting tree model in parallel based on the multi-dimensional feature vector of each door area.

[0014] Optionally, in some embodiments of this application, the method further includes the following step before S1:

[0015] S01. Obtain the training dataset; the training dataset includes platform screen door operation status data and a sequence of visual monitoring images that are time-aligned with the platform screen door operation status data; the training dataset is generated through simulation.

[0016] S02. Based on the historical monitoring video stream and historical operating status data, extract multi-dimensional historical feature vectors; the multi-dimensional historical feature vectors include time features with peak or off-peak periods, location features with platform screen door attribute identifiers, density features of passengers per unit area inside and outside the carriage, number of targets inside and outside the carriage with a number of fast-moving targets, speed features inside and outside the carriage with the average speed of fast-moving targets, special scene features inside and outside the carriage with a number of special targets, and physical operating status features with door closing process parameters;

[0017] S03. Based on the historical platform screen door operation status data, filter abnormal closing events, use the multi-dimensional historical feature vectors corresponding to the historical abnormal closing events as positive samples and the multi-dimensional historical feature vectors corresponding to the normal closing events as negative samples, train the gradient boosting tree model, and obtain a pre-trained gradient boosting tree model.

[0018] Optionally, in some embodiments of this application, the step S03 of filtering abnormal closing events based on the historical platform screen door operating status data includes:

[0019] The operating parameters in the historical operating status data are compared with preset anomaly detection thresholds. If any of the following conditions are met, it is determined to be a historical abnormal door closing event:

[0020] The time difference between the door closing trigger signal and the actual door closing start time exceeds the first preset duration threshold; the door closing duration exceeds the second preset duration threshold and no door arrival detection signal is received; the resistance value during the door closing process exceeds the third preset resistance threshold and the duration exceeds the fourth preset duration threshold; no door arrival detection signal is received after a single door closing action is completed, and the resistance change curve does not show obstacle removal characteristics;

[0021] Within the preset statistical time window, the number of times a single door closing fails exceeds the fifth preset frequency threshold.

[0022] Optionally, in some embodiments of this application, training the gradient boosting tree model in step S03 includes:

[0023] Using the mean of the risk label values ​​of all samples in the training dataset as the initial prediction value, an initial gradient boosting tree model is constructed, the expression of which is:

[0024] ;

[0025] Where N is the number of training samples, y i This represents the true risk value for the i-th sample.

[0026] For the m-th iteration, the predicted residual is calculated based on the current gradient boosting tree model, and the predicted residual is used as the training target of the m-th regression decision tree to train the m-th regression decision tree to fit the residual distribution.

[0027] A learning rate is introduced to control the contribution of each regression decision tree to the gradient boosting tree model. The output of the m-th regression decision tree is weighted and then incorporated into the current gradient boosting tree model to obtain the updated gradient boosting tree model. After M iterations, the weighted sum of all regression decision trees is the trained gradient boosting tree model.

[0028] Optionally, in some embodiments of this application, step S2 involves extracting a multi-dimensional feature vector for each vehicle door based on the acquired monitoring video stream and operating status data for each door, including:

[0029] In the surveillance video stream, a core ROI region is defined with the platform screen door corresponding to each vehicle door as the center; within the core ROI region, the number of passengers per unit area is obtained as the passenger density feature, the number and average speed of fast-moving targets are obtained as the passenger movement feature, and the number of special targets is obtained as the special scene feature.

[0030] Based on the operational status data, obtain the door closing process parameters as physical operational status features;

[0031] Obtain the time characteristics and platform screen door position characteristics at the current moment.

[0032] Optionally, in some embodiments of this application, S1 includes:

[0033] Control the visual monitoring cameras and platform screen door controllers to connect to the station's unified NTP time server; perform initial time calibration during the daily start-up phase of urban rail trains, and continuously monitor the time deviation of each device during operation;

[0034] The platform edge information processing device receives monitoring video streams from cameras and operational status data from platform screen door controllers. It pairs image frames and status data for the same door area based on door identification or platform screen door identification, and reads the paired image acquisition timestamp T. img and status acquisition timestamp T state ; Calculate the corresponding image acquisition timestamp T for each pair. img and status acquisition timestamp T state absolute value of the time difference between ;

[0035] Will Compared with the preset timing tolerance threshold Perform a comparison;

[0036] like ≤ If the visual monitoring images and platform screen door operation status data are determined to be valid data with consistent time sequence, the subsequent feature extraction steps will proceed.

[0037] like > If the data set is deemed to be time-series abnormal, it will be discarded, or a secondary time synchronization command will be triggered to send a forced synchronization signal to devices whose time deviation exceeds the threshold, and the data will be reacquired after synchronization is completed.

[0038] Optionally, in some embodiments of this application, S3 is followed by:

[0039] The central integrated controller receives the congestion risk values ​​of the platform screen doors in each door area uploaded by the platform edge information processing device; the central integrated controller generates corresponding equipment control commands based on the comparison results of the risk values ​​and preset risk thresholds, and sends them to the execution terminal.

[0040] Optionally, in some embodiments of this application, the step of generating differentiated equipment control instructions based on the comparison result of the platform screen door congestion risk value and the preset risk threshold includes:

[0041] Preset low-risk, medium-risk, and high-risk thresholds;

[0042] When the congestion risk value of the platform screen door is lower than the low-risk threshold, a regular door closing command and a regular voice prompt command are generated.

[0043] When the congestion risk value of the platform screen door reaches the medium risk threshold, a command to slow down the closing speed, a command to activate the enhanced detection mode of the anti-pinch sensor, and a cyclical evacuation broadcast command are generated.

[0044] When the congestion risk value of the platform screen doors reaches the high-risk threshold, instructions to keep the platform screen doors open, high-frequency evacuation broadcast instructions, display screen guidance instructions, and personnel dispatch instructions are generated.

[0045] Optionally, in some embodiments of this application, iterative optimization of the gradient boosting tree model is further included after S3:

[0046] When the congestion risk value of the platform screen door reaches the high-risk threshold or an abnormality is detected in the actual closing action, the monitoring video stream, operation status data, and multi-dimensional feature vector corresponding to the event are cached locally on the platform edge information processing device and marked as samples to be optimized; during non-operational periods, the locally cached samples to be optimized are uploaded to the cloud training platform.

[0047] The cloud-based training platform uses the samples to be optimized to perform incremental training or full retraining on the gradient boosting tree model to generate a new version of the model.

[0048] The central integrated controller receives the new version model file from the cloud and distributes it to the edge information processing devices at each station to replace the old version model.

[0049] Secondly, embodiments of this application provide a congestion risk prediction system for urban rail transit platform screen doors, including:

[0050] Multiple platform edge information processing devices are deployed on the platform side near the platform doors. Each platform edge information processing device covers multiple door areas and has a built-in memory and processor. The memory contains a pre-trained gradient boosting tree model.

[0051] Multiple visual monitoring cameras are installed in each door area to collect monitoring video streams and transmit them to the platform edge information processing device.

[0052] Multiple platform screen door controllers are connected to the corresponding platform screen doors, and are used to collect platform screen door operation status data and transmit it to the platform edge information processing device, as well as receive equipment control commands issued by the platform edge information processing device or the central integrated controller.

[0053] The central integrated controller, deployed in the station control room, is communicatively connected to the multiple platform edge information processing devices and is used to receive risk prediction results and instruction execution status feedback uploaded by each platform edge information processing device.

[0054] Thirdly, embodiments of this application provide a computing device, including: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program stored in the memory, specifically performing the method described in the above embodiments.

[0055] (III) Beneficial Effects

[0056] This application discloses a method and system for predicting congestion risks at platform screen doors in urban rail transit. It achieves millisecond-level time synchronization through a dedicated time server and establishes a data time sequence verification mechanism to ensure accurate matching of multi-source data in the time dimension, laying a reliable foundation for subsequent feature fusion and risk prediction. It deeply integrates visual monitoring images with platform screen door operation status data, extracting multi-dimensional features such as density, movement direction, special vehicles, and physical operation status to achieve comprehensive and refined perception of risks in the platform screen door area. Real-time inference through platform edge information processing devices ensures low-latency control, and high-risk data is transmitted back to the cloud for model iteration and optimization, enabling the system to have self-learning and self-optimization capabilities to adapt to dynamically changing passenger flow environments. Differentiated control commands are output based on risk levels, linking multiple systems such as platform screen doors, broadcasting, displays, and personnel scheduling to achieve closed-loop safety management from early warning to intervention, significantly improving emergency response efficiency. Attached Figure Description

[0057] Figure 1 This application presents a flowchart of a method for predicting congestion risks using platform screen doors in urban rail transit.

[0058] Figure 2 This is an architecture diagram of a congestion risk prediction system for urban rail transit platform screen doors in one embodiment of this application. Detailed Implementation

[0059] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0060] In existing technologies, methods for automatic filtering of airborne laser point clouds based on deep learning can be mainly categorized into the following two types:

[0061] The first category is based on passive protection methods, such as optimizing the mechanical structure design of platform screen doors, adding anti-pinch sensors, and setting up platform safety lines. These methods can only reduce the severity of harm through physical means after an accident occurs, and cannot predict or proactively intervene before people, objects, or congestion occur. They exhibit significant lag and limitations when facing special scenarios such as large passenger flows during peak hours, passengers carrying large luggage, strollers, or wheelchairs, making it difficult to meet the refined management needs of urban rail transit safety operations.

[0062] The second category is simple early warning methods based on a single data source, such as detecting crowd density solely through video surveillance or detecting the closing status of platform screen doors solely through sensors. While these methods introduce some monitoring mechanisms, their data sources are singular, their feature dimensions are limited, and they cannot integrate visual and physical status information for comprehensive risk assessment. They also lack the ability to identify special scenarios such as strollers and wheelchairs, and fail to consider spatiotemporal characteristics such as peak hours and platform screen door locations. When faced with complex and ever-changing platform passenger flow environments, they generally suffer from high false alarm rates, high false negative rates, and poor adaptability.

[0063] To address this, this application provides a method and system for predicting congestion risks at platform screen doors in urban rail transit. By deeply fusing monitoring video streams and platform screen door operation status data, it extracts multi-dimensional feature vectors containing passenger density, motion characteristics, special scenarios, and physical operating states. A pre-trained gradient boosting tree model is then used to achieve accurate quantitative assessment of risk values. The risk values ​​for each door area are calculated in parallel by a platform edge information processing device. A central integrated controller generates differentiated control commands based on the risk level, and the model is continuously iterated and optimized using a cloud server. This effectively overcomes the problems of passive response, insufficient feature utilization, poor scenario adaptability, and inability to evolve autonomously inherent in existing technologies.

[0064] To better explain and facilitate understanding of this application, a detailed description of its embodiments is provided below in conjunction with the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a clearer and more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0065] Example 1

[0066] Figure 1 According to this application, a method for predicting congestion risk of platform screen doors in urban rail transit is proposed, such as... Figure 1 As shown, the method for predicting congestion risk of urban rail transit platform screen doors includes:

[0067] S1. When the urban rail train arrives at the platform area, the platform edge information processing device acquires in real time the monitoring video stream of each door area, the operating status data reported by each door and the corresponding platform screen door; wherein, the monitoring video stream and the operating status data both have door identification or platform screen door identification and have the same timestamp.

[0068] S2. The platform edge information processing device extracts a multi-dimensional feature vector for each car door based on the acquired monitoring video stream and operating status data of each car door;

[0069] S3. The platform edge information processing device calculates the congestion risk value of the platform screen doors for each door area at the current moment by calling its built-in pre-trained gradient boosting tree model in parallel based on the multi-dimensional feature vector of each door area.

[0070] This embodiment ensures precise temporal matching between visual and physical status data by assigning the same timestamp to the monitoring video stream and operational status data, laying a reliable data foundation for subsequent fusion analysis and avoiding risk misjudgment due to time sequence misalignment. By assigning door or platform screen door identifiers to the monitoring video stream and operational status data of each door area, it ensures a one-to-one correspondence between the data and specific physical locations, supporting subsequent independent risk analysis for each door area. Data is directly acquired by the edge information processing device deployed on the platform side, reducing data transmission latency and ensuring real-time response capabilities in high-risk scenarios.

[0071] By integrating multidimensional visual data and physical state data and extracting multidimensional feature vectors, the problem of perception bias caused by existing technologies that rely on only a single data source (such as only pedestrian density or only the status of platform screen doors) is overcome, and a comprehensive assessment of the risk in the platform screen door area is achieved.

[0072] The parallel call model of the platform edge information processing device adopts a parallel call mechanism to calculate the risk values ​​of multiple door areas simultaneously, which significantly improves processing efficiency and meets the real-time requirements of short train stop time and a large number of doors. Through a pre-trained gradient boosting tree model, complex multidimensional features are mapped to continuous risk values ​​from 0 to 1, realizing the digitization and quantification of risks, and providing a scientific and accurate basis for the subsequent formulation of refined hierarchical control strategies.

[0073] Example 2

[0074] Figure 1 This is a flowchart illustrating a method for predicting congestion risk using platform screen doors in urban rail transit, according to this application. Figure 1 As shown, the method for predicting congestion risk of urban rail transit platform screen doors mainly consists of a model pre-training stage and a real-time online prediction and control stage.

[0075] First, before deploying the system, a high-precision pre-trained gradient boosting tree model needs to be built. The specific steps include:

[0076] S01. Obtain the training dataset; the training dataset includes platform screen door operation status data and a sequence of visual monitoring images that are time-aligned with the platform screen door operation status data.

[0077] Historical operational data was collected, with a focus on capturing the sequence of monitoring images inside and outside the train carriages and the corresponding operational status data 15 seconds before the platform screen door malfunctions and closes, which were used as positive samples (high-risk samples); the image sequences and data corresponding to the platform screen door malfunctions and closes normally were captured as negative samples (low-risk samples).

[0078] Given the scarcity of real-world anomaly samples, this embodiment employs a hybrid approach of digital twin simulation and real data. The simulation platform simulates hundreds of operating conditions, including morning and evening rush hours, inclement weather, passengers rushing to their destinations, and wheelchair congestion, generating video streams and status data with time-series alignment labels. Simultaneously, real-world anomaly sample door closing data (first 15 seconds sequence) from historical operations is collected. This hybrid enhancement of simulated and real data resolves the sample imbalance problem caused by the scarcity of real-world accident samples, giving the model strong generalization capabilities and enabling it to adapt to changes in passenger flow characteristics at different stations and time periods.

[0079] S02. Based on each set of samples in the historical monitoring video stream and historical operating status data, extract a 10-dimensional historical multidimensional feature vector X=[X1,X2...X10]. The 10-dimensional historical feature vector is shown in Table 1, including:

[0080]

[0081] Table 1 Quantitative Characteristics of Influencing Factors

[0082] Time Feature X1: Based on the peak passenger flow patterns of urban rail transit, the current time is divided into peak or off-peak periods (e.g., weekdays 7:00-9:00 are marked as peak and coded as 1; otherwise, it is 0).

[0083] Platform screen door location feature X2: Based on the platform screen door number, query the preset location attribute database to obtain the attribute identifier of whether the platform screen door is close to stairs, elevators or transfer passages (e.g., close to facilities with high passenger flow is marked as 1, otherwise it is marked as 0).

[0084] Specifically, before extracting the X3-X10 multidimensional features, the process also includes a vision-based core ROI region delineation and object detection step:

[0085] ROI Delineation: Based on the field of view and calibration parameters of the cameras inside and outside the carriage, a mapping relationship between pixels and physical scale is established. Taking the center line of the platform screen door seam as the reference, rectangular areas with a horizontal width of 4 meters and a vertical depth of 3 meters are delineated on both the inner and outer sides of the carriage as the core ROI areas to ensure coverage of the main waiting and passage spaces.

[0086] Target detection and tracking: The YOLOv8 target detection model is used to scan the core ROI region to identify pedestrians and large objects (strollers, wheelchairs, suitcases, etc.); the Deep SORT multi-target tracking algorithm is combined to complete cross-frame target matching. By calculating the pixel displacement and time interval of the target detection boxes in adjacent frames, the motion velocity vector of each target is calculated.

[0087] Passenger crowding level in the carriage X3: Based on the historical monitoring video stream, the total number of passengers in the core ROI area inside the carriage is counted, and the number of passengers per unit area is calculated, i.e., passenger density.

[0088] Passenger congestion outside the carriage X4: Based on the historical monitoring video stream, count the total number of passengers in the core ROI area outside the carriage and calculate the number of passengers per unit area;

[0089] Number of fast-moving targets inside the carriage X5: Count the number of passenger targets in the core ROI area inside the carriage whose direction of movement is towards the platform screen door and whose speed exceeds a preset threshold.

[0090] Number of fast-moving targets outside the carriage X6: Count the number of passenger targets in the core ROI area outside the carriage that are moving towards the platform screen door and whose speed exceeds the preset threshold.

[0091] Average speed of fast-moving targets inside the carriage X7: Calculates the average speed of all fast-moving targets within the core ROI area inside the carriage;

[0092] Average speed of fast-moving targets outside the carriage X8: Calculates the average speed of all fast-moving targets within the core ROI area outside the carriage;

[0093] X9 Special Scene Features Inside the Carriage: Identify at least one special target among strollers, wheelchairs, and large luggage within the core ROI area inside the carriage, and count their number;

[0094] X10 Special Scene Features Outside the Carriage: Identify at least one special target among strollers, wheelchairs, and large luggage within the core ROI area outside the carriage, and count their number;

[0095] For special scene features X9 / X10, if a wheelchair / stroller is detected and stays in the door gap area for more than 5 seconds, the feature value is non-linearly weighted and amplified to highlight its high-risk attributes.

[0096] S03. The 10-dimensional historical feature vector is standardized using the Z-score standardization method to obtain a standardized 10-dimensional historical feature vector to eliminate the difference in dimensions.

[0097] Specifically, Z-score standardization is used to standardize all quantized features (X1-X10) to obtain standardized feature values. ;

[0098] ;

[0099] Where μ is the mean of feature X, and σ is the standard deviation of feature X.

[0100] Abnormal closing events are screened based on the historical platform screen door operation status data. The screening criteria include: closing delay > 3 seconds, closing timeout > 15 seconds and not in place, closing resistance > 50 N and lasting > 2 seconds, closing not in place and no sudden drop in resistance, and opening and closing times ≥ 3 times within 1 minute.

[0101] The gradient boosting tree model is trained by using the standardized 10-dimensional historical feature vectors corresponding to the 15 seconds before the occurrence of the historical abnormal door closing event as positive samples and the standardized 10-dimensional historical feature vectors corresponding to the normal door closing event as negative samples.

[0102] Specifically, based on the historical platform screen door operation status data, abnormal closing events are screened, using the standardized 10-dimensional historical feature vectors corresponding to the historical abnormal closing events as positive samples, including:

[0103] The operating parameters in the historical operating status data are compared with preset anomaly detection thresholds. If any of the following conditions are met, it is determined to be a historical abnormal door closing event:

[0104] The time difference between the door closing trigger signal issuance time and the actual door closing start time exceeds the first preset duration threshold.

[0105] The door remains closed for a duration exceeding the second preset time threshold, and no door arrival detection signal is received.

[0106] The resistance change curve during the closing process shows that the resistance value at any time exceeds the third preset resistance threshold, and the duration exceeds the fourth preset duration threshold.

[0107] No door arrival detection signal was received after a single door closing action was completed, and the resistance change curve did not show the characteristic of a sudden drop in resistance indicating the removal of obstacles;

[0108] Within the preset statistical time window, the number of times a single door closing fails exceeds the fifth preset frequency threshold.

[0109] In this embodiment, the specific threshold for determining abnormal door closing events can be set as follows:

[0110] First threshold: The time difference between the door closing trigger signal issuance time and the actual door closing start time is greater than 3 seconds;

[0111] Second threshold: Door closing duration > 15 seconds;

[0112] Third threshold: Resistance during the closing process > 50 Newtons;

[0113] Fourth threshold: The duration for which the resistance value exceeds the third threshold is greater than 2 seconds;

[0114] Fifth threshold: Within a preset statistical time window (e.g., 1 minute), the number of times a single door fails to close is ≥3.

[0115] The above thresholds can be adjusted according to actual operating conditions.

[0116] The time point determined as a historical abnormal door closing event is traced back to a preset time period, and the historical monitoring video stream sequence and corresponding historical platform screen door operation status data within the preset time period are extracted as positive sample data; the preset time period for tracing back is 15 seconds.

[0117] Training the gradient boosting tree model includes:

[0118] The mean of the risk label values ​​(1 for positive samples and 0 for negative samples) of all samples in the training dataset is used as the initial prediction value for the gradient boosting tree model. The initial gradient boosting tree model is constructed as follows:

[0119] ;

[0120] Where N is the number of training samples, y i Let L be the true risk value of the i-th sample, and L be the loss function, using mean squared error loss. , These are the initial prediction constants;

[0121] The initial gradient boosting tree model is obtained as follows: ;

[0122] That is, the prediction value of the initial gradient boosting tree model is the arithmetic mean of the risk label values ​​of all samples.

[0123] For the m-th iteration (m=1,2,...,M), based on the current gradient boosting tree model... Calculate the prediction residual for the predicted value on the i-th sample. As the training target for the m-th regression decision tree;

[0124] When mean squared error loss is used, the prediction residuals are simplified to ;

[0125] With the predicted residual Train the m-th regression decision tree with the new training labels. The residual distribution is fitted using the node splitting criterion;

[0126] Introducing learning rate By controlling the contribution of each regression decision tree to the gradient boosting tree model, the weighted output of the m-th regression decision tree is incorporated into the current gradient boosting tree model to obtain the updated gradient boosting tree model: ;

[0127] After M iterations (e.g., 300 iterations), the weighted sum of all regression decision trees constitutes the trained gradient boosting tree model. ;

[0128] in This is the output value for the congestion risk of the platform screen doors.

[0129] Recall and accuracy are used as model evaluation metrics. The optimal combination of hyperparameters (such as tree depth of 6, learning rate of 0.05, minimum number of sample splits of 40, minimum number of leaf node samples of 20, subsampling rate of 0.8, feature sampling rate of 0.9, regularization coefficient of 0.01, etc.) is determined by grid search to avoid overfitting.

[0130] Meanwhile, a feature importance assessment mechanism is introduced to ensure that the model not only outputs risk values, but also identifies the dominant factors that lead to the risk (such as congestion caused by excessive passenger density or the presence of special targets), providing a basis for subsequent differentiated control and obtaining the final pre-trained model.

[0131] Then, after obtaining the trained gradient boosting tree model, the real-time online prediction and control steps are performed using the model.

[0132] S1. When the urban rail train arrives at the platform area, the platform edge information processing device acquires in real time the monitoring video stream of each door area, the operating status data reported by each door and the corresponding platform screen door; wherein, the monitoring video stream and the operating status data both carry door identification or platform screen door identification and have the same timestamp.

[0133] S11. Control the visual monitoring camera and platform screen door controller to connect to the station's unified NTP time server; perform initial time calibration during the early morning departure phase of the urban rail train, and continuously monitor the time deviation of each device during operation to keep the local clock of each device synchronized with the NTP time server.

[0134] In the video stream captured by the visual monitoring camera, each frame of the image is embedded with an image acquisition timestamp; in the operating status data collected by the shielded door controller, each data packet is embedded with a status acquisition timestamp.

[0135] S12. The platform edge information processing device receives real-time video streams from cameras and operational status data from platform screen door controllers via Ethernet. Based on the door identification or platform screen door identification, it pairs image frames from the video streams of the same door area with status data from the operational status data, and reads the image acquisition timestamp T carried by the paired image frames. img and the status acquisition timestamp T carried by the status data state ;

[0136] S13. Calculate the corresponding image acquisition timestamp T for the pairing. img and status acquisition timestamp T state absolute value of the time difference between ;

[0137] Will Compared with the preset timing tolerance threshold A comparison is performed; wherein, the preset timing tolerance threshold is... Set to 50 milliseconds;

[0138] S14, if ≤ If the visual monitoring images and platform screen door operation status data are determined to be valid data with consistent time sequence, the subsequent feature extraction steps will proceed.

[0139] like > If the data set is deemed to be time-series abnormal, one of the following operations will be performed:

[0140] Discard this set of time-series abnormal data and wait for the next set of data to be collected;

[0141] Alternatively, a secondary time synchronization command can be triggered to send a forced synchronization signal to devices whose time deviation exceeds a threshold, and then reacquire data after synchronization is completed.

[0142] Optionally, if the network load is high, the threshold can be dynamically adjusted to 80ms and the data confidence level can be marked.

[0143] S2. The platform edge information processing device extracts a multi-dimensional feature vector for each door based on the acquired monitoring video stream and operating status data of each door.

[0144] S21. In the monitoring video stream, taking the center line of the door gap of the shielding door corresponding to each car door as the reference, a rectangular area with a horizontal width of 4 meters and a vertical depth of 3 meters is delineated to the inside and outside of the car respectively, as the core ROI area.

[0145] S22. Within the core ROI area, a target detection algorithm is used to count the total number of passenger targets and calculate the number of passengers per unit area as a passenger density feature; wherein, the passenger density feature includes passenger density features inside the carriage and passenger density features outside the carriage;

[0146] Specifically, in this embodiment, a target detection algorithm is first used within the core ROI area to count the total number of passenger targets. As an example, the target detection algorithm can use the YOLO v8 model, which can effectively identify pedestrians and large objects (such as strollers, wheelchairs, suitcases, etc.).

[0147] Based on the definition of traffic flow density, the number of passengers per unit area is used as a quantitative indicator of passenger density characteristics, and the calculation formula is as follows: ;

[0148] in, This refers to the number of passengers in the area inside / outside the platform screen doors of the train carriage. This represents the effective passage area for the corresponding region. The density values ​​of the core ROI regions inside and outside the carriage are calculated separately to obtain the passenger density characteristics X3 inside the carriage and X4 outside the carriage.

[0149] S23. Use a multi-target tracking algorithm (such as Deep SORT algorithm) to continuously track passenger targets within the core ROI area, and calculate the motion velocity vector and motion direction of each passenger target;

[0150] Specifically, the instantaneous velocity is calculated by the displacement of the target detection box between adjacent frames, and the direction of motion is determined by combining the target trajectory.

[0151] Passenger targets whose movement direction is towards the platform screen door and whose movement speed exceeds a preset speed threshold v0 are selected and defined as fast-moving targets; in this embodiment, the preset speed threshold v0 is 1.2 m / s.

[0152] The number of fast-moving targets within the core ROI regions inside and outside the carriage is counted separately, serving as feature X5 for the number of fast-moving targets inside the carriage and feature X6 for the number of fast-moving targets outside the carriage. The calculation formula is as follows: ;

[0153] Where v is the speed of the passenger target and v0 is the preset speed threshold.

[0154] Calculate the average velocity of all fast-moving targets within the core ROI regions inside and outside the carriage, respectively, and use this as the average velocity feature X7 / X8 of fast-moving targets inside / outside the carriage. The calculation formula is as follows:

[0155] ;

[0156] Where n is the number of fast-moving targets, v i,k Let be the speed of the k-th fast-moving target.

[0157] S24. Use a target detection model (e.g., YOLO v8 model) to scan the core ROI region and identify at least one special target among strollers, wheelchairs, or large luggage; and count the number of the special targets as special scene features.

[0158] The special scene features include special scene feature X9 inside the carriage and special scene feature X10 outside the carriage, and the calculation formula is as follows:

[0159] ;

[0160] S25. Based on the operating status data, extract at least one of the following as physical operating status features: door closing trigger signal issuance time, actual door closing start time, door closing duration, resistance change curve during the door closing process, door positioning detection signal, number of single door closing failures, and interval time between consecutive opening and closing.

[0161] S26. Extract the time features and platform screen door location features of the current moment; the time features indicate the peak or off-peak period category of the current moment, and the platform screen door location features indicate whether the target platform screen door is close to stairs, elevators or transfer passages.

[0162] S3. The platform edge information processing device calculates the congestion risk value of the platform screen doors for each door area at the current moment by calling its built-in pre-trained gradient boosting tree model in parallel based on the multi-dimensional feature vector of each door area.

[0163] The congestion risk value of the platform screen door is a value between 0 and 1. The larger the value, the higher the probability of people or objects getting caught or congestion occurring in the current platform screen door area.

[0164] In this embodiment, after the train arrives at the station, the platform edge information processing device collects real-time monitoring video streams and operating status data, extracts multi-dimensional feature vectors, inputs them into a pre-trained gradient boosting tree model, and calculates a risk value between 0 and 1.

[0165] The risk value is uploaded to the central integrated controller, which presets a low-risk threshold of 0.3 and a medium-risk threshold of 0.7.

[0166] Furthermore, based on the comparison result between the congestion risk value of the platform screen door and the preset risk threshold, differentiated equipment control instructions are generated, including:

[0167] Preset low-risk, medium-risk, and high-risk thresholds;

[0168] When the congestion risk value of the platform screen door is lower than the low-risk threshold (risk value < 0.3), a regular control command is generated: a command to close the platform screen door according to a preset standard timing is sent to the platform screen door controller; and a regular passenger boarding and alighting prompt voice command is sent to the platform broadcasting system.

[0169] When the congestion risk value of the platform screen door reaches the medium-risk threshold but is lower than the high-risk threshold (0.3 ≤ risk value < 0.7), an early warning control command is generated: a command to slow down the closing speed (e.g., reduce to 80%) is issued to the platform screen door controller, and the enhanced detection mode of the anti-pinch sensor is activated; a passenger flow guidance command is issued to the platform broadcasting system; an early warning information is pushed to the station personnel dispatch terminal to remind station staff to pay attention to the platform screen door area; if the model determines that the main risk source is a "fast-moving target", then the "Beware of pinching your hand" prompt is played; if it is a "special target", then "Please wait for the next train" is played or the station staff are notified.

[0170] When the congestion risk value of the platform screen doors reaches the high-risk threshold (risk value ≥ 0.7), an emergency control command is generated: a command to keep the doors open or a command to delay closing is sent to the platform screen door controller to forcibly interrupt the current closing action until the risk value drops below the medium-risk threshold; a high-frequency evacuation command is sent to the platform broadcast system; a warning message or a guidance sign display command is sent to the platform display screen; and a high-risk alarm containing a specific location number and handling suggestions is pushed to the station personnel dispatch terminal to prompt station staff to handle the situation.

[0171] After a high-risk intervention, the system enters a continuous monitoring window, which can automatically determine whether the risk has been eliminated (after three consecutive frames of decline) and intelligently restore the door to its closed state, or escalate the alarm to request manual intervention if the risk persists.

[0172] After executing the device control command, the system receives the command execution status feedback from the execution terminal in real time; it also records the command issuance timestamp, command type, execution result, and changes in the platform screen door congestion risk value before and after execution.

[0173] In this embodiment, when the congestion risk value of the platform screen door reaches the high-risk threshold (risk value ≥ 0.7) or when an abnormality is detected in the actual closing action, the platform edge information processing device caches the monitoring video stream, operating status data, extracted multidimensional feature vector and timestamp corresponding to the event in the local storage area of ​​the platform edge information processing device, marks them as samples to be optimized, and stores them in the local black box.

[0174] During non-operational periods (such as after nighttime shutdown) or when the network is idle, the central controller responds to the data synchronization command issued by the cloud server and uploads the locally cached samples to be optimized to the cloud training platform through the station backbone network; wherein, the data synchronization command includes sample screening conditions, and only uploads samples that have passed the timestamp consistency verification.

[0175] When the cloud training platform reaches the preset iteration cycle or receives an emergency update instruction, it uses the uploaded samples to be optimized to perform incremental training or full retraining on the gradient boosting tree model, generating a new version model file and version verification code.

[0176] After the new model is generated, it is first tested in pilot stations to verify that there is no increase in false alarm rate, and then the central controller pushes the update to the whole network.

[0177] During non-working hours, the central integrated controller receives new version model files from the cloud. After verification, it distributes them to the edge information processing devices of each station via the local area network to replace the old version model and complete the online iterative optimization of the model.

[0178] This embodiment innovatively constructs a 10-dimensional multi-dimensional feature vector system that includes spatiotemporal context, visual passenger flow status, and physical operating parameters. By introducing the YOLOv8 target detection model and the Deep SORT multi-target tracking algorithm, it can not only accurately identify pedestrians and special objects such as strollers and wheelchairs, but also accurately calculate the motion vectors (speed and direction) of the targets. By defining a core ROI area of ​​4m×3m and establishing a pixel-physical scale mapping, combined with a fast motion threshold of 1.2m / s, it effectively distinguishes between normal waiting and rushing behavior. It overcomes the shortcomings of single video surveillance being easily affected by light occlusion and single sensors having blind spots, significantly improving the accuracy of congestion risk identification in complex passenger flow scenarios and greatly reducing the false alarm rate and false negative rate.

[0179] A rigorous timing verification mechanism was designed, with the platform edge information processing device acquiring the monitoring video stream and operation status data of each door area in real time. All data are marked with door / platform door identification and have the same timestamp, achieving precise alignment of visual data and physical status data in time and space dimensions, laying a reliable foundation for subsequent fusion analysis. At the same time, direct acquisition from the edge side reduces transmission latency, with a time consumption of less than 10ms, meeting real-time requirements and ensuring real-time response capability.

[0180] By integrating surveillance video streams and operational status data, multi-dimensional feature vectors containing passenger movement characteristics, passenger density characteristics, special scene characteristics, and physical operational status characteristics are extracted. This overcomes the shortcomings of a single data source in providing limited perception and enables comprehensive perception of passenger flow status, special scenes, and equipment operation status in the platform screen door area. Z-score normalization is used to process 10-dimensional heterogeneous features, eliminating the dimensional differences between different features (such as number of people, speed, and time encoding), accelerating model convergence, and improving the training stability of the gradient boosting tree model.

[0181] The platform edge information processing device calls the pre-trained gradient boosting tree model in parallel and calculates the risk values ​​of multiple door areas at the same time, meeting the real-time requirements of short train stopping time and a large number of doors; it outputs a quantitative risk value between 0 and 1, providing an accurate basis for subsequent differentiated control.

[0182] Through canary release and simulation verification mechanisms, the system possesses lifelong learning capabilities, enabling it to continuously optimize itself as operational data accumulates, thus solving the pain point that traditional fixed-rule systems struggle to adapt to long-tail scenarios and dynamically changing environments.

[0183] Example 3

[0184] This invention provides a congestion risk prediction system for urban rail transit platform screen doors, such as... Figure 2As shown, it includes: a perception and execution layer (visual monitoring cameras, platform screen door controllers), an edge computing layer (platform edge information processing devices), a central control layer (central integrated controller), and a cloud evolution layer (cloud training platform).

[0185] Multiple visual monitoring cameras are installed in each door area to collect monitoring video streams and transmit them to the platform edge information processing device.

[0186] Preferably, two cameras are installed for each platform screen door. The cameras are suspended from the platform ceiling, 2.5-3.0 meters above the ground, with the lens optical axis pointing vertically downwards or tilted at a 30-degree angle to ensure that the field of view (FOV) completely covers the door opening and closing area and the waiting area 1.5 meters in front and behind. They have a built-in hardware encoding chip supporting H.265 encoding; they also have OSD (On-Screen Display) functionality, which can embed a microsecond-level image acquisition timestamp and a unique camera ID (bound to the door number) into each frame of the video stream; and they push the video stream to the edge computing layer in real time via a gigabit Ethernet port (RJ45) or fiber optic interface using the RTSP / GB28181 protocol.

[0187] Multiple platform screen door controllers are connected to the corresponding platform screen doors, and are used to collect platform screen door operation status data and transmit it to the platform edge information processing device, as well as receive equipment control commands issued by the platform edge information processing device or the central integrated controller.

[0188] Each controller is physically connected to the corresponding single or multiple platform screen door drive motor and sensor group. It has dual command receiving interfaces, capable of receiving local low-latency control commands (such as emergency stop and speed adjustment) from the platform edge information processing device, as well as global scheduling commands from the central integrated controller; it has built-in command priority arbitration logic, with local emergency commands having higher priority than global scheduling commands by default to ensure safety.

[0189] Multiple platform edge information processing devices, using platform edge AI boxes, are deployed on the platform side near the platform doors and close to the platform door group they cover. They are directly connected to cameras and controllers via short-distance network cables to reduce transmission hops. Each platform edge information processing device covers multiple door areas and has a built-in memory and processor. The memory contains a pre-trained gradient boosting tree model. A single platform edge information processing device can simultaneously cover a group of platform doors on the same platform (e.g., 24 platform doors on one side of the platform, corresponding to 24 door areas) through multiple network ports or access switches.

[0190] The central integrated controller, deployed in the station control room, communicates with the multiple platform edge information processing devices to receive risk prediction results and instruction execution status feedback uploaded by each device. When multiple concurrent high-risk events or network-wide anomalies are detected, it generates global linkage instructions (such as station-wide broadcasts, PIS screen guidance, and gate flow restriction linkage) and sends them to relevant execution terminals. Acting as a bridge between the cloud and the edge, it receives new version model files from the cloud, performs digital signature verification, and then distributes them in batches to each platform edge information processing device via the station's local area network, supporting canary releases and version rollbacks. It records all risk events, control instructions, and execution feedback, forming an immutable operational safety log.

[0191] Specifically, a heat map of congestion risk for all train doors on the entire platform is displayed in real time through a graphical user interface (HMI), with the status of each door dynamically indicated by red, yellow, and green colors.

[0192] The cloud-based training platform is deployed in the private or hybrid cloud data center of the rail transit group; it receives "samples to be optimized" (i.e., raw data of high-risk or abnormal events) uploaded by each station at night; it uses distributed computing resources to perform incremental training or full retraining of the basic GBDT model based on the new samples, automatically adjusts hyperparameters, and solves the model drift problem; before the model is deployed, it uses historical data playback and new samples to conduct stress tests in a virtual environment to evaluate the accuracy, recall, and false positive rate of the new model, ensuring safe deployment.

[0193] Furthermore, the workflow and interaction logic of the urban rail transit platform screen door congestion risk prediction system include:

[0194] First, initial synchronization is performed: After the system is powered on, the platform edge information processing device calibrates the clocks of the cameras and platform screen door controllers through the IEEE 1588 protocol to establish a unified spatiotemporal reference.

[0195] The camera and controller continuously push a timestamped data stream to the information processing device at the edge of the platform.

[0196] The platform edge information processing device executes a method for predicting the congestion risk of urban rail transit platform screen doors as described in Example 2, and obtains the congestion risk value of the platform screen doors in each door area at the current moment, with the time taken controlled within 50ms.

[0197] If the congestion risk value of the platform screen doors is less than 0.3, the platform edge information processing device will transmit a standard door closing command to maintain normal operation.

[0198] If the risk value of platform screen door congestion is less than 0.7, the platform edge information processing device will issue a "decelerate" command and report to the central controller to trigger a voice prompt.

[0199] If the congestion risk value of the platform screen door is ≥0.7, the platform edge information processing device will first send a "keep open" command directly to the PEDC to forcibly interrupt the current closing action and prevent people or objects from being trapped; at the same time, the alarm will be reported to the central controller.

[0200] Furthermore, the system also includes procedures for handling fault scenarios:

[0201] Network interruption scenario: If the network connection between the platform edge information processing device and the central controller is interrupted, the platform edge information processing device will automatically switch to independent autonomous mode and continue to execute local risk prediction and control logic, only suspending data upload, to ensure that a single point of failure does not affect the safety function.

[0202] Equipment failure scenario: If a camera fails, the platform edge information processing device automatically identifies the missing data for that channel, performs dimensionality reduction inference based on the remaining physical state data and the correlation features of other adjacent doors, and outputs a risk warning of "reduced confidence" instead of directly reporting an error and shutting down.

[0203] This fault scenario handling process effectively avoids train delays caused by doors failing to close for extended periods due to momentary interference. It also prevents safety accidents caused by blindly retrying door closing under continuous congestion, significantly improving the system's robustness and operational efficiency.

[0204] Furthermore, every day at dawn, the platform edge information processing device encrypts and uploads high-value samples marked in the local SSD to the cloud.

[0205] After training is completed in the cloud and a v2.0 model is generated, it is sent to the central controller.

[0206] The central controller pushes update packages to the platform edge information processing device during non-operational hours (such as 04:00). After the platform edge information processing device verifies and passes the verification, the model is hot-replaced and takes effect the next day.

[0207] In this embodiment, the platform edge information processing device is deployed on the platform side near the platform door. It completes data acquisition, feature extraction and model inference locally without uploading the original video stream to the central server, which greatly reduces network bandwidth pressure, avoids risk omissions due to network delays or interruptions, and ensures real-time response capabilities in high-risk scenarios.

[0208] The central integrated controller is deployed in the station control room and communicates with all platform edge information processing devices. It receives risk prediction results from each door area, generates differentiated control instructions based on preset risk thresholds, and sends them to each execution terminal to achieve centralized scheduling and coordinated linkage of risk management of the entire station's platform screen doors.

[0209] The cloud server receives high-risk event data, uses the new samples to incrementally train or retrain the gradient boosting tree model, generates a new version of the model, and distributes it to the edge information processing devices of each station through the central controller, so that the system has the ability to learn and optimize itself, adapt to changes in passenger flow at different stations and at different times, and achieve continuous improvement in model performance.

[0210] The platform edge information processing device innovatively achieves spatiotemporal alignment and deep fusion of "visual micro-features" and "physical operating status." It overcomes the shortcomings of single video surveillance, which is susceptible to interference from lighting and occlusion, and the blind spots of single sensors. By using a gradient boosting tree model to jointly analyze multi-dimensional heterogeneous data such as passenger density, movement trajectory, special objects (wheelchairs / luggage), door resistance, and time-series parameters, it significantly improves the recognition accuracy of complex congestion scenarios (such as rushing on and off, and the stagnation of flexible objects) and greatly reduces the false alarm rate and missed alarm rate. By binding the channel number with the platform screen door ID, it can simultaneously access multiple monitoring video streams and multiple operating status data, with one device covering a group of platform screen doors on the same platform. Each module has clear functions and well-defined interfaces, supporting standardized deployment for new lines and facilitating the transformation and upgrading of existing lines.

[0211] This application provides a congestion risk prediction system for platform screen doors in urban rail transit, which can be widely applied to the safety monitoring of platform screen doors in urban rail transit stations. By deploying information processing devices at the platform edge, combined with a central integrated controller and cloud server, real-time prediction and proactive control of congestion risks in the platform screen door area can be achieved, significantly improving the safety level of urban rail operation. The system has a reasonable architecture, clear processing flow, and good real-time performance and reliability. It can effectively reduce the probability of safety accidents such as people or objects being trapped by platform screen doors, and has extremely high industrial application value.

[0212] Example 4

[0213] Finally, this application also proposes a computing device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the instructions stored in the memory so that the computer device performs the urban rail transit platform screen door congestion risk prediction method described in the above embodiments.

[0214] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0215] Furthermore, it should be noted that in the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0216] In this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first and second features are in direct contact, or that they are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0217] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0218] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting congestion risk of platform screen doors in urban rail transit, characterized in that, include: S1. When the urban rail train arrives at the platform area, the platform edge information processing device acquires in real time the monitoring video stream of each door area, the operating status data reported by each door and the corresponding platform screen door; wherein, the monitoring video stream and the operating status data both have door identification or platform screen door identification and have the same timestamp. S2. The platform edge information processing device extracts a multi-dimensional feature vector for each door based on the acquired monitoring video stream and operating status data of each door. S3. The platform edge information processing device calculates the congestion risk value of the platform screen doors for each door area at the current moment by calling its built-in pre-trained gradient boosting tree model in parallel based on the multi-dimensional feature vector of each door area.

2. The method according to claim 1, characterized in that, Before S1, it also includes: S01. Obtain the training dataset; the training dataset includes platform screen door operation status data and a sequence of visual monitoring images that are time-aligned with the platform screen door operation status data; the training dataset is generated through simulation. S02. Based on the historical monitoring video stream and historical operating status data, extract multi-dimensional historical feature vectors; the multi-dimensional historical feature vectors include time features with peak or off-peak periods, location features with platform screen door attribute identifiers, density features of passengers per unit area inside and outside the carriage, number of targets inside and outside the carriage with a number of fast-moving targets, speed features inside and outside the carriage with the average speed of fast-moving targets, special scene features inside and outside the carriage with a number of special targets, and physical operating status features with door closing process parameters; S03. Based on the historical platform screen door operation status data, filter abnormal closing events, use the multi-dimensional historical feature vectors corresponding to the historical abnormal closing events as positive samples and the multi-dimensional historical feature vectors corresponding to the normal closing events as negative samples, train the gradient boosting tree model, and obtain a pre-trained gradient boosting tree model.

3. The method according to claim 2, characterized in that, The step S03, which filters abnormal closing events based on the historical platform screen door operation status data, includes: The operating parameters in the historical operating status data are compared with preset anomaly detection thresholds. If any of the following conditions are met, it is determined to be a historical abnormal door closing event: The time difference between the door closing trigger signal and the actual door closing start time exceeds the first preset duration threshold; the door closing duration exceeds the second preset duration threshold and no door arrival detection signal is received; the resistance value during the door closing process exceeds the third preset resistance threshold and the duration exceeds the fourth preset duration threshold; no door arrival detection signal is received after a single door closing action is completed, and the resistance change curve does not show obstacle removal characteristics; Within the preset statistical time window, the number of times a single door closing fails exceeds the fifth preset frequency threshold.

4. The method according to claim 2, characterized in that, The training of the gradient boosting tree model in S03 includes: Using the mean of the risk label values ​​of all samples in the training dataset as the initial prediction value, an initial gradient boosting tree model is constructed, the expression of which is: ； Where N is the number of training samples, y i This represents the true risk value for the i-th sample. For the m-th iteration, the predicted residual is calculated based on the current gradient boosting tree model, and the predicted residual is used as the training target of the m-th regression decision tree to train the m-th regression decision tree to fit the residual distribution. A learning rate is introduced to control the contribution of each regression decision tree to the gradient boosting tree model. The output of the m-th regression decision tree is weighted and then incorporated into the current gradient boosting tree model to obtain the updated gradient boosting tree model. After M iterations, the weighted sum of all regression decision trees is the trained gradient boosting tree model.

5. The method according to claim 1, characterized in that, In step S2, based on the acquired monitoring video stream and operating status data for each vehicle door, a multi-dimensional feature vector is extracted for each door, including: In the surveillance video stream, a core ROI region is defined with the platform screen door corresponding to each vehicle door as the center; within the core ROI region, the number of passengers per unit area is obtained as the passenger density feature, the number and average speed of fast-moving targets are obtained as the passenger movement feature, and the number of special targets is obtained as the special scene feature. Based on the operational status data, obtain the door closing process parameters as physical operational status features; Obtain the time characteristics and platform screen door position characteristics at the current moment.

6. The method according to claim 1, characterized in that, S1 includes: Control the visual monitoring cameras and platform screen door controllers to connect to the station's unified NTP time server; perform initial time calibration during the daily start-up phase of urban rail trains, and continuously monitor the time deviation of each device during operation; The platform edge information processing device receives monitoring video streams from cameras and operational status data from platform screen door controllers. It pairs image frames and status data for the same door area based on door identification or platform screen door identification, and reads the paired image acquisition timestamp T. img and status acquisition timestamp T state ; Calculate the corresponding image acquisition timestamp T for each pair. img and status acquisition timestamp T state absolute value of the time difference between ; Will Compared with the preset timing tolerance threshold Perform a comparison; like ≤ If the visual monitoring images and platform screen door operation status data are determined to be valid data with consistent time sequence, the subsequent feature extraction steps will proceed. like > If the data set is deemed to be time-series abnormal, it will be discarded, or a secondary time synchronization command will be triggered to send a forced synchronization signal to devices whose time deviation exceeds the threshold, and the data will be reacquired after synchronization is completed.

7. The method according to claim 1, characterized in that, S3 is followed by: The central integrated controller receives the congestion risk values ​​of the platform screen doors in each door area uploaded by the platform edge information processing device; the central integrated controller generates corresponding equipment control commands based on the comparison results of the risk values ​​and preset risk thresholds, and sends them to the execution terminal.

8. The method according to claim 7, characterized in that, The step of generating differentiated equipment control commands based on the comparison between the platform screen door congestion risk value and a preset risk threshold includes: Preset low-risk, medium-risk, and high-risk thresholds; When the congestion risk value of the platform screen door is lower than the low-risk threshold, a regular door closing command and a regular voice prompt command are generated. When the congestion risk value of the platform screen door reaches the medium risk threshold, a command to slow down the closing speed, a command to activate the enhanced detection mode of the anti-pinch sensor, and a cyclical evacuation broadcast command are generated. When the congestion risk value of the platform screen doors reaches the high-risk threshold, instructions to keep the platform screen doors open, high-frequency evacuation broadcast instructions, display screen guidance instructions, and personnel dispatch instructions are generated.

9. The method according to claim 1, characterized in that, Following S3, iterative optimization of the gradient boosting tree model is also included: When the congestion risk value of the platform screen door reaches the high-risk threshold or an abnormality is detected in the actual closing action, the monitoring video stream, operating status data and multi-dimensional feature vector corresponding to the event are cached locally in the platform edge information processing device and marked as samples to be optimized. During non-operational periods, locally cached samples to be optimized are uploaded to the cloud training platform; The cloud-based training platform uses the samples to be optimized to perform incremental training or full retraining on the gradient boosting tree model to generate a new version of the model. The central integrated controller receives the new version model file from the cloud and distributes it to the edge information processing devices at each station to replace the old version model.

10. A congestion risk prediction system for urban rail transit platform screen doors, characterized in that, include: Multiple platform edge information processing devices are deployed on the platform side near the platform doors. Each platform edge information processing device covers multiple door areas and has a built-in memory and processor. The memory contains a pre-trained gradient boosting tree model. Multiple visual monitoring cameras are installed in each door area to collect monitoring video streams and transmit them to the platform edge information processing device. Multiple platform screen door controllers are connected to the corresponding platform screen doors, and are used to collect platform screen door operation status data and transmit it to the platform edge information processing device, as well as receive equipment control commands issued by the platform edge information processing device or the central integrated controller. The central integrated controller, deployed in the station control room, is communicatively connected to the multiple platform edge information processing devices and is used to receive risk prediction results and instruction execution status feedback uploaded by each platform edge information processing device.

Images