A regulatory environment risk prediction method and system based on spatiotemporal correlation

Abstract

The present invention relates to a method and system for predicting regulatory environment risks based on spatiotemporal correlation, which belongs to the field of artificial intelligence and includes acquiring data to obtain regulatory environment data information including time and space dimensions; data preprocessing; using the STGCN spatiotemporal correlation network to regularly train and learn online through the data information model acquired in real time, and continuously correcting and optimizing the model by updating the model weight parameters to reduce learning errors and improve accuracy; for prediction results exceeding the risk threshold, it is considered to be a risk area, and the system automatically issues a risk warning signal. The present invention captures the sequence information of the time dimension and the space dimension by establishing a spatiotemporal relationship model, and based on spatiotemporal correlation analysis, it can quickly and efficiently accurately predict the future risks of the regulatory environment, and intuitively display the risks through visualization technology to assist police in the safety management and risk prevention of the regulatory environment, and ensure the long-term stability of the regulatory environment.

Description

Technical Field

[0001] The present invention relates to a method and system for predicting regulatory environment risks based on spatiotemporal correlation, and belongs to the field of artificial intelligence technology. Background Art

[0002] As the national penal enforcement agency, the supervisory environment primarily carries out the detention of criminals. Its safety and stability significantly impact social stability. While the supervisory environment is currently at its best, numerous challenges remain, including a large number of detainees, a shortage of police officers, a complex demographic of supervised individuals, and significant management challenges. These challenges highlight the ongoing challenges facing supervisory security. In recent years, with the continued maturity of artificial intelligence (AI) technology, the support of massive amounts of big data, and the advancement of cloud computing capabilities, AI has begun to be widely applied across multiple sectors of social development. Smart supervisory environments are an innovative application of AI in this era.

[0003] At present, the current status of regulatory environment risk management in my country is mostly still at the stage of police patrols and camera surveillance and alarms. Manual operations still account for the vast majority, and the degree of informatization is relatively low. At present, information technology and network technology have entered all walks of life. Modern social security risks and risk warnings have also become more intelligent and covert. The in-depth application of big data intelligent technology can effectively improve the public security police's ability to combat crime and ensure social security. In order to improve the level of regulatory risk identification, the relevant technologies of machine learning are used to integrate relevant regulatory data, extract regulatory environment risk-related characteristics and risk assessment tables, and use multidisciplinary capabilities such as big data, database processing technology, computer software technology, geographic information system technology, and Internet technology to realize the informatization of regulatory environment management. Intelligent retrieval is particularly important. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a regulatory environment risk prediction method and system based on spatiotemporal correlation. By establishing a spatiotemporal relationship model, it captures the sequence information of the time dimension and the space dimension. Based on spatiotemporal correlation analysis, it can quickly and efficiently predict the future risks of the regulatory environment, and intuitively display the risks through visualization technology to assist police in the safety management and risk prevention of the regulatory environment, thereby ensuring the long-term stability of the regulatory environment.

[0005] Explanation of terms:

[0006] STGCN: Spatio-Temporal Graph Convolutional Networks (STGCN), which uses convolutional layers to perform spatiotemporal predictions on neural networks.

[0007] The present invention adopts the following technical solutions:

[0008] A regulatory environment risk prediction method based on spatiotemporal correlation includes the following steps:

[0009] (1) Data acquisition: Access the back-end interface of the regulatory environment big data system platform to upload data in real time, and obtain regulatory environment data information including time and space dimensions;

[0010] (2) Data preprocessing: preprocessing the acquired data information, including missing value filling, maximum and minimum, normalization, dimensionality reduction, data cleaning and other data preprocessing operations;

[0011] (3) Model training and optimization: The model adopts the STGCN spatiotemporal correlation network, and the regulatory environment big data system platform is embedded with the risk prediction model proposed in this invention. The model is regularly trained and learned online through real-time data information, and the model weight parameters are continuously updated to continuously correct and optimize the model to reduce learning errors and improve accuracy.

[0012] (4) Risk warning: Set a risk threshold. If the predicted results exceed the risk threshold, it is considered a risk area. The system will automatically send out a risk warning signal to remind the environmental police to conduct targeted inspections and provide awareness reminders to the staff in the relevant areas.

[0013] Preferably, the data information in step (1) includes the location information of criminals in the supervision environment, the entry and exit information of vehicles of outsiders, the operation information of equipment in the supervision environment, historical alarm information, and temperature changes in the supervision environment.

[0014] Preferably, the method for constructing the model in step (3) includes defining a regulatory environment network diagram and constructing a regulatory environment spatiotemporal correlation prediction model, wherein defining the regulatory environment network diagram includes:

[0015] A. Defining the spatial weight matrix of the regulatory environment

[0016] Assuming that the regions are surface elements and there are common edges between the regions, the adjacent regions are determined by constructing the adjacency space weight matrix based on the common edges;

[0017] The regulatory environment areas are marked as points, and the spatial weight matrix is ​​defined according to the geographical distribution map of the regulatory environment. The following formula is defined:

[0018]

[0019] Among them, a ij Indicates the adjacent relationship between regions i and j. If they are adjacent, then a ij =1, otherwise it is 0. The spatial weight matrix is ​​a quantitative representation of the adjacent relationship;

[0020] The spatial weight matrix is ​​constructed by measuring the spatial pattern through the definition of spatial adjacency. Traditional spatial adjacency is defined by the adjacency or distance between regions, and the spatial weight matrix is ​​constructed based on the distance. The specific operation is that the smaller the distance between spaces, the closer the correlation. The present invention uses adjacency to construct the spatial weight matrix, which can better reflect the adjacency relationship between spaces.

[0021] B. Define the regulatory environment undirected graph

[0022] First, the entire regulatory environment is represented as an undirected graph G(V, E, A), where each region in the regulatory environment can be regarded as a vertex, where V = {V 1, V 2, …V s} represents the set of regions, s represents the number of regions, E is the set of edges, representing the connectivity between regions, A∈R s*s It is the adjacency matrix of the undirected graph G, that is, the spatial weight matrix defined above, which represents the connection relationship between regions;

[0023] In the spatial dimension, the risk of each region is affected by the risk of the adjacent regions. In the temporal dimension, the risk of the region is affected by the risk at the historical moment. In the spatiotemporal dimension, the risk of the vertex in the undirected graph is affected by the neighborhood around the moment of the historical time record unit. Then the risk of region n at time t can be expressed as To express it, where n∈s, the risk of all regions at time t can be expressed as The entire regulatory environment can be represented by an undirected graph G, where each vertex has a feature vector consisting of historical risks, and graph convolution operations can be used to capture the irregular spatiotemporal dependencies in the regulatory environment.

[0024] Preferably, the process of constructing a spatiotemporal correlation prediction model for the regulatory environment is:

[0025] The core of the spatiotemporal correlation prediction model is the spatiotemporal convolutional network (STGCN), which consists of two spatiotemporal convolution blocks (ST-Conv Block) and an output fully connected layer (Output Layer). The core spatiotemporal convolution block is composed of two spatiotemporal gated convolutions (Temporal Gated-Conv) and a spatial graph convolution (Spatial Graph-Conv) in the middle. The structure is like a sandwich, with two temporal gated sequence convolution layers and a spatial graph convolution module in the middle. Temporal Gated-Conv is used to capture temporal correlation, and Spatial Graph-Conv is used to capture spatial correlation. The spatiotemporal gated convolution (Temporal Gated-Conv) is composed of a one-dimensional convolution (1-D Conv) and a gated linear unit (GLU).

[0026] For the extraction of temporal features, STGCN uses a gated convolutional neural network. The difference from the traditional spatiotemporal model is that it uses one-dimensional convolution to learn information in the time dimension. The advantage of the convolutional neural network is that it is not restricted by the prediction of data at previous time points, so it can better capture changes and has low parallel training consumption.

[0027] ①Use graph convolutional neural network to capture spatial characteristics

[0028] To extract the spatial characteristics of the regulatory environment, we use a graph convolutional neural network (GCN). A graph convolutional neural network applies convolution to graph-structured data. Conventional convolution operations are only suitable for processing Euclidean data, which has a regular spatial structure. For graph-structured data, the number of neighbor nodes (neighbor nodes are the adjacent nodes of each node in the graph structure) is uncertain, making it impossible to select a fixed-size convolution kernel to adapt to irregular graphs. Therefore, we use graph convolution.

[0029] The essence of graph convolution is to extract the spatial features of the graph structure and update the new node representation by aggregating the information of edges and nodes.

[0030]

[0031]

[0032]

[0033] Among them, A is the adjacency matrix, I N is the identity matrix, In this way, the information of the node itself and the neighboring nodes can be taken into account. It is used to normalize the nodes, so the left and right sides are multiplied by H(l) is the feature matrix of the node, W (l) is a parameter matrix, and σ is an activation function. Through the graph convolutional neural network, the information of the surrounding neighborhood can be effectively aggregated. In this way, the risk of a certain area in the regulatory environment is not only related to itself, but also affected by the surrounding areas, realizing the correlation of spatial dimensions;

[0034] ②Use gated convolutional neural network to extract temporal features

[0035] Among previous methods for extracting temporal features, recurrent neural networks are the most commonly used. However, given the ever-changing nature of risk prediction and the large amount of data involved, recurrent neural networks struggle to capture temporal characteristics. Furthermore, their complex gating mechanisms and multiple iterative training consume valuable computing power and time. Compared to recurrent neural networks, gated convolutional neural networks (GCNNs) can better adapt to ever-changing risk and can be trained in parallel, significantly reducing training time. Therefore, in the temporal dimension, this paper uses gated convolutional networks to capture temporal dependencies.

[0036] The gated convolutional neural network consists of a one-dimensional convolution and a gated linear unit (GLU), using causal convolution. For each vertex in the undirected graph G, the input X is a risk history observation of length M, and the output Y is:

[0037] Y=Γ*X=P⊙σ(Q)

[0038] Among them, Γ is the convolution kernel, P and Q are the inputs of the GLU gate respectively, ⊙ represents the Hadamard product, and σ is the nonlinear activation function. This formula represents the dot product of the output of a convolution layer without a nonlinear activation function and the output after the nonlinear activation function σ. The gated convolutional network can capture the long-term dependency of the historical risk data of each node, thereby realizing the association in the time dimension.

[0039] ③Build the entire spatiotemporal convolutional network

[0040] The entire spatiotemporal convolutional network consists of two temporal convolutional layers and a spatial graph convolutional layer in the middle. The spatial convolutional layer in the middle is connected to the two temporal convolutional layers in the front and back, which enables the spatial state to be quickly propagated between graph convolution and temporal convolution. Through the spatiotemporal convolutional layer, it captures both the correlation information in the time dimension and the correlation information in the spatial dimension, realizing the correlation between time and space. Regulatory environmental risks involve people, things and places, and are the result of the joint action of time and space. The use of a spatiotemporal correlation prediction model can capture risk characteristics at the largest scale and make accurate predictions.

[0041] Preferably, when defining the undirected graph of the supervision environment in step B, each area in the supervision environment refers to dividing the supervision environment into several areas according to the geographical structure, covering cells, corridors, restaurants, activity areas, factory workshops, meeting rooms, lounges, etc.

[0042] A regulatory environment risk prediction system based on spatiotemporal correlation, comprising a data acquisition and storage module, a model algorithm optimization module, a risk warning module, and a data display and query module;

[0043] The data acquisition and storage module is used to dynamically acquire data information within the regulatory environment, including time and space information, and perform data analysis and preprocessing;

[0044] The model algorithm optimization module is used to obtain risk prediction results. By reading changing data information in real time, online learning and updating model weights can reduce learning errors.

[0045] The risk warning module is used to issue risk warning signals based on the risk prediction results within the regulatory environment. For example, it can issue warning signals for medium and high risks. Police officers receive this information and pay special attention to it, taking corresponding measures to prevent dangerous events from occurring within the regulatory environment.

[0046] The data display and query module provides a display interface for real-time display and query of risk forecasts within the current regulatory environment. Risk forecasts are divided into three levels: high, medium, and low (the standards for high, medium, and low risks can be flexibly set according to actual needs), and are visualized in the form of a heat map, so that the risk situation within the regulatory environment is clearly visible.

[0047] Anything not described in detail in the present invention can be carried out using existing technologies.

[0048] The beneficial effects of the present invention are:

[0049] The present invention provides a regulatory environment risk prediction method based on spatiotemporal correlation, and proposes a prediction method based on spatiotemporal convolution blocks to capture regulatory environment risks. It performs spatiotemporal correlation analysis from the two dimensions of time and space, captures risk characteristics at the largest scale, and effectively improves the prediction accuracy of regulatory environment risks. At the same time, the use of spatiotemporal convolution can also reduce computational complexity and model complexity. This method provides theoretical support, basis and reference for risk prediction management and prevention and control work in the regulatory environment.

[0050] The present invention provides a regulatory environment risk prediction system based on spatiotemporal correlation, which includes a data acquisition and storage module, a model algorithm optimization module, a risk warning module, and a data display and query module. By combining the regulatory environment database, algorithm, and terminal, a complete regulatory environment risk prediction system is formed, which brings great convenience to the regulatory environment risk management work and effectively reduces the occurrence of various risk accidents in the regulatory environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of the regulatory environment risk prediction method based on spatiotemporal correlation according to the present invention;

[0052] Figure 2 This is a structural block diagram of the regulatory environment risk prediction system based on spatiotemporal correlation of the present invention;

[0053] Figure 3 To provide a geographical map of the regulatory environment;

[0054] Figure 4 To represent the regulatory environment as a diagram of a graph structure;

[0055] Figure 5 Schematic diagram of the STGCN structure;

[0056] Figure 6 Schematic diagram of spatial convolution structure;

[0057] Figure 7 Schematic diagram of the temporal convolution structure;

[0058] Figure 8 Schematic diagram of the spatiotemporal convolution block. DETAILED DESCRIPTION

[0059] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, they will be described in detail below with reference to the accompanying drawings and specific embodiments, but are not limited thereto. Matters not fully described in the present invention shall be based on conventional techniques in the art.

[0060] Example 1

[0061] A regulatory environment risk prediction method based on spatiotemporal correlation, such as Figure 1 As shown, the following steps are included:

[0062] (1) Data acquisition: Data is uploaded in real time through the back-end interface of the supervision environment big data system platform to obtain supervision environment data information with time and space dimensions. The data information includes the location information of criminals in the supervision environment, the entry and exit information of vehicles of outsiders, the operation information of equipment in the supervision environment, historical alarm information, and temperature changes in the supervision environment.

[0063] (2) Data preprocessing: preprocessing the acquired data information, including missing value filling, maximum and minimum, normalization, dimensionality reduction, data cleaning and other data preprocessing operations;

[0064] (3) Model training and optimization: The model adopts the STGCN spatiotemporal correlation network, and the regulatory environment big data system platform is embedded with the risk prediction model proposed in this invention. The model is regularly trained and learned online through real-time data information, and the model weight parameters are continuously updated to continuously correct and optimize the model to reduce learning errors and improve accuracy.

[0065] (4) Risk warning: Set a risk threshold. If the predicted results exceed the risk threshold, it is considered a risk area. The system will automatically send out a risk warning signal to remind the environmental police to conduct targeted inspections and provide awareness reminders to the staff in the relevant areas.

[0066] Example 2

[0067] A regulatory environment risk prediction method based on spatiotemporal correlation is as described in Example 1, except that the method for constructing a model in step (3) includes defining a regulatory environment network diagram and constructing a regulatory environment spatiotemporal correlation prediction model, wherein defining the regulatory environment network diagram includes:

[0068] A. Defining the spatial weight matrix of the regulatory environment

[0069] Assuming that the regions are surface elements and there are common edges between the regions, the adjacent regions are determined by constructing the adjacency space weight matrix based on the common edges;

[0070] like Figure 3 As shown, the regulatory environment areas are marked as points, and the spatial weight matrix is ​​defined according to the geographical distribution map of the regulatory environment. The following formula is defined:

[0071]

[0072] Among them, a ij Indicates the adjacent relationship between regions i and j. If they are adjacent, then a ij =1, otherwise it is 0. The spatial weight matrix is ​​a quantitative representation of the adjacent relationship;

[0073] The spatial weight matrix is ​​constructed by measuring the spatial pattern through the definition of spatial adjacency. Traditional spatial adjacency is defined by the adjacency or distance between regions, and the spatial weight matrix is ​​constructed based on the distance. The specific operation is that the smaller the distance between spaces, the closer the correlation. The present invention uses adjacency to construct the spatial weight matrix, which can better reflect the adjacency relationship between spaces.

[0074] B. Define the regulatory environment undirected graph

[0075] First, the entire supervision environment is represented as an undirected graph G(V, E, A), where each region in the supervision environment can be regarded as a vertex. Each region in the supervision environment refers to the supervision environment divided into several areas according to the geographical structure, covering cells, corridors, restaurants, activity areas, factory workshops, meeting rooms, lounges, etc., where V = {V 1, V 2, …V s} represents the set of regions, s represents the number of regions, E is the set of edges, representing the connectivity between regions, A∈R s*s It is the adjacency matrix of the undirected graph G, that is, the spatial weight matrix defined above, which represents the connection relationship between regions;

[0076] In the spatial dimension, the risk of each region is affected by the risk of the adjacent regions. In the temporal dimension, the risk of the region is affected by the risk at the historical moment. In the spatiotemporal dimension, the risk of the vertex in the undirected graph is affected by the neighborhood around the moment of the historical time record unit. Then the risk of region n at time t can be expressed as To express it, where n∈s, the risk of all regions at time t can be expressed as like Figure 4 As shown, the entire regulatory environment can be represented by an undirected graph G, where each vertex has a feature vector consisting of historical risks, and graph convolution operations can be used to capture the irregular spatiotemporal dependencies in the regulatory environment.

[0077] Preferably, the process of constructing a spatiotemporal correlation prediction model for the regulatory environment is:

[0078] like Figure 5 As shown in the figure, the core of the spatiotemporal correlation prediction model is the spatiotemporal convolutional network (STGCN), which includes two spatiotemporal convolution blocks (ST-Conv Block) and an output fully connected layer (Output Layer). The core spatiotemporal convolution block is composed of two spatiotemporal gated convolutions (Temporal Gated-Conv) and a spatial graph convolution (Spatial Graph-Conv) in the middle. The structure is like a sandwich, with two time-gated sequence convolution layers and a spatial graph convolution module in the middle. Temporal Gated-Conv is used to capture temporal correlation, and Spatial Graph-Conv is used to capture spatial correlation. The spatiotemporal gated convolution (Temporal Gated-Conv) is composed of a one-dimensional convolution (1-D Conv) and a gated linear unit (GLU).

[0079] For the extraction of temporal features, STGCN uses a gated convolutional neural network. The difference from the traditional spatiotemporal model is that it uses one-dimensional convolution to learn information in the time dimension. The advantage of the convolutional neural network is that it is not restricted by the prediction of data at previous time points, so it can better capture changes and has low parallel training consumption.

[0080] ①Use graph convolutional neural network to capture spatial characteristics

[0081] To extract the spatial characteristics of the regulatory environment, we use a graph convolutional neural network (GCN). A graph convolutional neural network applies convolution to graph-structured data. Conventional convolution operations are only suitable for processing Euclidean data, which has a regular spatial structure. For graph-structured data, the number of neighbor nodes (neighbor nodes are the adjacent nodes of each node in the graph structure) is uncertain, making it impossible to select a fixed-size convolution kernel to adapt to irregular graphs. Therefore, we use graph convolution.

[0082] The essence of graph convolution is to extract the spatial features of the graph structure and update the new node representation by aggregating the information of edges and nodes, such as Figure 6 ;

[0083]

[0084]

[0085]

[0086] Among them, A is the adjacency matrix, I N is the identity matrix, In this way, the information of the node itself and the neighboring nodes can be taken into account. It is used to normalize the nodes, so the left and right sides are multiplied by H (l) is the feature matrix of the node, W (l) is a parameter matrix, and σ is an activation function. Through the graph convolutional neural network, the information of the surrounding neighborhood can be effectively aggregated. In this way, the risk of a certain area in the regulatory environment is not only related to itself, but also affected by the surrounding areas, realizing the correlation of spatial dimensions;

[0087] ②Use gated convolutional neural network to extract temporal features

[0088] Among previous methods for extracting temporal features, recurrent neural networks are the most commonly used. However, given the ever-changing nature of risk prediction and the large amount of data involved, recurrent neural networks struggle to capture temporal characteristics. Furthermore, their complex gating mechanisms and multiple iterative training consume valuable computing power and time. Compared to recurrent neural networks, gated convolutional neural networks (GCNNs) can better adapt to ever-changing risk and can be trained in parallel, significantly reducing training time. Therefore, in the temporal dimension, this paper uses gated convolutional networks to capture temporal dependencies.

[0089] Gated convolutional neural networks such as Figure 7 As shown, it contains a one-dimensional convolution and a gated linear unit (GLU), using causal convolution. For each vertex in the undirected graph G, the input X is a risk history observation value of length M, and the output Y is:

[0090] Y=Γ*X=P⊙σ(Q)

[0091] Among them, Γ is the convolution kernel, P and Q are the inputs of the GLU gate respectively, ⊙ represents the Hadamard product, and σ is the nonlinear activation function. This formula represents the dot product of the output of a convolution layer without a nonlinear activation function and the output after the nonlinear activation function σ. The gated convolutional network can capture the long-term dependency of the historical risk data of each node, thereby realizing the association in the time dimension.

[0092] ③Build the entire spatiotemporal convolutional network

[0093] like Figure 8 As shown in the figure, the entire spatiotemporal convolutional network includes two temporal convolutional layers and a spatial graph convolutional layer in the middle. The spatial convolutional layer in the middle is connected to the two temporal convolutional layers, which enables the spatial state to be quickly propagated between graph convolution and temporal convolution. Through the spatiotemporal convolutional layer, both the correlation information in the time dimension and the correlation information in the spatial dimension are captured, realizing the correlation between time and space. Regulatory environmental risks involve people, things and places, and are the result of the joint action of time and space. The prediction model of spatiotemporal correlation can capture risk characteristics at the largest scale and make accurate predictions.

[0094] Example 3

[0095] A regulatory environment risk prediction system based on spatiotemporal correlation, such as Figure 2 As shown, it includes data acquisition and storage module, model algorithm optimization module, risk warning module and data display and query module;

[0096] The data acquisition and storage module is used to dynamically acquire data information within the regulatory environment, including time and space information, and perform data analysis and preprocessing;

[0097] The model algorithm optimization module is used to obtain risk prediction results. By reading changing data information in real time, online learning and updating model weights can reduce learning errors.

[0098] The risk warning module is used to issue risk warning signals based on the risk prediction results within the regulatory environment. For example, it can issue warning signals for medium and high risks. Police officers receive this information and pay special attention to it, taking corresponding measures to prevent dangerous events from occurring within the regulatory environment.

[0099] The data display and query module provides a display interface for real-time display and query of risk forecasts within the current regulatory environment. Risk forecasts are divided into three levels: high, medium, and low (the standards for high, medium, and low risks can be flexibly set according to actual needs), and are visualized in the form of a heat map, so that the risk situation within the regulatory environment is clearly visible.

[0100] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A regulatory environment risk prediction method based on spatiotemporal correlation, characterized in that: The method includes the following steps: (1) Data acquisition: Access the back-end interface of the regulatory environment big data system platform to upload data in real time, and obtain regulatory environment data information including time and space dimensions; (2) Data preprocessing: preprocessing the acquired data information, including missing value filling, maximum minimization, normalization, dimensionality reduction, and data cleaning; (3) Model training and optimization: The model adopts the STGCN spatiotemporal correlation network, and regularly trains and learns online through the data information model obtained in real time. The model is continuously corrected and optimized by updating the model weight parameters to reduce learning errors and improve accuracy; (4) Risk warning: Set a risk threshold. If the predicted result exceeds the risk threshold, it is considered a risk area. The system will automatically send out a risk warning signal to remind the environmental police to conduct targeted inspections and remind the staff in the area to be aware of the risk. The method for constructing the model in step (3) includes defining a regulatory environment network diagram and constructing a regulatory environment spatiotemporal correlation prediction model, wherein defining the regulatory environment network diagram includes: A. Defining the spatial weight matrix of the regulatory environment Assuming that the regions are surface elements and there are common edges between the regions, the adjacent regions are determined by constructing the adjacency space weight matrix based on the common edges; The regulatory environment areas are marked as points, and the spatial weight matrix is ​​defined according to the geographical distribution map of the regulatory environment. The following formula is defined: Among them, a ij Indicates the adjacent relationship between regions i and j. If they are adjacent, then a ij =1, otherwise it is 0. The spatial weight matrix is ​​a quantitative representation of the adjacent relationship; B. Define the regulatory environment undirected graph First, the entire regulatory environment is represented as an undirected graph G(V, E, A), where each region in the regulatory environment is considered as a vertex, where V = {V 1, V 2, …V s } represents the set of regions, s represents the number of regions, E is the set of edges, representing the connectivity between regions, A∈R s*s It is the adjacency matrix of the undirected graph G, that is, the spatial weight matrix, which represents the connection relationship between regions; In the spatial dimension, the risk of each region is affected by the risk of the adjacent regions. In the temporal dimension, the risk of the region is affected by the risk at the historical moment. In the spatiotemporal dimension, the risk of the vertex in the undirected graph is affected by the neighborhood around the moment of the historical time record unit. Then the risk of region n at time t is expressed as To express it, where n∈s, the risk of all regions at time t is expressed as The entire regulatory environment is represented by an undirected graph G, where each vertex has a feature vector consisting of historical risk; The core of the spatiotemporal correlation prediction model is the spatiotemporal convolutional network, which includes two spatiotemporal convolutional blocks and an output fully connected layer. The core spatiotemporal convolutional block is formed by two spatiotemporal gated convolutions and a spatial graph convolution in the middle. The spatiotemporal gated convolution is composed of a one-dimensional convolution and a gated linear unit.

2. The regulatory environment risk prediction method based on spatiotemporal correlation according to claim 1 is characterized in that: The data information in step (1) includes the location information of criminals in the supervision environment, the entry and exit information of vehicles of outsiders, the operation information of equipment in the supervision environment, historical alarm information, and the temperature change information in the supervision environment.

3. The regulatory environment risk prediction method based on spatiotemporal correlation according to claim 2 is characterized in that: The process of building a spatiotemporal correlation prediction model for the regulatory environment is as follows: ①Use graph convolutional neural network to capture spatial characteristics Graph convolutional neural networks are used for extraction. Graph convolutional neural networks apply convolution to graph-structured data. The essence of graph convolution is to extract the spatial features of the graph structure and update new node representations by aggregating information from edges and nodes. Among them, A is the adjacency matrix, I N is the identity matrix, It is used to normalize the nodes, so the left and right sides are multiplied by H (l) is the feature matrix of the node, W (l) is the parameter matrix, σ is the activation function; ②Use gated convolutional neural network to extract temporal features The gated convolutional neural network consists of a one-dimensional convolution and a gated linear unit, using causal convolution. For each vertex in the undirected graph G, the input X is a risk history observation of length M, and the output Y is: Y=Γ*X=P⊙σ(Q) Where Γ is the convolution kernel, P and Q are the inputs of the GLU gate, ⊙ represents the Hadamard product, and σ is the nonlinear activation function; ③Build the entire spatiotemporal convolutional network The entire spatiotemporal convolutional network consists of two temporal convolutional layers and a spatial graph convolutional layer in the middle. The spatial convolutional layer in the middle is connected to the two temporal convolutional layers in front and behind.

4. The regulatory environment risk prediction method based on spatiotemporal correlation according to claim 3 is characterized in that: When defining the undirected graph of the supervision environment in step B, each area in the supervision environment refers to dividing the supervision environment into several areas according to the geographical structure, covering cells, corridors, restaurants, activity areas, factory workshops, meeting rooms, and lounges.

5. A regulatory environment risk prediction system based on spatiotemporal correlation, characterized in that: Used to implement the regulatory environment risk prediction method based on spatiotemporal correlation as described in claim 1, including a data acquisition and storage module, a model algorithm optimization module, a risk warning module and a data display and query module; The data acquisition and storage module is used to dynamically acquire data information within the regulatory environment, including time and space information, and perform data analysis and preprocessing; The model algorithm optimization module is used to obtain risk prediction results. By reading changing data information in real time, online learning and updating model weights can reduce learning errors. The risk warning module is used to issue risk warning signals for risk prediction results within the regulatory environment; The data display and query module provides a display interface for real-time display and query of risk predictions within the current regulatory environment. Risk predictions are divided into three levels: high, medium, and low, and are visualized in the form of heat maps.

Images