Safety diversion control method for smart park

Abstract

The invention belongs to the technical field of smart city safety management, and discloses a smart park safety shunting control method, which comprises the steps of collecting data through a multi-modal sensor network to construct a space-time tensor, predicting people flow distribution through tensor decomposition and calculating a residual error, generating a path strategy by adopting a multi-agent game, and performing intelligent park safety shunting control. A shunting instruction is dynamically adjusted in combination with Lyapunov optimization, and closed-loop control is realized by cooperatively updating model parameters through residual feedback; the system comprises a multi-modal sensor network module, a spatio-temporal data fusion module, a tensor decomposition and prediction module, a multi-agent game planning module, a Lyapunov optimization control module and a parameter collaborative updating module. According to the method, the space-time tensor is constructed through multi-modal sensing and data fusion, people flow prediction is realized in combination with tensor decomposition, individual and system targets are balanced and coordinated by using a game, and based on a Lyapunov dynamic optimization shunting strategy, parameters are adaptively adjusted through a residual feedback closed loop, so that the stability of park safety management and control is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of smart city security management, and in particular to a security diversion control method for a smart park. Background Art

[0002] With the advancement of smart city construction, parks, as an important part of urban public spaces, face increasingly complex demands for crowd management. Traditional crowd monitoring technologies mostly rely on a single type of sensor to count the number of people in a local area, which makes it difficult to ensure data integrity in complex lighting, occlusion or equipment failure scenarios.

[0003] At the same time, existing prediction models often use independent time series analysis methods to simplify spatial distribution characteristics into state superpositions of discrete areas, ignoring the dynamic correlation of human flows in the spatial and temporal dimensions, resulting in prediction results that cannot accurately reflect the migration trends and aggregation patterns of large-scale crowds. At the level of diversion control, mainstream methods are usually based on preset static path planning rules or simple threshold trigger mechanisms, which fail to effectively balance the travel efficiency of individual tourists and the overall safety goals of the system, and are prone to cause local congestion or imbalance in resource utilization.

[0004] In addition, existing systems mostly use fixed parameter models and lack the ability to adaptively adjust to long-term evolution of passenger flow patterns and environmental disturbances. As a result, problems such as decreased prediction accuracy and delayed control instructions gradually become more prominent as the operating time increases. Summary of the invention

[0005] The purpose of the present invention is to provide a safe diversion control method for a smart park, which solves the problems in the prior art of poor real-time performance, low prediction accuracy and high safety risk of park crowd diversion control caused by one-sided data perception, discrete spatiotemporal modeling, unbalanced decision-making strategies and insufficient long-term system adaptability.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A safe diversion control method for a smart park, comprising the following steps: S1. Collect the crowd flow data in the park in real time through a multimodal sensor network and construct a spatiotemporal tensor; S2, performing tensor decomposition on the space-time tensor, generating predicted distribution of passenger flow in future time periods, and calculating the residual between the predicted value and the actual monitored value; S3, inputting the predicted distribution of passenger flow into a multi-agent game model to solve a path planning strategy; S4, according to the real-time queue status under the path planning strategy, using Lyapunov optimization to dynamically adjust the diversion strategy; S5. Coordinately update the tensor decomposition parameters and the game model parameters based on the prediction residual to form a closed-loop control.

[0007] Preferably, the step S1 includes: Deploy sensor nodes according to a preset grid across the entire park area, with each node integrating a binocular vision camera, a millimeter-wave radar, and a multi-band wireless signal transceiver; Normalize the collected raw data and construct a three-dimensional spatio-temporal tensor, where the first and second dimensions represent the spatial grid coordinates and the third dimension represents the time slice sequence.

[0008] Preferably, the step S2 includes: Perform Tucker decomposition on the spatio-temporal tensor : ; Among them, is the core tensor, is the preset rank parameter; , , are the spatial, temporal, and feature factor matrices respectively; is the residual tensor; Based on the temporal factor matrix extrapolate the time slices of the future , and generate a prediction tensor: ; Calculate the prediction residual : ; Among them, is the spatio-temporal tensor constructed from the actual monitoring data; represents the Frobenius norm.

[0009] Preferably, the rank parameter of the core tensor is dynamically adjusted according to historical pedestrian flow data, and the adjustment rule is: ; Among them, is the learning rate; is the residual change rate; The update of the spatial factor matrix satisfies: ; Among them, is the update step size; is the gradient of the system risk entropy function; is the grid occupancy matrix for path planning; represents singular value decomposition to maintain matrix orthogonality.

[0010] Preferably, the step S3 includes: Define the path selection cost function of individual tourists as follows: ; where, represents the path selection vector of tourist ; represents the length of path segment ; is the average walking speed of tourists; represents the maximum capacity of area ; represents the real-time number of people in area at time ; is the congestion penalty factor; Establish the system global risk entropy function: ; where, and are preset exponential parameters; Solve the Nash equilibrium point satisfying the following conditions by the gradient projection method: ; where, is the system optimization weight coefficient.

[0011] Preferably, the path selection cost function further includes a spatio-temporal coupling term: ; where, is the spatio-temporal coupling coefficient; represents the matrix trace operation.

[0012] Preferably, the step S4 includes: Define the real-time queue length of area , and its update rule is: ; where, represents the number of people entering area during time period ; represents the number of people leaving area during time period ; Construct the Lyapunov function: ; Design the optimization control objective: ; where, is the optimized weight coefficient for dynamic adjustment; Apply a shunt constraint: ; wherein, is the preset minimum output ratio coefficient.

[0013] Preferably, the optimized weight coefficient is dynamically adjusted according to the predicted residual : ; wherein, is the initial weight; is the preset residual threshold.

[0014] Preferably, the step S5 includes: When the predicted residuals of consecutive time slices exceed the threshold , perform the following operations: Reduce the rank parameter of the core tensor : ; Adjust the congestion penalty factor in the game model ; Perform singular value decomposition correction on the spatial factor matrix : ; wherein, is the correction step size.

[0015] The present invention also provides a safety shunt control system for a smart park, and the system includes: A multi-modal sensor network module, deployed at grid nodes throughout the park, for real-time collection of data on crowd density, movement speed, and signal strength; A spatio-temporal data fusion module, connected to the sensor network module, for normalizing multi-source heterogeneous data and constructing a three-dimensional spatio-temporal tensor; A tensor decomposition and prediction module, configured to perform high-order decomposition on the spatio-temporal tensor, generate a predicted distribution of the crowd flow in the future period, and calculate the predicted residual; A multi-agent game planning module, receiving the predicted distribution data, and solving the equilibrium strategy of the tourist path selection and the system safety goal; A Lyapunov optimization control module, for real-time monitoring of the regional queue status and dynamically adjusting the shunt instruction to meet the preset capacity constraint; The parameter collaborative update module synchronously adjusts the tensor decomposition rank parameter and the game penalty factor according to the prediction residual to achieve closed-loop control.

[0016] In summary, the present invention includes at least one of the following beneficial technical effects: 1. The present invention uses a multimodal sensor network to collaboratively collect visual, radar and wireless signal data, breaking through the environmental adaptability limitations of a single sensor and achieving comprehensive perception of crowd density, movement speed and distribution location in complex scenarios. It integrates multi-source heterogeneous data to construct a spatiotemporal tensor, eliminates data dimension differences, provides high-reliability input for subsequent modeling, and significantly improves the integrity and anti-interference ability of crowd state representation.

[0017] 2. This invention extracts the spatiotemporal evolution characteristics of passenger flow based on a high-order tensor decomposition algorithm, dynamically integrates historical laws with real-time data, and breaks through the local limitations of traditional time series prediction. Through the joint extrapolation of the core tensor and the factor matrix, multi-step prediction of passenger flow distribution in future periods is achieved, providing forward-looking guidance for diversion decisions and effectively reducing control blind spots caused by prediction lags.

[0018] 3. The present invention constructs a multi-agent game model, and incorporates tourists' path selection preferences and the system's global safety goals into a unified optimization framework. By defining a joint gradient descent mechanism of the individual cost function and the system risk entropy function, an optimal equilibrium strategy is generated that not only meets tourists' travel efficiency needs but also suppresses regional congestion risks, solving the problem of conflict between individual rationality and collective interests in traditional methods.

[0019] 4. The present invention adopts Lyapunov optimization theory to design the diversion control strategy, and jointly models the regional queue length fluctuation, the flow change trend and the system risk entropy. Through the dynamic weight adjustment mechanism, the short-term queue stability and long-term risk prevention needs are balanced in real time, ensuring that the physical feasibility and control stability of the diversion instruction can be maintained in abnormal situations such as sudden surge in flow or equipment failure.

[0020] 5. The present invention uses the prediction residual as the trigger signal of the parameter collaborative update mechanism, and realizes the dynamic adaptation of model parameters and environmental changes through the closed-loop linkage of core tensor rank parameter dimension reduction, congestion penalty factor enhancement and spatial factor matrix correction. This mechanism breaks through the performance degradation problem caused by the solidification of static model parameters, allowing the system to maintain prediction accuracy and decision reliability in long-term operation, and adapt to the gradual evolution of park traffic patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the method flow of the present invention; Figure 2 Schematic diagram of the system architecture of the present invention. DETAILED DESCRIPTION

[0022] The following is a further detailed description of the present invention in conjunction with the attached Figure 1 - attached Figure 2 drawings,

[0023] The present invention provides a safety diversion control method for a smart park, including the following steps: S1. Real-time collect the pedestrian flow data in the park through a multi-modal sensor network and construct a spatio-temporal tensor; In this embodiment, the specific implementation manner of step S1 for real-time collecting the pedestrian flow data in the park through a multi-modal sensor network and constructing a spatio-temporal tensor is as follows: Deploy sensor nodes according to a preset grid layout in the entire park area, and each node integrates multi-modal sensing devices. Preferably, the sensor node includes a binocular vision camera, a millimeter-wave radar, and a multi-band wireless signal transceiver. Among them, the binocular vision camera extracts the pedestrian flow density data in the grid area through a stereo vision algorithm; the millimeter-wave radar detects the moving speed and direction angle of pedestrians based on the frequency-modulated continuous wave (FMCW) technology; the multi-band wireless signal transceiver estimates the spatial distribution of mobile devices in the area by receiving the signal strength of WiFi, Bluetooth, and UWB and combining the triangulation method. Each sensor node transmits the original data to the data processing center through a wired / wireless communication network.

[0024] For the heterogeneous data collected by different sensors, perform data normalization processing respectively: Visual density data: Map the original pixel-level pedestrian flow count result to a preset grid coordinate system and perform standardization through the following formula: ; where is the original density value; and are respectively the mean and standard deviation of historical data statistics; is the density feature after normalization.

[0025] Moving speed data: Perform dynamic range compression on the instantaneous speed detected by the millimeter-wave radar: ; where is the preset maximum walking speed threshold for eliminating outlier interference.

[0026] Signal strength data: Normalize the received signal strength indication (RSSI) of the wireless signal: ; where and are respectively the minimum and maximum sensitivity thresholds of the device.

[0027] Organize the normalized multi-modal data into a three-dimensional tensor in the spatio-temporal dimension . Preferably, the construction rules are as follows: Spatial dimension: the first dimension and the second dimension respectively represent the number of grid rows and columns in the park area. Each grid cell corresponds to an actual geographical coordinate range, for example, 10 meters × 10 meters; Temporal dimension: the third dimension represents a sequence of time slices. Each slice corresponds to a preset duration (such as 30 seconds) and is stacked in chronological order to form a continuous observation window; Definition of tensor elements: each element stores the multi-modal feature vector in the th time slice .

[0028] Preferably, the construction of the spatio-temporal tensor further includes data alignment and missing compensation: Temporal alignment: Unify the sampling timestamps of each sensor through interpolation to ensure data synchronization within the same time slice; Spatial compensation: For grid data in signal coverage blind spots, use the weighted average of neighboring node observations for filling, and the weight is inversely proportional to the distance.

[0029] S2. Perform tensor decomposition on the spatio-temporal tensor to generate the predicted distribution of the pedestrian flow in the future period, and calculate the residual between the predicted value and the actual monitored value; In this embodiment, the specific implementation of step S2 for performing tensor decomposition on the spatio-temporal tensor to generate the predicted distribution of the pedestrian flow and calculate the prediction residual is as follows: Based on the spatio-temporal tensor constructed in step S1 , use the Tucker decomposition algorithm to extract spatio-temporal features. The decomposition model is defined as: ; where is the core tensor, used to characterize the coupling relationship between space, time, and feature modalities; , , are the spatial factor matrix, the temporal factor matrix, and the feature factor matrix respectively; is the decomposition residual tensor. Preferably, the rank parameter is initialized through cross-validation of historical data and dynamically adjusted in subsequent steps to adapt to the non-steady characteristics of the pedestrian flow pattern.

[0030] Based on the decomposed temporal factor matrix , predict the time evolution pattern of the future time period through the extrapolation algorithm. Preferably, use the autoregressive integrated moving average (ARIMA) model to perform multi-step extrapolation on the column vectors of the time factor matrix to generate the extended time factor matrix of the future number of time slices . Substitute the extrapolation result into the tensor reconstruction formula to generate the pedestrian flow prediction tensor for the future time period: ; Each element of the predicted tensor represents the predicted values of the pedestrian flow density, speed, and signal strength corresponding to the grid and time slice.

[0031] Obtain the actual monitoring data of the sensor network in real time, construct the measured spatio-temporal tensor , and calculate the prediction residual through the Frobenius norm: ; where represents the Frobenius norm; The residual is used to quantify the accuracy of the prediction model and serve as the feedback signal for closed-loop control. Preferably, the residual calculation is performed in a sliding window manner, and the window length is consistent with the prediction time period to ensure time series alignment.

[0032] Dynamically adjust the rank parameter of the core tensor according to the residual change rate , and the adjustment rule is: ; where is the learning rate. When the residual increases ( ), increase the rank parameter to improve the model complexity; when the residual decreases ( ), decrease the rank parameter to avoid overfitting; is the residual change rate.

[0033] Combine the path planning result to correct the spatial factor matrix , and the update formula is: ; where is the update step size; is the gradient of the system risk entropy function , which reflects the spatial distribution of the regional congestion risk; is the grid occupancy matrix of the path planning. When the path passes through the grid at , otherwise it is 0; The orthogonality of the matrix is ​​maintained through singular value decomposition to avoid the loss of low-rank constraint of the updated factor matrix.

[0034] S3, inputting the predicted distribution of passenger flow into a multi-agent game model to solve a path planning strategy; In this embodiment, the specific implementation method of step S3 to solve the path planning strategy through the multi-agent game model is as follows: Visitors in the park are modeled as independent decision-making agents, whose path selection behaviors constitute a non-cooperative game. The decision variable is the path selection vector , represents the tourist's preference in the candidate path set. Preferably, the path selection vector is a binary vector, where Indicates tourists Select paths, and the rest of the elements are 0. The joint strategy space of all agents constitutes the search domain of the equilibrium solution of the game.

[0035] In order to balance the travel efficiency of individual tourists and system security, the individual cost function is defined: ; in, Indicates tourists Path selection vector of Represents a path segment Length; is the average walking speed of tourists, used to quantify the travel time cost; Indicates area Maximum capacity; Indicates area In time Real-time number of people; is the congestion penalty factor, which is used to characterize the sensitivity of tourists to the degree of congestion. The congestion penalty term in the form of a logarithmic function is expressed by The nonlinear characteristics of ) will dramatically increase the cost, thereby guiding tourists to actively avoid high-risk areas.

[0036] In order to quantify the overall safety risk of the park, a system global risk entropy function is established: ; in, and are preset index parameters; they are used to adjust the weights of static congestion and dynamic passenger flow changes. The risk entropy function describes the system risk through the following mechanism: Static congestion risk: Item reflects area Instantaneous load rate, Apply an exponential penalty to the overload area at Dynamic aggregation risk: One item captures the changing trend of the pedestrian flow density. When (pedestrian flow accelerates and aggregates), the risk value increases significantly.

[0037] Solve the equilibrium strategy that satisfies the individual optimal and system risk constraints through the gradient projection method. Specifically, for each agent , its strategy update needs to satisfy the following conditions: ; Among them, is the system optimization weight coefficient, which is used to adjust the trade-off intensity between the individual goal and the global goal. Preferably, the iterative projection gradient descent algorithm is used for solving: Gradient calculation: Parallelly calculate the gradient of the cost function of each agent and the system risk gradient ; Strategy update: Update the path selection vector along the negative gradient direction , where is the learning rate; Projection operation: Project the updated vector into the feasible strategy space (that is, satisfy and constraints).

[0038] To enhance the coordination between path planning and spatio-temporal prediction, a spatio-temporal coupling term is introduced into the individual cost function: ; Among them, is the spatio-temporal coupling coefficient; represents the matrix trace operation. The spatio-temporal coupling term quantifies the correlation between the path and the congestion pattern characterized by the spatial factor matrix, guiding tourists to choose paths with a lower correlation with historical congestion.

[0039] S4. According to the real-time queue state under the path planning strategy, use Lyapunov optimization to dynamically adjust the diversion strategy; In this embodiment, the specific implementation manner of dynamically adjusting the diversion strategy based on the real-time queue state and Lyapunov optimization in step S4 is as follows: Define the real-time queue length of the area , and its update rule is: ; Among them, represents the time period entering the area The number of people; Indicates a time period Who leave the area within The number of people is regulated by the flow diversion control instruction. The queue model ensures the physical meaning of the queue length through a non - negative truncation operation ( ), avoiding model distortion caused by negative values.

[0040] Construct a quadratic Lyapunov function to quantify the stability of the system queue state: ; The function reflects the congestion degree of the overall system by accumulating the sum of the squares of the queue lengths in each area. When Decreases over time, it indicates that the queue fluctuations tend to be gentle and the system stability is enhanced; otherwise, it indicates a potential congestion risk.

[0041] Jointly optimize the queue stability and the system security risk, and design a dynamic control objective: ; Among them, Is the system global risk entropy function defined in step S3, Is the optimized weight coefficient for dynamic adjustment. The objective function realizes multi - objective trade - off through the following mechanism: Queue stability term: Characterizes the instantaneous change of the queue length, and minimizes this term to suppress queue fluctuations; System risk term: Incorporates the regional congestion risk into the optimization objective, and the weight coefficient Is dynamically adjusted according to the prediction residual to balance the short - term control response and the long - term risk prevention.

[0042] Apply the constraint on the proportion of the output of the flow of people to ensure the flow diversion efficiency: ; Among them, Is the preset minimum output proportion coefficient. The constraint ensures that each area diverts at least Within the time period The proportion of the incoming tourists, avoiding persistent congestion in local areas due to insufficient output.

[0043] The optimized weight coefficient Is dynamically adjusted according to the prediction residual In step S2: ; Among them, Is the initial weight, Is the preset residual threshold. The adjustment rule relates the prediction accuracy and the optimized weight through an exponential function: when the prediction residual When it increases (model accuracy decreases), reduce to enhance the response ability to the immediate queue status; when When it decreases, increase to increase the attention to the long-term risks of the system.

[0044] S5. Coordinately update the tensor decomposition parameters and the game model parameters according to the prediction residuals to form a closed-loop control; In this embodiment, the specific implementation manner of step S5 for coordinately updating the tensor decomposition parameters and the game model parameters based on the prediction residuals to achieve closed-loop control is as follows: When the prediction residual exceeds the preset threshold within consecutive time slices, it is determined that the current model parameters mismatch the real-time pedestrian flow pattern, and a parameter coordinated update process is triggered. Preferably, the triggering condition is detected through a sliding window mechanism: within a time window of length , if the residuals of all time slices , the parameter update is started. The threshold

[0045] is set according to historical data statistics and reflects the upper tolerance limit of the model prediction error.

[0045] Dynamically reduce the rank parameter of the core tensor in the Tucker decomposition of step S2 for the core tensor of , and the adjustment rule is: ; where is the learning rate; represents the relative amplitude of the residual exceeding the threshold. The adjustment reduces the rank parameter of the core tensor, reduces the model degrees of freedom to suppress the overfitting phenomenon, and thus improves the generalization ability for the dynamic pedestrian flow pattern. When (i.e., ), the rank parameter is reduced proportionally; when the residual falls back within the threshold, the rank parameter remains stable.

[0046] Exponentially adjust the congestion penalty factor of the game model in step S3: ; When the prediction residual exceeds the threshold , causes It increases, thereby enhancing the penalty intensity for congested areas in the path selection cost function. The adjustment mechanism dynamically corrects the risk preference through the feedback prediction error: when the model prediction accuracy decreases, it forces tourists to choose more conservative paths to reduce the actual congestion risk.

[0047] Perform singular value decomposition (SVD) correction on the spatial factor matrix in step S2: ; ; where is the correction step size; is the gradient matrix of the system risk entropy function , and is the path planning space occupancy matrix generated in step S3. The correction process is implemented through the following steps: Gradient-path association calculation: Calculate , spatially align the system risk gradient with the path occupancy pattern, and extract the correction direction of the risk-sensitive area; Orthogonal basis update: Perform singular value decomposition on the association matrix, extract the main eigen-directions, and ensure that the updated maintains column orthogonality; Step size control: Adjust the correction amplitude through to avoid excessive deviation from the spatially learned pattern in the past.

[0048] A safety diversion control system for a smart park described below can be correspondingly referred to the safety diversion control method for a smart park described above.

[0049] The present invention also provides a safety diversion control system for a smart park, and the system includes: A multi-modal sensor network module, which is composed of sensing devices distributed on the grid nodes throughout the park, and each node integrates multiple heterogeneous sensors. Preferably, the sensors include binocular vision cameras, millimeter-wave radars, and multi-band wireless signal transceivers, which are respectively used to collect the pedestrian flow density, pedestrian movement speed, and mobile device signal strength data in the grid area. Through the grid deployment strategy, spatial continuous coverage and time-synchronized sampling are achieved to ensure the global perception of the pedestrian dynamics. Each node transmits the original data to the central processing unit through a wired or wireless communication protocol, forming a multi-source heterogeneous data acquisition network; ​​​​Spatio-temporal data fusion module. This module receives multi-modal data streams from the sensor network and performs data cleaning, format alignment, and normalization processing. For heterogeneous data output by different sensors, independent channels are used for standardization conversion to eliminate dimension differences and unify the numerical range. Preferably, the normalized data is organized into a three-dimensional spatio-temporal tensor in three dimensions: spatial grid coordinates, time slice sequence, and feature mode. Among them, the spatial dimension corresponds to the grid division of the park geographical area, the time dimension represents the sampling time window, and the feature dimension fuses multi-modal indicators such as density, speed, and signal strength to form a structured data representation; Tensor decomposition and prediction module. This module extracts potential feature patterns in the spatio-temporal tensor based on the high-order tensor decomposition algorithm, including periodic pedestrian flow fluctuations, spatial correlations, and multi-modal coupling relationships. Through the core tensor and factor matrices obtained by decomposition, the pedestrian flow distribution trend in future time periods is extrapolated to generate prediction results. At the same time, the residual calculation is performed between the predicted data and the actual monitoring data to quantify the prediction accuracy of the model. Preferably, the residual data is input into the downstream control module as a feedback signal to trigger dynamic adjustment of the model parameters; Multi-agent game planning module. This module inputs the predicted pedestrian flow distribution data into the game decision-making model to construct a collaborative optimization problem of individual path selection of tourists and global system safety. By defining the individual cost function and the system risk entropy function, the needs of tourists for travel efficiency and the control objectives of the park for congestion risks are characterized. The gradient projection algorithm is used to solve the Nash equilibrium strategy to generate a path planning scheme that meets individual rationality and system constraints. Preferably, the path scheme is sent in real time through the electronic guide device to guide tourists to dynamically adjust their travel routes; Lyapunov optimization control module. This module monitors the change of queue length in each area in real time and constructs a dynamic optimization model based on the Lyapunov stability theory. By minimizing the comprehensive cost function of queue fluctuation and system risk, a flow control instruction is generated. Preferably, the instructions include the adjustment of the passing rate of each exit gate, the real-time information update of the path guiding screen, and the enabling strategy of the emergency channel to ensure that the number of people in the area is always lower than the safety capacity threshold; Parameter collaborative update module. This module synchronously adjusts the parameter configuration of the upstream model according to the predicted residual data. When the residual continuously exceeds the limit, operations such as reducing the rank parameter of the core tensor, strengthening the congestion penalty factor, and correcting the spatial factor matrix are triggered. Preferably, the update amplitude is controlled by the weight coefficient and the learning rate parameter during the adjustment process to avoid parameter oscillation. The adjusted parameters are re-input into the tensor decomposition module and the game model to form a closed-loop control loop of perception-decision-feedback.

[0050] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A safe diversion control method for a smart park, characterized in that: The following steps are involved: S1. Collect the crowd flow data in the park in real time through a multimodal sensor network and construct a spatiotemporal tensor; S2, performing tensor decomposition on the space-time tensor, generating predicted distribution of passenger flow in future time periods, and calculating the residual between the predicted value and the actual monitored value; S3, inputting the predicted distribution of passenger flow into a multi-agent game model to solve a path planning strategy; S4, according to the real-time queue status under the path planning strategy, using Lyapunov optimization to dynamically adjust the diversion strategy; S5. Coordinately update the tensor decomposition parameters and the game model parameters based on the prediction residual to form a closed-loop control.

2. A safety diversion control method for a smart park according to claim 1, characterized in that: The step S1 comprises: Sensor nodes are deployed in a preset grid throughout the park, with each node integrating a binocular vision camera, millimeter-wave radar, and a multi-band wireless signal transceiver; The collected raw data are normalized to construct a three-dimensional space-time tensor, in which the first and second dimensions represent the spatial grid coordinates and the third dimension represents the time slice sequence.

3. The safety diversion control method of a smart park according to claim 1 is characterized in that: The step S2 comprises: For the space-time tensor Perform Tucker decomposition: ； in, is the core tensor, is the preset rank parameter; , , are space, time, and characteristic factor matrices respectively; is the residual tensor; Based on the time factor matrix Extrapolating into the future time slice , generate the prediction tensor: ； Calculate prediction residuals : ； in, A space-time tensor constructed for actual monitoring data; represents the Frobenius norm.

4. A safe diversion control method for a smart park according to claim 3, characterized in that: The rank parameter of the core tensor Dynamically adjust based on historical passenger flow data, the adjustment rules are: ； in, is the learning rate; is the residual change rate; Spatial factor matrix The update satisfies: ； in, is the update step size; is the gradient of the system risk entropy function; The grid occupancy matrix for path planning; represents the singular value decomposition to maintain matrix orthogonality.

5. The safety diversion control method of a smart park according to claim 1 is characterized in that: The step S3 comprises: Defining individual tourists The path selection cost function is: ； in, Indicates tourists Path selection vector of Represents a path segment Length; is the average walking speed of tourists; Indicates area Maximum capacity; Indicates area In time Real-time number of people; is the congestion penalty factor; Establish the system global risk entropy function: ； in, and is the preset index parameter; The Nash equilibrium point that satisfies the following conditions is solved by the gradient projection method: ； in, Optimize weight coefficients for the system.

6. A safety diversion control method for a smart park according to claim 5, characterized in that: The path selection cost function further includes a spatiotemporal coupling term: ； in, is the space-time coupling coefficient; Represents a matrix trace operation.

7. The safety diversion control method of a smart park according to claim 1 is characterized in that: The step S4 comprises: Define the area Real-time queue length , and its update rule is: ； in, Indicates time period Inner entry area Number of people; Indicates time period Leave the area Number of people; Construct a Lyapunov function: ； Design optimization control objectives: ； in, is the optimization weight coefficient for dynamic adjustment; Apply flow splitting constraints: ； in, The preset minimum output scale factor.

8. A safety diversion control method for a smart park according to claim 7, characterized in that: The optimization weight coefficient According to the prediction residual Dynamic Adjustment: ； in, is the initial weight; is the preset residual threshold.

9. The safety diversion control method of a smart park according to claim 1 is characterized in that: The step S5 comprises: When continuous The prediction residuals for time slices Exceeding the threshold , do the following: Lower core tensor The rank parameter : ； Adjust the congestion penalty factor in the game model : ； For the spatial factor matrix Perform singular value decomposition correction: ； in, is the calibration step length.

10. A safety diversion control system for a smart park, applied to the method according to any one of claims 1 to 9, characterized in that: The system comprises: Multimodal sensor network modules are deployed in the park's global grid nodes to collect real-time data on crowd density, movement speed, and signal strength; A spatiotemporal data fusion module, connected to the sensor network module, is used to normalize multi-source heterogeneous data and construct a three-dimensional spatiotemporal tensor; A tensor decomposition and prediction module, configured to perform high-order decomposition on the spatiotemporal tensor, generate a predicted distribution of passenger flows for future time periods, and calculate the prediction residual; A multi-agent game planning module receives the predicted distribution of passenger flow and solves the equilibrium strategy between the tourist path selection and the system safety goal; Lyapunov optimized control module monitors the regional queue status in real time and dynamically adjusts the diversion instructions to meet the preset capacity constraints; The parameter collaborative update module synchronously adjusts the tensor decomposition rank parameter and the game penalty factor according to the prediction residual to achieve closed-loop control.

Images