Intelligent social risk early warning method and system based on deep learning hybrid model

Abstract

The present application relates to the technical field of risk early warning, in particular to a smart social risk early warning method and system based on a deep learning hybrid model. The present application obtains the crowd flow state of multiple monitoring areas in a large public place, determines the abnormal crowd flow increment; predicts the time period of abnormal crowd flow reaching the downstream bottleneck and the number of each time slice according to the place passing topology and the deep learning hybrid model; calculates the backlog state and recovery result of independent action and cumulative action on the common prediction time axis in combination with the normal crowd flow baseline and the bottleneck expected number of people that can pass; when each independent action is recoverable and the cumulative action is not recoverable, determines the unstable bottleneck and outputs the personnel gathering risk early warning information. The present application can identify the persistent gathering risk formed by the spatiotemporal evolution of multi-source crowd flow.

Description

Technical Field

[0001] This invention relates to the field of risk warning technology, specifically to a smart society risk warning method and system based on a deep learning hybrid model. Background Technology

[0002] As large-scale public venues such as convention centers, stadiums, airport terminals, transportation hubs, and commercial complexes continue to expand, these venues typically contain multiple functional areas, connecting passageways and escalators, turnstiles, security checkpoints, and narrow exits—locations with limited access. During event dispersal, concentrated arrivals of transportation, changes in equipment status, or temporary control measures, the flow of people in different areas can change significantly over time and spread spatially to other areas. Existing crowd gathering risk early warning technologies can utilize data such as video surveillance, traffic counting, and wireless positioning to obtain data on the number of people, density, speed, dwell time, and inflow / outflow volume, and predict future crowd flow patterns or congestion propagation trends through time series models and graph neural networks.

[0003] However, existing methods still largely rely on the current number of people, density, risk indicators, or future overall pedestrian flow in a single area as the primary evaluation criteria, determining the warning result based on preset or dynamic thresholds. Although some technologies have further considered the traffic relationships between areas, shared bottleneck facilities, and future traffic changes, their evaluation focus remains primarily on whether the area's condition, channel load, or bottlenecks have reached risk conditions. During the normal operation of large public venues, multiple spatially separated areas may experience pedestrian flow changes that have not yet reached independent warning conditions. Each area may appear normal or low-risk when evaluated individually, but these local pedestrian flow states propagate downstream along the traffic path. When the arrival volume overlaps in the same or adjacent time slots of the same downstream limited traffic area, it may still gradually lead to a persistent accumulation of people in that downstream limited traffic area that is difficult to dissipate on its own. Existing methods of evaluating independently by area or predicting overall pedestrian flow cannot identify the persistent clustering risk formed by this subsequent spatiotemporal evolution in advance when the local condition has not yet reached the warning conditions, thus limiting the advance warning capability of the population gathering risk.

[0004] In summary, how to identify the risk of sustained crowd gathering caused by the subsequent spatiotemporal evolution of large public places when multiple local crowd flow conditions have not reached the independent warning conditions is a technical problem that urgently needs to be solved in this field.

[0005] To address this, a smart society risk early warning method and system based on a deep learning hybrid model is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a smart society risk early warning method and system based on a deep learning hybrid model. It acquires the pedestrian flow status of multiple monitoring areas in a large public place to determine abnormal pedestrian flow increments; predicts the time period and number of people in each time slot when abnormal pedestrian flow reaches the downstream bottleneck based on the venue's traffic topology and a deep learning hybrid model; combines the normal pedestrian flow baseline and the expected number of people passing through the bottleneck to calculate the backlog status and recovery results of independent and cumulative effects on a common prediction time axis; when each independent effect is recoverable but the cumulative effect is unrecoverable, it identifies the unstable bottleneck and outputs a pedestrian gathering risk early warning information. This invention can identify the persistent gathering risk formed by the spatiotemporal evolution of multi-source pedestrian flow.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A smart society risk early warning method based on a deep learning hybrid model includes: Acquire the pedestrian flow status of multiple monitoring areas in a large public place, and determine the abnormal pedestrian flow increments corresponding to each monitoring area; Based on the location topology, the propagation path of each abnormal increase in pedestrian flow to the downstream bottleneck is determined, and the corresponding expected arrival time and the expected number of people arriving in each time slice are obtained through a deep learning hybrid model. Based on the normal passenger flow baseline, determine the baseline expected arrival number for each time slot of the downstream bottleneck. Based on the current queuing status and actual passage status of the downstream bottleneck, determine the expected passage number for each time slot. Map the baseline expected arrival number and the expected arrival number of each abnormal passenger flow increment to the common prediction time axis. Use the remaining backlog of the previous time slot as the initial backlog of the current time slot to determine the backlog status when each abnormal passenger flow increment acts independently and when multiple abnormal passenger flow increments act cumulatively. Based on the backlog status, determine whether each independent action and cumulative action meets the recovery conditions within their respective recovery intervals; if each independent action meets the conditions but the cumulative action does not, the corresponding downstream bottleneck is identified as an unstable bottleneck and a personnel gathering risk warning is output.

[0008] Preferably, determining the abnormal pedestrian flow increment for each monitoring area includes: determining the corresponding normal pedestrian flow baseline from historical pedestrian flow status according to activity stage, date type, and regional operation status; counting the number of people moving along the downstream direction in the current pedestrian flow status according to a unified time slice, determining the new number of people exceeding the normal pedestrian flow baseline as the abnormal pedestrian flow increment for the corresponding monitoring area, and forming an abnormal pedestrian flow increment sequence according to time sequence.

[0009] Preferably, determining the propagation path includes: constructing a site access topology based on the actual connectivity between the monitoring area, connecting channels, and converging nodes; identifying restricted access nodes based on the effective access width of the nodes, the number of open channels, and the historically stable number of people passing through; identifying restricted access nodes with upstream reachable paths from two or more different monitoring areas as candidate downstream bottlenecks; and determining the corresponding propagation path based on the connectivity between the monitoring area corresponding to each abnormal increase in pedestrian flow and the candidate downstream bottleneck.

[0010] Preferably, the deep learning hybrid model includes a spatial propagation sub-model and a temporal evolution sub-model. The spatial propagation sub-model determines the expected number of people diverted into each subsequent channel of the corresponding propagation path based on the location topology, abnormal pedestrian flow increment, and propagation path. The temporal evolution sub-model determines the expected number of people arriving at each time slot of the propagation to the downstream bottleneck based on the expected number of people diverted and the pedestrian flow speed sequence, density sequence, and travel time of the corresponding channel. The time slots in which the expected number of people arrives continuously are determined as the corresponding expected arrival time slots.

[0011] Preferably, determining the expected number of people passing through each future time slot includes: obtaining the number of people queuing at the downstream bottleneck at the evaluation time and the actual number of people passing through the exit in multiple consecutive time slots before the evaluation time; determining the actual number of people passing through the downstream bottleneck per unit time slot based on the actual number of people passing through the exit and the bottleneck opening status; and determining the expected number of people passing through each future time slot based on this. When the bottleneck opening status changes, the expected number of people passing through subsequent time slots is updated based on the changed opening status.

[0012] Preferably, determining the backlog status when each abnormal flow increment acts independently and when multiple abnormal flow increments act cumulatively includes: when acting independently, mapping the baseline expected arrival number and the expected arrival number for each time slot corresponding to an abnormal flow increment to a common prediction time axis; when acting cumulatively, mapping the baseline expected arrival number and the expected arrival number for each time slot corresponding to multiple abnormal flow increments leading to the same downstream bottleneck to a common prediction time axis, and including the expected arrival number within the same time slot in that time slot; and determining the remaining backlog amount for each time slot based on the expected arrival number, the expected number of people who can pass through, and the remaining backlog amount in the previous time slot.

[0013] Preferably, determining the recovery result includes: when acting independently, the end time of the last expected arrival time of the corresponding abnormal flow increment is taken as the recovery starting point; when acting cumulatively, the latest end time of the last expected arrival time of each abnormal flow increment participating in the cumulative action is taken as the recovery starting point; a recovery interval is determined based on the recovery starting point and a preset recovery duration; within the recovery interval, the backlog is updated based on the expected number of people who can pass; when the backlog decreases to a queuing recovery threshold that matches the venue's operating conditions at the evaluation time and is maintained for a preset number of time slices, it is determined that the corresponding action meets the recovery conditions within the recovery interval; otherwise, it is determined that the corresponding action does not meet the recovery conditions within the recovery interval.

[0014] Preferably, the method further includes: expanding the cumulative source set in batches according to the expected arrival start time, with the same start time constituting the same batch, and using the scenario containing only the initial backlog and the baseline expected arrival number as the initial source set; when the previous source set meets the recovery conditions but the source set after adding the current batch does not, the start time of the batch is taken as the recovery state flip trigger time, and the source set after adding the batch is the flip scenario; the earliest unstable bottlenecks at the trigger time are combined into the first unstable bottleneck set; keeping the evaluation time, recovery start point, and recovery interval of the flip scenario unchanged, the expected arrival number of abnormal flow increments is removed one by one and the recovery result is re-determined, and the monitoring area corresponding to the abnormal flow increments that make the flip scenario meet the recovery conditions after removal is determined as the key abnormal source area set; when there are no abnormal flow increments that meet the conditions, the single source cancellation cannot be determined flag is output.

[0015] A smart social risk early warning system based on a deep learning hybrid model includes: Anomaly Increment Module: Acquires the pedestrian flow status of multiple monitoring areas in a large public place and determines the abnormal pedestrian flow increment corresponding to each monitoring area; Propagation prediction module: Based on the location traffic topology, determine the propagation path of each abnormal increase in pedestrian flow to the downstream bottleneck, and obtain the corresponding expected arrival time and the expected number of people arriving in each time slice through a deep learning hybrid model; Backlog Recovery Module: Based on the normal passenger flow baseline, determine the baseline estimated number of arrivals for each time slot of the downstream bottleneck; based on the current queuing status and actual passage status of the downstream bottleneck, determine the estimated number of passages for each time slot; map the baseline estimated number of arrivals and the estimated number of arrivals for each abnormal passenger flow increment to a common prediction time axis; use the remaining backlog amount of the previous time slot as the initial backlog amount of the current time slot; determine the backlog status when each abnormal passenger flow increment acts independently and when multiple abnormal passenger flow increments act cumulatively. Risk warning module: Based on the backlog status, determine whether each independent action and cumulative action meets the recovery conditions within their respective recovery intervals; if each independent action meets the conditions but the cumulative action does not, the corresponding downstream bottleneck is identified as an unstable bottleneck and personnel gathering risk warning information is output.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention predicts the expected arrival time of abnormal pedestrian flow increments in each monitoring area to downstream bottlenecks and the expected number of people arriving in each time slot based on the location traffic topology. This allows changes in pedestrian flow from different sources to participate in subsequent analysis according to the actual propagation path and time process, which is beneficial to improving the ability of early warning of the risk of crowd gathering in large public places to characterize the spatial propagation process and the temporal evolution process.

[0017] 2. This invention maps the baseline estimated arrival number of people at the downstream bottleneck and the estimated arrival number of abnormal pedestrian flows leading to the bottleneck to a common prediction time axis. It then continuously updates the backlog status by combining the estimated passage capacity of people in each future time slot, ensuring that the passage capacity occupied by normal pedestrian flow, the new arrivals from abnormal pedestrian flow, and the remaining backlog from the previous time slot are all calculated under a unified population metric. Based on this, by comparing the recovery results when each abnormal pedestrian flow increment acts independently and cumulatively, changes in the recovery status caused by the spatiotemporal superposition of multiple pedestrian sources can be identified, improving the ability to identify the risk of persistent crowd gatherings in advance.

[0018] 3. This invention can determine the first unstable bottleneck based on the recovery state flip trigger time of different unstable bottlenecks, and under the condition of keeping the evaluation time and recovery interval of the original cumulative effect scenario unchanged, it can determine the set of key abnormal source areas or cancel the markers of single sources that cannot be determined by removing the expected number of people corresponding to the abnormal flow increment one by one, so that the early warning information can be associated with the unstable location, flip trigger time and corresponding upstream source data. Attached Figure Description

[0019] Figure 1 A flowchart of a smart society risk early warning method based on a deep learning hybrid model provided in an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a smart social risk early warning system based on a deep learning hybrid model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the process for determining the initial instability bottleneck and key anomaly source region provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1 to 3This invention provides a smart society risk early warning method and system based on a deep learning hybrid model, the technical solution of which is as follows: A smart society risk early warning method based on deep learning hybrid models includes: Acquire the pedestrian flow status of multiple monitoring areas in a large public place, and determine the abnormal pedestrian flow increments corresponding to each monitoring area; Based on the location topology, the propagation path of each abnormal increase in pedestrian flow to the downstream bottleneck is determined, and the corresponding expected arrival time and the expected number of people arriving in each time slice are obtained through a deep learning hybrid model. Based on the normal passenger flow baseline, determine the baseline expected arrival number for each time slot of the downstream bottleneck. Based on the current queuing status and actual passage status of the downstream bottleneck, determine the expected passage number for each time slot. Map the baseline expected arrival number and the expected arrival number of each abnormal passenger flow increment to the common prediction time axis. Use the remaining backlog of the previous time slot as the initial backlog of the current time slot to determine the backlog status when each abnormal passenger flow increment acts independently and when multiple abnormal passenger flow increments act cumulatively. Based on the backlog status, determine whether each independent action and cumulative action meets the recovery conditions within their respective recovery intervals; if each independent action meets the conditions but the cumulative action does not, the corresponding downstream bottleneck is identified as an unstable bottleneck and a personnel gathering risk warning is output.

[0022] Example 1: This embodiment is applied to the risk warning of crowd gathering during the exit and transition periods of a large convention center. The convention center includes multiple exhibition halls, a public lobby, connecting corridors, and exit areas. Crowds from different exhibition halls can converge at the same turnstile, escalator entrance, and narrow connecting passages through different channels. The monitoring system continuously acquires data on the number of people in each area, movement direction, passage speed, number of people entering and leaving the area, and number of people passing through the exit. When multiple upstream areas generate abnormal crowd flows that have not yet formed a sustained accumulation, this embodiment does not directly determine the risk based on the current number of people in each area. Instead, it predicts the time it takes for each abnormal crowd flow to reach the downstream bottleneck along the site's topology, and continuously calculates the future backlog process based on the actual passage status of the bottleneck to determine whether the combined effect of multiple abnormal crowd flows will cause the downstream bottleneck to recover and reverse its state.

[0023] As one embodiment of the present invention, refer to Figure 1 Flowchart of a smart society risk early warning method based on a deep learning hybrid model.

[0024] The video surveillance equipment, turnstile counting equipment, and regional pedestrian flow statistics are time-aligned according to a unified time slice. The monitored area is used as the upstream node of the venue's traffic topology, and the passage relationships between nodes are formed by connecting channels and confluence points. For each monitored area, the abnormal pedestrian flow increment corresponding to the current time slice is first determined. The propagation path is determined based on the connectivity between the area where the abnormal pedestrian flow increment is located and the candidate downstream bottlenecks. A deep learning hybrid model is then used to predict the expected arrival time of the abnormal pedestrian flow along the propagation path to each downstream bottleneck and the expected number of people arriving in each time slice.

[0025] During time alignment, each device uses the venue's unified clock to record the collection time and device number, and collects data according to the time slices of left-closed and right-open. The number of people and inflow / outflow counts are summed within the time slice, while speed and density are calculated using a weighted average based on the effective sampling duration. The bottleneck opening status uses the effective status at the end of the time slice. Data reported repeatedly by the same device number and collection time is retained only once. Data delayed by no more than one time slice is backfilled to the original time slice; data exceeding this range retains the delayed status marker but does not overwrite the already frozen online judgment result. If a time slice lacks the number of people, directions, or exit passages, and consistency verification cannot be completed by another counting device in the same location, that time slice is marked as incomplete, and no definitive recovery result is output based on this.

[0026] For the same downstream bottleneck, the actual throughput is determined based on the current queuing status, the recent measured number of people passing through the exit, and the bottleneck's open status. Based on this, the expected throughput for each future time slot is determined. The baseline expected arrival number for the bottleneck and the expected arrival numbers corresponding to each abnormal flow increment are mapped to the same prediction time axis according to their respective expected arrival times. When calculating the backlog status for a certain time slot, the total backlog number not released in the previous time slot is used as the initial backlog number for the current time slot. The baseline expected arrival number for the current time slot and the expected arrival number of all participating abnormal flows are then included, and the total expected throughput for that time slot is deducted. This continuously yields the backlog status for each future time slot.

[0027] For each abnormal flow of people, an independent action scenario is established, along with a cumulative action scenario encompassing multiple abnormal flows, and the recovery outcome is determined for each scenario. When the independent actions of each abnormal flow are recoverable, but the cumulative action of multiple abnormal flows is unrecoverable, the corresponding downstream bottleneck is identified as an instability bottleneck, and a crowd gathering risk warning is generated. If multiple instability bottlenecks exist, the first instability bottleneck is determined according to the recovery state reversal trigger time of the cumulative source set (described later).

[0028] This embodiment establishes the early warning results based on the actual traffic structure of the location, the time of pedestrian flow, the estimated number of people arriving in the time slice, and the continuous backlog of bottlenecks. It can further determine whether the subsequent combined effect of upstream areas will cause downstream bottlenecks to lose their recovery ability within a predetermined time before a sustained clustering has formed in each upstream area. This improves the ability to identify the risk of sustained clustering caused by the spatiotemporal evolution of pedestrian flow in multiple areas in advance.

[0029] Furthermore, determining the abnormal population increase corresponding to each monitoring area includes: determining the corresponding normal population baseline from historical population status according to activity stage, date type and regional operation status; counting the number of people moving along the downstream direction in the current population status according to a unified time slice, determining the new number of people exceeding the normal population baseline as the abnormal population increase of the corresponding monitoring area, and forming an abnormal population increase sequence according to time order.

[0030] When determining abnormal pedestrian flow increases for each monitoring area, historical normal operation data are grouped according to activity stage, date type, and area operation status. For example, historical sample sets are established for the exhibition opening stage, conference closing stage, and venue closing stage respectively; for the same activity stage, corresponding historical samples are formed based on weekday status and venue open area status. For each monitoring area, historical samples within the same activity stage are aligned according to relative time position. From the valid samples where the activity stage, date type, and area operation status are consistent with the evaluation time, the median of the number of people moving downstream in each time slot is taken as the normal pedestrian flow baseline; in this embodiment, at least 5 valid historical samples are used for each time slot. The actual number of people arriving at each downstream bottleneck in these historical samples is simultaneously retained. Samples where the venue topology and bottleneck opening status are consistent with the evaluation time are aligned with the same time position, and the median is taken as the baseline expected arrival number for the corresponding bottleneck. If there are fewer than 5 valid samples, the nearest normal sample from the same activity stage with the same bottleneck opening status is added; if it is still insufficient after adding, a baseline data insufficiency indicator is output, and no further definitive recovery result is given. After the activity confirmed to be normal operation ends, the corresponding historical sample set is updated. When the number of people moving downstream in the current time slice exceeds the normal flow baseline, the excess portion is identified as an abnormal flow increment for that time slice; if the current time slice does not exceed the normal flow baseline, no abnormal flow increment is generated for that time slice. Abnormal flow increments formed in consecutive time slices are recorded according to region number and time sequence, enabling subsequent propagation models to handle flow changes generated in different monitoring areas separately.

[0031] To avoid generating duplicate abnormal population increments when the same group of people passes through adjacent monitoring areas, priority is given to retaining the source area identifier of the first abnormal increment for the same person or the same counting batch, based on cross-regional anonymous trajectory identifiers and departure / entry records at channel boundaries. Subsequent observations in downstream areas are only used to update position, speed, and arrival time, without generating new abnormal increments. When the equipment cannot provide continuous trajectory identifiers, batch matching is performed according to adjacent area boundaries, movement direction, passage time range, and the number of people leaving and entering, with the original source identifier retained for the matched portion; if source differentiation cannot be completed, an uncertain source association identifier is output, and the key abnormal source area is not determined based on this. Here, abnormal population increments are uniformly recorded using "time slice number of people". When the original equipment output is a rate of change in the number of people, it is first converted into the corresponding number of people in the time slice according to the time slice length before entering the subsequent propagation calculation, so that the abnormal population increment, the expected number of people arriving, the expected number of people passing through, and the remaining backlog of people maintain a consistent data caliber.

[0032] This invention establishes a baseline of normal crowd flow based on the activity stage and regional operation status, and forms abnormal crowd flow increments based on the number of people in a uniform time slot. This distinguishes between crowd flow changes caused by normal venue changes and planned departures and crowd flow increments that deviate from normal operation status. At the same time, it ensures that abnormal crowd flow increments can be directly included in subsequent propagation prediction and backlog calculation, thereby improving the comparability of crowd flow status between different stages of the venue.

[0033] Furthermore, determining the propagation path includes: constructing the site access topology based on the actual connectivity between the monitoring area, connecting channels, and confluence nodes; identifying restricted access nodes based on the effective access width of the nodes, the number of open channels, and the historically stable number of people passing through; identifying restricted access nodes with upstream reachable paths from two or more different monitoring areas as candidate downstream bottlenecks; and determining the corresponding propagation path based on the connectivity between the monitoring area corresponding to each abnormal increase in pedestrian flow and the candidate downstream bottleneck.

[0034] The venue access topology is established based on the venue's architectural plan data and actual open access relationships. Corresponding topology nodes are set for monitoring areas, passage entrances, passage exits, escalator entrances, turnstiles, and connecting corridor entrances. Directed connections are established between nodes where personnel movement can occur. When a passage is temporarily closed, the corresponding connections cease to participate in propagation path calculations.

[0035] For nodes that may create passage constraints, the available net width after deducting fixed obstacles and closed areas is recorded as the effective passage width. The number of currently open passages is obtained from the status records of turnstiles, escalators, or passageways. The number of people exiting per unit time slice during two or more consecutive time slices when there are queues and the equipment status remains stable is selected as the historical stable passage sample. The number of people passing through the effective passage width is converted using the historical stable passage number per unit net width of the same type of passageway. The number of people passing through the open passages is the sum of the historical stable passage numbers of each open passageway under the same operating state. The historical stable passage number is the median of the effective sample under the current open state.

[0036] The smaller of the following values ​​is used as the effective passage capacity per unit time slice: the number of people passing through the effective passage width, the number of people passing through the open passage, and the historical stable number of people passing through under the current open state. When upstream reachable paths from two or more different monitoring areas converge on the same node, and the upper quartile of the historical upstream arrivals under the same conditions reaches the effective passage capacity, or when the node has queues in two or more consecutive time slices under the same open state, the node is identified as a candidate downstream bottleneck. For each monitoring area that generates abnormal pedestrian flow increases, connecting paths that can reach each candidate downstream bottleneck are determined from the site's passage topology; closed passages and currently prohibited passage directions are not included in the propagation paths.

[0037] This invention establishes a pedestrian flow propagation topology based on actual open access relationships and uses the status of access facilities to determine candidate downstream bottlenecks. This limits abnormal pedestrian flow prediction to objectively existing pedestrian movement paths, reducing the possibility of including pedestrian flows from areas that cannot reach the same bottleneck in the same cumulative effect scenario. It also provides a clear spatial effect boundary for subsequent bottleneck backlog state calculations.

[0038] Furthermore, the deep learning hybrid model includes a spatial propagation sub-model and a temporal evolution sub-model. The spatial propagation sub-model determines the expected number of people diverted into each subsequent channel of the corresponding propagation path based on the location topology, abnormal pedestrian flow increment, and propagation path. The temporal evolution sub-model determines the expected number of people arriving at each time slot of the propagation to the downstream bottleneck based on the expected number of people diverted and the pedestrian flow speed sequence, density sequence, and travel time of the corresponding channel. The time slots in which the expected number of people arrives continuously are determined as the corresponding expected arrival time slots.

[0039] In this embodiment, the deep learning hybrid model consists of a spatial propagation sub-model and a temporal evolution sub-model. The spatial propagation sub-model employs a graph attention network based on location traffic topology, while the temporal evolution sub-model employs a gated recurrent network. The aforementioned network types are used to illustrate one possible implementation and do not constitute a limitation on the model structure. Training samples are generated according to a unified time slice. Node inputs include at least the abnormal increase in pedestrian flow in the monitoring area, the number of people in the area, the average moving speed, and the density; channel inputs include channel connectivity, travel distance, current open status, and travel time for the corresponding time slice. For historical data, the source areas of people are determined based on the spatiotemporal continuity of video trajectories in adjacent monitoring areas, and the actual number of people arriving at the candidate downstream bottleneck from each source area is counted according to the time slice, serving as the model training labels.

[0040] The spatial propagation sub-model employs a two-layer graph attention network. Each layer has four attention heads and 64 hidden features. Node features include abnormal pedestrian flow increment, number of people in the area, average movement speed, and density. Edge features include travel distance, openness, and travel time. The second layer outputs a branch score for each currently passable outgoing edge and normalizes the scores across the currently passable outgoing edges at the same branch node to a non-negative, summing ratio of 1. The spatial propagation sub-model outputs the estimated number of people to be diverted for each abnormal pedestrian flow source, with source and channel identifiers. The temporal evolution sub-model employs a two-layer gated recurrent network with 64-dimensional hidden states in each layer. The channel velocity, density, and travel time of the 30 most recent time slices before the evaluation time are input chronologically to obtain the terminal hidden state at the evaluation time. This terminal hidden state is then concatenated with the expected number of people diverted to each channel from the corresponding source, the travel distance, and the current open state. This concatenation is passed through a non-negative fully connected layer with an output dimension equal to the preset number of prediction time slices to generate the expected number of people arriving at the target bottleneck in each future time slice. Each output position corresponds sequentially to a prediction time slice after the evaluation time. The output of the temporal evolution sub-model retains the source identifier and the target bottleneck identifier. When the same source arrives at the same target bottleneck via multiple paths, the expected number of people arriving is summarized according to the prediction time slices, and different sources form separate sequences of expected arrival numbers. The above layer and dimension configuration is a representative configuration of this embodiment and can be adjusted using a validation set based on the venue size. The estimated number of people diverted is an intermediate model quantity connecting the spatial propagation sub-model and the temporal evolution sub-model. It is constrained by the loss of people arriving at the target bottleneck and the conservation deviation of people at the bifurcation nodes. It is used to generate the estimated number of people arriving at the target bottleneck and is not used as an independent measurement result of the actual number of people passing through the corresponding channel.

[0041] During model training, the model is divided into training, validation, and test sets based on the number of activity sessions, with proportions of 70%, 15%, and 15%, respectively. Data from the same activity session does not cross sets. Continuous features are standardized using the mean and standard deviation of the training set. The spatial propagation sub-model and the temporal evolution sub-model are jointly trained. During training, the expected number of people diverted from each source and each channel, output by the spatial propagation sub-model, is used as input to the temporal evolution sub-model. The arrival loss is calculated based on the actual number of people arriving at the candidate downstream bottleneck in each time slice from each source, and this loss is then backpropagated from the temporal evolution sub-model to the spatial propagation sub-model. The loss term includes the mean squared error between the expected number of people arriving at the target bottleneck in each time slice from each source and the training label, as well as the conserved bias between the number of people entering from each fork node from the same source and the sum of the expected number of people diverted from the currently passable outgoing edge. The weight of the conserved bias term is 0.1. The Adam optimizer is used with an initial learning rate of 0.001, a batch size of 32, and a maximum training time of 200 epochs. Training stops when the validation set loss does not decrease for 10 consecutive epochs, and the parameter with the lowest validation set loss is retained. During model prediction, each monitoring area generating an abnormal increase in pedestrian flow retains an independent source identifier. The spatial propagation sub-model determines the estimated number of people diverted into each subsequent channel of the reachable propagation path based on topological connectivity and current traffic status. The estimated number of people diverted into each currently passable subsequent channel at a bifurcation node is non-negative, and their total number is consistent with the number of people entering that bifurcation node. Branches corresponding to closed channels and prohibited directions are not included in the normalization; branches that are passable but cannot reach the current target bottleneck participate in the normalization of the estimated diversion ratio, but their estimated diversion numbers are not included in the estimated number of people arriving at the target bottleneck.

[0042] The temporal evolution sub-model, following the aforementioned method, outputs the non-negative real number of people arriving at the target bottleneck along each propagation path of each abnormal flow source and its corresponding time slot. When the total output value for each time slot is greater than 0, it is corrected proportionally to ensure the total number matches the expected number of people arriving at the target bottleneck along the propagation path, as determined by the spatial propagation sub-model. When the total output value is 0 and the expected number is greater than 0, the prediction is marked as a temporal evolution prediction anomaly. The arrival time slot determined by the median historical travel time can be used as a backup arrival time slot, but the recovery result formed using the backup arrival time slot is marked as uncertain, and the recovery state reversal is not determined based on this. The total number of people is then rounded down, and the number of people in each time slot is rounded down, with the remaining number padded in descending order of the decimal part, so that the total number of people in the discretized time slots matches the total number of people after rounding and that no negative numbers are found. The time slices in which the number of people expected to arrive is continuously greater than 0 after discretization are divided into one expected arrival period. When there are multiple discontinuous positive time slices, multiple expected arrival periods are formed, and the start and end points of the corresponding expected arrival period are determined by the first and last positive time slices of each segment. When there are multiple expected arrival periods, the expected start time of the abnormal flow increase is the start time of the earliest expected arrival period, and the expected end time is the end time of the last expected arrival period. The preset number of prediction time slices is determined before model training and deployment based on the maximum passage time, preset recovery time, and stable duration in historical samples, so that the common prediction time axis covers the recovery time and stable duration after the latest expected end time. When the required prediction range exceeds the preset number of prediction time slices due to temporary changes in passage status, an insufficient prediction range indicator is output, and no definite recovery result is given.

[0043] In this application, the deep learning hybrid model does not directly generate risk levels, but instead generates the estimated number of people arriving in each time slice and the estimated arrival time period required for subsequent backlog calculations. Therefore, the model's prediction results can continue to participate in the bottleneck queuing state calculation, forming a continuous data relationship between spatial propagation prediction and risk assessment, and preserving the time differences in the arrival of abnormal pedestrian flows at the same bottleneck in different monitoring areas.

[0044] Furthermore, determining the expected number of people who can pass through each time slot in the future includes: obtaining the number of people queuing at the downstream bottleneck at the evaluation time and the actual number of people passing through the exit in multiple consecutive time slots before the evaluation time; determining the actual number of people passing through the downstream bottleneck per unit time slot based on the actual number of people passing through the exit and the bottleneck opening status; and determining the expected number of people who can pass through each time slot in the future based on this; when the bottleneck opening status changes, updating the expected number of people who can pass through in subsequent time slots based on the changed opening status.

[0045] For each candidate downstream bottleneck, continuously record the actual number of people passing through the exit, the current number of people in the queue, and the bottleneck's open status. When there is a continuous queue, the recent actual number of people passing through the exit can reflect the actual throughput under that open status. Use continuous time slice samples with consistent status to determine the actual throughput status.

[0046] When the number of people queuing recently is insufficient to continuously occupy the bottleneck, the actual number of people passing through the exit may be affected by upstream demand. In this case, valid passing samples are supplemented from historical time slices that have been continuously queuing under the same open state. If the total number of current samples and historical samples is still less than 10, the actual passing status of the bottleneck is marked as insufficient data, and no definite backlog recovery result is output based on this.

[0047] If there are no known changes in the open status during the future forecast period, the actual number of people passing through per unit time slice determined by the current actual passage status is first used as the candidate expected number of people who can pass. Based on this, the initial expected number of people queuing and the remaining backlog number for each forecast time slice are initially calculated. If the number of open channels, passage direction, and facility operation status of the forecast time slice are consistent with the effective passage sample, and the initial expected number of people queuing for each forecast time slice is within the range of the initial queuing number of the effective passage sample time slice, the candidate expected number of people who can pass is used to form the backlog result. If any forecast time slice exceeds the coverage of the effective passage sample, an insufficient actual passage status data is output, and no definite backlog recovery result is given based on this. When there is a pre-determined plan for gate opening, channel closure, or traffic direction adjustment, the change time and the changed opening status are written into the traffic topology of the corresponding time slice: when only the target bottleneck's opening status is changed, the estimated number of people who can pass through in subsequent time slices is updated; when the change also affects the upstream reachable path, the estimated number of people arriving at the affected baseline, the abnormal flow propagation path, and the estimated number of people arriving at the abnormal flow are re-determined from the changed time slice, and the updated estimated number of people arriving and the estimated number of people who can pass through are used together for backlog calculation. The throughput capacity per unit time slice of newly opened channels is determined using historical stable throughput records under the same equipment state. When performing the re-determination, the current position of each source, the number of people who have not yet reached the target bottleneck, and the queuing contribution already formed are recorded in the last complete time slice before the change as the inherited state; only for the number of people who have not yet reached the target bottleneck, the propagation path and the estimated number of people arriving are re-determined according to the changed location traffic topology, and the queuing contribution already formed continues to be included in the initial backlog of the corresponding scenario.

[0048] Therefore, the future backlog calculation uses the "expected number of people who can pass", rather than the actual number of people who have not yet occurred in the future.

[0049] This invention uses the actual passage status and bottleneck operation status to form the expected number of people who can pass through in future time slices, so that the predicted entry volume and the predicted release volume have a consistent time slice scale. At the same time, it avoids directly equating the facility's designed passage capacity with the actual release capacity, and allows the backlog status calculation to be updated as the gates open, the channel status and the recent actual passage situation.

[0050] Based on the above implementation method, further, time slots in which downstream bottlenecks continue to have queues and whose open status is consistent with the current time are selected from the time slots before the evaluation time. The number of people passing through the corresponding exit is determined as the valid passing sample. The median of the valid passing samples is used to determine the actual number of people passing through per unit time slot, and the expected number of people passing through in each future time slot is determined accordingly. If there are less than 10 valid passing samples at present, samples are supplemented from historical time slots in the same open status where queues continue to exist.

[0051] Specifically, the actual number of people passing through the exits, the number of people queuing at the start of the time slot, and the bottleneck opening status are recorded according to the same time slots as the predicted flow of people. Time slots with a starting queue number greater than the queue recovery threshold matching the operating conditions at the evaluation time, and where the number of open channels, traffic direction, and facility operating status are consistent with the evaluation time, are selected. The actual number of people passing through their exits is included in the valid passing sample. If the current sample is less than 10, historical samples in the same state are added. If the sample is still less than 10 after adding, an insufficient actual passing status data flag is output. When the future opening status changes as known, the valid passing samples are re-matched based on the changed opening status, and the predicted passing number for subsequent time slots is updated.

[0052] This invention extracts valid passage samples from time slices where there is a continuous queue and the operating status is consistent. This allows the estimated number of people who can pass to mainly reflect the actual passage level when the bottleneck is in a state of continuous waiting for people to pass. This reduces the impact of insufficient demand at the entrance on the judgment of the backlog recovery, thereby improving the consistency between the recovery status calculation and the actual passage status of the bottleneck on site.

[0053] Furthermore, determining the backlog status when each abnormal flow increment acts independently and when multiple abnormal flow increments act cumulatively includes: when acting independently, mapping the baseline expected arrival number and the expected arrival number for each time slot corresponding to an abnormal flow increment to a common prediction time axis; when acting cumulatively, mapping the baseline expected arrival number and the expected arrival number for each time slot corresponding to multiple abnormal flow increments leading to the same downstream bottleneck to a common prediction time axis, and including the expected arrival number within the same time slot in that time slot; and determining the remaining backlog amount for each time slot based on the expected arrival number, the expected number of people who can pass through, and the remaining backlog amount in the previous time slot.

[0054] When establishing independent action scenarios, the evaluation time is preferentially set before the earliest expected arrival of the abnormal pedestrian flow participating in the analysis to the target bottleneck, and the actual total queuing number at that time is used as the common baseline initial backlog number for each scenario. Each independent action scenario retains the baseline expected arrival number, and only the abnormal pedestrian flow increment generated by one monitoring area is added each time. If the abnormal pedestrian flow participating in the analysis has already arrived at the evaluation time, all queuing contributions already formed by the participating sources are removed from the actual queuing number according to the source identifier to obtain the common baseline initial backlog number, and only the queuing contribution already formed by the current source is added back in each independent action scenario; if source decomposition cannot be completed, an initial state indivisible identifier is output, and no comparison of the recovery state of independent action and cumulative action is performed. When the abnormal pedestrian flow has already arrived at the target bottleneck at the evaluation time, the cross-regional anonymous trajectory identifier or batch matching result is used to count the number of people who have not yet passed through the target bottleneck queuing area for each source identifier, and the number of people is used as the queuing contribution already formed by the corresponding source; when the corresponding person or counted batch is detected to have passed through the bottleneck exit, it is deducted from the queuing contribution of that source. If the source identifier cannot be maintained until the target bottleneck queuing area or the batch matching result cannot be uniquely determined, the initial state indivisible identifier is output, and no comparison of the recovery state with independent action and cumulative action is performed.

[0055] When establishing a cumulative effect scenario, the baseline expected arrival numbers and common baseline initial backlog numbers are retained, identical to those in the independent effect scenario. If abnormal passenger flows already participating in the analysis have arrived at the evaluation time, the queuing contributions from each source within the cumulative effect scenario are added back. Multiple abnormal passenger flows leading to the same downstream bottleneck are then mapped to the same prediction time axis according to their actual expected arrival times. When multiple abnormal passenger flows are expected to arrive in the same time slice, the expected arrival numbers from each source are included in that time slice without artificially adjusting the actual arrival order.

[0056] Each time slot uses the remaining backlog from the previous time slot as the initial backlog, then adds the baseline estimated arrival number for the current time slot and the estimated arrival number of all abnormal flows involved, and subtracts the estimated passable number for that time slot. If the result is less than 0, the remaining backlog for that time slot is recorded as 0. Therefore, the baseline estimated arrival number and the total estimated passable number for the bottleneck use the same total metric, and the increase in abnormal flows is only added to the total arrival number. For example, using a 1-minute time slot, if the current queue number for a bottleneck is 10, the baseline estimated arrival number for each future time slot is 10, and the estimated passable number is 25. Abnormal flow A is expected to arrive in time slots 1-3 with 10, 20, and 10 people respectively; abnormal flow B is expected to arrive in time slots 2-4 with 8, 18, and 8 people respectively.

[0057] The remaining backlog of people due to the effects of A alone, B alone, and the cumulative effects of A+B are as follows: Time slice A acting alone B acting alone The cumulative effect of A+B Time Capsule 5 0 5 Second Time Slice 10 0 18 Time Capsule 3 5 3 31 4th Time Slice 0 0 24 5th Time Slice 0 0 9 6th Time Slice 0 0 0 In the second and third time slices, A and B actually arrive at the same time. Therefore, in the cumulative effect scenario, they enter the corresponding time slice calculation at the same time, instead of completing all the effects of A first and then calculating B.

[0058] This invention maintains the expected arrival sequence of normal baseline pedestrian flow and different abnormal pedestrian flow by jointly predicting the time axis. This allows normal pedestrian flow to occupy the bottleneck capacity and abnormal pedestrian flow with overlapping arrival times to participate in the backlog calculation in the same time slot. At the same time, the number of people not released in the previous time slot is continuously transferred to the next time slot, thereby characterizing the superposition of multiple sources of pedestrian flow in time and the resulting continuous queuing state.

[0059] Further, determining the recovery result includes: for independent actions, the end time of the last expected arrival time of the corresponding abnormal passenger flow increment is used as the recovery starting point; for cumulative actions, the latest end time of the last expected arrival time among the abnormal passenger flow increments participating in the cumulative action is used as the recovery starting point; a recovery interval is determined based on the recovery starting point and a preset recovery duration; within the recovery interval, the backlog is updated based on the expected number of people who can pass; when the backlog decreases to a queuing recovery threshold that matches the venue's operating conditions at the evaluation time and remains at a preset number of time slices, it is determined that the corresponding action meets the recovery conditions within the recovery interval; otherwise, it is determined that the corresponding action does not meet the recovery conditions within the recovery interval. In this embodiment, "recoverable state" or "able to recover" both mean that the corresponding action meets the recovery conditions within its recovery interval, and "unrecoverable state" or "cannot recover" both mean that the corresponding action does not meet the recovery conditions within its recovery interval.

[0060] When determining the backlog recovery result, recovery starting points are set for both independent and cumulative scenarios. For independent scenarios, the recovery starting point is the end time of the last expected arrival time slot of the current abnormal pedestrian flow; for cumulative scenarios, the recovery starting point is the latest end time slot among the last expected arrival time slots of all abnormal pedestrian flows currently participating in the cumulative recovery. This setting ensures that recovery judgments begin after all abnormal inputs in the current scenario have ended.

[0061] The queue recovery threshold and preset recovery duration are determined before deployment by grouping according to venue area, activity stage, and bottleneck opening status, and remain unchanged throughout the same evaluation process. When determining the queue recovery threshold, for each normally operating activity in the same group, excluding time slots with equipment malfunctions, temporary control measures, and abnormal passenger flow increases, the 75th percentile of the starting queue number in the remaining time slot is taken as the normal queue representative value for that activity. Then, the 75th percentile of the representative values ​​from at least 20 activities is taken as the queue recovery threshold. When determining the preset recovery duration, from at least 20 activities in the same group that included abnormal passenger flow increases and ultimately returned to normal, the starting point is the last moment when the abnormal passenger flow actually reached the target bottleneck, and the ending point is the moment when the queue number first drops to the queue recovery threshold and remains stable for the shortest possible duration. The 95th percentile of the duration obtained for each activity is taken as the preset recovery duration. The shortest stable duration is taken as 2 minutes, and converted into the number of consecutively maintained time slots according to the time slot length; in this embodiment, a 1-minute time slot is used, and two consecutive time slots are maintained. If there are fewer than 20 samples for any type of activity, the output parameter will indicate that there are not enough samples, and no definite recovery result will be given.

[0062] Within the recovery interval, the backlog status continues to be updated according to the expected number of people who can pass. When the number of people in the backlog drops below the queue recovery threshold and remains within a preset number of time slices, the scenario is determined to be recoverable; if the above conditions are not met by the end of the recovery interval, it is determined to be unrecoverable. The 95th percentile of the absolute error of the number of people expected to arrive in the model validation set is used as the error amount of the expected number of people arriving in abnormal flow, which is added to or subtracted from the model output, and the result after subtraction is not less than 0; the upper and lower quartiles of the baseline expected arrival sample and the effective passing sample are used as the boundaries of the baseline arrival number and the expected passing number, respectively. Unfavorable results are calculated using the upper boundary of abnormal arrival, the upper boundary of baseline arrival, and the lower boundary of the expected passing number, and favorable results are calculated using the opposite boundary; when the two conclusions are consistent, the corresponding recovery result is output; when the conclusions are inconsistent, an uncertain recovery result is output, and the recovery status is not determined to be reversed based on this. For example, using the above example, assuming the queue recovery threshold is 5 people, the recovery interval includes 2 time slices after the abnormal input ends, and it is required that the number of people in the backlog in 2 consecutive time slices does not exceed 5 people. Player A arrives at the end of time slice 3, and the backlog is 0 in time slices 4 and 5, so it is determined that player A's independent action can be recovered. Player B arrives at the end of time slice 4, and the backlog is 0 in time slices 5 and 6, so it is determined that player B's independent action can be recovered. The last arrival time of the A+B cumulative scenario is time slice 4. There are 9 players remaining in time slice 5 and 0 players remaining in time slice 6. Therefore, the recovery conditions are not met for two consecutive time slices within the corresponding recovery interval, and the cumulative action scenario is determined to be unrecoverable. This example is only used to illustrate the recursive and recovery determination steps; the actual recovery parameters are determined according to the aforementioned historical samples.

[0063] This results in the following: A acting independently can be recovered; B acting independently can be recovered; the cumulative effect of A+B cannot be recovered under the same recovery determination rule. This change is the recovery state reversal described in this application.

[0064] This invention distinguishes between general queue increases and changes in recovery characteristics caused by the combined effects of multiple sources of people by adopting a unified recovery judgment rule for independent and cumulative action scenarios. It uses "independent action can recover, while cumulative action cannot recover" as a criterion for continuous clustering risk, which is beneficial for identifying in advance the risk of personnel backlog caused by individual normal local areas but subsequent joint transmission.

[0065] Building upon the above implementation, a further step is to use the same target bottleneck as the boundary for grouping abnormal pedestrian flows. For each independent impact scenario of an abnormal pedestrian flow increment, the backlog status is determined from its expected arrival start time, and the backlog amount continues to be updated after the expected arrival ends. When the recovery condition is met for the first time, the end time of the corresponding time slice is taken as the end of the effective impact period. If recovery is still not possible until the end of the common prediction timeline, the effective impact period is extended to the end of that timeline, and the abnormal pedestrian flow is marked as an independent impact that cannot be recovered. This source does not participate in the reversal judgment of "all independent impacts can be recovered", but is output separately as an independent continuous backlog risk.

[0066] Within the same target bottleneck, when the expected arrival time of one abnormal flow overlaps with the effective period of another abnormal flow, an overlap relationship is established between them, and cumulative effect groups are formed according to the transitive closure of the overlap relationship. Therefore, when A overlaps with B and B overlaps with C, even if A and C do not directly overlap, A, B, and C are still grouped into the same group. Multiple abnormal flows corresponding to different target bottlenecks are grouped separately and do not merge across bottlenecks. For example, if the expected arrival time of abnormal flow A is 10:00-10:10, and the backlog formed by its independent effect meets the recovery condition at 10:15, then the effective period of A is 10:00-10:15; when abnormal flow B begins to arrive at 10:12, A and B are grouped into the same group. If the effective effect of B continues until 10:20, and abnormal flow C begins to arrive at 10:18, then C is merged into the group containing A and B through the overlap relationship of B. Once the cumulative effect scenario is entered, each abnormal flow of people is still mapped to the common prediction time axis according to the original expected arrival time, and the number of expected arrivals that do not exist is not added according to the effective effect time period.

[0067] This invention incorporates the backlog recovery process after abnormal pedestrian flow stops arriving into its effective scope. Even when the original arrival times of different abnormal pedestrian flows do not directly overlap but the preceding backlog has not yet been recovered, it can still identify their temporal coupling relationship in the bottleneck state, thereby improving the ability to identify the risk of continuous backlog caused by continuous arrival of multi-source pedestrian flows.

[0068] Based on the above implementation method, further, the abnormal pedestrian flow increments within the same cumulative effect group are sorted according to the expected start time of arrival at the target bottleneck. Abnormal pedestrian flows with the same start time are considered as the same batch, and the cumulative source set is expanded sequentially. The scenario that does not contain abnormal pedestrian flow increments and only includes the common initial backlog number at the evaluation time and the baseline expected arrival number is used as the initial source set; when adding the first batch, the initial source set is used as the previous source set. Each cumulative source set uses the same evaluation time, the common initial backlog number after source decomposition, the baseline expected arrival number, and the expected passable number; if the previous cumulative source set can be recovered, but cannot be recovered after being added to the current batch, the expected arrival start time of the current batch is determined as the recovery state reversal trigger time, and the cumulative source set after being added is determined as the reversal scenario. The trigger times are compared for multiple unstable bottlenecks, and one or more bottlenecks with the earliest trigger time are determined as the first unstable bottleneck set.

[0069] Before being added to the first batch, the scenario containing only the initial backlog of people at the common baseline and the baseline expected arrival number at the evaluation time is designated as the previous cumulative source set. This previous cumulative source set uses the evaluation time as the recovery starting point and determines the recovery result according to the same preset recovery duration, queue recovery threshold, and number of continuous hold time slices. If the baseline scenario does not meet the preset recovery conditions, the recovery state reversal caused by abnormal population flow increments is not considered. The cumulative source set that first fails to meet the preset recovery conditions after being added to a batch is identified as the reversal cumulative effect scenario, and subsequent single-source removals will use the abnormal population flow increments contained in the reversal cumulative effect scenario as the object to be examined. When multiple unstable bottlenecks have the same earliest recovery state reversal trigger time, they are collectively identified as the first unstable bottleneck set.

[0070] Specifically, assuming the abnormal pedestrian flows involved in the analysis are A, B, and C in sequence, cumulative source sets corresponding to A, A+B, and A+B+C are established respectively. Sources with the same expected arrival start time are added in the same batch, without artificially setting the order. Each expansion starts from the common initial backlog of people at the same evaluation time, keeping the original expected arrival time and expected number of people for each abnormal flow unchanged. The latest expected arrival end time in each cumulative source set is used as the corresponding recovery start point, and the same preset recovery duration, queue recovery threshold, and number of continuous time slices are used to determine the recovery result of each source set. If A can be recovered, A+B can still be recovered, but A+B+C cannot be recovered, the expected arrival start time of the batch containing C is determined as the recovery state reversal trigger time, and C is determined as the recovery state reversal trigger source. When the same batch contains multiple sources, the entire batch is recorded as the trigger source batch, without forcibly distinguishing the order when there is no physical order. The trigger time represents the time when the cumulative source set first changes the recovery result in the model evaluation, and is not equivalent to the actual time when the risk of personnel gathering occurs, nor is it used to directly determine the unique key abnormal source.

[0071] This invention gradually expands the cumulative source set while keeping the actual expected arrival time of each abnormal flow unchanged. This allows multi-source flows with overlapping arrival times to still participate in the backlog calculation according to the actual time relationship. Furthermore, it can further identify the abnormal flows that cause the bottleneck recovery properties to change for the first time, providing more specific identification results for the formation stage and triggering source of the first instability risk.

[0072] Furthermore, it also includes: expanding the cumulative source set in batches according to the expected arrival start time, with the same start time constituting the same batch, and using scenarios containing only the initial backlog and the baseline expected arrival number as the initial source set; when the previous source set meets the recovery conditions but the source set after adding the current batch does not, the start time of this batch is used as the recovery state flip trigger time, and the source set after adding it is the flip scenario; the earliest unstable bottleneck at the trigger time is formed into the first unstable bottleneck set; keeping the evaluation time, recovery start point, and recovery interval of the flip scenario unchanged, the expected arrival number of abnormal flow increments is removed one by one and the recovery result is re-determined, and the monitoring area corresponding to the abnormal flow increment that makes the flip scenario meet the recovery conditions after removal is determined as the key abnormal source area set; when there are no abnormal flow increments that meet the conditions, the single source cancellation cannot be determined flag is output.

[0073] Once a downstream bottleneck has been identified as having reversed its recovery state, obtain all abnormal pedestrian flow increments that form the cumulative effect scenario and their corresponding monitoring areas.

[0074] Reference Figure 3 This document illustrates the process for determining the initial instability bottleneck and key anomaly source regions. For each abnormal flow, the evaluation time, common initial backlog number, baseline expected arrival number, expected passable number, recovery start point, and recovery interval of the original cumulative effect scenario are fixed. The expected arrival time periods and expected arrival numbers for each time slot of the remaining abnormal flows remain unchanged. Only the contribution of all expected arrival numbers corresponding to the current abnormal flow to be examined is removed. Then, based on the expected arrival numbers, expected passable numbers, and remaining backlog amount of the previous time slot, the remaining backlog amount is re-determined time-by-time, and the recovery result is re-determined according to the original recovery interval. When canceling a source, the time reference used for comparison is not changed due to the latest arrival time of the remaining source. After removing the expected arrival number contribution of the current abnormal flow, if the recovery result changes from unrecoverable to recoverable, the monitoring area corresponding to that abnormal flow is added to the key anomaly source region set. If multiple abnormal flows meet the conditions after removal, all corresponding monitoring areas are written into the set. If the recovery result cannot be changed after removing any single abnormal flow, the key anomaly source region is not left empty; instead, an indicator indicating that it could not be determined by single-source cancellation is output.

[0075] In the aforementioned A+B example, the recovery starting point after the end of time slice 4 and the recovery intervals of time slices 5 and 6 in the original A+B cumulative effect scenario are used. After removing the contribution of A's estimated arrivals, only the baseline estimated arrivals and B's contribution are retained, and the backlogs in time slices 5 and 6 are both 0; after removing B, only the baseline estimated arrivals and A's contribution are retained, and the backlogs in time slices 5 and 6 are also both 0. Therefore, both the monitoring areas where A and B are located are added to the set of key anomaly source regions corresponding to the first instability bottleneck. The set of key anomaly source regions represents the single-source removal of sensitive results while keeping other predictions unchanged. It is used to identify the source region whose estimated arrivals contribution can change the recovery result of the current calculation scenario after removal, and does not indicate that the corresponding region is the only actual cause of the gathering of people.

[0076] The final risk warning information for crowd gathering should include at least the location identifier of the initial instability bottleneck, the trigger time of the recovery state reversal, the set of key anomaly source areas or the identifier of areas that could not be determined by a single source, and the data status identifier corresponding to the baseline data, actual passage status, or insufficient prediction range. The warning information is output to the venue risk management terminal to display the physical instability location, time marker, upstream source area, and result reliability status.

[0077] Under the condition of fixing the original cumulative effect scenario time base, this invention removes the expected arrival number contribution of abnormal flow one by one and re-executes the backlog and recovery calculation, so that the early warning information is further associated with one or more upstream source areas from the unstable position; when it cannot be determined by a single source removal, a clear identifier is output, thereby avoiding ambiguity due to changes in the comparison interval or empty results of key sources.

[0078] Example 2: This embodiment provides a smart social risk early warning system based on a deep learning hybrid model, including a crowd flow monitoring data interface, a memory, and a processor. The crowd flow monitoring data interface is used to receive the number of people, movement direction, speed, density, inflow and outflow of people, and the number of people passing through bottleneck exits in various monitoring areas of large public places; the memory is used to store the location's traffic topology, normal crowd flow baseline, deep learning hybrid model, and computer program; the processor calls the above data and model to execute crowd gathering risk early warning.

[0079] Reference Figure 2 A schematic diagram of the structure of a smart society risk early warning system based on a deep learning hybrid model.

[0080] When the processor executes the computer program, it forms an anomaly increment module, a propagation prediction module, a backlog recovery module, and a risk warning module. The anomaly increment module determines the abnormal pedestrian flow increment exceeding the normal pedestrian flow state according to the pedestrian flow status and normal pedestrian flow baseline in each monitoring area, and maintains the source relationship of the abnormal pedestrian flow in different monitoring areas, providing input for subsequent source propagation prediction.

[0081] The propagation prediction module determines the propagation path of each abnormal pedestrian flow increment to the downstream bottleneck based on the location's traffic topology, and calls a deep learning hybrid model for propagation prediction. Specifically, the spatial propagation sub-model determines the estimated number of people diverted into each subsequent channel based on topological connections, while the temporal evolution sub-model combines the pedestrian flow speed, density, and travel time of the corresponding channel to obtain the estimated arrival time of each abnormal pedestrian flow to the downstream bottleneck and the estimated number of people arriving in each time slot.

[0082] The backlog recovery module determines the baseline estimated arrival number for each time slot of the downstream bottleneck based on the normal pedestrian flow baseline, and determines the estimated arrival number for each future time slot based on the actual number of people passing through the bottleneck exit and its opening status. For the same downstream bottleneck, the baseline estimated arrival number and the estimated arrival number corresponding to each abnormal pedestrian flow increment are mapped to a common prediction time axis according to the actual estimated arrival time; for independent action scenarios, only one abnormal pedestrian flow increment is added, while for cumulative action scenarios, multiple abnormal pedestrian flow increments leading to the same bottleneck are added. When multiple abnormal pedestrian flows arrive in the same time slot, they are all included in that time slot, and the remaining backlog amount of the previous time slot is used as the initial backlog amount of the next time slot, forming a backlog state time by time slot.

[0083] Furthermore, the backlog recovery module determines the recovery results for independent and cumulative impact scenarios. When the independent impact of each abnormal flow can be recovered, but the cumulative impact of multiple abnormal flows cannot be recovered, the corresponding downstream bottleneck is identified as an unstable bottleneck. For multiple unstable bottlenecks, the risk warning module determines the first unstable bottleneck based on the trigger time of the recovery state reversal of the cumulative source set; and while keeping the original cumulative impact scenario evaluation time, recovery start point, and recovery interval unchanged, it removes the contribution of the expected arrival number of abnormal flows one by one and re-determines the recovery results. The monitoring area corresponding to the abnormal flows that can enable the cumulative impact scenario to recover is identified as the key abnormal source area, and finally outputs personnel gathering risk warning information containing unstable bottlenecks and corresponding source information.

[0084] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart society risk early warning method based on a deep learning hybrid model, characterized in that, include: Acquire the pedestrian flow status of multiple monitoring areas in a large public place, and determine the abnormal pedestrian flow increments corresponding to each monitoring area; Based on the location topology, the propagation path of each abnormal increase in pedestrian flow to the downstream bottleneck is determined, and the corresponding expected arrival time and the expected number of people arriving in each time slice are obtained through a deep learning hybrid model. Based on the normal passenger flow baseline, determine the baseline expected arrival number for each time slot of the downstream bottleneck. Based on the current queuing status and actual passage status of the downstream bottleneck, determine the expected passage number for each time slot. Map the baseline expected arrival number and the expected arrival number of each abnormal passenger flow increment to the common prediction time axis. Use the remaining backlog of the previous time slot as the initial backlog of the current time slot to determine the backlog status when each abnormal passenger flow increment acts independently and when multiple abnormal passenger flow increments act cumulatively. Based on the backlog status, determine whether each independent action and cumulative action meets the recovery conditions within their respective recovery intervals; if each independent action meets the conditions but the cumulative action does not, the corresponding downstream bottleneck is identified as an unstable bottleneck and a personnel gathering risk warning is output.

2. The intelligent society risk early warning method based on a deep learning hybrid model according to claim 1, characterized in that, Determining the abnormal population increase for each monitoring area includes: determining the corresponding normal population baseline from historical population status according to activity stage, date type, and regional operation status; counting the number of people moving downstream in the current population status according to a unified time slice, identifying the new number of people exceeding the normal population baseline as the abnormal population increase for the corresponding monitoring area, and forming an abnormal population increase sequence according to time order.

3. The intelligent society risk early warning method based on a deep learning hybrid model according to claim 1, characterized in that, Determining the propagation path includes: constructing the site access topology based on the actual connectivity between the monitoring area, connecting channels, and converging nodes; identifying restricted access nodes based on the effective access width of the nodes, the number of open channels, and the historically stable number of people passing through; identifying restricted access nodes with upstream reachable paths from two or more different monitoring areas as candidate downstream bottlenecks; and determining the corresponding propagation path based on the connectivity between the monitoring area corresponding to each abnormal increase in pedestrian flow and the candidate downstream bottleneck.

4. The intelligent society risk early warning method based on a deep learning hybrid model according to claim 1, characterized in that, The deep learning hybrid model includes a spatial propagation sub-model and a temporal evolution sub-model; The spatial propagation sub-model determines the expected number of people diverted into each subsequent channel in the corresponding propagation path based on the location topology, abnormal pedestrian flow increment, and propagation path. The temporal evolution sub-model determines the expected number of people arriving at each time slot of the propagation to the downstream bottleneck based on the expected number of people diverted and the pedestrian flow speed sequence, density sequence, and travel time of the corresponding channel. The time slots in which the expected number of people arrive continuously are determined as the corresponding expected arrival time slots.

5. The intelligent society risk early warning method based on a deep learning hybrid model according to claim 1, characterized in that, Determining the expected number of people who can pass through each time slot in the future includes: obtaining the number of people queuing at the downstream bottleneck at the evaluation time and the actual number of people passing through the exit in multiple consecutive time slots before the evaluation time; determining the actual number of people passing through the downstream bottleneck per unit time slot based on the actual number of people passing through the exit and the bottleneck opening status; and determining the expected number of people who can pass through each time slot in the future based on this; when the bottleneck opening status changes, updating the expected number of people who can pass through in subsequent time slots based on the changed opening status.

6. The intelligent society risk early warning method based on a deep learning hybrid model according to claim 5, characterized in that, Determining the backlog status when each abnormal passenger flow increment acts independently and when multiple abnormal passenger flow increments act cumulatively includes: when acting independently, mapping the baseline expected arrival number and the expected arrival number for each time slot corresponding to an abnormal passenger flow increment to a common prediction time axis; when acting cumulatively, mapping the baseline expected arrival number and the expected arrival number for each time slot corresponding to multiple abnormal passenger flow increments leading to the same downstream bottleneck to a common prediction time axis, and including the expected arrival number within the same time slot in that time slot; and determining the remaining backlog amount for each time slot based on the expected arrival number, the expected number of people who can pass through, and the remaining backlog amount in the previous time slot.

7. The intelligent society risk early warning method based on a deep learning hybrid model according to claim 6, characterized in that, Determining the recovery result includes: when acting independently, the end time of the last expected arrival time of the corresponding abnormal flow increment is used as the recovery starting point; when acting cumulatively, the latest end time of the last expected arrival time of each abnormal flow increment participating in the cumulative action is used as the recovery starting point; the recovery interval is determined based on the recovery starting point and the preset recovery duration, and the backlog is updated according to the expected number of people who can pass within the recovery interval. When the backlog decreases to the queuing recovery threshold that matches the venue's operating conditions at the evaluation time and is maintained for a preset number of time slices, it is determined that the corresponding action meets the recovery conditions within the recovery interval; otherwise, it is determined that the corresponding action does not meet the recovery conditions within the recovery interval.

8. The intelligent society risk early warning method based on a deep learning hybrid model according to claim 7, characterized in that, Also includes: The cumulative source set is expanded in batches according to the expected arrival start time. The same start time is the same batch. The initial source set is the scenario containing only the initial backlog and the baseline expected number of arrivals. If the previous source set meets the recovery conditions but the source set after the current batch does not, the start time of the batch is taken as the recovery state flip trigger time, and the source set after the batch is taken as the flip scenario. The earliest unstable bottlenecks at the trigger time are formed into the first unstable bottleneck set. Keeping the evaluation time, recovery start time and recovery interval of the flip scenario unchanged, the expected number of arrivals of abnormal flow increments are removed one by one and the recovery results are re-determined. The monitoring area corresponding to the abnormal flow increment that makes the flip scenario meet the recovery conditions after removal is determined as the key abnormal source area set. If there are no abnormal flow increments that meet the conditions, the single source cancellation cannot be determined flag is output.

9. A smart social risk early warning system based on a deep learning hybrid model, characterized in that, include: Anomaly Increment Module: Acquires the pedestrian flow status of multiple monitoring areas in a large public place and determines the abnormal pedestrian flow increment corresponding to each monitoring area; Propagation prediction module: Based on the location traffic topology, determine the propagation path of each abnormal increase in pedestrian flow to the downstream bottleneck, and obtain the corresponding expected arrival time and the expected number of people arriving in each time slice through a deep learning hybrid model; Backlog Recovery Module: Based on the normal passenger flow baseline, determine the baseline estimated number of arrivals for each time slot of the downstream bottleneck; based on the current queuing status and actual passage status of the downstream bottleneck, determine the estimated number of passages for each time slot; map the baseline estimated number of arrivals and the estimated number of arrivals for each abnormal passenger flow increment to a common prediction time axis; use the remaining backlog amount of the previous time slot as the initial backlog amount of the current time slot; determine the backlog status when each abnormal passenger flow increment acts independently and when multiple abnormal passenger flow increments act cumulatively. Risk warning module: Based on the backlog status, determine whether each independent action and cumulative action meets the recovery conditions within their respective recovery intervals; if each independent action meets the conditions but the cumulative action does not, the corresponding downstream bottleneck is identified as an unstable bottleneck and personnel gathering risk warning information is output.

Images