A large-scale event venue-oriented distributed adaptive closed-loop crowd safety control method and system

Abstract

The application discloses a large-scale event venue-oriented distributed adaptive closed-loop crowd safety control method and system, belongs to the technical field of crowd safety monitoring and intelligent control, and is used for crowd safety control of a large-scale event venue. The method comprises the following steps: acquiring original physical signals of heterogeneous sensors and preprocessing the original physical signals, constructing an input function, and calculating an original prediction flow field; a linear extended state observer is used to calculate a corrected flow field, and a crush index is calculated; a regional comprehensive emergency degree is constructed, a balanced distribution vector of each region is constructed by using the comprehensive emergency degree, and a reconfiguration trigger criterion is set; an instantaneous regional evacuation and receiving capacity is calculated, a safety evacuation target function is constructed, and calculation and feedback are performed. Through construction of a differentiable closed-loop control framework with a normalized crush index as a core hub, real-time, robustness and collaboration of group behavior recognition and congestion early warning in a large-scale dense crowd venue are realized.

Description

Technical Field

[0001] This invention discloses a distributed adaptive closed-loop crowd safety control method and system for large event venues, belonging to the field of crowd safety monitoring and intelligent control technology. Background Technology

[0002] A sudden surge of dense crowds, especially if localized overcrowding or erratic movement occurs, can easily trigger pushing, falls, and even stampedes. Real-time monitoring of crowd behavior, early warning of risks, and organized evacuation accordingly are crucial for ensuring public safety. Existing technologies can be broadly categorized into three types, each with its own significant limitations.

[0003] The first category is crowd monitoring technology based on video analytics. This type of method acquires images through cameras and uses target detection or optical flow to estimate crowd density and movement trends. Its problems are: recognition accuracy drops sharply when there is severe high-density occlusion; real-time processing of multiple high-resolution videos from large venues requires extremely high computing power, making it difficult to complete in real time at the edge; and cameras are greatly affected by lighting, smoke, and blind spots, with single-modality cameras lacking reliability in sudden scenarios.

[0004] The second category is crowd dynamics prediction technology based on physical simulation. This type of method uses social force models and fluid dynamics equations to numerically simulate crowd flow, offering strong physical interpretability. However, its problems include: the computational cost of numerically solving partial differential equations is enormous, making it difficult to provide real-time predictions for large-scale venues within millisecond delays; model parameters rely on manual calibration, lacking online adaptability to real-world noise and sudden disturbances; and boundary conditions (such as exit opening / closing and channel guidance) must be re-solved once they change, making it unable to quickly respond to dynamic control commands.

[0005] The third category is evacuation scheduling techniques based on rules or centralized optimization. These methods pre-set fixed evacuation plans or perform global centralized optimization at a control center. The problems are: fixed plans cannot adapt to real-time changes in risk distribution; centralized optimization faces combinatorial explosion when there are many areas, and the solution speed cannot meet real-time requirements; independent decisions by each area are prone to conflict, for example, multiple areas simultaneously directing pedestrian flow to the same exit can create new congestion.

[0006] A deeper problem lies in the fact that the aforementioned stages are often fragmented: there is a lack of unified intermediate metrics and closed-loop feedback between the perception, prediction, decision-making, and execution modules. This manifests in four main deficiencies. First, perception data (ground pressure arrays, WiFi positioning, camera visual flow, etc.) contains varying degrees of noise and intermittent faults. Existing methods lack online correction mechanisms, allowing sensor errors to propagate directly and accumulate continuously in prediction. Second, the prediction model and observation correction are independent. Traditional extended state observers, while perturbation-resistant, are not differentiable and cannot be embedded in end-to-end training, preventing prediction errors from being corrected through a closed-loop feedback mechanism. Third, the computing power of edge computing nodes at venues is limited and typically statically allocated. During peak hours, high-risk areas are prone to prediction and decision-making delays due to insufficient computing power, causing early warnings to lose their timeliness. Fourth, the execution results of evacuation decisions (opening gates, closing gates, guiding diversions) cannot be fed back into the prediction model. The system cannot predict the evolution of the crowd flow "after measures are taken," making it difficult to determine the effectiveness of the decisions and to form a continuously optimizing closed loop. Summary of the Invention

[0007] The purpose of this invention is to provide a distributed adaptive closed-loop crowd safety control method and system for large event venues, in order to solve the problems in the existing technology where the perception, prediction, correction, decision-making and execution links are isolated from each other, errors accumulate, computing power allocation is rigid and the evacuation effect cannot be verified, resulting in insufficient real-time performance, robustness and coordination.

[0008] A distributed adaptive closed-loop crowd safety control method for large event venues includes: S1. Obtain the original physical signals from heterogeneous sensors and preprocess them to obtain the denoised and fused regional density field and velocity field. Construct an input function and use the input function and basis function to calculate the original predicted flow field. Transform the linear expansion state observer into a gain-learnable discrete differentiable form and calculate the corrected flow field. Use the corrected flow field to construct the squeezing exponent. Construct the total loss function and a two-stage training strategy. S2. Introduce risk change rate and latency margin, use squeeze index to construct regional comprehensive urgency and map it to computing power demand. Construct a game with each region as a follower and the reconfigurable computing array as the leader. Use comprehensive urgency to construct balanced allocation vector for each region. Construct reconfiguration trigger criteria based on balanced allocation vector and current actual allocation vector, and set continuity guarantee during switching period. S3. Calculate the instantaneous evacuation or acceptance capacity of the area based on the set of exits or passages of the activity venue, the effective width of the exit, and the free flow velocity of the exit, and calculate the acceptance margin. Construct a safe evacuation objective function based on the crowding index and the flow rate, and solve it using a distributed iterative auction algorithm. Construct execution instructions based on the solution results and the flow rate. Map the flow rate to a boundary condition function and feed it back to step S1 to construct the input function. After each iteration, calculate the sum of the global crowding indices as the global risk potential energy. If the global risk potential energy has not decreased compared to the previous round, revert to the previous round and recalculate the safe evacuation objective function.

[0009] S1 includes, S1.1, acquiring the original physical signals from the heterogeneous sensors, performing preprocessing, and obtaining the denoised and fused region density field. With velocity field ; Construct the input function: ; In the formula, The spatial coordinates of the sensor sampling points, Spatial location The input function, For the input function, For the noise-reducing density field, For denoising the velocity field, These are boundary condition functions; The DeepONet-based operator mapping outputs the original predicted flow field: ; In the formula, To predict the time offset, For the branch network to the input function The encoded result of the first One coefficient, The backbone network at the query point The output of the first basis functions The number of basis functions. For bias terms, For the query point The original predicted flow field several steps in the future. For the original predicted flow field several steps in the future, , To predict the density field, To predict the velocity field; calculate Physical residuals: ; In the formula, For physical residuals, This is the gradient operator.

[0010] S1 includes S1.2, which transforms the linearly extended state observer into a gain-learnable, discretely differentiable form, including the residual driving equation: ; In the formula, for Regional observation residuals for Measured flow field in the region for The estimated disturbance or compensation amount for the region at the current moment; Disturbance compensation update: ; In the formula, For the next time step disturbance estimation, For the long walk, It is a small, micro-sized neural observation device. Corrected flow field: ; In the formula, for Flow field after regional correction , for Density field after regional correction for Velocity field after regional correction; Define the regional compression index: ; ; In the formula, For the region exist The compression index at any moment, for Density field after time correction for The velocity field after time correction For variance, For the normalized extrusion function, For the region The compression index, This is the normalization constant; S1 includes S1.3, constructing the total loss function: ; In the formula, This is the total loss value. For the true value flow field, To correct the flow field, To correct the physical residuals of the flow field, To observe the residuals, These are the trainable parameters for ESO. Weights for each loss item; Constructing a two-stage training strategy: ; In the formula, Phase One, Phase Two For DeepONet parameters, The learning rate for DeepONet. For ESO learning rate, For data fitting loss, This represents the loss due to physical constraints.

[0011] S2 includes S2.1, which introduces risk change rate and time margin to construct a regional comprehensive urgency level: ; In the formula, For the region exist The overall urgency at any given moment, The rate of change of the extrusion index. For the region The remaining time margin of the current round of inference, To impose hard constraints on early warning latency, For the region Urgency adjustment item, Weights for trend and time delay terms; Mapping overall urgency to computing power requirements: ; In the formula, For the region Required PE quantity This is the lower limit for the number of PEs in a single zone. This is the maximum number of PEs in a single zone. This is the urgency normalization constant. This is for rounding down.

[0012] S2 includes S2.2, which is built based on comprehensive urgency. The utility function of the region is used to calculate the analytical solution of the optimal response of the followers; based on the comprehensive urgency and the analytical solution of the optimal response, the leader is constructed to maximize the total risk and return of the system, and the equilibrium unit computing power price is calculated. The equilibrium PE allocation for each region is calculated based on the comprehensive urgency and the equilibrium unit computing power price, resulting in the regional... Equal distribution of PE number ; S2 includes S2.3, and the criteria for triggering reconfiguration: ; In the formula, Assign vectors to all PEs. , Assign a vector to the currently executing processing unit. It is an L1 norm. To reconfigure the dead zone threshold; Continuity assurance during the switchover period: ; In the formula, For a flow field that is continuously output to the outside, To switch the indicator gating, The output of the normal inference model is configured with the target number of processing units. Output a simplified standby inference model that requires the least amount of processing units; Switch indicator gating: ; In the formula, To switch the start time, This refers to the time for loading the bitstream in a single DPR cycle.

[0013] S3 includes, S3.1, the instantaneous evacuation or reception capacity of the calculation area: ; In the formula, For the region The ability to evacuate or receive people instantly. For the region A collection of exits or passages. For export The effective width, For export free flow velocity, This is the export efficiency coefficient. In gate state, For indicator functions; Calculate the acceptable margin: ; In the formula, For the region Acceptable margin This indicates that a non-negative value is taken.

[0014] S3 includes S3.2, constructing the objective function for safe evacuation. : ; In the formula, For congested areas Guided to the health area foot traffic rate for Regional compression index Congested area Guided to the health area The transfer cost; right Set constraints: ; ; In the formula, For the region Maximum possible flow of people for Regional acceptance margin; Solve using a distributed iterative approach with an auction algorithm: ; In the formula, For the receiving party The shadow price The step size.

[0015] S3 includes S3.3, and the build execution instructions: ; ; In the formula, For exit or gate opening / closing commands. For the region The main guiding direction. For export The corresponding receiving area To set the threshold, The threshold value is the closing threshold. Will The inference load required for the next round of execution is fed back to step S2.1 as an urgency adjustment term. : ; S3 includes S3.4, constructing mapping rules and feeding them back to step S1.1, for the exit in the open state: ; In the formula, The exit normal unit vector, The density field at the boundary; For channels that are closed: ; ; In the formula, The baseline density set at the boundary The velocity vector at the boundary; For steering instructions: ; In the formula, For the region The applied flow source term, For boundary or region spatial coordinates, For the receiving area Representative coordinates, It is the Sigmoid smoothing function. This is the steepness coefficient. Let be the Dirac impulse function.

[0016] S3 includes S3.5, which defines the global risk potential: ; In the formula, for The global risk potential energy at any given moment Set a closed-loop stability criterion: ; In the formula, This represents the overall risk potential following the implementation of the strategy in this round. This represents the risk potential following the previous strategy. This represents the minimum effective descent.

[0017] A distributed adaptive closed-loop crowd safety control system for large event venues includes: The sensing and differentiable correction prediction module is used to execute step S1, acquire signals from multiple heterogeneous sensors and fuse them to generate regional density and velocity fields, input the current density field, velocity field and boundary conditions into the prediction model to obtain the original predicted flow field, and use the differentiable observer to drive perturbation compensation with the residual between the measured flow field and the original predicted flow field to obtain the corrected flow field. Then, the product of the corrected density and velocity turbulence is used as the regional risk measure, and the normalized squeezing index is output after normalization. The resource game allocation module is used to execute step S2. It constructs a comprehensive urgency based on the normalized squeeze index and its changing trend of each region and maps it to computing resource demand. It determines the computing resource allocation of each region based on the game between each region and the reconfigurable computing array. When the allocation deviates from the current allocation, it triggers the dynamic reconfiguration of the reconfigurable computing array. During the reconfiguration, the prediction output is maintained by the backup computing model. The evacuation decision-making and closed-loop feedback module is used to execute step S3. It calculates the acceptability margin based on the physical evacuation capacity and normalized congestion index of each area. It optimizes the transfer of people from areas with high normalized congestion index to areas with low normalized congestion index by means of outflow constraints and acceptance constraints. It solves the joint evacuation strategy through distributed auction. It generates physical execution instructions based on the strategy and outputs them to the field equipment. At the same time, it maps the flow rate of people in the strategy into boundary condition functions and feeds them back to the prediction model input of the sensing and differentiable correction prediction module to form closed-loop control. This module is also used to calculate the sum of the normalized congestion index of the whole field as the global risk potential after each control cycle. If the risk potential of the current round does not achieve an effective decrease compared with the previous round, it reverts to the evacuation strategy of the previous round.

[0018] Compared to existing technologies, this invention has the following advantages: By constructing a differentiable closed-loop control architecture with the normalized squeeze index as its core, this invention connects multi-source heterogeneous sensing fusion, physical constraint deep operator prediction, online correction of differentiable extended state observers, risk-driven Stackelberg game computing power scheduling, and bilateral auction evacuation decision-making into an end-to-end closed-loop adaptive system. Compared to existing technologies where sensing, prediction, decision-making, and execution are fragmented, this invention, on the one hand, achieves joint optimization by backpropagating the prediction residuals to the prediction model through a differentiable observer, fundamentally addressing the issue. It solves the problem of prediction errors accumulating continuously due to the inability to close the loop for correction; on the other hand, it enables computing resources to flow dynamically with the risk of the crowd through risk-driven computing power game, avoiding the defect of delayed early warning in high-risk areas caused by static computing power allocation; furthermore, it feeds back the prediction model with the evacuation decision results in the form of differentiable boundary conditions, enabling the system to predict the effect of measures and ensure stable convergence with the monotonically decreasing global risk potential energy as the criterion. Thus, in the real venue environment where sensor noise, local faults and sudden disturbances coexist, it significantly improves the real-time performance, robustness and coordination of dense crowd behavior recognition and congestion early warning. Attached Figure Description

[0019] Figure 1 This is a flowchart of the technology of this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] A distributed adaptive closed-loop crowd safety control method for large event venues includes: S1. Obtain the original physical signals from heterogeneous sensors and preprocess them to obtain the denoised and fused regional density field and velocity field. Construct an input function and use the input function and basis function to calculate the original predicted flow field. Transform the linear expansion state observer into a gain-learnable discrete differentiable form and calculate the corrected flow field. Use the corrected flow field to construct the squeezing exponent. Construct the total loss function and a two-stage training strategy. S2. Introduce risk change rate and latency margin, use squeeze index to construct regional comprehensive urgency and map it to computing power demand. Construct a game with each region as a follower and the reconfigurable computing array as the leader. Use comprehensive urgency to construct balanced allocation vector for each region. Construct reconfiguration trigger criteria based on balanced allocation vector and current actual allocation vector, and set continuity guarantee during switching period. S3. Calculate the instantaneous evacuation or acceptance capacity of the area based on the set of exits or passages of the activity venue, the effective width of the exit, and the free flow velocity of the exit, and calculate the acceptance margin. Construct a safe evacuation objective function based on the crowding index and the flow rate, and solve it using a distributed iterative auction algorithm. Construct execution instructions based on the solution results and the flow rate. Map the flow rate to a boundary condition function and feed it back to step S1 to construct the input function. After each iteration, calculate the sum of the global crowding indices as the global risk potential energy. If the global risk potential energy has not decreased compared to the previous round, revert to the previous round and recalculate the safe evacuation objective function.

[0022] S1 includes, S1.1, acquiring the original physical signals from the heterogeneous sensors, performing preprocessing, and obtaining the denoised and fused region density field. With velocity field ; Construct the input function: ; In the formula, The spatial coordinates of the sensor sampling points, Spatial location The input function, For the input function, For the noise-reducing density field, For denoising the velocity field, These are boundary condition functions; The DeepONet-based operator mapping outputs the original predicted flow field: ; In the formula, To predict the time offset, For the branch network to the input function The encoded result of the first One coefficient, The backbone network at the query point The output of the first basis functions The number of basis functions. For bias terms, For the query point The original predicted flow field several steps in the future. For the original predicted flow field several steps in the future, , To predict the density field, To predict the velocity field; calculate Physical residuals: ; In the formula, For physical residuals, This is the gradient operator.

[0023] S1 includes S1.2, which transforms the linearly extended state observer into a gain-learnable, discretely differentiable form, including the residual driving equation: ; In the formula, for Regional observation residuals for Measured flow field in the region for The estimated disturbance or compensation amount for the region at the current moment; Disturbance compensation update: ; In the formula, For the next time step disturbance estimation, For the long walk, It is a small, micro-sized neural observation device. Corrected flow field: ; In the formula, for Flow field after regional correction , for Density field after regional correction for Velocity field after regional correction; Define the regional compression index: ; ; In the formula, For the region exist The compression index at any moment, for Density field after time correction for The velocity field after time correction For variance, For the normalized extrusion function, For the region The compression index, This is the normalization constant; S1 includes S1.3, constructing the total loss function: ; In the formula, This is the total loss value. For the true value flow field, To correct the flow field, To correct the physical residuals of the flow field, To observe the residuals, These are the trainable parameters for ESO. Weights for each loss item; Constructing a two-stage training strategy: ; In the formula, Phase One, Phase Two For DeepONet parameters, The learning rate for DeepONet. For ESO learning rate, For data fitting loss, This represents the loss due to physical constraints.

[0024] S2 includes S2.1, which introduces risk change rate and time margin to construct a regional comprehensive urgency level: ; In the formula, For the region exist The overall urgency at any given moment, The rate of change of the extrusion index. For the region The remaining time margin of the current round of inference, To impose hard constraints on early warning latency, For the region Urgency adjustment item, Weights for trend and time delay terms; Mapping overall urgency to computing power requirements: ; In the formula, For the region Required PE quantity This is the lower limit for the number of PEs in a single zone. This is the maximum number of PEs in a single zone. This is the urgency normalization constant. This is for rounding down.

[0025] S2 includes S2.2, which is built based on comprehensive urgency. The utility function of the region is used to calculate the analytical solution of the optimal response of the followers; based on the comprehensive urgency and the analytical solution of the optimal response, the leader is constructed to maximize the total risk and return of the system, and the equilibrium unit computing power price is calculated. The equilibrium PE allocation for each region is calculated based on the comprehensive urgency and the equilibrium unit computing power price, resulting in the regional... Equal distribution of PE number ; S2 includes S2.3, and the criteria for triggering reconfiguration: ; In the formula, Assign vectors to all PEs. , Assign a vector to the currently executing processing unit. It is an L1 norm. To reconfigure the dead zone threshold; Continuity assurance during the switchover period: ; In the formula, For a flow field that is continuously output to the outside, To switch the indicator gating, The output of the normal inference model is configured with the target number of processing units. Output a simplified standby inference model that requires the least amount of processing units; Switch indicator gating: ; In the formula, To switch the start time, This refers to the time for loading the bitstream in a single DPR cycle.

[0026] S3 includes, S3.1, the instantaneous evacuation or reception capacity of the calculation area: ; In the formula, For the region The ability to evacuate or receive people instantly. For the region A collection of exits or passages. For export The effective width, For export free flow velocity, This is the export efficiency coefficient. In gate state, For indicator functions; Calculate the acceptable margin: ; In the formula, For the region Acceptable margin This indicates that a non-negative value is taken.

[0027] S3 includes S3.2, constructing the objective function for safe evacuation. : ; In the formula, For congested areas Guided to the health area foot traffic rate for Regional compression index Congested area Guided to the health area The transfer cost; right Set constraints: ; ; In the formula, For the region Maximum possible flow of people for Regional acceptance margin; Solve using a distributed iterative approach with an auction algorithm: ; In the formula, For the receiving party The shadow price The step size.

[0028] S3 includes S3.3, and the build execution instructions: ; ; In the formula, For exit or gate opening / closing commands. For the region The main guiding direction. For export The corresponding receiving area To set the threshold, The threshold value is the closing threshold. Will The inference load required for the next round of execution is fed back to step S2.1 as an urgency adjustment term. : ; S3 includes S3.4, constructing mapping rules and feeding them back to step S1.1, for the exit in the open state: ; In the formula, The exit normal unit vector, The density field at the boundary; For channels that are closed: ; ; In the formula, The baseline density set at the boundary The velocity vector at the boundary; For steering instructions: ; In the formula, For the region The applied flow source term, For boundary or region spatial coordinates, For the receiving area Representative coordinates, It is the Sigmoid smoothing function. This is the steepness coefficient. Let be the Dirac impulse function.

[0029] S3 includes S3.5, which defines the global risk potential: ; In the formula, for The global risk potential energy at any given moment Set a closed-loop stability criterion: ; In the formula, This represents the overall risk potential following the implementation of the strategy in this round. This represents the risk potential following the previous strategy. This represents the minimum effective descent.

[0030] A distributed adaptive closed-loop crowd safety control system for large event venues includes: The sensing and differentiable correction prediction module is used to execute step S1, acquire signals from multiple heterogeneous sensors and fuse them to generate regional density and velocity fields, input the current density field, velocity field and boundary conditions into the prediction model to obtain the original predicted flow field, and use the differentiable observer to drive perturbation compensation with the residual between the measured flow field and the original predicted flow field to obtain the corrected flow field. Then, the product of the corrected density and velocity turbulence is used as the regional risk measure, and the normalized squeezing index is output after normalization. The resource game allocation module is used to execute step S2. It constructs a comprehensive urgency based on the normalized squeeze index and its changing trend of each region and maps it to computing resource demand. It determines the computing resource allocation of each region based on the game between each region and the reconfigurable computing array. When the allocation deviates from the current allocation, it triggers the dynamic reconfiguration of the reconfigurable computing array. During the reconfiguration, the prediction output is maintained by the backup computing model. The evacuation decision-making and closed-loop feedback module is used to execute step S3. It calculates the acceptability margin based on the physical evacuation capacity and normalized congestion index of each area. It optimizes the transfer of people from areas with high normalized congestion index to areas with low normalized congestion index by means of outflow constraints and acceptance constraints. It solves the joint evacuation strategy through distributed auction. It generates physical execution instructions based on the strategy and outputs them to the field equipment. At the same time, it maps the flow rate of people in the strategy into boundary condition functions and feeds them back to the prediction model input of the sensing and differentiable correction prediction module to form closed-loop control. This module is also used to calculate the sum of the normalized congestion index of the whole field as the global risk potential after each control cycle. If the risk potential of the current round does not achieve an effective decrease compared with the previous round, it reverts to the evacuation strategy of the previous round.

[0031] The following description, in conjunction with the accompanying drawings and embodiments, further illustrates the process of this invention. Figure 1 As shown, the system first collects raw data from various areas within the venue using multi-source sensor signals. After processing by the S1 perception and correction prediction module, a normalized crowding index is output as the core risk measure. Subsequently, this index is driven in parallel by two paths: one path is sent to the S2 dynamic computing power allocation module to achieve risk-driven allocation of computing resources, and the other path is sent to the S3 game-theoretic evacuation decision module to generate a joint evacuation strategy. At the same time, S2 provides necessary computing resources to S3 to ensure that the inference load of evacuation decision-making can be completed in a timely manner. After generating the evacuation strategy, S3 outputs physical execution instructions to the field equipment and maps the flow rate of people in the strategy to a boundary condition function, which is then fed back to the input of the prediction model of S1 to complete the closed loop. After each iteration, it is determined whether the global risk potential energy has decreased. If it has decreased, the next cycle continues; if it has not decreased, it reverts to the previous safety strategy and re-executes S3.

[0032] The first step in this embodiment of the invention is to obtain the original physical signals from heterogeneous sensors through a complete differentiable link from the multimodal raw signals to the intermediate variable "compression index," and to perform preprocessing, including converting the pressure array into a region density field. ; in For the first A block of pressure floor tiles Force reading at time (input, unit N); The average weight per person is a constant (the value is set at 60 to 70 kg based on the statistics of people in the venue). acceleration due to gravity ; For the region The physical area; Belongs to the region A set of tile numbers; The population density of the region estimated by the pressure channel (output, in people / m²). Visual flow velocity vector field (Horn–Schunck): ; in The gray-level gradient of the image in the spatial and temporal directions (input); Pixel-level velocity field (output); To smooth out regularization weights (value selection logic: larger values ​​are used in densely populated scenes to suppress local jitter); Butterworth low-pass filter for noise reduction: ; in The signal angular frequency; The cutoff angular frequency (value logic: set to the highest effective frequency of crowd behavior, approximately 0.5 to 1 Hz). The filter order (value selection logic:) (Taking into account both steepness and phase distortion); the output is the denoised version. and ; Multimodal reliability weighted fusion: ; ; ; in It is divided into three modalities: pressure, WiFi positioning, and camera; For modality In the region Near-window observation variance (input, as an indication of fault or noise); The reliability weights are outputs; if the failure mode variance is large, the weights automatically approach zero to achieve soft switching. The sensitivity coefficient is denoted as .

[0033] The mapping from "current state function + boundary condition function" to "future flow field function" is modeled as operator G. The input function is: ; in This is the boundary condition function re-fed to the game theory module in Chapter 3 (set to the default open state in the first round). DeepONet output is: ; in To the sensor sampling point The input function (input) at the location; Spatial coordinates; For the predicted time offset; The k-th coefficient is obtained by encoding the input function in the BranchNet (multimodal signals are encoded by CNN or MLP). At the query point of the backbone network (TrunkNet) The k-th basis function output at point k; p is the number of basis functions (value range: 50 to 100, the larger the value, the stronger the expressive power but the higher the computational power); The original predicted flow field for several future steps (output); Physical constraints (continuity or mass conservation residuals) are written as soft constraints: ; Theoretically, the population is conserved. This serves as a penalty for non-physical predictions as a loss term.

[0034] The classic linear ESO (which exists and can be represented in continuous form) unifies "unmodeled dynamics + sensor failure + sudden disturbances" into an extended state; ; ; in For the estimation of true density; This is an estimate of the total disturbance. For observer gain; Differentiable modification of the present invention: replacing the fixed gain with a trainable gain function. Discrete residual correction is performed based on DeepONet predictions: ; ; The corrected flow field is as follows: ; in The measured flow field (input); This is the original prediction (input); The residual signal between the measured value and the "prediction + compensation" (an intermediate variable, whose sole source is the difference between the outputs of the previous two steps, and the chain is clear); For small, micro-sized neural observers (parameters) Output disturbance compensation increment; For a long walk; This is the disturbance estimate or compensation amount (output, the propagation loss function is used for backpropagation). This is for the corrected flow field.

[0035] Hazard quantification based on fluid pressure (Helbing crowd pressure modified formula): ; ; in For the corrected density and velocity fields (input, source is unique and clearly defined); The variance of the velocity vector within the region characterizes the precursory signs of a "crowded and chaotic" stampede. The output is the regional congestion index or congestion risk level. This is a normalization constant (calibrated according to historical trampling thresholds); Normalized compression index (output, 0 for safe, 1 for extremely dangerous).

[0036] The total loss is: ; in For the true flow field (training labels, from simulation or historical playback); to Loss weights; data term constraints to approximate the true value using "prediction + compensation"; physical terms; ensuring conservation; observation consistency terms to suppress residuals. ; ESO parameters for final regularization ; Training strategy (to resolve gradient coupling oscillations): ; Phase I freezes ESO (order) Separate pre-trained DeepONet parameters Phase II thawing Joint fine-tuning, DeepONet learning rate Much smaller than the ESO learning rate This avoids error oscillations caused by the coupling of two targets.

[0037] Key variable handover from the first step: Corrected flow field Disturbance compensation Normalized compression index .in It will serve as the common input for both the second and third steps.

[0038] The second step follows the output of the first step: the normalized extrusion index. This paper presents a complete chain from "risk to bidding to pricing to PE resource reallocation." Each region, as a "follower" in the game, must first transform abstract risks into bidding-based computing power demands. This invention does not directly use... Instead, it introduces the risk change rate and time margin to construct the urgency level. : ; Mapping urgency to computing power requirements (number of PEs required): ; in The normalized extrusion index (input, with a unique and clearly defined source); To squeeze the rate of change of the index and depict the trend of "deteriorating" risk (taking... (Only included in the upward count), used to preemptively secure computing power; For the region The remaining latency margin of the previous round of DeepONet or ESO inference; A hard constraint is set for the warning delay (e.g., 100ms); Weights for trend and time delay terms (value selection logic: critical venues) (Take a broad approach to enhance foresight). For the overall urgency of the region (intermediate variable or output); This is the normalization constant; For the region Required PE quantity (output, serving as a demand benchmark for follower bids); These are the upper and lower limits of PE for a single zone.

[0039] area The price per unit of computing power given by CGRA Next, decide on your own bid (i.e., request PE ratio). Maximize utility: ; By first-order optimality conditions Obtain the follower's optimal response (analytical solution): ; in For the region The number of PEs requested (decision variables or outputs); The unit computing power price set for CGRA (input, from step 2.3); The urgency level (input) serves as the marginal weight for risk and return; the higher the risk, the more PE ratio is required to "bid". To prevent diminishing marginal returns and to prevent a single region from infinitely demanding computing power; Indicates truncation to interval The best bid for the follower.

[0040] As the "leader," CGRA anticipates the optimal response from all its followers. In total PE resources Setting prices under constraints to maximize the system's total risk-reward ratio (equivalent to directing computing power to high-risk areas): ; The optimal price satisfies the condition of exactly clearing resources (KKT condition, taking equal-rate when resources are scarce): ; in To balance the unit computing power price (output, feedback, and balanced distribution are achieved). The total number of schedulable PEs for CGRA nodes (input, hardware constant); this formula is about The monotonic equation can be solved in milliseconds using the bisection method. .

[0041] Will Substituting the follower response, we obtain the balanced PE distribution across regions: ; ; in Output the number of PEs evenly distributed in region k. Assign vectors to all PEs (output, as the target configuration for the DPR reconfiguration in step 2.5); This represents the total number of regions. In nature... Constantly satisfying (resources are not over-issued), and The bigger The larger (the risk priority can be proven).

[0042] When the target configuration With current configuration Reconfiguration is triggered when the difference exceeds a threshold. To ensure that "no warning signal is lost during handover," a ping-pong dual-zone system with low-precision backup is used. The reconfiguration trigger criterion is as follows: ; in The reconfiguration dead zone threshold is set (value logic: if it is too small, frequent reconfigurations will increase overhead; if it is too large, the response will be sluggish; it is calibrated according to the single DPR overhead). Continuity assurance during handover (core): Handover area New position flow During loading, a low-precision backup model is used. Take over the output, and seamlessly switch back after the switch is complete; the total effective inference output is: ; ; in Output for a fully configured DeepONet or ESO instance; Output a resident low-precision backup model (ensuring no empty windows); To switch the indicator gating (input provided by the configuration controller); The time for a single DPR bitstream loading (in milliseconds); For the continuous output flow field (output, backfeeding Chapter 1, next round) or (Channel, closed loop, continuous flow); Throughput Guarantee After Reconfiguration: High-Risk Area After one PE, its inference frequency increases by approximately This will ensure that the next round of DeepONet or ESO in this region is [times], thus guaranteeing [performance]. Completed within the specified time.

[0043] The key variable handover from the second step output: urgency. Equilibrium price PE allocation vector Continuous output during switching period .in Determine the upper limit of computing power in each region (constraining the executable computing power of the third step of the game). Sharing the same risk source with the third step This ensures that the allocation of computing power is consistent with the risks of evacuation decisions.

[0044] The third step follows from the first step: normalized extrusion index. Computing power allocation vector in Chapter 2 This completes the closed loop of "risk quantification, joint evacuation game, physical execution command, and boundary condition reinjection".

[0045] Before engaging in the game, the physical evacuation capacity of each area needs to be quantified into a tradable "acceptance capacity": ; in For the region A collection of exits or passageways; For export Effective width (input, units) ); For export Free flow velocity (input); This is the export efficiency coefficient (congestion reduction, with a logical value of 0.6 to 0.9). The gate state (input, and also a subsequent decision variable); For indicator functions; For the region Instantaneous evacuation or reception capacity (output, in person / s); Transfer margin (how many outsiders can the health zone accommodate): ; in For the region Acceptable margin (output), risk The lower the threshold and the greater the traffic capacity, the higher the margin. This indicates taking the non-negative form.

[0046] Congested areas (high) As the "requester," the health zone (high-risk area) issued a transfer request. As "receiving parties," they bid to accept the flow of people. Solve for the joint evacuation decision variables. Depend on Towards (Transfer of people) ; The constraints are: ; ; in For congested areas Guided to the health area Passenger flow rate (decision variable or output); The output of Chapter 1 is the two-zone compression index (input, with a unique and clearly defined source); for Transfer cost (path length or congested route, value logic: longer routes or routes through high-risk areas are more costly). For the region Maximum possible flow of people; For acceptance margin (input); The globally optimal joint evacuation strategy (output); This problem is a linear programming or assignable problem, solved using the Auction Algorithm in a distributed iterative manner, with the price updated as follows: ; in For the receiving party The shadow price; The step size is used. When supply and demand are out of balance, the price adjusts adaptively, iterating until the bidding converges in each region.

[0047] Continuous evacuation strategy Discretized into three types of outputs, the physical world executes instructions (thresholding): ; ; in For output or gate opening / closing commands (output, sent to audio-visual signs or broadcasts); For the region The main guiding direction (output); Threshold for determining whether to open or close (with dead zone to prevent instruction jitter); CGRA computing power bidding signal: The next round of inference load required for execution is fed back as the urgency adjustment term of the second step (the area to be evacuated needs to be predicted and verified more frequently), which closes the "decision-to-computing power" branch.

[0048] The discrete evacuation instructions are mapped to continuous boundary condition functions that DeepONet can accept in the first step. This enables decision-making and injection prediction. The mapping rules include an open outlet transitioning to high flow velocity + low resistance boundary (Neumann type): ; Closing the channel to zero flow boundary (Dirichlet type): ; ; Guided redirection to traffic source item injection: ; in For boundary or region spatial coordinates; For export direction; For the free flow velocity at the outlet; Human flow rate (input, determines boundary strength, source is unique and definite); It is the Sigmoid smoothing function; This is the steepness coefficient. The key is to use smoothing. Instead of a hard step, it makes right Differentiable, the decision gradient can be backpropagated to the game parameters; The boundary resistance coefficient decreases exponentially with the pedestrian flow rate when the area is open. As the reference resistance; The guiding direction is the unit vector; For traffic source items; The output is the DeepONet input function. Boundary components in the closed-loop recharge.

[0049] Define the overall risk potential and require the closed loop to decrease monotonically in each round to prevent "the more sparse, the more chaotic": ; in For global risk potential (input from each region) Summation); if in a certain round Unused If the risk decreases, the strategy will revert to the previous round of safety measures and be re-auctioned to avoid oscillations and divergence. The minimum effective decrease is preferably set at 1% to 5% of the current global risk potential energy. The specific value will be determined through simulation experiments based on the venue size and sensor refresh rate.

[0050] The overall system forms a closed loop based on the following data and control flows: The original multimodal signals from the sensors are denoised and reconstructed into density and velocity fields; PI-DeepONet receives the current density and velocity fields and the boundary conditions from the previous recharge cycle. Output the original predicted flow field }; Differentiable ESO: Comparison of measured and predicted residuals Output disturbance compensation Generate the corrected flow field Based on this, the core intermediate variable, the squeezing index, is quantified. .

[0051] Compression Index Divide into two routes: one route is switched to emergency level. The Stackelberg game is used to obtain the PE allocation vector. Then, DPR zero-interruption reconfiguration ensures computing power in high-risk areas; another approach uses capacity modeling and auction game to obtain a joint evacuation strategy. Discretization into physical execution instructions With guiding direction Finally, the evacuation command is differentiable and mapped to a boundary condition function. The reinjection of PI-DeepONe triggers the next round of predictions, and is driven by global risk potential. The criterion for monotonic decline is used for verification.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed adaptive closed-loop crowd safety control method for large event venues, characterized in that, include: S1. Obtain the original physical signals from heterogeneous sensors and preprocess them to obtain the denoised and fused regional density field and velocity field. Construct an input function and use the input function and basis function to calculate the original predicted flow field. Transform the linear expansion state observer into a gain-learnable discrete differentiable form and calculate the corrected flow field. Use the corrected flow field to construct the squeezing exponent. Construct the total loss function and a two-stage training strategy. S2. Introducing risk change rate and latency margin, a regional comprehensive urgency is constructed using a squeeze index and mapped to computing power demand. A game is constructed with each region as a follower and a reconfigurable computing array as the leader. An equilibrium allocation vector for each region is then built using the comprehensive urgency. Based on... The reconfiguration trigger criterion is constructed by comparing the balanced allocation vector with the current actual allocation vector, and continuity is guaranteed during the switching period; S3. Calculate the instantaneous evacuation or acceptance capacity of the area based on the set of exits or passages of the activity venue, the effective width of the exit, and the free flow velocity of the exit, and calculate the acceptance margin. Construct a safe evacuation objective function based on the crowding index and the flow rate, and solve it using a distributed iterative auction algorithm. Construct execution instructions based on the solution results and the flow rate. Map the flow rate to a boundary condition function and feed it back to step S1 to construct the input function. After each iteration, calculate the sum of the global crowding indices as the global risk potential energy. If the global risk potential energy has not decreased compared to the previous round, revert to the previous round and recalculate the safe evacuation objective function.

2. The distributed adaptive closed-loop crowd safety control method for large event venues according to claim 1, characterized in that, S1 includes, S1.1, acquiring the original physical signals from the heterogeneous sensors, performing preprocessing, and obtaining the denoised and fused region density field. With velocity field ; Construct the input function: ； In the formula, The spatial coordinates of the sensor sampling points, Spatial location The input function, For the input function, For the noise-reducing density field, For denoising the velocity field, These are boundary condition functions; The DeepONet-based operator mapping outputs the original predicted flow field: ； In the formula, To predict the time offset, For the branch network to the input function The encoded result of the first One coefficient, The backbone network at the query point The output of the first basis functions The number of basis functions. For bias terms, For the query point The original predicted flow field several steps in the future. For the original predicted flow field several steps in the future, , To predict the density field, To predict the velocity field; calculate Physical residuals: ； In the formula, For physical residuals, This is the gradient operator.

3. The distributed adaptive closed-loop crowd safety control method for large event venues according to claim 2, characterized in that, S1 includes S1.2, which transforms the linearly extended state observer into a gain-learnable, discretely differentiable form, including the residual driving equation: ； In the formula, for Regional observation residuals for Measured flow field in the region for The estimated disturbance or compensation amount for the region at the current moment; Disturbance compensation update: ； In the formula, For the next time step disturbance estimation, For the long walk, It is a small, micro-sized neural observation device. Corrected flow field: ； In the formula, for Flow field after regional correction , for Density field after regional correction for Velocity field after regional correction; Define the regional compression index: ； ； In the formula, For the region exist The compression index at any moment, for Density field after time correction for The velocity field after time correction For variance, For the normalized extrusion function, For the region The compression index, This is the normalization constant; S1 includes S1.3, constructing the total loss function: ； In the formula, This is the total loss value. For the true value flow field, To correct the flow field, To correct the physical residuals of the flow field, To observe the residuals, These are the trainable parameters for ESO. Weights for each loss item; Constructing a two-stage training strategy: ； In the formula, Phase One, Phase Two For DeepONet parameters, The learning rate for DeepONet. For ESO learning rate, For data fitting loss, This represents the loss due to physical constraints.

4. A distributed adaptive closed-loop crowd safety control method for large event venues according to claim 3, characterized in that, S2 includes S2.1, which introduces risk change rate and time margin to construct a regional comprehensive urgency level: ； In the formula, For the region exist The overall urgency at any given moment, The rate of change of the extrusion index. For the region The remaining time margin of the current round of inference, To impose hard constraints on early warning latency, For the region Urgency adjustment item, Weights for trend and time delay terms; Mapping overall urgency to computing power requirements: ； In the formula, For the region Required PE quantity This is the lower limit for the number of PEs in a single zone. This is the maximum number of PEs in a single zone. This is the urgency normalization constant. This is for rounding down.

5. A distributed adaptive closed-loop crowd safety control method for large event venues according to claim 4, characterized in that, S2 includes S2.2, which is built based on comprehensive urgency. The utility function of the region is used to calculate the analytical solution of the optimal response of the follower. Based on the comprehensive urgency and the optimal response analytical solution, a leader is constructed to maximize the total risk and return of the system, and the equilibrium unit computing power price is calculated. The equilibrium PE allocation for each region is calculated based on the comprehensive urgency and the equilibrium unit computing power price, resulting in the regional... Equal distribution of PE number ; S2 includes S2.3, and the criteria for triggering reconfiguration: ； In the formula, Assign vectors to all PEs. , Assign a vector to the currently executing processing unit. It is an L1 norm. To reconfigure the dead zone threshold; Continuity assurance during the switchover period: ； In the formula, For a flow field that is continuously output to the outside, To switch the indicator gating, The output of the normal inference model is configured with the target number of processing units. Output a simplified standby inference model that requires the least amount of processing units; Switch indicator gating: ； In the formula, To switch the start time, This refers to the time for loading the bitstream in a single DPR cycle.

6. A distributed adaptive closed-loop crowd safety control method for large event venues according to claim 5, characterized in that, S3 includes, S3.1, the instantaneous evacuation or reception capacity of the calculation area: ； In the formula, For the region The ability to evacuate or receive people instantly. For the region A collection of exits or passages. For export The effective width, For export free flow velocity, This is the export efficiency coefficient. In gate state, For indicator functions; Calculate the acceptable margin: ； In the formula, For the region Acceptable margin This indicates that a non-negative value is taken.

7. A distributed adaptive closed-loop crowd safety control method for large event venues according to claim 6, characterized in that, S3 includes S3.2, constructing the objective function for safe evacuation. : ； In the formula, For congested areas Guided to the health area foot traffic rate for Regional compression index Congested area Guided to the health area The transfer cost; right Set constraints: ； ； In the formula, For the region Maximum possible flow of people for Regional acceptance margin; Solve using a distributed iterative approach with an auction algorithm: ； In the formula, For the receiving party The shadow price The step size.

8. A distributed adaptive closed-loop crowd safety control method for large event venues according to claim 7, characterized in that, S3 includes S3.3, and the build execution instructions: ； ； In the formula, For exit or gate opening / closing commands. For the region The main guiding direction. For export The corresponding receiving area To set the threshold, The threshold value is the closing threshold. Will The inference load required for the next round of execution is fed back to step S2.1 as an urgency adjustment term. : ； S3 includes S3.4, constructing mapping rules and feeding them back to step S1.1, for the exit in the open state: ； In the formula, The exit normal unit vector, The density field at the boundary; For channels that are closed: ； ； In the formula, The baseline density set at the boundary The velocity vector at the boundary; For steering instructions: ； In the formula, For the region The applied flow source term, For boundary or region spatial coordinates, For the receiving area Representative coordinates, It is the Sigmoid smoothing function. This is the steepness coefficient. Let be the Dirac impulse function.

9. A distributed adaptive closed-loop crowd safety control method for large event venues according to claim 8, characterized in that, S3 includes S3.5, which defines the global risk potential: ； In the formula, for The global risk potential energy at any given moment Set a closed-loop stability criterion: ； In the formula, This represents the overall risk potential following the implementation of the strategy in this round. This represents the risk potential following the previous strategy. This represents the minimum effective descent.

10. A distributed adaptive closed-loop crowd safety control system for large event venues, characterized in that, include: The sensing and differentiable correction prediction module is used to execute step S1, acquire signals from multiple heterogeneous sensors and fuse them to generate regional density and velocity fields, input the current density field, velocity field and boundary conditions into the prediction model to obtain the original predicted flow field, and use the differentiable observer to drive perturbation compensation with the residual between the measured flow field and the original predicted flow field to obtain the corrected flow field. Then, the product of the corrected density and velocity turbulence is used as the regional risk measure, and the normalized squeezing index is output after normalization. The resource game allocation module is used to execute step S2. It constructs a comprehensive urgency based on the normalized squeeze index and its changing trend of each region and maps it to computing resource demand. It determines the computing resource allocation of each region based on the game between each region and the reconfigurable computing array. When the allocation deviates from the current allocation, it triggers the dynamic reconfiguration of the reconfigurable computing array. During the reconfiguration, the prediction output is maintained by the backup computing model. The evacuation decision-making and closed-loop feedback module is used to execute step S3. It calculates the acceptability margin based on the physical evacuation capacity and normalized congestion index of each area. It optimizes the transfer of people from areas with high normalized congestion index to areas with low normalized congestion index by means of outflow constraints and acceptance constraints. It solves the joint evacuation strategy through distributed auction. It generates physical execution instructions based on the strategy and outputs them to the field equipment. At the same time, it maps the flow rate of people in the strategy into boundary condition functions and feeds them back to the prediction model input of the sensing and differentiable correction prediction module to form closed-loop control. This module is also used to calculate the sum of the normalized congestion index of the whole field as the global risk potential after each control cycle. If the risk potential of the current round does not achieve an effective decrease compared with the previous round, it reverts to the evacuation strategy of the previous round.

Images