A police system optimization method based on big data analysis

Abstract

The application relates to the technical field of big data analysis, and provides a police service system optimization method based on big data analysis, which comprises the following steps: dividing grids, deploying edge nodes to obtain sensing data, taking each grid as a graph node, comprehensively calculating edge weights according to the distance between grid centers and the historical case co-occurrence frequency, applying graph convolution operation to obtain a fused global state vector; taking the global state vector as environment observation, defining an instant reward item for each grid at the grid level, calculating reward and penalty coefficients at the regional level according to a period, and combining the two-layer rewards into a total reward function according to a configurable weight; periodically distributing police forces in the police force pool as an initial constraint according to the risk summary of the region at the regional level, and adjusting the police forces in the grid in real time through reinforcement learning at the grid level, and triggering a linkage signal of adjusting the police force resources according to the grid risk threshold, so that the regional distribution and the grid dynamic scheduling form a closed-loop cooperation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data analysis, and particularly relates to a police service system optimization method based on big data analysis. BACKGROUND

[0002] The police service system refers to a management and work platform integrating informatization, digitization and intelligentization specially designed for police agencies. Through the integration of personnel management, case management, information query, data analysis and other functions, the police service system can realize real-time dynamic monitoring, support decision-making and improve the detection capability.

[0003] In Chinese application No. CN202411336440.4, a police service system optimization method based on big data analysis is disclosed. The configuration method comprises: collecting multi-element police service data including police situation reports, video monitoring and social media data, and pre-processing data from different sources to integrate them into a unified data set to establish a strategy network model. Based on the established strategy network model, the optimization target is set as the cumulative return, which represents the total reward obtained by the model in the long term during task execution. An award function including police utilization rate, response time and crime rate change is set. The pre-processed data is used for training of the strategy network model, and the model parameters are optimized to maximize the return. The trained strategy network model is used to predict the collected and integrated multi-element police service data. According to the model prediction result and the optimization target, the police service resource allocation and command strategy are adjusted.

[0004] In the technical field of big data analysis, although there is a technical solution combining strategy gradient algorithm and big data analysis to improve the efficiency of responding to emergencies and urgent situations, in the context of police service resource scheduling and configuration, the existing technology has the following problems:

[0005] 1. The traditional police service static scheduling method fails to fully consider real-time police situation changes and resource requirements, and lacks a scheduling scheme that can combine the current remaining resources according to the event risk level, resulting in over or under allocation of police forces in some areas, affecting police service efficiency and response speed.

[0006] 2. The existing police service resource scheduling is usually carried out in a fixed grid, which is difficult to consider different scheduling granularities, and there is a time mismatch between short-term prediction and police scheduling. SUMMARY

[0007] The present application provides a police service system optimization method based on big data analysis, which aims to solve the problems of lack of event risk level and remaining resource combination for scheduling of police service resources, and time mismatch between different scheduling granularities.

[0008] The technical solution adopted by the present application to solve the above technical problems is to provide a police service system optimization method based on big data analysis, comprising:

[0009] Multi-modal data acquisition processing, extracting feature fusion graph structure, dividing grid deploying edge node to obtain video, GPS, environmental sound data, taking each grid as a graph node, calculating edge weight according to the distance between grid centers and the historical case co-occurrence frequency, applying graph convolution operation to the feature matrix and adjacency matrix of the graph node to obtain the fused global state vector;

[0010] Constructing a reinforcement learning scheduling environment, designing a hierarchical reward function, taking the global state vector as the environment observation, defining an immediate reward item for each grid at the grid level, including high-risk event handling positive reward, mis-scheduling penalty, and response time reward proportional to the penalty, calculating the average response time, police utilization rate and police allocation variance at the regional level, and giving corresponding periodic reward and penalty coefficients, and combining the two-layer rewards into the overall reward function according to the configurable weight;

[0011] Multi-scale hierarchical scheduling coordination, closed-loop feedback and online correction, periodically allocating police pool as initial constraint based on regional risk summary at regional level, adjusting police in grid in real time through reinforcement learning at grid level and triggering linkage signal of adjusting police resources according to grid risk threshold, forming closed-loop coordination between regional allocation and grid dynamic scheduling.

[0012] As a preferred embodiment, the specific steps of the multi-modal data acquisition processing and feature fusion graph structure extraction are as follows:

[0013] Deploying multi-modal acquisition units at the grid positions divided by the police command center, including video acquisition units, GPS acquisition units, and environmental sound acquisition units; installing local computing edge agents for each acquisition unit, responsible for preliminary preprocessing, maintaining time synchronization with the server of the police command center through NTP protocol, for video data, performing YOLO model inference on each frame of image to obtain the statistics of vehicle and pedestrian number, for GPS data, removing abnormal points with speed mutation greater than a preset threshold, and simultaneously using cubic spline interpolation to complete missing data points, for environmental sound data, only retaining data with frequency between 300 and 3000 Hz, and then dividing into short-time data blocks for Fourier transform;

[0014] Establishing a regular rectangular grid with a range of 50m for the multi-modal acquisition unit, combining the grid number with the geometric features and motion features in the video data, the average speed and stay time features in the GPS data, and the energy features in the environmental sound data, and then sequentially integrating the features to form a complete feature vector of the grid at the current timestamp; extracting the feature vectors of all grids to construct an undirected graph G=(V,E), where the node set V corresponds to the grid number, and the edge set E is established according to the adjacent relationship between the grids, and the weight of each edge is calculated by combining the distance from the grid center and the historical case co-occurrence frequency, the formula is:

[0015]

[0016] where i, j represent the grid number respectively, w i,j represents the weight of edge i, j, exp represents the exponential function, d center represents the distance scaling constant, d i,j represents the Euclidean distance between edge i, j and the center of the grid, σ represents the historical case frequency index, f i,j represents the number of historical case co-occurrences of the grid pair i, j, f max represents the highest co-occurrence number among all grid pairs;

[0017] An N*N matrix is constructed with the number of nodes N of the undirected graph, the edge weight of each edge is assigned to the corresponding position of the matrix, the non-existent edge is set to 0 to obtain the adjacency matrix, all values in each row are added, and the diagonal line is assigned to obtain the diagonal matrix, and the fused graph structure features are obtained by applying multiple graph convolution operations.

[0018] As a preferred embodiment, the specific steps of constructing the reinforcement learning scheduling environment and designing the hierarchical reward function are as follows:

[0019] The police records of the past month are obtained from the historical police record database, for each police record, the latitude and longitude are mapped to the corresponding grid, and the initial risk score is calculated according to the total number of cases in the past month and the average response time of the grid; the real-time scheduling system interface of the police command center is called to obtain the current police resources, and an initial environment resource vector is formed; the fused graph structure features of each grid are spliced in node order to form a global state vector, which is input into the policy network model to generate a scheduling action, and the immediate reward is calculated by the reward function, which is composed of the grid level and the regional level;

[0020] For the reward function of the grid level, according to the risk score of grid i at time t, if the risk score is higher than the preset high risk threshold and the police force is successfully dispatched to handle the police case, a positive reward is given If the risk score of grid i is less than the preset low risk threshold and no police case occurs within the subsequent Δt, it is considered as misdispatching and a negative punishment is given Where the preset high risk threshold and low risk threshold are obtained by analyzing the past historical cases, taking the 25th percentile of the risk scores of each grid in the past 30 days as the high risk threshold and the 75th percentile as the low risk threshold; the response delay of the police force from receiving the case to the scene is recorded and converted into a delay reward The rewards of the above three items for grid i are added to obtain the immediate reward of the grid at time t, and the immediate rewards of all M grids are summed to obtain the micro-level reward of this step:

[0021]

[0022] wherein R t represents the micro layer reward at time t, i represents the grid, M represents the total number of grids, r i,t represents the reward value of the grid i at time t;

[0023] For the reward function of the regional level, first, the evaluation period length of the regional level is set, in each evaluation period, all the police events dispatched to the high-risk grid and successfully handled are collected, and the average response time is calculated, the total working time and online time of all the police officers in the evaluation period are counted, the utilization rate is calculated, the allocation variance of the police force of each grid is counted, the average response time, the police force utilization rate, and the allocation variance are compared with the corresponding data obtained from the historical police event database, the average response time less than or equal to the candidate threshold is set to a reward coefficient D p , less than the candidate threshold is set to a penalty coefficient D n , the police force utilization rate greater than or equal to the minimum utilization rate target is set to a reward coefficient P g , less than the minimum utilization rate target is set to a penalty coefficient P n , the allocation variance less than or equal to the mean variance is set to a reward coefficient A p , greater than the mean variance is set to a penalty coefficient A n The above coefficients are summarized to obtain the reward function of the regional level, and the above two layer rewards are combined into the overall reward function according to the configurable weight.

[0024] As a preferred embodiment, the specific steps of the multi-scale hierarchical scheduling, closed-loop feedback and online correction are as follows:

[0025] First, the initial allocation of police resources is performed at the regional level, according to the historical high-risk event distribution and the average response time and police force utilization rate target in the regional level reward function, the allocation proportion of the police resources of each regional level is calculated, and the police resources are allocated in proportion to meet the requirements of the positive reward of the regional level;

[0026] At the beginning of the scheduling period, the latest risk score of each grid is extracted, and the grid risk is accumulated according to the pre-defined regional division to obtain the risk sum vector of each region, the regional weight vector is pre-set according to the historical data analysis, the risk sum is multiplied by the regional weight to obtain the weighted risk value, and the sum of the weighted risk values of all regions is summed to obtain the allocation proportion of the police resources of each region, the formula is as follows:

[0027]

[0028] wherein a represents the region, PR aa represents the police resource allocation proportion of region a, S represents the sum of all region weighted risk values, a represents the weighted risk value of region a;

[0029] The total number of currently schedulable police resources is multiplied by the police resource allocation proportion to calculate the allocation amount of police resources in each region, and then the optimal grid allocation action is selected in real time according to the grid level instant reward function, high-risk disposal, mis-scheduling penalty and response time incentive, to improve the grid level return. In the grid level scheduling process, when the instant reward of any grid is lower than the preset threshold, a police resource reallocation signal is sent to the region level, and the region level recalculates and adjusts the size of the police resources in each region according to the latest region level reward function, to ensure that high-risk regions obtain additional resources;

[0030] After completing a round of scheduling period, the region level reward return and the grid level cumulative instant reward of this period are summarized, the indicators are corrected, the police resource allocation proportion, the threshold of the reward coefficient and the grid level reward coefficient in the region level are fine-tuned through least squares regression, and the graph convolution feature fusion parameters are updated online, to ensure that the system can be continuously and adaptively optimized at each level.

[0031] The beneficial effects of the present application are:

[0032] 1. The present application can perform overall police pool allocation at half-day level and realize dynamic adjustment at grid level within hour level through the hierarchical cooperation and triggering mechanism of region level and grid level, thereby solving the problem of mismatch between short-term prediction and scheduling timeliness.

[0033] 2. The present application gives positive incentive to rapid disposal and low response time based on the hierarchical reward function, and the strategy learning will preferentially optimize the arrival speed and disposal success rate of high-risk events, thereby reducing the average response time and improving the disposal rate of high-risk events.

[0034] Legend

[0035] Figure 1 A flowchart of a police service system optimization method based on big data analysis.

[0036] Figure 2 An effect comparison chart of a police service system optimization method based on big data analysis. DETAILED DESCRIPTION

[0037] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application is further described below in combination with specific embodiments, but the following embodiments are only preferred embodiments of the present application, not all. Based on the embodiments in the embodiments, other embodiments obtained by those skilled in the art without creative labor also belong to the protection scope of the present application.

[0038] Embodiment 1, as Figure 1 The application discloses a police system optimization method based on big data analysis, comprising the following steps:

[0039] Multi-modal data acquisition processing, feature extraction and graph structure fusion;

[0040] Construct a reinforcement learning scheduling environment and design a hierarchical reward function;

[0041] Multi-scale hierarchical scheduling coordination, closed-loop feedback and online correction.

[0042] The following are specific implementation steps: a police system optimization method based on big data analysis, wherein the specific steps of multi-modal data acquisition processing, feature extraction and graph structure fusion are as follows:

[0043] Deploy multi-modal acquisition units at selected positions such as traffic lights, police car roofs and community sensing poles in the police command center, including video acquisition units, GPS acquisition units and environmental sound acquisition units; install local computing edge agents for each acquisition unit to perform preliminary preprocessing, keep time synchronization with the server of the police command center through the NTP protocol, perform YOLO model inference on each frame of image for video data to obtain the statistics of the number of vehicles and pedestrians, the YOLO model is a single-stage real-time target detection model that simultaneously completes target positioning and category prediction in one forward inference, divides the input image into fixed grids, and each grid is responsible for predicting the target bounding box in the region; for GPS data, remove abnormal points with a speed mutation greater than a preset threshold, and simultaneously use 3-order spline interpolation to complete the missing data points, for example, the GPS speed preset threshold is set to 50 m / s, and for environmental sound data, only data with a frequency of 300 to 3000 Hz is retained, and then divided into short-time data blocks for Fourier transform;

[0044] A regular rectangular grid with a range of 50 m is established for the multi-modal acquisition unit, the geometric features and motion features in the video data, the average speed and stay time features in the GPS data, and the energy features in the environmental sound data are combined with the grid number, and the features are sequentially spliced and integrated to form a complete feature vector of the grid at the current timestamp; extract the feature vectors in all grids to construct an undirected graph G=(V, E), wherein the node set V corresponds to the grid number, and the edge set E is established according to the adjacent relationship between the grids, and a weight is assigned to each edge, which is calculated by combining the distance from the grid center and the historical case co-occurrence frequency, and the formula is as follows:

[0045]

[0046] Wherein i, j represent the grid number respectively, w i,jdenotes the weight of edge i,j, exp denotes the exponential function, d center denotes the distance scaling constant, d i,j denotes the Euclidean distance between edge i,j and the grid center, σ denotes the historical case frequency index, f i,j denotes the historical case co-occurrence number of grid pair i,j, f max denotes the highest co-occurrence number among all grid pairs;

[0047] An N*N matrix is constructed with the number of nodes N of the undirected graph, and the edge weight of each edge is assigned to the corresponding position of the matrix. The non-existent edge is set to 0 to obtain the adjacency matrix. For example, the value of the i-th row and the j-th column in the matrix is w i,j , all values in each row are added, and the diagonal matrix is obtained by assigning values along the diagonal. The fused graph structure features are obtained by applying multiple graph convolution operations. The graph convolution operation specifically symmetrical normalizes the adjacency matrix with the diagonal matrix, then performs a weighted average of the features of each node and its neighbors in the adjacency matrix, multiplies it with the trainable weight matrix, and passes it through the activation function to obtain the new features of the fused nodes.

[0048] The specific steps of constructing a reinforcement learning scheduling environment and designing a hierarchical reward function are as follows:

[0049] The recent one-month police records are obtained from the historical police record database. For each police record, the latitude and longitude are mapped to the corresponding grid, and the initial risk score is calculated based on the total number of cases in the past month and the average response time of the grid. The real-time scheduling system interface of the police command center is called to obtain the current police resources to form the initial environment resource vector. The fused graph structure features of each grid are spliced in node order to form a global state vector, which is input into the policy network model to generate a scheduling action. The immediate return is calculated by the reward function, which is composed of the grid level and the regional level.

[0050] For the reward function of the grid level, if the risk score of grid i at time t is higher than the preset high risk threshold, and the police force is successfully dispatched to handle the police record, a positive reward is given If the risk score of grid i is less than the preset low risk threshold, and no police record occurs within the subsequent Δt, it is considered as misdispatching, and a negative punishment is given The preset high risk threshold and low risk threshold are determined by analyzing the past historical cases. The 25th percentile of the risk scores of each grid in the past 30 days is taken as the high risk threshold, and the 75th percentile is taken as the low risk threshold. The response delay of the police force from receiving the case to the scene is recorded and converted into a delay reward. Specifically, the ratio of response delay to maximum acceptable delay is calculated, and this ratio is subtracted from 1; the rewards for grid i are summed from the above three items to obtain the instantaneous reward for that grid at time t; the instantaneous rewards for all M grids are summed to form the micro-level reward for this step:

[0051]

[0052] Where R t Let i represent the micro-level reward at time t, M represent the total number of grid cells, and r represent the micro-level reward at time t. i,t This represents the reward value of grid i at time t;

[0053] For the reward function at the regional level, firstly, the evaluation period length at the regional level is set, for example, 3 hours. In each evaluation period, all incidents dispatched to high-risk grids and successfully handled are collected, and the average response time is calculated. The total working time and online time of all officers during the evaluation period are statistically analyzed, and the utilization rate is calculated. The variance of police force allocation for each grid is statistically analyzed. The average response time, police force utilization rate, and allocation variance are compared with corresponding data obtained from the historical incident database. Specifically, the 50th percentile of the response time data is used as the candidate threshold, and the 25th percentile of the police force utilization data is used as the minimum utilization rate target. The allocation variance is compared with the mean variance of historical data. A reward coefficient D is set if the average response time is less than or equal to the candidate threshold. p The penalty coefficient D is set if it is less than the candidate threshold. n A reward coefficient P is set for police force utilization rate that is greater than or equal to the minimum utilization rate target. g The penalty coefficient P is set for values ​​less than the minimum utilization target. n The reward coefficient A is set when the variance of the allocation is less than or equal to the mean of the variance. p The penalty coefficient A is set to be greater than the mean variance. n The above coefficients are summarized to obtain the regional level reward function, and the above two levels of rewards are combined into the overall reward function according to the configurable weights.

[0054] The specific steps of multi-scale hierarchical scheduling coordination, closed-loop feedback, and online correction are as follows:

[0055] First, initial allocation of police resources is carried out at the regional level. Based on the distribution of historical high-risk events and the average response time and police utilization rate targets in the regional level reward function, the allocation ratio of police resources at each regional level is calculated, and police resources are allocated proportionally to meet the requirements of positive rewards at the regional level.

[0056] Specifically, the calculation of the distribution ratio of each regional police resource, at the beginning of the scheduling period, extracts the latest risk score of each grid, and accumulates the grid risk according to the pre-defined regional division such as administrative district, patrol area, etc. to obtain the risk total vector of each region. According to the historical data analysis, the regional weight vector is preset, the risk total is multiplied by the regional weight to obtain the weighted risk value, and the sum of the weighted risk values of all regions is summed to obtain the police resource distribution ratio of each region, and the formula is:

[0057]

[0058] Where a represents the region, PR a represents the police resource distribution ratio of region a, S represents the sum of the weighted risk values of all regions, represents the weighted risk value of region a.

[0059] The total number of currently schedulable police resources is multiplied by the police resource distribution ratio to calculate the allocation of police resources in each region, and then the grid level real-time reward function is used to select the optimal grid allocation action in real time according to the high-risk disposal, mis-scheduling punishment and response time incentive, to improve the grid level return. During the grid level scheduling process, when the real-time reward of any grid is lower than the preset threshold, a police resource redistribution signal is sent to the regional level. After receiving the signal, the regional level recalculates and adjusts the size of the police resource in each region according to the latest regional level reward function, to ensure that high-risk areas obtain additional resources.

[0060] After completing a round of scheduling period, the regional level reward return and the grid level cumulative real-time reward of this period are summarized, and the indicators such as disposal success rate, mis-scheduling rate, average response time and police utilization rate are corrected. The least squares regression is used to fine-tune the police resource distribution ratio in the regional level, the threshold of the reward coefficient, and the grid level reward coefficient, and the graph convolution feature fusion parameter is updated online to ensure that the system can continuously and adaptively optimize at each level.

[0061] As Figure 2 An effect comparison chart of a central air conditioning energy-saving management control system based on artificial intelligence, the horizontal axis lists the key performance indicators, and the vertical axis is the performance score of the example, ranging from 0 to 100, and the higher the value, the better the performance, which aims to show the expected improvement of the invention compared with typical existing technology in key capabilities in an intuitive way.

[0062] Embodiment 2, based on the above embodiment 1, a police system optimization method based on big data analysis in the police resource scheduling and configuration scenario, specifically the following scheme:

[0063] Step one, divide the police area into equal-sized grids, deploy edge nodes in each grid, deploy lightweight agents that can execute federated learning protocols in each edge node, specifically extract target count and trajectory embedding vectors of the video via MobileNetV3 model, calculate short-time energy and MFCC features based on FPGA module parallel computing, and acoustic anomaly probability under adaptive threshold, align and unify the timestamps of each modality feature obtained by the agent of the edge node to generate local modality features, use time series graph attention network to generate dynamic graph adjacency matrix according to the spatio-temporal correlation and historical case co-occurrence frequency between nodes, and under the action of multi-head attention mechanism and trainable mapping matrix, weighted aggregation is performed on node fusion representation to obtain enhanced graph features;

[0064] Step two, instantiate micro-agent objects according to the schedulable police resources, instantiate macro-agents according to the number of region divisions, and according to the region risk summary and historical response in each macro period, automatically optimize the police pool distribution ratio through the conditional value at risk index, and the micro-agent adjusts the grid distribution in real time in each micro step by combining high-risk disposal reward, mis-scheduling penalty, response time delay reward and grid balance penalty; For the macro layer, a periodic reward function is designed, the average response time delay threshold is used to give a reward coefficient to the overall performance, and the police utilization rate deviation triggers positive and negative weighted returns, and for the micro layer, high-risk disposal, remote grid Manhattan distance penalty and resource reserve constraint are combined to build an instant reward, and the two-layer rewards are combined into a total return according to the preset weight;

[0065] Step three, first, the Bayesian confidence estimation of the risk sum of each region is performed according to the event severity weighting on the macro level, to dynamically adjust the initial size of the regional police pool, and a grid patrol agent based on meta-reinforcement learning is introduced on the micro level, which constructs a state representation based on real-time disposal performance and environmental uncertainty, and adaptively optimizes the patrol path in the sub-grid space; By comparing the predicted risk of each dispatch round with the actual disposal result, the micro and macro reward coefficients are automatically recalculated, and the graph convolution fusion parameters and high and low risk thresholds are real-time fine-tuned based on the incremental Bayesian optimization algorithm, so that small batch feedback can drive global policy parameter update.

[0066] The above describes the embodiments of the present application, without departing from the embodiments of the present application and its broader aspects, in specific operation, those skilled in the art can make data modifications and mode changes based on this place, the appended claims are for all such data modifications and mode changes that do not deviate from the embodiments of the present application.

Claims

1. A method for optimizing a policing system based on big data analysis, characterized in that, Comprise: Multimodal data acquisition processing, extracting feature fusion graph structure, dividing grid deploying edge node to obtain video, GPS, environmental sound data, taking each grid as a graph node, calculating the edge weight according to the distance between the grid centers and the historical case co-occurrence frequency, applying graph convolution operation to the feature matrix and adjacency matrix of the graph node to obtain the fused global state vector; Constructing a reinforcement learning scheduling environment, designing a hierarchical reward function, taking the global state vector as the environment observation, defining an immediate reward item for each grid at the grid level, including high-risk event handling positive reward, mis-scheduling penalty, response time reward proportional to the response time, calculating the average response time, police utilization rate and police allocation variance at the regional level, and giving corresponding periodic reward and penalty coefficients, and combining the two-layer rewards into a total reward function according to the configurable weight; Multiscale hierarchical scheduling coordination, closed-loop feedback and online correction, periodically allocating police pools as initial constraints based on the risk summary of the region at the regional level, and adjusting the police in the grid in real time through reinforcement learning at the grid level, and triggering the linkage signal of adjusting police resources according to the grid risk threshold, so that the regional allocation and grid dynamic scheduling form a closed-loop coordination.

2. The police system optimization method based on big data analysis according to claim 1, characterized in that: The specific steps of the multimodal data acquisition processing, extracting feature fusion graph structure are: Deploying multimodal acquisition units at the grid positions divided by the police command center, including: video acquisition unit, GPS acquisition unit, environmental sound acquisition unit; Install local computing edge agent for each acquisition unit, responsible for preliminary preprocessing, keep time synchronization with the server of the police command center through NTP protocol, for video data, execute YOLO model inference on each frame of image to get the statistics of vehicle and pedestrian number, for GPS data, remove abnormal points with speed mutation greater than a preset threshold, and simultaneously use 3-spline interpolation to complete the missing data points, for environmental sound data, only keep the data with frequency between 300 and 3000Hz, and then divide it into short-time data blocks for Fourier transform.

3. The method of claim 2, wherein: The specific steps of the multimodal data acquisition processing, extracting feature fusion graph structure further comprise: Establishing a regular rectangular grid with a range of 50m for the multimodal acquisition unit, combining the grid number, and integrating the geometric features and motion features in the video data, the average speed and stay time features in the GPS data, and the energy features in the environmental sound data in order to form a complete feature vector of the grid at the current timestamp; Extract the feature vector in all grids to construct an undirected graph G=(V,E), wherein the node set V corresponds to the grid number, and the edge set E is established according to the adjacent relationship between the grids, and the weight of each edge is calculated by combining the distance from the grid center and the historical case co-occurrence frequency, and the formula is: where i, j represent the grid numbers respectively, w i,j represents the weight of edge i, j, exp represents the exponential function, d center represents the distance scaling constant, d i,j represents the Euclidean distance between edge i, j and the grid center, σ represents the historical case frequency index, f i,j represents the historical case co-occurrence number of grid pair i, j, f max represents the highest co-occurrence number in all grid pairs.

4. The police system optimization method based on big data analysis according to claim 2, characterized in that: The specific steps of the multimodal data acquisition processing, extracting feature fusion graph structure further comprise: An N*N matrix is constructed with the number of nodes N of the undirected graph, and the edge weight of each edge is assigned to the corresponding position of the matrix. The adjacency matrix is obtained by setting the non-existent edge to 0, and the diagonal matrix is obtained by adding all values in each row and assigning according to the diagonal line. The fused graph structure features are obtained by applying multiple graph convolution operations.

5. The method of claim 1, wherein: The specific steps of constructing a reinforcement learning scheduling environment and designing a hierarchical reward function are as follows: The police service command center's real-time scheduling system interface is called to obtain the current police resources, and an initial environment resource vector is formed. The fused graph structure features of each grid are spliced in node order to form a global state vector, which is input into the policy network model to generate a scheduling action. The immediate reward is calculated by the reward function, which is composed of the grid level and the regional level. The specific steps of constructing a reinforcement learning scheduling environment and designing a hierarchical reward function also include:

6. The method of claim 5, wherein: The specific steps of constructing a reinforcement learning scheduling environment and designing a hierarchical reward function also include: For the reward function of the grid level, according to the risk score of the grid i at time t, if the risk score is higher than the preset high-risk threshold and the police force successfully handles the alarm, a positive reward is given If the police force is dispatched when the risk score of the grid i is less than the preset low-risk threshold, and no alarm occurs within the subsequent Δt, it is considered as misdispatching, and a negative punishment is given Wherein the preset high-risk threshold and low-risk threshold are obtained by analyzing past historical cases, taking the 25th percentile of the risk scores of each grid from high to low within 30 days as the high-risk threshold, and the 75th percentile as the low-risk threshold; the response time delay of the police force from receiving the case to the scene is recorded and converted into a time delay reward The rewards of the above three items for the grid i are added to obtain the instant reward of the grid at time t, and the instant rewards of all M grids are summed to construct the micro-level reward of this step: where R t represents the micro-layer reward at time t, i represents the grid, M represents the total number of grids, r i,t represents the reward value of the grid i at time t.

7. The method of claim 5, wherein: The specific steps of multi-scale hierarchical scheduling coordination, closed-loop feedback and online correction are as follows: For the reward function at the regional level, firstly, the evaluation period length at the regional level is set. In each evaluation period, all incidents dispatched to high-risk grids and successfully handled are collected, and the average response time is calculated. The total working hours and online time of all officers during the evaluation period are statistically analyzed, utilization rate is calculated, and the variance of police force allocation in each grid is statistically analyzed. The average response time, police force utilization rate, and allocation variance are compared with corresponding data obtained from the historical incident database. A reward coefficient D is set if the average response time is less than or equal to the candidate threshold. p The penalty coefficient D is set if it is less than the candidate threshold. n A reward coefficient P is set for police force utilization rate that is greater than or equal to the minimum utilization rate target. g The penalty coefficient P is set for values ​​less than the minimum utilization target. n The reward coefficient A is set when the variance of the allocation is less than or equal to the mean of the variance. p The penalty coefficient A is set to be greater than the mean variance. n The above coefficients are summarized to obtain the regional level reward function, and the above two levels of rewards are combined into the overall reward function according to the configurable weights.

8. The police system optimization method based on big data analysis according to claim 1, characterized in that: First, the initial allocation of police resources is performed at the regional level. According to the historical distribution of high-risk events and the average response time and police utilization rate targets in the regional level reward function, the allocation proportion of police resources at the regional level is calculated, and the police resources are allocated in proportion to meet the requirements of the regional level positive reward. The specific steps of multi-scale hierarchical scheduling coordination, closed-loop feedback and online correction also include: 9.The police system optimization method based on big data analysis of claim 8, wherein: At the beginning of the scheduling period, the latest risk score of each grid is extracted, and the grid risk is accumulated according to the pre-defined regional division to obtain the risk total vector of each region. According to the historical data analysis, the regional weight vector is pre-set, and the risk total is multiplied by the regional weight to obtain the weighted risk value. The weighted risk values of all regions are summed to obtain the police resource allocation proportion of each region, which is calculated as follows: The specific steps of multi-scale hierarchical scheduling coordination, closed-loop feedback and online correction also include: where a represents a region, PR a represents the proportion of police resources allocated to region a, S represents the sum of the weighted risk values for all regions, represents the weighted risk value for region a.

10. The method of claim 8, wherein: The total number of currently schedulable police resources is multiplied by the police resource allocation proportion to calculate the allocation amount of police resources in each region. According to the grid level immediate reward function, the optimal grid allocation action is selected in real time to improve the grid level return. During the grid level scheduling process, when the immediate reward of any grid is lower than the pre-set threshold, a police resource reallocation signal is sent to the regional level. The regional level receives the signal and recalculates and adjusts the size of the police resources in each region according to the latest regional level reward function to ensure that high-risk regions obtain additional resources; After completing a scheduling period, the regional level reward return and the grid level cumulative immediate reward of this period are summarized, and the indicators are corrected. The police resource allocation proportion, reward coefficient threshold in the regional level, and grid level reward coefficient are fine-tuned through least squares regression, and the graph convolution feature fusion parameters are updated online to ensure that the system can continuously and adaptively optimize at each level. ​

Images