An urban public security risk intelligent early warning and decision support system

Abstract

The application relates to the technical field of urban public safety, and discloses an intelligent early warning and decision support system for urban public safety risks, which comprises the following modules: a data acquisition and preprocessing module, which acquires and generates standardized data flow; an urban state vectorization module, which generates high-dimensional state vectors for urban space grids; a risk dynamics modeling and evolution engine, which adopts a physical information neural network to solve partial differential dynamic equations to predict a risk field, and adaptively corrects the model according to the actual and predicted residual error; an intervention strategy optimization module, which searches and determines an optimal intervention strategy through counterfactual reasoning; and an early warning generation and visualization module, which is used for generating early warnings and comprehensively visually presenting. The risk modeling based on physical laws, the forward-looking strategy optimization and the closed-loop adaptive correction capability are combined, the accuracy of risk prediction, the scientificity of decision and the adaptive capability of the system are significantly improved, and a complete intelligent decision closed loop is constructed.

Description

Technical Field

[0001] This invention relates to the field of urban public safety technology, specifically to an intelligent early warning and decision support system for urban public safety risks. Background Technology

[0002] With the acceleration of urbanization, the complexity and uncertainty of urban operations are increasing, posing unprecedented challenges to public safety risk management. To enhance urban safety governance capabilities, modern urban management systems have widely adopted information technology, constructing various intelligent monitoring and early warning platforms based on big data and the Internet of Things. These platforms integrate video surveillance, sensor networks, social media, and government department data to achieve real-time monitoring of urban pedestrian flow, traffic, and the environment. Through data analysis, they identify potential risks and trigger alarms when specific indicators exceed preset thresholds, providing urban managers with crucial situational awareness tools.

[0003] Most current risk prediction models rely primarily on statistical methods or traditional machine learning algorithms. While these methods excel at extracting correlations from historical data, they often lack a deep understanding of the intrinsic mechanisms of risk evolution. They treat the occurrence and evolution of risk as isolated statistical events, ignoring the propagation, diffusion, and growth patterns of risk as a continuous dynamic process across time and space. Consequently, when faced with novel risk patterns or complex coupling effects, the predictive accuracy and reliability of these models significantly decrease, making it difficult to meet the high-confidence prediction requirements of critical decision-making.

[0004] Furthermore, existing systems exhibit significant limitations in decision support. Their function largely remains at the level of risk notification—identifying current or predicting impending risks—but they fail to provide quantitative and forward-looking guidance on how to optimally address this core issue. The decision-making process still heavily relies on the experience and intuition of those in command. Simultaneously, these systems generally lack an effective closed-loop feedback mechanism. Once deployed, the models are typically static, unable to incorporate real-world feedback after intervention—the actual effectiveness of the strategy—into their learning and optimization processes. This prevents the models from learning from successful or unsuccessful intervention cases, and their performance gradually degrades over time and with dynamic changes in the city. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent early warning and decision support system for urban public safety risks. This system solves the problems of insufficient data fusion dimensions, inadequate modeling depth of risk evolution mechanisms, lack of forward-looking optimal intervention decision-making capabilities, and the inability of models to learn from actual intervention effects and adaptively optimize when dealing with urban public safety risks.

[0006] To achieve the above objectives, this invention provides the following technical solution: an intelligent early warning and decision support system for urban public safety risks. This system includes: a data acquisition and preprocessing module, an urban state vectorization module, a risk dynamics modeling and evolution engine, an intervention strategy optimization module, and an early warning generation and visualization module.

[0007] The output of the data acquisition and preprocessing module is connected to the input of the city state vectorization module; the output of the city state vectorization module is connected to the input of the risk dynamics modeling and evolution engine; the output of the risk dynamics modeling and evolution engine outputs information to the intervention strategy optimization module and the early warning generation and visualization module respectively; the output of the intervention strategy optimization module is connected to the input of the early warning generation and visualization module.

[0008] The risk dynamics modeling and evolution engine employs a physically-informed neural network to predict risk by solving a partial differential dynamic equation describing the spatiotemporal evolution of risk probability density. This method combines a data-driven neural network with a kinetic equation with explicit physical meaning, enabling the prediction model to not only fit historical data but also follow the inherent laws of risk propagation. The specific form of this partial differential dynamic equation is as follows: ; in: It is the partial derivative with respect to time; It is a location In time The risk probability density scalar; It is a state vector field provided by the city state vectorization module 200; It is the Hamiltonian operator, representing the gradient with respect to spatial coordinates; It depends on the state. The convective velocity field vector represents the directional propagation trend of risk, and its specific functional form needs to be studied. It depends on the state. The diffusion tensor represents the random diffusion rate of risk, and its specific functional form needs to be studied. It depends on risk density and state The reaction source term represents the local generation or decay rate of risk, and its specific functional form needs to be studied.

[0009] The system learns the unknown functions in the equations through a physical information neural network. , and The parameter set of this network By minimizing a composite loss function To optimize, its form is as follows: ; in; It is used to measure the data loss that deviates between network predictions and historical real risk event data. The physical loss is constructed by calculating the residuals of the dynamic equations at a large number of points in the spatiotemporal domain, in order to ensure that the learned function solution satisfies the physical constraint. and These are the corresponding weighting coefficients.

[0010] The intervention strategy optimization module, based on the Monte Carlo tree search algorithm, achieves a prospective search for the optimal intervention strategy through deep coupling with risk dynamics modeling and the evolution engine. For any candidate intervention strategy... The system obtains the risk field under this strategy by calling the evolution engine to perform counterfactual inference. and state field And based on a comprehensive cost function Its effectiveness is evaluated. The comprehensive cost function takes the following form: ; in: It is an execution strategy The total cost incurred; and These are the start and end times of the evaluation time window, respectively. It refers to the entire urban geographical area; In strategy The risk field is obtained through counterfactual reasoning; In strategy The state field is obtained through counterfactual deduction; It is a risk penalty function that assigns higher cost values ​​to high-risk probability densities and adverse states; It is an execution strategy The necessary resource costs, such as human and material resources, are obtained directly from the intervention operator library.

[0011] The system constructs an intelligent closed loop from prediction and decision-making to feedback, possessing adaptive correction capabilities that learn from prediction errors. After intervention measures are implemented, the risk dynamics modeling and evolution engine calculates the actual observed risk field. Compared with previously predicted risk fields The counterfactual-reality residual field ; in, It is the actual observed risk field after intervention, generated in real time through the data acquisition and preprocessing module 100 and the city state vectorization module 200. It is the risk evolution field previously predicted by the system based on this intervention.

[0012] The system performs dual-time-scale assimilation correction based on the residual field: in the short term, the residual is introduced as a correction forcing term into the dynamic equation to adjust the predicted trajectory in real time; in the long term, the parameters of the physical information neural network are fine-tuned and updated using accumulated historical residual data. Through the above technical solution, this invention organically integrates risk evolution modeling based on physical laws, strategy optimization based on counterfactual inference, and closed-loop adaptive correction capabilities based on real-world feedback, thereby improving the intelligence, foresight, and accuracy of urban public safety risk management.

[0013] This invention provides an intelligent early warning and decision support system for urban public safety risks. It has the following beneficial effects: 1. This invention utilizes a risk dynamics modeling and evolution engine, employing a physical information neural network to solve the partial differential dynamic equations describing the spatiotemporal evolution of risk. This approach combines data-driven learning capabilities with the physical laws of risk propagation, transforming risk prediction from a simple statistical fit into a dynamic deduction based on intrinsic mechanisms, thereby significantly improving the accuracy and reliability of future risk field predictions.

[0014] 2. This invention utilizes an intervention strategy optimization module, employing a Monte Carlo tree search algorithm to invoke an evolutionary engine to perform multiple counterfactual simulations. This design, based on forward-looking simulations, automatically searches and evaluates the future effects of numerous intervention strategy combinations, and determines an optimal intervention strategy that balances risk reduction and resource consumption based on a comprehensive cost function. This elevates decision support from a passive response to a scientific, forward-looking, and quantitative proactive optimization.

[0015] 3. This invention calculates the residual between the actual risk and the predicted risk after the intervention measures are implemented, and adaptively corrects the risk dynamics model based on this residual. This closed-loop feedback mechanism enables the system to continuously learn from prediction errors and evolve, constantly approximating the dynamics of the real world, thereby ensuring the accuracy of the model and its adaptability to environmental changes during long-term operation. Attached Figure Description

[0016] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is an internal framework diagram of the data acquisition and preprocessing module of the present invention; Figure 3 This is an internal framework diagram of the city state vectorization module of the present invention; Figure 4 This is a diagram of the internal framework of the risk dynamics modeling and evolution engine of the present invention; Figure 5 This is an internal framework diagram of the intervention strategy optimization module of the present invention; Figure 6 This is a diagram of the internal framework of the early warning generation and visualization module of the present invention.

[0017] The module includes: 100, Data Acquisition and Preprocessing; 200, Urban State Vectorization; 300, Risk Dynamics Modeling and Evolution Engine; 400, Intervention Strategy Optimization; and 500, Early Warning Generation and Visualization. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See attached document Figure 1 , Figure 1 This is a functional module diagram of an intelligent early warning and decision support system for urban public safety risks according to an embodiment of the present invention. The present invention provides an intelligent early warning and decision support system for urban public safety risks, comprising: The data acquisition and preprocessing module 100 is configured to acquire raw data related to city operation from multi-source heterogeneous data sources, and perform spatiotemporal alignment and standardization operations on the raw data to generate a standardized data stream; The city state vectorization module 200 is configured to receive the standardized data stream and map the data stream to a preset city spatial grid, generating a high-dimensional state vector containing observable and implicit states for each grid cell. The risk dynamics modeling and evolution engine 300 is configured to predict the risk field in the future period by solving a partial differential dynamics equation describing the spatiotemporal evolution of the risk probability density based on the high-dimensional state vector. The risk dynamics modeling and evolution engine 300 is also configured to calculate the residual between the actual risk state and the predicted risk field when the actual risk state is received after the implementation of external intervention measures, and adaptively correct the parameters of the partial differential dynamics equation based on the residual. The intervention strategy optimization module 400 is configured to, when the system predicts future risks, call the risk dynamics modeling and evolution engine 300 to perform multiple counterfactual simulations to simulate the impact of different intervention measures on the evolution of the risk field, and search for and determine an optimal intervention strategy based on a preset comprehensive cost function. The early warning generation and visualization module 500 is configured to generate structured early warning information when the predicted risk field exceeds a preset threshold, and to comprehensively visualize the predicted risk field, its evolution trend, and the optimal intervention strategy on the human-computer interaction interface.

[0020] The various modules described above in the embodiments of the present invention will be explained in detail below.

[0021] See attached document Figure 2 , Figure 2 This is a schematic diagram of the internal functional units of a data acquisition and preprocessing module 100 according to an embodiment of the present invention.

[0022] In one specific implementation, the data acquisition and preprocessing module 100 is configured as the system's unified data access layer, responsible for collecting raw data from diverse and structurally varied data sources distributed throughout the city, and converting this raw data into standardized data with a unified spatiotemporal reference and data structure that can be directly processed by subsequent modules. The implementation of the data acquisition and preprocessing module 100 may specifically include the following steps: S101: The data acquisition and preprocessing module 100 establishes communication connections with external data systems and receives data through multiple data interface adapters configured internally. For example, it accesses a video surveillance network via Real-time Streaming Protocol (RTSP) to obtain real-time video stream data; it subscribes to an Internet of Things (IoT) platform via Message Queuing Telemetry Transport (MQTT) protocol to obtain sensor data uploaded by devices such as traffic flow sensors and environmental noise monitors; it executes Structured Query Language (SQL) commands via Open Database Connectivity (ODBC) or Java Database Connectivity (JDBC) interfaces to batch pull or incrementally obtain event records from public safety event databases (such as 110 alarm records and 119 fire alarm records); or it obtains publicly available text data with geotags from social media or news aggregation services by calling dedicated application programming interfaces (APIs).

[0023] S102: The data acquisition and preprocessing module 100 performs data cleaning operations on the received heterogeneous raw data to improve data quality. For missing values ​​in the data stream, the system adopts different filling strategies according to the data type. For example, for continuously changing time series data (such as temperature), linear interpolation between previous and subsequent time points is used for filling; for discrete count data, the mean or median of the data source within the past time window is used for filling. For outliers in the data, the system identifies and filters them according to preset statistical rules, for example, filtering out data points that exceed three standard deviations from the mean.

[0024] S103: The data acquisition and preprocessing module 100 performs a spatial reference alignment operation on the cleaned data. Because the location information of the original data is in various formats, such as WGS-84 coordinates, local coordinate systems, administrative division names, or base station cell IDs, the data acquisition and preprocessing module 100 applies a preset geocoding conversion function. Map the location information of all data records to a unified global geographic coordinate system (e.g., WGS-). ; in: The converted longitude coordinates conforming to the WGS-84 standard; The converted latitude coordinates conforming to the WGS-84 standard; For geocoding conversion functions related to device type or data source, the specific implementation can be a lookup table storing coordinate mapping relationships or an API interface of an external geocoding service. A unique device identifier or data source identifier for reporting this data; The localized or unstructured raw location information reported by the device or data source.

[0025] S104: The data acquisition and preprocessing module 100 further performs a time base alignment operation on the data. The data acquisition and preprocessing module 100 sets a globally unified discrete time step. (For example, 60 seconds), and normalize all data streams into a discrete time sequence. superior; in: It is the first in the sequence A point in time; It is the preset system start time or reference time; It is a non-negative integer ( ), representing the ordinal number of the time step; This refers to the globally unified discrete time step. For event-based data (such as an alarm), its original occurrence timestamp is quantized to the start time of its corresponding time interval. Above. For continuously acquired sensor data, the data falling within the time window... Aggregate all data points within the time frame, for example, by taking their arithmetic mean or the last value, and use this as the reference for that time point. The representative value.

[0026] Through the above steps, the data acquisition and preprocessing module 100 ultimately outputs a series of structured data records that are fully aligned in the spatiotemporal dimensions and whose data quality is guaranteed. Each record contains a timestamp in a uniform format, geographic coordinates, and corresponding business data payload, and is transmitted to the city status vectorization module 200 for subsequent status feature extraction and analysis.

[0027] See attached document Figure 3 , Figure 3 This is a schematic diagram of the internal structure of a city state vectorization module 200 according to an embodiment of the present invention.

[0028] In one specific implementation, the input of the city state vectorization module 200 is connected to the output of the data acquisition and preprocessing module 100. The city state vectorization module 200 is configured to spatially discretize a spatiotemporally aligned standardized data stream into a grid cell covering the entire city. On a fixed grid, and for each grid cell at each time step Generate a high-dimensional feature vector that can characterize its overall state. The implementation of the city state vectorization module 200 can specifically include the following steps: S201: The city state vectorization module 200 receives standardized data records from the data acquisition and preprocessing module 100. For each grid cell... and current time step The city status vectorization module 200 will include all spatial data. Within range, timestamp is Data records within intervals are aggregated. Statistical calculations (such as averaging, summing, or density calculation) are then performed on this aggregated data to generate an observable state vector. The components of the observable state vector are direct physical measurements or statistics, such as the pedestrian density within the grid, average vehicle speed, the number of occurrences of a specific type of incident, and the number of posts with specific keywords from social media.

[0029] S202: The city state vectorization module 200 is further configured to infer and generate implicit state features that cannot be directly measured but have a significant impact on risk evolution based on the observable state vector and other relevant data. These features together constitute the implicit state vector. This is a specific implementation of the "implicit state" in the high-dimensional state vector. In a lower-level embodiment, the generation of the implicit state vector includes calculating the regional psychological stress index. The regional psychological stress index is calculated using a pre-trained fusion function. To calculate this, the fusion function integrates observable states from multiple dimensions: ; in: It is a grid cell In time Regional psychological stress index; It is the average noise level extracted by the grid cells from the noise sensor data; It is a crowd congestion index extracted from video or signaling data; It is the intensity value of negative emotions extracted from social media texts belonging to this region using natural language processing technology; The fusion function can be implemented as a multi-layer linear regression model or a small feedforward neural network.

[0030] In another lower-level embodiment, the generation of implicit state vectors also includes quantizing the social connection strength between grid cells. For two adjacent grid cells... and They are in time Connection weights Defined by the flow of people between them: ; in, Indicates the time interval Inside, from the unit Move to unit The number of anonymous individuals. Based on grid cells. social connectivity Defined as the sum of its connection weights with all adjacent cells: ; in, Represent all and The set of adjacent grid cells, It is an index variable representing the identifier of a grid cell. This social connectivity will serve as a key input parameter in the risk dynamics model to describe the spatial diffusion of risk.

[0031] S203: The city state vectorization module 200 will convert the observable state vector generated in the preceding steps into an observable state vector. and implicit state vector The vectors are concatenated to form the complete state vector of the grid cell at that time step. ; in, This represents the transpose operation of a vector.

[0032] Finally, the city state vectorization module 200 sets the state vectors of all grid cells at the current time step. (in, The set of all grid cells. It is a universal quantifier. The entire grid cell is output to the Risk Dynamics Modeling and Evolution Engine 300 to drive subsequent risk prediction and analysis.

[0033] See attached document Figure 4 , Figure 4 This is a schematic diagram of the internal structure of a risk dynamics modeling and evolution engine 300 according to an embodiment of the present invention.

[0034] In one specific implementation, the risk dynamics modeling and evolution engine 300 has its input connected to the output of the city state vectorization module 200. The risk dynamics modeling and evolution engine 300 is configured as the core computing unit of the system. It predicts the future spatiotemporal evolution of urban public safety risks through a data-driven and physically constrained dynamic model, and possesses the ability to learn from prediction errors to achieve adaptive model correction. The implementation of the risk dynamics modeling and evolution engine 300 may specifically include the following steps: S301: In the initial stage of system deployment, the risk dynamics modeling and evolution engine 300 performs a model initialization training to learn and solidify a partial differential dynamic equation describing the spatiotemporal evolution of risk probability density. This is achieved by solving a specific technical implementation of the partial differential dynamic equation describing the spatiotemporal evolution of risk probability density. In a lower-level embodiment, the equation takes the following form: ; in: It is the partial derivative with respect to time; It is a location In time The risk probability density scalar; It is a state vector field provided by the city state vectorization module 200; It is the Hamiltonian operator, representing the gradient with respect to spatial coordinates; It depends on the state. The convective velocity field vector represents the directional propagation trend of risk, and its specific functional form needs to be studied. It depends on the state. The diffusion tensor represents the random diffusion rate of risk, and its specific functional form needs to be studied. It depends on risk density and state The reaction source term represents the local generation or decay rate of risk, and its specific functional form needs to be studied.

[0035] To learn the unknown function in the above equation , and The system is trained using a Physical Information Neural Network (PINN). The parameter set of the Physical Information Neural Network... By minimizing a composite loss function To optimize: ; in; It is used to measure the data loss that deviates between network predictions and historical real risk event data. The physical loss is constructed by calculating the residuals of the dynamic equations at a large number of points in the spatiotemporal domain, in order to ensure that the learned function solution satisfies the physical constraint. and These are the corresponding weighting coefficients.

[0036] S302: During the real-time operation phase of the system, the risk dynamics modeling and evolution engine 300 receives the current moment provided by the city state vectorization module 200. City-wide state vector field Using this state field as input, the Risk Dynamics Modeling and Evolution Engine 300 calculates the initial risk field using a pre-trained physical information neural network. Subsequently, numerical methods (such as the finite element method or the finite difference method) are used to perform forward integration on the dynamic partial differential equation, thereby calculating the risk probability density field over the next continuous time period. (in The predicted risk field evolution sequence is output to the intervention strategy optimization module 400 and the early warning generation and visualization module 500.

[0037] One specific numerical solution implementation uses the finite difference method. In this method, the geographic space of the entire city is divided into a regular grid (this grid can be consistent with the grid used by the city state vectorization module 200). The time partial derivatives in the equation Both spatial partial derivatives are replaced with difference approximations based on function values ​​at grid nodes. For example, the time forward difference can be expressed as... ,in, Representing grid points At time step The risk value, Representing grid points At time step The risk value. Through this substitution, the original partial differential equation is transformed into a large system of algebraic equations, which the system can iteratively calculate from the current time step. Given the risk field distribution, solve for the next time step. By progressively extending the risk field distribution in this way, the risk probability density field for consecutive future time periods can be obtained. (in .

[0038] Another alternative numerical solution is the finite element method. This method divides the urban space into multiple smaller, which can be irregularly shaped "units" (such as triangles or quadrilaterals). The resulting mesh, or "finite element mesh." Within each element, the unknown risk probability density function... It is used with a simple, node-value-weighted basis function (e.g., a linear or quadratic polynomial). We approximate the solution by substituting this approximate solution into the weak form (integral form) of the original partial differential equation. By integrating over the entire region, a large system of algebraic equations with unknowns at all nodes can be obtained. Solving this system of equations yields the risk value of each node at different times. The finite element method is particularly suitable for handling geographical areas with complex or irregular boundaries.

[0039] S303: The risk dynamics modeling and evolution engine 300 is further configured to achieve online closed-loop adaptive correction of the model after external intervention measures are implemented. A specific technical implementation is provided for adaptively correcting the parameters of the partial differential dynamics equations based on the residuals. This process first calculates the counterfactual-reality residual field. ; in, It is the actual observed risk field after intervention, generated in real time through the data acquisition and preprocessing module 100 and the city state vectorization module 200. It is the risk evolution field previously predicted by the system based on this intervention.

[0040] Counterfactual-Real Residual Field It is used to perform dual-timescale assimilation correction. On the short-term timescale, it is introduced as a correction forcing term into the dynamic equations to adjust the predicted trajectory in real time, at which point the equations are updated to: ; in, It is a pre-defined or adaptively calculated Kalman gain matrix. Over a long timescale, accumulated historical residual field data is used to optimize the parameter set of the physical information neural network. Fine-tuning updates can be performed, and the update rules can be expressed as follows: ; in, and These are the model parameters before and after the update, respectively. It's the learning rate. It is based on residual field The constructed assimilation loss function, It's about the parameter set. The gradient operator. Through this dual-timescale assimilation mechanism, the internal model of the Risk Dynamics Modeling and Evolution Engine 300 can continuously learn from prediction errors, constantly approximating the dynamics of the real world.

[0041] See attached document Figure 5 , Figure 5 This is a schematic diagram of the internal structure of an intervention strategy optimization module 400 according to an embodiment of the present invention.

[0042] In one specific implementation, the intervention strategy optimization module 400 receives predicted risk information from the risk dynamics modeling and evolution engine 300 at its input end, and connects its output end to the early warning generation and visualization module 500. When the system predicts that a risk triggers an early warning condition, the intervention strategy optimization module 400 is activated. The function of the intervention strategy optimization module 400 is to automatically search for and determine an optimal combination of intervention measures to mitigate or eliminate potential risks based on forward-looking simulations. The implementation of the pre-strategy optimization module 400 may specifically include the following steps: S401: The intervention strategy optimization module 400 internally constructs and maintains an intervention operator library. The intervention operator library stores a set of standardized intervention measures, each real-world intervention measure being encoded as an intervention operator with a defined mathematical definition. For example, measures such as deploying security forces, setting up traffic control, and issuing evacuation guidance information each correspond to an independent operator. Each operator... It clearly defines the specific way in which it acts on the dynamic system, namely, by selectively modifying one or more variables within the internal model of the Risk Dynamics Modeling and Evolution Engine 300 to intervene in the evolution of risk. This modification can affect the state vector of a specific region. Assign a value to a certain component, or adjust the convective velocity field in the dynamic equation. or reaction source term The function parameters.

[0043] S402: The intervention strategy optimization module 400 will optimize an intervention strategy Defined as one or more time-ordered interference budget subsequences ; in: This represents a complete intervention strategy; Represents a budgetary variable that executes at a specific time; It is a unique identifier for an intervention operator, corresponding to a specific intervention in the intervention operator library; This is the set execution time point for the intervention. It is used to evaluate any candidate intervention strategy. To assess overall effectiveness, the intervention strategy optimization module 400 defines a comprehensive cost function. This is a specific technical implementation based on a preset comprehensive cost function.

[0044] The pre-defined comprehensive cost function takes the following form: ; in: It is an execution strategy The total cost incurred; and These are the start and end times of the evaluation time window, respectively. It refers to the entire urban geographical area; In strategy The risk field is obtained through counterfactual reasoning; In strategy The state field is obtained through counterfactual deduction; It is a risk penalty function that assigns higher cost values ​​to high-risk probability densities and adverse states; It is an execution strategy The necessary resource costs, such as human and material resources, are obtained directly from the intervention operator library.

[0045] The goal of the intervention strategy optimization module 400 is to solve the following optimization problem in order to find the optimal strategy. ; S403: To solve the above optimization problem within a large policy combination space, the intervention policy optimization module 400 employs the Monte Carlo Tree Search (MCTS) algorithm. In each iteration, the Monte Carlo Tree Search algorithm selects a candidate policy. Subsequently, the intervention strategy optimization module 400 invokes the risk dynamics modeling and evolution engine 300 to request the use of this strategy. Perform a counterfactual simulation under the given conditions to obtain the future risk field evolution under this strategic intervention. Next, the comprehensive cost function is used. The results of this deduction are evaluated, and the evaluation value is used as feedback to update the internal statistics of Monte Carlo tree search, so as to guide the subsequent search direction and make it more inclined to explore the strategy branches with lower costs. The Monte Carlo tree search algorithm approximates the optimal solution by iteratively following these four core steps: The first step is the selection step. The algorithm starts from the root node of the tree representing an empty policy and recursively selects child nodes based on an internal criterion of balancing exploration (trying policy branches that are not fully evaluated) and exploitation (going deeper into policy branches that are known to perform well) until a leaf node is reached.

[0046] The second step is the expansion step, where the algorithm generates one or more new child nodes for the leaf nodes reached in the selection step. Each new child node represents an additional available intervention from the intervention operator library that has not been previously tried, based on the current policy sequence.

[0047] The next step is the simulation step, which is crucial for the deep coupling between the Monte Carlo tree search algorithm and other modules in this invention. Starting from a newly created child node in the extension step, the algorithm performs a complete "counterfactual deduction." Specifically, the intervention strategy optimization module 400 calls the risk dynamics modeling and evolution engine 300, requesting that, using the incomplete intervention strategy represented by the child node as initial conditions, the algorithm deduce the complete risk field evolution result under this strategy up to the end of the evaluation time window. .

[0048] Finally, there is the backtracking step, where the system uses the comprehensive cost function after the simulation is complete. The result of this counterfactual deduction is evaluated. This evaluated cost value is then propagated back to the root node from bottom to top along the path taken by the selection steps, and used to update the internal statistics of all nodes on the path (e.g., the number of times a node is visited, the cumulative evaluation cost, etc.).

[0049] S404: After completing a preset number of iterations or when the computation time is exhausted, the Monte Carlo tree search algorithm converges and outputs the currently discovered optimal intervention strategy sequence. The intervention strategy optimization module 400 will select this optimal intervention strategy sequence. Output to the early warning generation and visualization module 500 for subsequent visualization and decision support.

[0050] See attached document Figure 6 , Figure 6 This is a schematic diagram of the internal structure of an early warning generation and visualization module 500 according to an embodiment of the present invention.

[0051] In one specific implementation, the early warning generation and visualization module 500 is configured as the system's decision support and human-computer interaction terminal, responsible for monitoring predicted risks, generating alerts, and providing decision-makers with a comprehensive view integrating the risk situation and optimal response strategies. The implementation of this module may specifically include the following steps: S501: The risk warning generation unit of the warning generation and visualization module 500 continuously receives and monitors the predicted risk field output from the risk dynamic character modeling and evolution engine 300. The risk warning generation unit determines whether to generate a warning based on preset warning triggering rules. In one lower-level embodiment, the rule is defined as: When at any spatial location and some point in the future (in An alert will be generated when the following conditions are met: ; in: In position and time The probability density of predicted risk; It is a pre-set safety threshold related to the type of risk; It refers to the current moment; It is the maximum prediction time range set by the system.

[0052] When the above conditions are met, the risk warning generation unit immediately generates a structured warning message. The warning message is a data object containing fields such as warning identifier, generation time, risk location, predicted peak time, risk level, and risk type. This warning message is used to activate the intervention strategy optimization module 400 and is simultaneously transmitted to the situation presentation and decision-making interaction unit within the module.

[0053] S502: The situation presentation and decision-making interaction unit of the early warning generation and visualization module 500 is configured to comprehensively visualize the received multi-channel information. The information it receives includes: the risk field evolution sequence output by the risk dynamics modeling and evolution engine 300. And the optimal intervention strategy output by the intervention strategy optimization module 400. In one further embodiment, the visualization is implemented through a Geographic Information System (GIS) based human-computer interaction interface.

[0054] S503: On this human-computer interaction interface, the predicted risk field The data is rendered as a dynamic heatmap and overlaid on an electronic map. The heatmap's color intensity corresponds to the risk probability density value. A positive correlation exists, for example, a smooth transition from green, representing low risk, to red, representing high risk. The interface provides a timeline control, which decision-makers can drag to visually observe the spatial extent and intensity evolution of risk areas over future time periods.

[0055] S504: Optimal Intervention Strategy Upon receiving the data, it is parsed and converted into visual primitives on a map. For example, instructions to deploy police forces are represented by specific icons displayed at designated locations; instructions to set up traffic barriers or implement traffic control are achieved by highlighting the corresponding road segments and indicating the control direction; instructions to issue evacuation broadcasts are represented by a semi-transparent circular area centered on the broadcast point and covering its area of ​​influence. Simultaneously, the specific steps and text descriptions of the strategy are listed in the sidebar of the interface and interact with the corresponding primitives on the map. In this way, the early warning generation and visualization module 500 transforms abstract risk data and strategy instructions into intuitive, contextualized situation maps, providing decision-makers with an immediate and comprehensive understanding of "what the risk is," "how the risk will develop," and "what the optimal response is," thereby supporting rapid and accurate decision-making responses.

[0056] See attached document Figure 1 , Figure 1 This is a schematic diagram of the functional modules of an intelligent early warning and decision support system for urban public safety risks according to an embodiment of the present invention. To illustrate how the various functional modules of the system work together to form a complete intelligent decision-making closed loop, the overall workflow of the system is described below using a security scenario of a large-scale public event (e.g., a New Year's Eve concert held in a city center square) as an example.

[0057] S601: During the concert, the data acquisition and preprocessing module 100 continuously collects data from multiple data sources deployed in the square and surrounding areas. This data includes: real-time video streams from high-definition cameras to analyze crowd density and flow; anonymous signaling data from mobile communication base stations to estimate the total number of people and their dwell time in the area; public posts with geotagging from social media platforms to gauge public sentiment; and real-time traffic data from traffic management departments. The data acquisition and preprocessing module 100 performs spatiotemporal alignment on this heterogeneous data, unifying it into a time series with a 1-minute step, and outputs a standardized data stream containing timestamps, geographic coordinates, and data payload.

[0058] S602: The city state vectorization module 200 receives the standardized data stream and divides the square area into 10m × 10m grid units. For each grid unit, the city state vectorization module 200 calculates its observable state vector at each time step. The observable state vector includes components such as crowd density, average movement speed, and the number of posts with negative sentiment. Simultaneously, based on these observable data, the city state vectorization module 200 infers and generates an implicit state vector. For example, by fusing indicators such as high density, low movement speed, and high negative sentiment, a quantified "stampede risk factor" is calculated. The observable state vector and the implicit state vector are concatenated to form a high-dimensional state vector representing the overall state of the grid unit at the current moment.

[0059] S603: The Risk Dynamics Modeling and Evolution Engine 300 receives a state vector field covering the entire plaza area. The engine's internal physical information neural network, trained using historical data, calculates the initial risk field at the current moment based on the current state field and preset dynamic partial differential equations, and extrapolates the spatiotemporal evolution of the risk probability density over the next 30 minutes. At a certain moment, the extrapolation results show that due to excessive crowd concentration in front of the stage, the risk probability density in area A in front of the stage is expected to exceed the preset safety threshold in 15 minutes.

[0060] S604: The early warning generation and visualization module 500 monitors the above-mentioned predicted risks exceeding the threshold, immediately generates a structured early warning, indicating that there is a high risk of stampede in area A, and activates the intervention strategy optimization module 400.

[0061] S605: The intervention strategy optimization module 400 initiates a Monte Carlo tree search to find the optimal response strategy from its internal library of intervention operators. It simulates various strategy combinations, such as "Strategy A: Deploy police forces around area A to manage pedestrian flow," "Strategy B: Guide the crowd to the more sparsely populated areas B and C via plaza broadcasts," and "Strategy C: Combine police force management with broadcast guidance." For each strategy, the intervention strategy optimization module 400 calls the risk dynamics modeling and evolution engine 300 to perform a counterfactual simulation, simulating the evolution of the risk field after implementing the strategy. By comparing the comprehensive cost functions under different strategies (considering both risk reduction effects and intervention resource consumption), the intervention strategy optimization module 400 ultimately determines "Strategy C" as the optimal intervention strategy.

[0062] S606: The human-computer interaction interface of the early warning generation and visualization module 500 pushes early warning information to decision-makers in the command center. The interface dynamically displays the predicted risk evolution trend of area A on the electronic map with highlighted red. At the same time, the interface clearly presents the recommended optimal intervention strategy "Strategy C", marks the suggested police deployment points and broadcast coverage on the map, and includes a detailed list of action instructions.

[0063] S607: Command personnel confirm and issue execution instructions. The system does not cease operation after the intervention measures are implemented. The data acquisition and preprocessing module 100 continuously acquires real-time data from the site, showing that the population density in area A has begun to decrease. The risk dynamics modeling and evolution engine 300 compares the actual risk state after the intervention measures are implemented with the risk state previously predicted based on "Strategy C," calculating the residual between the two. This residual is used, on the one hand, to immediately correct the current prediction model; on the other hand, its accumulated value is used to fine-tune the parameters of the physical information neural network in the background.

[0064] Through the above steps, the system not only completes the entire process from risk perception, prediction, early warning to decision-making, but more importantly, it forms an intelligent closed loop that can be continuously optimized by incorporating real-world feedback after decision execution into the model.

Claims

1. An intelligent early warning and decision support system for urban public safety risks, characterized in that, include: The data acquisition and preprocessing module is configured to acquire raw data related to city operation from multi-source heterogeneous data sources, and perform spatiotemporal alignment and standardization operations on the raw data to generate a standardized data stream; The city state vectorization module, connected to the output of the data acquisition and preprocessing module, is configured to receive the standardized data stream and map the standardized data stream onto a preset city spatial grid, generating a high-dimensional state vector containing observable and implicit states for all grid cells. A risk dynamics modeling and evolution engine, connected to the output of the city state vectorization module, is configured to predict the risk field for future periods by solving the partial differential dynamics equations describing the spatiotemporal evolution of risk probability density based on the received high-dimensional state vectors. The risk dynamics modeling and evolution engine is also configured to calculate the residual between the actual risk state and the predicted risk field when receiving the actual risk state after the implementation of external intervention measures, and adaptively correct the parameters of the partial differential dynamics equations based on the residuals. The intervention strategy optimization module, whose input is connected to the output of the risk dynamics modeling and evolution engine, is configured to, when the system predicts future risks, call the risk dynamics modeling and evolution engine to perform multiple counterfactual simulations to simulate the impact of different intervention measures on the evolution of the risk field, and search for and determine the optimal intervention strategy based on a preset comprehensive cost function. The early warning generation and visualization module has its input end connected to the output end of the risk dynamics modeling and evolution engine and the output end of the intervention strategy optimization module, respectively. It is configured to generate structured early warning information when the predicted risk field exceeds a preset threshold, and to comprehensively visualize the predicted risk field, its evolution trend and the optimal intervention strategy on the human-computer interaction interface.

2. The intelligent early warning and decision support system for urban public safety risks according to claim 1, characterized in that, The city state vectorization module is specifically configured as follows: Data records falling into all grid cells are aggregated, and observable state vectors are generated through statistical calculations. Based on the observable state vector, an implicit state vector is inferred and generated; The observable state vector and the implicit state vector are concatenated along the vector dimension to form the high-dimensional state vector.

3. The intelligent early warning and decision support system for urban public safety risks according to claim 2, characterized in that, The implicit state vector includes a regional psychological stress index, which is calculated by a fusion function that integrates the average noise level extracted from sensor data, the crowd density index extracted from video or signaling data, and the negative emotion intensity value extracted from social media text.

4. The intelligent early warning and decision support system for urban public safety risks according to claim 1, characterized in that, The risk dynamics modeling and evolution engine learns and solves the partial differential dynamics equations through a physical information neural network, which is optimized by minimizing the composite loss function.

5. The intelligent early warning and decision support system for urban public safety risks according to claim 4, characterized in that, The composite loss function includes: a data loss term for measuring the deviation between network predictions and historical real risk event data, and a physical loss term constructed by calculating the residuals of the partial differential dynamic equations at placement points in the spatiotemporal domain.

6. The intelligent early warning and decision support system for urban public safety risks according to claim 1, characterized in that, The risk dynamics modeling and evolution engine adaptively corrects the parameters of the partial differential dynamics equations based on the residuals, which is achieved by performing dual-timescale assimilation correction, including: On a short timescale, the residual is introduced as a correction forcing term into the partial differential dynamics equation to adjust the predicted trajectory in real time. Over a long timescale, the parameters of the physical information neural network are fine-tuned and updated using accumulated historical residual field data.

7. The intelligent early warning and decision support system for urban public safety risks according to claim 1, characterized in that, The intervention strategy optimization module uses the Monte Carlo tree search algorithm to search and determine the optimal intervention strategy in a preset intervention measure operator library.

8. The intelligent early warning and decision support system for urban public safety risks according to claim 7, characterized in that, The Monte Carlo tree search algorithm is specifically configured in its simulation step to: invoke the risk dynamics modeling and evolution engine, request to perform counterfactual inference with candidate intervention strategies as conditions, so as to obtain the future risk field evolution results under the intervention of the strategy.

9. The intelligent early warning and decision support system for urban public safety risks according to claim 1, characterized in that, The comprehensive cost function includes: a risk cost term obtained by integrating the risk penalty function over a preset spatiotemporal domain, and a resource cost term required to implement the intervention strategy itself.

10. The intelligent early warning and decision support system for urban public safety risks according to claim 1, characterized in that, The comprehensive visualization presentation of the early warning generation and visualization module on the human-computer interaction interface includes: The predicted risk field is rendered as a dynamic heatmap overlaid on an electronic map, wherein the color intensity of the heatmap is positively correlated with the risk probability density value. The optimal intervention strategy is parsed and converted into visual primitives on a map and an accompanying list of action instructions.

Images