Regional risk identification method, graph neural network training method, system and medium

Abstract

This disclosure provides a method for identifying regional risks, a graph neural network training method, a device for identifying regional risks, a computer system, a readable storage medium, and a computer program product, relating to the field of smart city technology. The method for identifying regional risks includes: generating a regional map including the region to be identified, wherein the regional map includes multiple nodes and multiple connecting edges, nodes are used to represent regions, regions are areas where transport vehicles used for transporting goods stop along their routes, and connecting edges are used to represent the association relationship between two nodes; extracting features from the regional map to obtain a target feature vector; and obtaining the risk identification result of the region to be identified in the regional map based on the target feature vector.

Description

Technical Field

[0001] This disclosure relates to the field of smart city technology, and more particularly to the field of urban governance technology, specifically to a method for identifying regional risks, a graph neural network training method, an apparatus, a computer system, a readable storage medium, and a computer program product. Background Technology

[0002] Hazardous chemicals (HCCs) are widely used and have become an indispensable part of people's daily lives, such as acids used in battery manufacturing, ammonium nitrates used in fertilizer production, and liquefied petroleum gas used in cooking. However, due to their flammable, explosive, and toxic properties, HCCs require strict supervision throughout their entire lifecycle by relevant departments; otherwise, they could pose a threat to public safety in cities. Therefore, areas used for the production and storage of HCCs, referred to as Hazardous Chemical-related Locations (HCLs), will become key areas of focus.

[0003] In realizing the present invention, the inventors discovered that the related technologies have at least the following problems: low efficiency and low accuracy in identifying hazardous chemical sites. Summary of the Invention

[0004] In view of this, the present disclosure provides a method for identifying regional risks, a method for training graph neural networks, an apparatus, a computer system, a readable storage medium, and a computer program product.

[0005] One aspect of this disclosure provides a method for identifying regional risks, comprising: generating a regional map including a region to be identified, wherein the regional map includes multiple nodes and multiple connecting edges, the nodes being used to represent regions, the regions being areas where transport vehicles for transporting goods stop along their routes, and the connecting edges being used to represent the association relationship between two of the nodes; extracting features from the regional map to obtain a target feature vector; and obtaining a risk identification result for the region to be identified based on the target feature vector.

[0006] Another aspect of this disclosure provides a graph neural network training method, comprising: generating a sample region map including a target sample region, wherein the sample region map includes multiple sample nodes and multiple sample connection edges, the sample nodes are used to represent sample regions, the sample regions are areas where transport vehicles for transporting goods stop along their routes, the sample connection edges are generated based on the driving trajectory of the transport vehicles, and the sample connection edges are used to represent the association relationship between two sample nodes; determining sample labels related to the target sample region; and training a graph neural network using the sample region map and the sample labels to obtain a trained graph neural network.

[0007] Another aspect of this disclosure provides a regional risk identification device, comprising: a first generation module for generating a regional map including a region to be identified, wherein the regional map includes multiple nodes and multiple connecting edges, the nodes are used to represent regions, the regions are areas where transport vehicles for transporting goods stop along their routes, and the connecting edges are used to represent the association relationship between two of the nodes; a feature extraction module for performing feature extraction on the regional map to obtain a target feature vector; and an identification module for obtaining a risk identification result for the region to be identified based on the target feature vector.

[0008] Another aspect of this disclosure provides a graph neural network training apparatus, comprising: a second generation module for generating a sample region map including a target sample region, wherein the sample region map includes a plurality of sample nodes and a plurality of sample connection edges, the sample nodes being used to represent sample regions, the sample regions being areas where transport vehicles for transporting goods stop along their routes, the sample connection edges being generated based on the travel trajectory of the transport vehicles, and the sample connection edges being used to represent the association relationship between two sample nodes; a determination module for determining sample labels related to the target sample region; and a training module for training a graph neural network using the sample region map and the sample labels to obtain a trained graph neural network.

[0009] Another aspect of this disclosure provides a computer system comprising:

[0010] One or more processors;

[0011] Memory, used to store one or more programs.

[0012] When the above one or more programs are executed by the above one or more processors, the above one or more processors implement the above method.

[0013] Another aspect of this disclosure provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the above-described method.

[0014] Another aspect of this disclosure provides a computer program product including computer-executable instructions that, when executed, implement the method described above.

[0015] According to embodiments of this disclosure, a regional risk identification method is employed, comprising: generating a regional map including a region to be identified, wherein the regional map includes multiple nodes and multiple connecting edges, nodes representing regions, regions being areas where transport vehicles for transporting goods stop along their routes, and connecting edges representing the relationships between two nodes; extracting features from the regional map to obtain a target feature vector; and obtaining a risk identification result for the region to be identified based on the target feature vector. This method utilizes the areas where transport vehicles for transporting goods stop along their routes to generate a regional map including multiple stopping areas. On one hand, information such as the travel trajectory and stopping points of the transport vehicles is readily available. On the other hand, the relationships between multiple stopping areas can also be reflected through the regional map. Consequently, the target feature vector obtained from extracting features from the regional map contains semantic information about the stopping areas where the transport vehicles stop along their routes, as well as information about the relationships between multiple stopping areas in the regional map. Therefore, the risk identification result for the region to be identified based on the regional map is accurate and effective. Therefore, it at least partially overcomes the technical problems of low efficiency and low accuracy in the existing manual identification of hazardous chemicals, thereby achieving the technical effect of improving the efficiency and accuracy of risk identification related to hazardous chemicals. Attached Figure Description

[0016] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0017] Figure 1 An exemplary system architecture for which the regional risk identification method and apparatus of this disclosure can be applied is illustrated schematically;

[0018] Figure 2 A flowchart illustrating a regional risk identification method according to an embodiment of the present disclosure is shown schematically;

[0019] Figure 3 A flowchart illustrating a regional risk identification method according to another embodiment of this disclosure is shown schematically;

[0020] Figure 4 A schematic diagram illustrating the determination of a dwell point according to another embodiment of the present disclosure is shown.

[0021] Figure 5A A schematic diagram illustrating the determination of a hazardous materials-related location according to an embodiment of this disclosure is shown.

[0022] Figure 5B A schematic diagram illustrating the determination of a hazardous materials-related location according to an embodiment of this disclosure is shown.

[0023] Figure 6 A flowchart illustrating a graph neural network training method according to another embodiment of the present disclosure is shown schematically.

[0024] Figure 7 The schematic diagram illustrates a flowchart of a graph neural network training method according to another embodiment of the present disclosure;

[0025] Figure 8 A block diagram of a regional risk identification device according to an embodiment of the present disclosure is shown schematically;

[0026] Figure 9 A block diagram of a graph neural network training apparatus according to an embodiment of the present disclosure is schematically shown; and

[0027] Figure 10 A block diagram of a computer system suitable for implementing a regional risk identification method according to an embodiment of the present disclosure is shown schematically. Detailed Implementation

[0028] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0029] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0030] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0031] When using expressions such as "at least one of A, B, and C," the expression should generally be interpreted in accordance with the meaning commonly understood by a person skilled in the art (e.g., "a system having at least one of A, B, and C" should include, but is not limited to, systems having A alone, having B alone, having C alone, having A and B, having A and C, having B and C, and / or having A, B, and C, etc.). Similarly, when using expressions such as "at least one of A, B, or C," the expression should generally be interpreted in accordance with the meaning commonly understood by a person skilled in the art (e.g., "a system having at least one of A, B, or C" should include, but is not limited to, systems having A alone, having B alone, having C alone, having A and B, having A and C, having B and C, and / or having A, B, and C, etc.).

[0032] Hazardous chemicals (HCCs) are widely used and have become an indispensable part of people's daily lives, such as acids used in battery manufacturing, ammonium nitrates used in fertilizer production, and liquefied petroleum gas (LPG) used in cooking. However, due to their flammable, explosive, and toxic properties, HCCs require strict regulation throughout their entire lifecycle; otherwise, they can threaten urban public safety. While improving urban emergency management measures is crucial for handling such incidents, the fundamental solution lies in proactively identifying risks and shifting the focus from emergency response to prevention. If high-risk HCCs can be identified in advance, limited resources can be better deployed to prevent such tragedies, thereby ensuring urban safety.

[0033] Research related to hazardous chemical risk management typically focuses on risk analysis or route planning during transportation. Much of this work is based on the assumption that management possesses information on all hazardous chemical liability records (HCLs). However, in practice, over 90% of HCLs may remain unknown to management, posing a significant threat to cities. Identifying these hidden HCLs and determining their risk levels is a crucial task for urban safety management.

[0034] Current work on HCL classification includes: 1) traditional methods, such as manual processing; 2) data-driven methods that divide the city into grids to identify risky locations.

[0035] Traditional methods rely heavily on manual processing, such as random checks, to determine whether an area belongs to an HCL (Highly Critical Category). However, this is a labor-intensive and inefficient approach. Another method involves dividing the city into multiple grids and identifying each grid individually. For example, multiple chemical industrial parks could be grouped into a single grid, which would then be designated as an HCL. However, this method is less accurate and prone to misidentification.

[0036] According to embodiments of this disclosure, a regional risk identification method is employed, comprising: generating a regional map including a region to be identified, wherein the regional map includes multiple nodes and multiple connecting edges, nodes representing regions, regions being areas where transport vehicles for transporting goods stop along their routes, and connecting edges representing the relationships between two nodes; extracting features from the regional map to obtain a target feature vector; and obtaining a risk identification result for the region to be identified based on the target feature vector. This method utilizes the areas where transport vehicles for transporting goods stop along their routes to generate a regional map including multiple stopping areas. On one hand, information such as the travel trajectory and stopping points of the transport vehicles is readily available. On the other hand, the relationships between multiple stopping areas can also be reflected through the regional map. Consequently, the target feature vector obtained from extracting features from the regional map contains semantic information about the stopping areas where the transport vehicles stop along their routes, as well as information about the relationships between multiple stopping areas in the regional map. Therefore, the risk identification result for the region to be identified based on the regional map is accurate and effective. Therefore, it at least partially overcomes the technical problems of low efficiency and low accuracy in the existing manual identification of hazardous chemicals, thereby achieving the technical effect of improving the efficiency and accuracy of risk identification in hazardous chemicals.

[0037] Figure 1 An exemplary system architecture 100, illustrating an embodiment of the present disclosure, is shown schematically, in which a regional risk identification method and apparatus can be applied. It should be noted that... Figure 1 The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, in order to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other devices, systems, environments or scenarios.

[0038] like Figure 1 As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, and 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the terminal devices 101, 102, and 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0039] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as regional risk identification applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (for example only).

[0040] Terminal devices 101, 102, and 103 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0041] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using terminal devices 101, 102, and 103 (for example only). The backend management server can analyze and process the received cooking video data, such as user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0042] It should be noted that the regional risk identification method provided in this disclosure embodiment can generally be executed by server 105. Correspondingly, the regional risk identification device provided in this disclosure embodiment can generally be installed in server 105. The regional risk identification method provided in this disclosure embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105. Correspondingly, the regional risk identification device provided in this disclosure embodiment can also be installed in a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105. Alternatively, the regional risk identification method provided in this disclosure embodiment can also be executed by terminal devices 101, 102, or 103, or by other terminal devices different from terminal devices 101, 102, or 103. Accordingly, the regional risk identification device provided in this embodiment of the present disclosure may also be installed in terminal devices 101, 102, or 103, or in other terminal devices different from terminal devices 101, 102, or 103.

[0043] For example, the risk identification model composed of a graph neural network and a classifier can be originally stored in any one of terminal devices 101, 102, or 103 (e.g., terminal device 101, but not limited thereto), or it can be stored on an external storage device and imported into terminal device 101. Then, terminal device 101 can send the risk identification model to other servers or server clusters, and the other servers or server clusters that receive the risk identification model can execute the regional risk identification method provided in the embodiments of this disclosure.

[0044] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0045] Figure 2A flowchart illustrating a regional risk identification method according to an embodiment of the present disclosure is shown schematically.

[0046] like Figure 2 As shown, the method includes operations S201 to S203.

[0047] In operation S201, a region map including the region to be identified is generated.

[0048] According to embodiments of this disclosure, the region map includes multiple nodes and multiple connecting edges. Nodes are used to represent regions, which are the areas where transport vehicles used to transport goods stop along their routes. Connecting edges are used to represent the association relationship between two nodes.

[0049] In operation S202, feature extraction is performed on the region map to obtain the target feature vector.

[0050] In operation S203, based on the target feature vector, the risk identification result of the region to be identified in the regional map is obtained.

[0051] According to embodiments of this disclosure, the risk identification result may include the identification result of risk category, but is not limited thereto. It may also include the identification result of risk level, or the identification result of a combination of risk category and risk level.

[0052] For example, risk identification results can be categorized according to risk type, such as a risk identification result from production, storage, gas station, consumption, disposal, commerce, transportation, and others. Risk identification results can also be categorized according to risk level, such as a risk identification result from level 0 to 7. Alternatively, risk identification results can be categorized according to a combination of risk type and risk level, such as a risk identification result of level 7 in the production category.

[0053] According to embodiments of this disclosure, a region map including multiple stopping areas is generated using the stopping areas traversed by transport vehicles used for transporting goods. Furthermore, the goods transported by the transport vehicles can be hazardous chemicals. On one hand, information such as the travel trajectory and stopping points of the transport vehicles is readily available. On the other hand, the relationships between multiple stopping areas can also be reflected in the region map. Consequently, the target feature vector obtained by extracting features from the region map contains semantic information about the stopping areas traversed by the transport vehicles, as well as information about the relationships between multiple stopping areas in the region map. Therefore, the risk identification results for the region to be identified based on the region map are accurate and effective. The region risk identification method provided by embodiments of this disclosure can be used to quickly determine the risk identification results for the region to be identified, thereby improving the safety of the region to be identified.

[0054] In summary, the regional risk identification method provided by the embodiments of this disclosure can identify potential HCLs in cities in advance and accurately locate high-risk HCLs, providing decision support for urban hazardous chemical risk management.

[0055] The following is for reference. Figure 3 Figure 5 illustrates a specific embodiment. Figure 2 The method shown will be further explained.

[0056] Figure 3 A flowchart illustrating a regional risk identification method according to another embodiment of this disclosure is shown.

[0057] like Figure 3 As shown, a hazardous chemical risk site (also known as a region) mining framework can be used to perform a region risk identification method. This framework includes a data preprocessing module 301, a region map generation module 302, and a region classification module 303. The data preprocessing module 301 can process the driving trajectories 3011 of transport vehicles used for hazardous chemical transportation to determine a set of stopping points 3012. This process includes removing noisy trajectory points and detecting stopping points. The region map generation module 302 uses the stopping point set 3012 as input to generate a region map 3021. This process includes identifying nodes based on the stopping point set 3012 generated by the data preprocessing module 301, denoted as node set N; using POI (Point of Interest) data and stopping points, determining the dynamic and static data of multiple nodes and multiple connecting edges, and generating a region map representation matrix 3021. The classification module 303 of the region to be identified can take the region map representation matrix 3021 as input and output the risk identification result 3031 of the region to be identified.

[0058] According to embodiments of this disclosure, such as Figure 2 The operation S201 shown generates a region map including the area to be identified, which may include: determining a set of stopping points for transport vehicles based on their travel trajectories; determining multiple nodes in the region map based on the set of stopping points; determining multiple connecting edges between the multiple nodes based on their travel trajectories; and generating the region map based on the multiple nodes and multiple connecting edges.

[0059] According to embodiments of this disclosure, determining a set of stopping points for a transport vehicle based on its driving trajectory includes: determining multiple target trajectory points from the vehicle's driving trajectory; and determining the set of stopping points for the transport vehicle based on the dwell time and location of each of the multiple target trajectory points.

[0060] According to embodiments of this disclosure, the driving trajectory of a transport vehicle may include a sequence of location points generated by the transport vehicle, which can be represented as trajectory points arranged in chronological order. Wherein, trajectory point p = (vid, lng, lat, t), vid is the vehicle ID, (lng, lat) represents the longitude and latitude positions, and t is a timestamp. Multiple trajectory points form a trajectory point set, which can be represented as T.

[0061] According to embodiments of this disclosure, the travel trajectories of transport vehicles used in the hazardous chemical site excavation framework can originate from hazardous chemical transport vehicles. These transport vehicles are mandated by the government to be equipped with GPS (Global Positioning System) devices to report their real-time locations to the local government, and the resulting trajectory point set T is aggregated.

[0062] According to embodiments of this disclosure, the travel trajectories of transport vehicles are typically filled with noise and uncertainty. Extracting stable trajectory points from this set of noisy and uncertain trajectories is difficult. To address this issue, the data preprocessing module in the hazardous chemical risk site mining framework employs a noise filter to reduce random noise in the trajectory point set and extracts stable stopping points of transport vehicles from a large set of uncertain trajectory points.

[0063] According to embodiments of this disclosure, a stop point can be understood as a trajectory point where the duration of a transport vehicle's stay exceeds a predetermined time threshold, such as 30 minutes. Compared to regular trajectory points, stop points contain specific semantic information, such as the location where the vehicle loads or unloads hazardous chemicals.

[0064] According to embodiments of this disclosure, the driving trajectory of a transport vehicle can be preprocessed to obtain a set of stopping points. Noise filtering can be performed on the driving trajectory, i.e., abnormal trajectory points are removed from the trajectory point set T, resulting in multiple target trajectory points. The target trajectory points are then detected to identify a set of stopping points S from the multiple target trajectory points.

[0065] According to embodiments of this disclosure, considering variable atmospheric conditions and signal congestion, GPS-generated hazardous chemical transport trajectory data will not be completely accurate and will contain noisy trajectory points. These noisy trajectory points can affect the quality of stop point detection and subsequent tasks. Heuristic methods, for example, can be used to automatically filter out noisy trajectory points from the trajectory point set. The travel speed of each trajectory point in the trajectory point set can be calculated sequentially based on historical trajectory points and the current trajectory point. If the calculated speed of the current trajectory point is greater than a speed threshold, the current trajectory point is considered noise and should be removed from the trajectory point set.

[0066] According to embodiments of this disclosure, a dwell point detection algorithm can be used to extract dwell points from multiple target trajectory points. For example, the target trajectory points on the trajectory point set are traversed to check if the distance between the current anchor point and its successor point is less than a distance threshold dmax. For the current anchor point whose distance is less than the distance threshold dmax, it is further checked whether the time interval between the current anchor point and the last successor point is greater than a time duration threshold tmin. The successor points whose distance to the current anchor point is less than the distance threshold and whose time interval is greater than the time duration threshold, along with the current anchor point, are collectively referred to as the dwell point set corresponding to the current anchor point.

[0067] Figure 4 A schematic diagram illustrating the determination of a dwell point according to another embodiment of the present disclosure is shown.

[0068] like Figure 4 As shown, assuming trajectory point p3 is the current anchor point, {trajectory point p4, trajectory point p5, trajectory point p6} are successor points whose distance from the current anchor point p3 is less than the distance threshold. Furthermore, the time interval between the current anchor point p3 and its last successor point p6 is greater than the duration threshold tmin, i.e., the dwell time timestamp p6.t - timestamp p3.t > tmin. The center coordinates (sp.lng, sp.lat) of the dwell point sp are generated by formula (1):

[0069]

[0070] Where q is the trajectory point number, sp.lng is the longitude position of the dwell point sp, and sp.lat is the latitude position of the dwell point sp.

[0071] According to an embodiment of this disclosure, the timestamp sp.tp of the dwell point sp is as shown in formula (2).

[0072]

[0073] According to an embodiment of this disclosure, the anchor point is moved to the next trajectory point p7 after the current dwell point, and the above process is repeated until the anchor point moves to the end of the driving trajectory. This ultimately yields the set S of dwell points for the transport vehicle.

[0074] According to embodiments of this disclosure, the set of stopping points of transport vehicles is determined based on the driving trajectory of transport vehicles using the above method, making the stopping points more accurate and further facilitating the generation of subsequent regional maps.

[0075] According to embodiments of this disclosure, the set of depots is a set of depots for multiple transport vehicles. A larger number of transport vehicles is more conducive to the generation of the region map.

[0076] According to embodiments of this disclosure, determining multiple nodes in a region map based on a set of dwell points may include: for each of a plurality of transport vehicles, determining multiple initial dwelling areas for the transport vehicle based on the set of dwell points, thus obtaining an initial dwelling area set for the plurality of transport vehicles. Determining at least one set of target initial areas that overlaps from the initial dwelling area set. Each set of target initial areas includes multiple target initial areas. For each set of target initial areas, merging the multiple target initial areas to obtain a region. Based on the multiple regions, determining multiple nodes in the region map.

[0077] Figure 5A A schematic diagram illustrating the determination of a hazardous chemical-related location according to an embodiment of this disclosure is shown.

[0078] like Figure 5A As shown, HCL is a geographical area that contains various associated docking points for hazardous materials transport vehicles. HCL may consist of multiple boundaries 501 and multiple docking points 502 {sp1, ..., spM}.

[0079] According to embodiments of this disclosure, based on a set of dwelling points S, multiple dwelling points in the set of dwelling points whose distance is less than a predetermined distance threshold can be divided into the same region according to the conventional density clustering method, generating hazardous chemical related areas with boundaries and dwelling points, that is, regions corresponding to nodes in the region map.

[0080] According to embodiments of this disclosure, Figure 5A The method shown can be applied to scenarios where multiple hazardous chemical sites are geographically dispersed. In scenarios with a highly skewed spatial distribution of HCLs (Hazardous Chemical Clusters), a small area within a city may contain multiple HCLs (e.g., over 80% of HCLs are located in several small chemical industrial parks). This characteristic presents a significant challenge to accurately identifying individual HCLs, as traditional clustering algorithms easily merge several adjacent HCLs into a single HCL, for example... Figure 5A The two hazardous chemical-related sites HCL1 and HCL2 have been combined into one hazardous chemical-related site.

[0081] Figure 5B A schematic diagram illustrating the determination of hazardous material areas according to embodiments of the present disclosure is shown.

[0082] like Figure 5B As shown, an HCL is enclosed by multiple boundaries 501 and multiple dwell points 502 {sp1, ..., spM}. A set of HCLs is represented as {HCL1, ..., HCLi, ..., HCLI}. Dwell points can be divided based on the identification information of the transport vehicles. For example, the identification shape of dwell point 502 for transport vehicle A is triangular, and the identification shape of dwell point 502 for transport vehicle B is square.

[0083] According to embodiments of this disclosure, the HCL (Hazardous Goods Container) identification technology, which integrates the identification information of transport vehicles, can be understood as follows: HCLs are identified by clustering the stopping points along the travel trajectories of hazardous goods transport vehicles. Each transport vehicle's stopping point is individually density-clustered, resulting in multiple sets of potential HCLs after separate clustering, i.e., initial stopping area sets. Multiple overlapping initial target areas in the potential HCL sets are merged to obtain hazardous goods-related areas, i.e., areas in the area map. This strategy avoids clustering stopping points from different HCLs into a single cluster. Integrating vehicle identification information, such as ID information, that can indicate the stopping point affiliation into the clustering process solves the problem of small spacing between multiple HCLs during HCL identification, which can easily lead to erroneous merging of multiple HCLs into one HCL, thus achieving accurate HCL identification.

[0084] According to embodiments of this disclosure, a regional risk identification model can be used to process a regional map to obtain risk identification results for the region to be identified. The regional risk identification model may include a graph neural network (GNN) as the encoder.

[0085] According to embodiments of this disclosure, such as Figure 2 The operation S202 shown, which extracts features from the region map to obtain the target feature vector, can further include: using a graph neural network to extract features from the region map to obtain the target feature vector.

[0086] According to embodiments of this disclosure, a graph neural network (GNN) is used as the encoder to extract and encode features for each HCL. Specifically, a region map is constructed using all driving trajectories. Using a GNN encoder f θ The representation matrix Z' of the generated region map is shown in formula (3).

[0087]

[0088] Where Z'∈R N×D Represents a matrix, with node n i The representation vector z' i ∈R D .

[0089] According to embodiments of this disclosure, the GNN encoder f θ Common graph neural networks, such as two-layer graph convolutional networks, can be used to construct them.

[0090] According to embodiments of this disclosure, a region map (also known as an HCL map) describes the relationships between multiple HCLs (corresponding one-to-one with multiple nodes) in the travel trajectory of vehicles transporting hazardous chemicals, and can be represented by connecting edges. An HCL map can be represented as follows: Where, node N = {n'1, ..., n'} N}, It is a set of connection edges representing the relationships between multiple HCLs, X∈R N×F It is the node matrix corresponding to the node set, where x i ∈R F The eigenvector n of HCL i A'∈{0,1} N×N It is the connection edge matrix corresponding to the connection edge set, if there exists a connection edge from n' i to n' j The trajectory (i.e. e) ij ∈E), a' ij =1, otherwise a' ij =0.

[0091] According to embodiments of this disclosure, a graph neural network is used to extract features from a region map to obtain a target feature vector. This includes: for each of a plurality of nodes, determining the dynamic and static data of the node, thus obtaining the dynamic and static data for each of the plurality of nodes. The static data is data related to the region. The dynamic data is data related to transport vehicles. Based on the dynamic and static data of each of the plurality of nodes, a node matrix X related to the plurality of nodes is determined. For a plurality of connecting edges, a connecting edge matrix related to the connecting edges is determined. The connecting edge matrix and the node matrix are input into the graph neural network to obtain the target feature vector.

[0092] According to embodiments of this disclosure, static and dynamic data are extracted for each HCL using POI data and dwell points. The region-related static data may include the number of POIs within each HCL; statistical information for each HCL, such as area, the number of polygons along its boundaries, and the number of Geohash (an address code) entries. The transport vehicle-related dynamic data may include the dwell time of dwell points within each HCL, arrival / departure times (48 time slices per day), weekly arrival dates, and transport vehicle ownership. Transport vehicle ownership can be categorized into three types: city, other cities within the province, and other provinces.

[0093] According to embodiments of this disclosure, the target feature vector is used to characterize at least one of the following feature vectors: the number of connecting edges connected to the node, and the contextual results between the node and its neighboring nodes.

[0094] According to embodiments of this disclosure, in a region graph, the category of a region to be identified can be determined based on the number of connecting edges, also known as degree, connected to a node. For example, the degree of a region that is a hazardous chemical warehouse differs significantly from the degree of a region that is a gas station, because a hazardous chemical warehouse transports gasoline to multiple gas stations, while gas stations typically obtain gasoline from a fixed warehouse. The risk level of a hazardous chemical warehouse is higher than that of a gas station. Therefore, if a graph neural network can extract features related to the degree of the node to be identified, the accuracy of the risk identification result for the region to be identified can be improved.

[0095] According to embodiments of this disclosure, in a region map, for multiple nodes of different categories, the categories of their adjacent nodes, which correspond one-to-one with each node, are typically different. In some cases, the implicit feature information of adjacent nodes is more discriminative than the original feature information of the nodes. For example, in a chemical industrial park, regions designated as oil refineries (production type) and regions designated as warehouses (storage type) may be spatially close, resulting in similar original feature information. However, in a region map, oil refineries are mostly adjacent to gas stations, while warehouses are adjacent to various production plants. Therefore, the contextual information corresponding to adjacent nodes of oil refineries (production type) and adjacent nodes of warehouses (storage type) exhibits significant differences. Therefore, if a graph neural network can extract contextual results related to the region to be identified, it can improve the accuracy of risk identification results for the region to be identified.

[0096] According to embodiments of this disclosure, K is the total number of risk categories for a region, and the nth... i The context result c' of the k-th element of each node i，k Formula (4) is shown below.

[0097]

[0098] Among them, B′ i Indicate n' i The number of connecting edges, B′ i，k This represents the number of neighboring nodes belonging to the k-th category.

[0099] According to embodiments of this disclosure, by using a graph neural network as an encoder, it is possible to learn node-related features and the contextual results of adjacent nodes, resulting in rich extracted features and strong implicit features, which in turn helps to improve the accuracy of subsequent risk identification results.

[0100] According to embodiments of this disclosure, for Figure 2 The operation S203 shown, which obtains the risk identification result of the area to be identified based on the target feature vector, may further include: inputting the target feature vector into a classifier to obtain the risk identification result of the area to be identified.

[0101] According to embodiments of this disclosure, the regional risk identification model may further include a classifier. The classifier may include sequentially connected fully connected layers and activation functions. The classifier can be used to process the target feature vector output by the graph neural network to obtain the risk identification result.

[0102] According to embodiments of this disclosure, since the target feature vector includes at least one feature vector from the number of connecting edges connected to the node corresponding to the region to be identified and the contextual results between the node and its neighboring nodes, the obtained risk identification result is accurate and effective.

[0103] Figure 6 A flowchart illustrating a graph neural network training method according to an embodiment of the present disclosure is shown schematically.

[0104] like Figure 6 As shown, the method includes operations S601 to S603.

[0105] In operation S601, a sample area map including the target sample area is generated.

[0106] According to embodiments of this disclosure, the sample region map includes multiple sample nodes and multiple sample connection edges. The sample nodes are used to represent sample regions, which are the areas where transport vehicles used to transport goods stop along the way. The sample connection edges are used to represent the association relationship between two sample nodes.

[0107] In operation S602, determine the sample labels associated with the target sample region.

[0108] In operation S603, a graph neural network is trained using sample area maps and sample labels to obtain a trained graph neural network.

[0109] According to embodiments of this disclosure, sample labels may include labels used to characterize the risk identification results of the target sample region. However, this is not a limitation. Sample labels may also include labels used to characterize the features of the target sample nodes corresponding to the target sample region, such as labels used to characterize the association relationship between the target sample node and its neighboring nodes, labels used to characterize the number of connecting edges connected to the target sample node, etc. Sample labels may also include labels used to characterize the feature vector of the target sample region. Any sample label related to the target sample region is acceptable.

[0110] According to embodiments of this disclosure, a graph neural network can be used to process a sample region map to obtain sample feature results corresponding to sample labels. The graph neural network is then trained using the prediction results and sample labels to obtain a trained graph neural network.

[0111] According to embodiments of this disclosure, the sample region refers to the areas where transport vehicles traveling to transport goods stop along their routes. Multiple stop areas are used to form multiple sample nodes in the sample region map. On one hand, information such as the travel trajectory and stop points of the transport vehicles is readily available. On the other hand, the relationships between multiple stop areas can also be reflected in the sample region map. This allows the sample region map to contain semantic information about the stop areas traversed by the transport vehicles, as well as the relationship information between multiple stop areas in the sample region map. A graph neural network can be trained based on the sample region map and sample labels, enabling the trained graph neural network to fully learn the features of the semantic information about the stop areas and the relationship information between multiple stop areas in the sample region map. This allows the trained graph neural network to fully extract features from the region map during application, improving the accuracy of risk identification results for the region to be identified.

[0112] The following describes specific embodiments and references. Figure 7 ,right Figure 6 The method shown will be further explained.

[0113] According to embodiments of this disclosure, for Figure 6 Operation S603 trains a graph neural network using a sample region map and sample labels to obtain a trained graph neural network. This process may include: inputting the sample region map into the graph neural network and outputting a sample target feature vector; determining the sample feature results of the target sample region based on the sample target feature vector; determining the loss function value based on the sample feature results and sample labels; and training the graph neural network using the loss function value to obtain the trained graph neural network.

[0114] According to embodiments of this disclosure, a graph neural network can be used as an encoder for feature extraction. The sample region map is input into the graph neural network to obtain a sample target feature vector. Based on the sample target feature vector, the sample feature result of the target sample region is determined. The sample feature result can be a risk identification result, but is not limited to this; it can also be the feature result of other non-risk identification results of the target sample region. The key is that the sample feature result matches the sample label.

[0115] According to embodiments of this disclosure, sample feature results and sample labels are input into a loss function to obtain a loss function value. Training a graph neural network using the loss function value to obtain a trained graph neural network may include: adjusting the parameters of the graph neural network based on the loss function value until the loss function value reaches a convergence condition. The graph neural network whose loss function value has reached a convergence condition is then used as the trained graph neural network.

[0116] According to other embodiments of this disclosure, considering the scarcity of risk level labels in practical applications, it is difficult to train a high-accuracy model using a small number of training samples. This disclosure utilizes a self-supervised approach to train graph neural networks, for example, by using sample feature results and labels related to the latent features of sample nodes to generate training samples, so that the trained graph neural network can fully mine the latent feature information in the sample area map, thereby enabling the trained graph neural network to achieve equivalent accuracy.

[0117] According to an optional embodiment of this disclosure, the sample label may include at least one of a degree label and a context label for the target sample node corresponding to the target sample region. The degree label may be a label representing the number of edges connected to the target sample node. The context label may be a label representing the risk level or risk category of sample nodes adjacent to the target sample node.

[0118] According to embodiments of this disclosure, the sample label includes a degree label and a context label of the target sample node corresponding to the target sample region.

[0119] According to embodiments of this disclosure, training a graph neural network using a sample region map and sample labels to obtain a trained graph neural network includes: inputting the sample region map into the graph neural network and outputting a sample target feature vector; determining the number of sample connection edges connected to the target sample node based on the sample target feature vector; determining the context result between the target sample node and its neighboring nodes based on the sample target feature vector; determining a first loss function value based on the number recognition result and degree labels; determining a second loss function value based on the context result and context labels; and training the graph neural network using the first and second loss function values ​​to obtain a trained graph neural network.

[0120] According to embodiments of this disclosure, training tasks regarding degree training samples and context training samples can be set using a self-supervised approach.

[0121] For example, consider the training task for degree training samples. In a sample region graph, the degree of a sample node can refer to the number of edges connecting it. The category of a target sample region can be determined based on the degree of the sample nodes. For instance, the degree of a sample region that is a hazardous materials warehouse differs significantly from the degree of a sample region that is a gas station, because the hazardous materials warehouse transports gasoline to multiple gas stations, while gas stations typically obtain gasoline from a fixed warehouse. The risk level of the hazardous materials warehouse is higher than that of the gas station. Therefore, if the graph neural network can predict the degree recognition result for the target sample nodes, the learned features can be helpful in classifying the risk level of the region to be identified.

[0122] According to embodiments of this disclosure, a graph neural network f can be given...θ (·) Any sample node n generated i eigenvectors z ∈ N i Here, N represents the set of sample nodes. The goal of the training task on degree training samples is to minimize the number of sample nodes n. i The mean squared error (MSE) loss between the true degree (i.e., degree label) and the predicted degree (i.e., quantity identification result) L dp That is, formula (5).

[0123]

[0124] Among them, g φ (·) is a degree predictor implemented using linear regression, d i It is sample node n i The degree. Based on the sample area map, it can be determined by d. i =∑ j a ji +∑ j a ij Calculate each HCLn i The degree of the sample nodes n j Pointing to sample node n i a ji Sample node n i Pointing to sample node n j a ij These are elements of the sample connection edge matrix A of the sample region map.

[0125] For example, consider training tasks using contextual training samples. In some cases, the implicit feature information of neighboring sample nodes is more discriminative than the original feature information of the sample nodes. For instance, in a chemical industrial park, sample areas that are oil refineries (production type) and sample areas that are warehouses (storage type) may be spatially close, resulting in similar original feature information. However, in the sample area map, oil refineries are mostly adjacent to gas stations, while warehouses are adjacent to various production plants. Therefore, the contextual information corresponding to the neighboring sample nodes of oil refineries (production type) and the contextual information corresponding to the neighboring sample nodes of warehouses (storage type) show significant differences. Therefore, using contextual training samples enables graph neural networks to better learn features related to the contextual information of sample nodes, making the learned features helpful for risk level classification of the area to be identified.

[0126] According to embodiments of this disclosure, sample node n can be... i Construct a context label c i Assume K is the total number of risk categories in the sample area, and the nth... i The context label c of the k-th element of a sample node ikFormula (6) is shown below.

[0127]

[0128] Among them, B i Indicates sample node n i The number of sample connection edges, B i，k This represents the number of neighboring sample nodes belonging to the k-th category.

[0129] According to embodiments of this disclosure, based on context tag c ik The goal is to minimize the MSE loss function L. cp , as in formula (7).

[0130]

[0131] Among them, g ω (·) is a context predictor implemented by SoftMax regression, where |N| represents the number of sample nodes.

[0132] According to embodiments of this disclosure, in order to ensure that all adjacent sample nodes are in the calculation of B i，k Each sample node has a category, and a label propagation algorithm can be used to assign a category to each sample node.

[0133] According to embodiments of this disclosure, the training tasks described above regarding degree training samples and context training samples can be used to train graph neural networks. However, this is not limited to these methods. Multiple sample region maps with the same sample nodes but different sample connection edges can also be used, i.e., comparative training samples can be used for comparative training tasks. The generation method of multiple sample region maps with the same sample nodes but different sample connection edges can refer to generating multiple sample region maps according to different predetermined durations. For example, multiple sample region maps can be divided into sample region sub-maps (also known as sample local region maps) and sample region maps (also known as sample global region maps) according to the length of their duration. Sample region sub-maps can be region maps generated from short-term (e.g., January 1st to January 15th) transport vehicle trajectories, while sample region maps can be region maps generated from long-term (e.g., January 1st to March 1st) transport vehicle trajectories. Both sample region maps and sample region sub-maps include target sample nodes. Thus, the consistency between the sample target feature vectors and sample target feature sub-vectors between the sample region maps and sample region sub-maps can be used to perform self-supervised training of the graph neural network.

[0134] According to embodiments of this disclosure, comparative training of a graph neural network using contrastive training samples can improve the quality of the graph neural network's representation of sample nodes using data augmentation methods. Traditional data augmentation typically obtains sample node representations by randomly perturbing connection edges or node attributes, but this method may lead to information loss and introduce unnecessary noise. This deficiency can be overcome by introducing the concept of a sample region subgraph. A sample region subgraph records the relationships between multiple sample nodes in the short term, which are constituted by short-term (e.g., one day) driving trajectories. Since the sample region subgraph records the relationships between multiple sample nodes in the long term, it is a subgraph of the sample region map. Many sample region subgraphs can be generated from trajectory data as augmentation data for the sample region map. Compared with random perturbation methods, sample region subgraphs have explicit physical meaning. The category of the target sample node is invariant in both the sample region subgraph and the sample region map, therefore the representation of the same target sample node in both the sample region subgraph and the sample region map is consistent. Based on this principle, training the graph neural network using contrastive training samples can be more accurate.

[0135] According to other embodiments of this disclosure, graph neural networks can also be trained by combining degree training samples, context training samples, and global-local contrast training samples.

[0136] According to embodiments of this disclosure, a graph neural network is trained using a sample region map and sample labels to obtain a trained graph neural network. The process further includes: inputting the sample region map into the graph neural network to output a sample target feature vector; determining the number of sample connection edges connected to the target sample node based on the sample target feature vector; determining the context result between the target sample node and its neighboring nodes based on the sample target feature vector; determining a first loss function value based on the number of connection edges and degree labels; determining a second loss function value based on the context result and context labels; determining a sample region sub-map from the sample region map; the sample region sub-map including the target sample region; inputting the sample region sub-map into the graph neural network to output a sample target feature vector; determining a third loss function value based on the sample target feature vector and the sample target feature vector; and training the graph neural network using the first, second, and third loss function values ​​to obtain a trained graph neural network.

[0137] According to other embodiments of this disclosure, graph neural networks can also be trained by combining degree training samples, context training samples, global-local contrast training samples, and local-local contrast training samples.

[0138] According to embodiments of this disclosure, the global-local contrast training samples can be a sample region map and a first sample region submap. The local-local contrast training samples can be a first sample region submap and a second sample region submap. The first and second sample region submaps are similar in that they both include target sample nodes, but differ in that they are region maps generated from the travel trajectories of transport vehicles within different historical time periods. For example, the first sample region submap can be a region map generated from the travel trajectories of transport vehicles in a short period (e.g., January 1st to January 15th), and the second sample region submap can be a region map generated from the travel trajectories of transport vehicles in a short period (e.g., February 1st to February 15th). The sample region map is a region map generated from the travel trajectories of transport vehicles in a long period (e.g., January 1st to March 1st).

[0139] According to embodiments of this disclosure, training is performed using local-local contrastive training samples. The local-local contrastive training task ensures that the representation vectors between different local maps are consistent. The first sample region sub-map... Second sample region submap The input is fed into a graph neural network to generate a representation matrix Z. (1) and Z (2) , and Let Z represent the matrix respectively. (1) and Z (2) Medium sample node n i The representation vectors of the nodes are used. Representation vectors of nodes with the same data can be used as positive sample pairs, and representation vectors of nodes with different data can be used as negative sample pairs. A nonlinear projection function g is employed. ξ The loss is calculated by mapping the representation vector to another latent space. and Based on InfoNCE loss, pairwise local-local contrastive loss L llc (u i v i ), as in formula (8).

[0140]

[0141] According to embodiments of this disclosure, h(·) is a criterion function defined as h(u, v) = exp(sim(u, v) / τ), where sim(·, ·) is the cosine similarity and τ is an adjustable temperature parameter. g can be implemented using a two-layer MLP (Multilayer Perceptron). ξ The final local-local contrast loss can be expressed as shown in Equation (9).

[0142]

[0143] Where, N (llc) =N (1) ∩N (2) N (1) N represents the number of sample nodes in the first sample region map. (2) This represents the number of sample nodes in the second sample region map.

[0144] According to embodiments of this disclosure, the local-to-local contrastive loss function described above involves a cosine similarity function used to calculate the similarity between the representation vectors of two sample nodes. However, it is not limited to this; any function capable of calculating the similarity between the representation vectors of two sample nodes is acceptable.

[0145] According to embodiments of this disclosure, a graph neural network is trained using local-global contrastive training samples. The training task focuses on achieving consistency in representation vectors between sample region submaps and sample region maps. A nonlinear projection function g is employed. ξ Sample node n i The corresponding identifier z i Map it to another latent space and calculate the loss, i.e., w. i =g ξ (z i Based on InfoNCE loss, the pairwise local-total contrastive loss L lgc (u i w i As shown in formula (10).

[0146]

[0147] According to embodiments of this disclosure, the final local-global contrast loss L lgc As shown in formula (11).

[0148]

[0149] in,

[0150] According to embodiments of this disclosure, the local-global contrastive loss function described above involves a cosine similarity function used to calculate the similarity between the representation vectors of two sample nodes. However, it is not limited to this; any function capable of calculating the similarity between the representation vectors of two sample nodes is acceptable.

[0151] Figure 7 The schematic diagram illustrates a flowchart of a graph neural network training method according to another embodiment of the present disclosure.

[0152] like Figure 7 As shown, the sample area map can be... Input to graph neural network f θ(·), output the target feature vector Z. Based on the target feature vector Z, determine the number of connection edges connected to the target sample node. Based on the target feature vector Z, determine the context between the target sample node and its neighboring nodes. Based on the number recognition result and degree label, determine the value of the first loss function. For example, input the number recognition result and degree label into the first loss function L. dp The first loss function value is obtained. Based on the context result and context label, the second loss function value is determined. The context result and context label are then input into the second loss function L. cp The second loss function value is obtained from the sample region map. The first sample region submap was determined in the middle. Second sample region submap First Sample Region Submap Second sample region submap Each includes the target sample region and the first sample region sub-map. It is generated based on the historical trajectories of transport vehicles in the first historical period, and the second sample area submap. It is generated based on the historical trajectory of the transport vehicle in the second historical period, which is different from the first and second historical periods.

[0153] like Figure 7 As shown, this is used to submap the first sample region. The input is fed into a graph neural network, and the output is the first sample target feature vector Z. (1) Based on the sample target feature vector Z and the sample target feature sub-vector Z (1) Determine the value of the third loss function. For example, combine the sample target feature vector Z and the sample target sub-feature vector Z. (1) Input to the third loss function L lgc In the process, the third loss function value is obtained. The second sample region submap is then used. The input is fed into a graph neural network, and the output is the second sample target feature vector Z. (2) Based on the first sample target feature vector Z (1) The second sample target feature vector Z (2) Determine the value of the fourth loss function. For example, take the target feature vector Z of the first sample. (1) The second sample target feature vector Z (2) Input to the fourth loss function L llc The fourth loss function value is obtained from this. The sum of the first, second, third, and fourth loss functions is taken as the overall loss function L. all The graph neural network is trained using the first, second, third, and fourth loss function values ​​to obtain the trained graph neural network.

[0154] According to embodiments of this disclosure, parameters can be updated using the backpropagation algorithm by minimizing the overall loss. The overall loss function is defined as L all As shown in formula (12).

[0155] L all =L dp +L cp +L llc +L lgc ; Formula (12)

[0156] Among them, L dp L cp L llc and L lgc These are the first loss function, the second loss function, the third loss function, and the fourth loss function.

[0157] According to embodiments of this disclosure, a trained graph neural network can be used as the encoder f of a regional risk identification model. θ (·) is used to generate the target feature vector, and its parameters are frozen. Then, the classifier of the regional risk identification model is trained.

[0158] According to embodiments of this disclosure, the sample nodes n output by the trained graph neural network can be... i The representation vector z i The input data to the classifier to be trained is used to predict the sample nodes n with known outcomes using a linear transformation with SoftMax activation. i Predictive risk identification results As shown in formula (13).

[0159]

[0160] in, It predicts the outcome of risk identification, W c ∈R K×D and b c ∈R K These are learnable parameters.

[0161] According to embodiments of this disclosure, cross-entropy loss is used as the target loss function L. c Learnable parameters are trained, as shown in formula (14).

[0162]

[0163] Among them, y i ∈R K It is sample node n i The true risk identification results and y iAll are represented using one-hot encoding.

[0164] Figure 8 A block diagram of a regional risk identification device according to an embodiment of the present disclosure is shown schematically.

[0165] like Figure 8 As shown, the regional risk identification device 800 includes a first generation module 810, a feature extraction module 820, and an identification module 830.

[0166] The first generation module 810 is used to generate a region map including the region to be identified. The region map includes multiple nodes and multiple connecting edges. Nodes are used to represent regions, and regions are the areas where transport vehicles used to transport goods stop along the way. Connecting edges are used to represent the relationship between two nodes.

[0167] The feature extraction module 820 is used to extract features from the region map to obtain the target feature vector.

[0168] The identification module 830 is used to obtain the risk identification result of the area to be identified based on the target feature vector.

[0169] According to embodiments of this disclosure, the first generation module includes: a first determining unit, a second determining unit, a third determining unit, and a generation unit.

[0170] The first determining unit is used to determine the set of stopping points of the transport vehicles based on their driving trajectories.

[0171] The second determining unit is used to determine multiple nodes in the region map based on the set of resident points.

[0172] The third determining unit is used to determine the connection edges between multiple nodes based on the driving trajectory of the transport vehicle, thus obtaining multiple connection edges.

[0173] The generation unit is used to generate a region map based on multiple nodes and multiple connection edges.

[0174] According to embodiments of this disclosure, the set of rest points is a set of rest points for multiple transport vehicles.

[0175] According to embodiments of this disclosure, the second determining unit includes: a first determining subunit, a second determining subunit, a merging subunit, and a third determining subunit.

[0176] The first determining subunit is used to determine multiple initial dwelling areas for each of the multiple transport vehicles based on the set of dwelling points, thereby obtaining a set of initial dwelling areas for the multiple transport vehicles.

[0177] The second determining subunit is used to determine at least one set of target initial regions that overlap from the initial set of residence regions, wherein each set of target initial regions includes multiple target initial regions.

[0178] The merge sub-unit is used to merge multiple target initial regions for each target initial region set to obtain a region.

[0179] The third determining sub-unit is used to determine multiple nodes in the region map based on multiple regions.

[0180] According to embodiments of this disclosure, the first determining unit includes a fourth determining subunit and a fifth determining subunit.

[0181] The fourth determination subunit is used to determine multiple target trajectory points from the driving trajectory of the transport vehicle.

[0182] The fifth determining sub-unit is used to determine the set of stopping points for transport vehicles based on the dwell time and location of each of the multiple target trajectory points.

[0183] According to embodiments of this disclosure, the feature extraction module includes a feature extraction unit.

[0184] The feature extraction unit is used to extract features from the region map using a graph neural network to obtain a target feature vector, wherein the target feature vector is used to represent at least one of the following feature vectors: the number of connecting edges connected to a node, and the contextual results between a node and its neighboring nodes.

[0185] According to embodiments of this disclosure, the feature extraction unit includes: a sixth determining subunit, a seventh determining subunit, a connecting edge determining subunit, and an extraction subunit.

[0186] The sixth determining sub-unit is used to determine the dynamic and static data of each of the multiple nodes, thus obtaining the dynamic and static data for each node. The static data is data related to the region; the dynamic data is data related to the transport vehicles.

[0187] The seventh determination sub-unit is used to determine the node matrix related to multiple nodes based on their respective dynamic and static data.

[0188] The connection edge determination sub-unit is used to determine the connection edge matrix associated with the connection edges based on multiple connection edges.

[0189] Extract sub-units to input the connection edge matrix and node matrix into the graph neural network to obtain the target feature vector.

[0190] According to embodiments of this disclosure, the identification module includes an identification unit.

[0191] The identification unit is used to input the target feature vector into the classifier to obtain the risk identification result of the area to be identified. The classifier includes a fully connected layer and an activation function connected in sequence.

[0192] Figure 9 A block diagram of a regional risk identification device according to an embodiment of the present disclosure is shown schematically.

[0193] like Figure 9 As shown, the graph neural network training device 900 includes: a second generation module 910, a determination module 920, and a training module 930.

[0194] The second generation module 910 is used to generate a sample region map including the target sample region. The sample region map includes multiple sample nodes and multiple sample connection edges. The sample nodes are used to represent the sample region, which is the area where the transport vehicle used to transport goods stops. The sample connection edges are generated based on the driving trajectory of the transport vehicle and are used to represent the relationship between two sample nodes.

[0195] The determination module 920 is used to determine the sample labels associated with the target sample region.

[0196] Training module 930 is used to train a graph neural network using sample area maps and sample labels to obtain a trained graph neural network.

[0197] According to embodiments of this disclosure, the training module includes: a first input unit, a fourth determination unit, a fifth determination unit, and a first training unit.

[0198] The first input unit is used to input the sample area map into the graph neural network and output the sample target feature vector.

[0199] The fourth determining unit is used to determine the sample feature results of the target sample region based on the sample target feature vector.

[0200] The fifth determining unit is used to determine the loss function value based on the sample feature results and sample labels.

[0201] The first training unit is used to train the graph neural network using the loss function value, resulting in a trained graph neural network.

[0202] According to embodiments of this disclosure, the sample label includes a degree label and a context label of the target sample node corresponding to the target sample region.

[0203] According to embodiments of this disclosure, the training module includes: a first input unit, a sixth determination unit, a seventh determination unit, an eighth determination unit, a ninth determination unit, a tenth determination unit, a second input unit, an eleventh determination unit, a third input unit, a twelfth determination unit, and a second training unit.

[0204] The first input unit is used to input the sample area map into the graph neural network and output the sample target feature vector.

[0205] The sixth determining unit is used to determine the number of sample connection edges connected to the target sample node based on the sample target feature vector.

[0206] The seventh determining unit is used to determine the context result between the target sample node and its neighboring nodes based on the sample target feature vector.

[0207] The eighth determining unit is used to determine the value of the first loss function based on the quantity recognition result and the degree label.

[0208] The ninth determining unit is used to determine the value of the second loss function based on the context result and the context label;

[0209] The tenth determining unit is used to determine a first sample area sub-map and a second sample area sub-map from the sample area map. The first sample area sub-map and the second sample area sub-map each include the target sample area. The first sample area sub-map is generated based on the historical trajectory of the transport vehicle in a first historical period, and the second sample area sub-map is generated based on the historical trajectory of the transport vehicle in a second historical period. The first historical period and the second historical period are different.

[0210] The second input unit is used to input the first sample region sub-map into the graph neural network and output the first sample target feature sub-vector.

[0211] The eleventh determining unit is used to determine the value of the third loss function based on the sample target feature vector and the sample target feature sub-vector.

[0212] The third input unit is used to input the second sample region sub-map into the graph neural network and output the second sample target feature sub-vector.

[0213] The twelfth determining unit is used to determine the value of the fourth loss function based on the first sample target feature vector and the second sample target feature vector.

[0214] The second training unit is used to train the graph neural network using the first loss function value, the second loss function value, the third loss function value, and the fourth loss function value, to obtain the trained graph neural network.

[0215] Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure, or at least part of the functions of any one or more of them, can be implemented in one module. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be implemented by dividing them into multiple modules. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be at least partially implemented as hardware circuitry, such as Field Programmable Gate Arrays (FPGAs), Programmable Logic Arrays (PLAs), Systems-on-Chip, Systems-on-Substrate, Systems-on-Package, Application-Specific Integrated Circuits (ASICs), or implemented in hardware or firmware by any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be at least partially implemented as computer program modules, which, when run, can perform corresponding functions.

[0216] For example, any and more of the first generation module 810, feature extraction module 820, recognition module 830, second generation module 910, determination module 920, and training module 930 can be combined into one module / unit / subunit, or any one of these modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least some of the functions of one or more of these modules / units / subunits can be combined with at least some of the functions of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of this disclosure, at least one of the first generation module 810, feature extraction module 820, recognition module 830, second generation module 910, determination module 920, and training module 930 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or any other reasonable means of integrating or packaging circuits, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the first generation module 810, feature extraction module 820, recognition module 830, second generation module 910, determination module 920, and training module 930 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0217] It should be noted that the regional risk identification device part in the embodiments of this disclosure corresponds to the regional risk identification method part in the embodiments of this disclosure. The description of the regional risk identification device part is specifically referred to in the regional risk identification method part, and will not be repeated here.

[0218] Figure 10 A block diagram of a computer system suitable for implementing the methods described above, according to embodiments of the present disclosure, is illustrated schematically. Figure 10 The computer system shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0219] like Figure 10 As shown, a computer system 1000 according to an embodiment of the present disclosure includes a processor 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage portion 1008 into a random access memory (RAM) 1003. The processor 1001 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 1001 may also include onboard memory for caching purposes. The processor 1001 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0220] RAM 1003 stores various programs and data required for the operation of system 1000. Processor 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Processor 1001 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 1002 and / or RAM 1003. It should be noted that the programs may also be stored in one or more memories other than ROM 1002 and RAM 1003. Processor 1001 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in said one or more memories.

[0221] According to embodiments of this disclosure, system 1000 may further include an input / output (I / O) interface 1005, which is also connected to bus 1004. System 1000 may also include one or more of the following components connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. A removable medium 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 1010 as needed so that computer programs read from it can be installed into storage section 1008 as needed.

[0222] According to embodiments of this disclosure, the method flow according to embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1009, and / or installed from removable medium 1011. When the computer program is executed by processor 1001, it performs the functions defined in the system of embodiments of this disclosure. According to embodiments of this disclosure, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0223] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0224] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0225] For example, according to embodiments of this disclosure, a computer-readable storage medium may include one or more memories other than the ROM 1002 and / or RAM 1003 described above.

[0226] Embodiments of this disclosure also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of this disclosure. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the methods provided in the embodiments of this disclosure.

[0227] When the computer program is executed by the processor 1001, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0228] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1009, and / or installed from a removable medium 1011. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0229] According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can execute entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0230] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features recited in the various embodiments and / or claims of this disclosure can be combined and / or combined in various ways, even if such combinations or combinations are not expressly described in this disclosure. In particular, the features described in the various embodiments and / or claims of this disclosure may be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0231] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The scope of this disclosure is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A training method for a graph neural network for the relevance identification of hazardous chemicals, comprising: Generate a sample region map including the target sample region, wherein the sample region map includes multiple sample nodes and multiple sample connection edges, the sample nodes are used to characterize the sample region, the sample region is the stopping area of ​​the transport vehicle used to transport hazardous chemicals, and the sample connection edges are generated based on the driving trajectory of the transport vehicle and are used to characterize the association relationship between two sample nodes. Determine sample labels associated with the target sample region, the sample labels including the degree label and context label of the target sample node corresponding to the target sample region; and A graph neural network is trained using the sample region map and the sample labels to obtain the trained graph neural network. The step of training a graph neural network using the sample region map and the sample labels to obtain a trained graph neural network includes: The sample region map is input into the graph neural network, which outputs the sample target feature vector. Based on the target feature vector of the sample, the number of sample connection edges connected to the target sample node is determined as an identification result; Based on the target feature vector of the sample, the context result between the target sample node and its neighboring nodes is determined; Based on the quantity identification result and the degree label, the value of the first loss function is determined; Based on the context results and the context labels, determine the value of the second loss function; A first sample area submap and a second sample area submap are determined from the sample area map, wherein the first sample area submap and the second sample area submap include the target sample area, the first sample area submap is generated based on the historical trajectory of the transport vehicle in a first historical period, and the second sample area submap is generated based on the historical trajectory of the transport vehicle in a second historical period, wherein the first historical period and the second historical period are different; The first sample region sub-map is input into the graph neural network, and the first sample target feature sub-vector is output; The value of the third loss function is determined based on the target feature vector of the sample and the sub-vector of the target sample. The second sample region sub-map is input into the graph neural network, and the second sample target feature sub-vector is output; Based on the first sample target feature vector and the second sample target feature sub-vector, determine the value of the fourth loss function; and The graph neural network is trained using the first loss function value, the second loss function value, the third loss function value, and the fourth loss function value to obtain the trained graph neural network.

2. A method for relevant identification of hazardous chemicals, comprising: Generate a region map including the region to be identified, wherein the region map includes multiple nodes and multiple connecting edges, the nodes are used to represent regions, the regions are the stops of transport vehicles used to transport hazardous chemicals, and the connecting edges are generated according to the driving trajectory of the transport vehicles and are used to represent the association relationship between two nodes; The graph neural network trained using the training method of the graph neural network according to claim 1 is used to extract features from the region map to obtain a target feature vector, wherein the target feature vector is used to characterize at least one of the following feature vectors: the number of connecting edges connected to the node, the context result between the node and its neighboring nodes; and Based on the target feature vector, the risk identification result of the region to be identified in the region map is obtained. The generation of the region map including the region to be identified includes: Based on the driving trajectory of the transport vehicle, determine the set of the transport vehicle's stopping points; Based on the set of outposts, the plurality of nodes in the region map are determined; Based on the travel trajectory of the transport vehicle, the connection edges between the plurality of nodes are determined, thus obtaining the plurality of connection edges; and The region map is generated based on the multiple nodes and the multiple connecting edges; The set of stopping points for the transport vehicles, determined based on their travel trajectories, includes: Traverse the target trajectory points on the trajectory point set to check whether the distance between the current anchor point and the successor point is less than the distance threshold; For the current anchor point whose distance is less than the distance threshold, further check whether the interval between the current anchor point and the last successor point is greater than the duration threshold. The successor points whose distance from the current anchor point is less than the distance threshold and whose time interval is greater than the duration threshold, together with the current anchor point, are collectively referred to as the set of dwell points corresponding to the current anchor point. Move the anchor point to the next trajectory point after the current dwell point, and repeat the above process until the anchor point moves to the end of the driving trajectory, thereby finally obtaining the set of dwell points of the transport vehicle.

3. The method according to claim 2, wherein, The set of rest points is a set of rest points for multiple transport vehicles; The step of determining the plurality of nodes in the region map based on the set of dwelling points includes: For each of the plurality of transport vehicles, based on the set of rest points, a plurality of initial rest areas for the transport vehicle are determined, resulting in a set of initial rest areas for the plurality of transport vehicles. From the initial set of settlement areas, at least one set of target initial areas that overlaps is determined, wherein each set of target initial areas includes multiple target initial areas; For each set of target initial regions, the multiple target initial regions are merged to obtain the region; and Based on the multiple regions, the multiple nodes in the region map are determined.

4. The method according to claim 2, wherein, The determination of the set of stopping points for the transport vehicles based on their driving trajectories includes: Multiple target trajectory points are determined from the driving trajectory of the transport vehicle; and Based on the dwell time and location of each of the multiple target trajectory points, the set of dwell points of the transport vehicle is determined.

5. The method according to claim 2, wherein, The step of using a graph neural network to extract features from the region map to obtain the target feature vector includes: For each of the plurality of nodes, dynamic data and static data of the node are determined to obtain the dynamic data and static data of each of the plurality of nodes, wherein the static data is data related to the region; and the dynamic data is data related to the transport vehicle. Based on the dynamic and static data of each of the multiple nodes, a node matrix related to the multiple nodes is determined; Based on the plurality of connecting edges, determine the connecting edge matrix associated with the connecting edges; and The connection edge matrix and the node matrix are input into the graph neural network to obtain the target feature vector.

6. The method according to claim 2, wherein, The process of obtaining the risk identification result of the region to be identified based on the target feature vector includes: The target feature vector is input into a classifier to obtain the risk identification result of the region to be identified. The classifier includes a sequentially connected fully connected layer and an activation function.

7. A training device for a graph neural network for the relevance identification of hazardous chemicals, comprising: The second generation module is used to generate a sample region map including the target sample region. The sample region map includes multiple sample nodes and multiple sample connection edges. The sample nodes are used to characterize the sample region, which is the stopping area of ​​the transport vehicle used to transport hazardous chemicals. The sample connection edges are generated based on the driving trajectory of the transport vehicle and are used to characterize the association relationship between two sample nodes. The determination module is used to determine sample labels related to the target sample region, wherein the sample labels include degree labels and context labels of target sample nodes corresponding to the target sample region; and The training module is used to train a graph neural network using the sample region map and the sample labels to obtain a trained graph neural network. The training module uses the sample region map and the sample labels to train a graph neural network, and the operation to obtain the trained graph neural network includes: The sample region map is input into the graph neural network, which outputs the sample target feature vector. Based on the target feature vector of the sample, the number of sample connection edges connected to the target sample node is determined as an identification result; Based on the target feature vector of the sample, the context result between the target sample node and its neighboring nodes is determined; Based on the quantity identification result and the degree label, the value of the first loss function is determined; Based on the context results and the context labels, determine the value of the second loss function; A first sample area submap and a second sample area submap are determined from the sample area map, wherein the first sample area submap and the second sample area submap include the target sample area, the first sample area submap is generated based on the historical trajectory of the transport vehicle in a first historical period, and the second sample area submap is generated based on the historical trajectory of the transport vehicle in a second historical period, wherein the first historical period and the second historical period are different; The first sample region sub-map is input into the graph neural network, and the first sample target feature sub-vector is output; The value of the third loss function is determined based on the target feature vector of the sample and the sub-vector of the target sample. The second sample region sub-map is input into the graph neural network, and the second sample target feature sub-vector is output; Based on the first sample target feature vector and the second sample target feature sub-vector, determine the value of the fourth loss function; and The graph neural network is trained using the first loss function value, the second loss function value, the third loss function value, and the fourth loss function value to obtain the trained graph neural network.

8. An apparatus for identifying hazardous chemicals, comprising: The first generation module is used to generate a region map including the region to be identified. The region map includes multiple nodes and multiple connecting edges. The nodes are used to represent regions, which are the areas where transport vehicles for transporting hazardous chemicals stop along their routes. The connecting edges are generated according to the driving trajectory of the transport vehicles and are used to represent the relationship between two nodes. The feature extraction module is used to extract features from the region map using a graph neural network trained according to the graph neural network training method of claim 1, to obtain a target feature vector, wherein the target feature vector is used to characterize at least one of the following feature vectors: the number of connecting edges connected to the node, the context result between the node and its neighboring nodes; and The identification module is used to obtain the risk identification result of the area to be identified based on the target feature vector. The operation of the first generation module in generating a region map including the region to be identified includes: Based on the driving trajectory of the transport vehicle, determine the set of the transport vehicle's stopping points; Based on the set of outposts, the plurality of nodes in the region map are determined; Based on the travel trajectory of the transport vehicle, the connection edges between the plurality of nodes are determined, thus obtaining the plurality of connection edges; and The region map is generated based on the multiple nodes and the multiple connecting edges; The set of stopping points for the transport vehicles, determined based on their travel trajectories, includes: Traverse the target trajectory points on the trajectory point set to check whether the distance between the current anchor point and the successor point is less than the distance threshold; For the current anchor point whose distance is less than the distance threshold, further check whether the interval between the current anchor point and the last successor point is greater than the duration threshold. The successor points whose distance from the current anchor point is less than the distance threshold and whose time interval is greater than the duration threshold, together with the current anchor point, are collectively referred to as the set of dwell points corresponding to the current anchor point. Move the anchor point to the next trajectory point after the current dwell point, and repeat the above process until the anchor point moves to the end of the driving trajectory, thereby finally obtaining the set of dwell points of the transport vehicle.

9. A computer system, comprising: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 6.

10. A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 6.

11. A computer program product, comprising: Computer-executable instructions, when executed, are used to implement the method of any one of claims 1 to 6.

Images