A slow traffic emergency early warning method and system based on artificial intelligence

Abstract

The application belongs to the technical field of traffic emergency early warning, and discloses a slow traffic emergency early warning method and system based on artificial intelligence, which constructs a knowledge graph of a slow traffic area, takes static characteristics, dynamic characteristics of each sub-area and spatial correlation and functional correlation between sub-areas as node and edge attributes, then encodes the characteristics of each sub-area by using a graph neural network to obtain embedded representation, further, combines event characteristics and time dimension characteristics of an emergency, introduces an attention network to calculate regional attention and time attention weights, weights the sub-area embedded representation and the time dimension characteristics, and finally outputs an early warning level and a disposal scheme according to the fused characteristics; thereby effectively solving the problems that the early warning decision lacks scene adaptability and fails to fully consider the differences in regional and time resilience in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic emergency warning, in particular to a slow traffic emergency warning method and system based on artificial intelligence. BACKGROUND

[0002] In the urban slow traffic system, the safety of pedestrians and non-motor vehicles is of great importance. Currently, the slow traffic management department usually deploys a multi-source heterogeneous sensor network to collect a large amount of slow traffic data in real time, and uses artificial intelligence algorithms to deeply analyze these data to identify and predict potential slow traffic emergencies. When the intelligent analysis platform identifies a potential emergency, its core task is to quickly generate warning information according to the nature, severity, location, impact range, and potential risks of the event, and intelligently select the most appropriate decision-making scheme for disposal.

[0003] However, in the prior art, the warning decision selection usually adopts preset rules or simple decision tree models. This fixed rule-based decision-making method has significant limitations, especially in complex and variable slow traffic environments, for example, in leisure areas such as park trails, the nature of a pedestrian falling event is completely different from that of a city trunk road, and its emergency level may be lower. However, existing systems often cannot distinguish between such scene differences and treat them all as the highest level, resulting in excessive response of emergency resources or unnecessary waste of resources.

[0004] In addition, the warning decision of slow traffic emergencies needs to consider the "resilience" differences of different regions and time periods. For example, in a densely populated business district, a minor shared bicycle parking event can quickly cause local traffic congestion or pedestrian traffic obstruction, which has a much greater impact than the same event occurring in a remote park area. Similarly, under night or adverse weather conditions, even a relatively minor event will significantly increase the difficulty of disposal and potential risks. Existing systems often fail to fully integrate these regional characteristics and time dimension "resilience" parameters into the decision-making process, resulting in insufficient universality of the decision-making scheme. This lack of perception and adaptability to regional and temporal "resilience" differences makes the decision-making scheme less precise and efficient in specific scenarios.

[0005] In view of the above problems, the prior art needs to be improved. SUMMARY

[0006] In order to solve the problems in the prior art, the application provides a slow traffic emergency early warning method and system based on artificial intelligence, aiming at solving the problems of the existing slow traffic emergency early warning decision, that is, the limitations of preset rules or simple decision tree models, the inability to distinguish scene differences leading to over-response or waste of emergency resources, and the inability to fully integrate regional characteristics and time dimension "resilience" parameters, resulting in insufficient universality of the decision scheme.

[0007] In a first aspect, the application provides a slow traffic emergency early warning method based on artificial intelligence, and the steps of the method include:

[0008] A1. Obtain a knowledge graph of a slow traffic area; the knowledge graph takes the static characteristics and dynamic characteristics of each sub-area of the slow traffic area as node attributes, and takes the spatial correlation and functional correlation between the sub-areas as edge attributes;

[0009] A2. According to the knowledge graph, encode the characteristics of each sub-area by using a graph neural network to obtain an embedding representation of each sub-area;

[0010] A3. Obtain event characteristics and time dimension characteristics of an emergency event;

[0011] A4. According to the event characteristics, the embedding representation of the sub-area, and the time dimension characteristics, calculate the regional attention weight of each sub-area for the current event decision and the time attention weight of different time dimension characteristics for the current event decision by using an attention network, so as to weight the embedding representation of the sub-area and the time dimension characteristics, and obtain a weighted sub-area characteristic embedding and a weighted time characteristic embedding;

[0012] A5. According to the event characteristics, the weighted sub-area characteristic embedding, and the weighted time characteristic embedding, output an early warning level and a disposal scheme.

[0013] In a second aspect, the application provides a slow traffic emergency early warning system based on artificial intelligence, and the system includes:

[0014] A knowledge graph obtaining module is configured to obtain a knowledge graph of a slow traffic area; the knowledge graph takes the static characteristics and dynamic characteristics of each sub-area of the slow traffic area as node attributes, and takes the spatial correlation and functional correlation between the sub-areas as edge attributes;

[0015] An encoding module is configured to encode the characteristics of each sub-area by using a graph neural network according to the knowledge graph to obtain an embedding representation of each sub-area;

[0016] An event information obtaining module is configured to obtain event characteristics and time dimension characteristics of an emergency event;

[0017] an attention calculation module configured to calculate, by using an attention network, a regional attention weight of each sub-region on a current event decision and a time attention weight of different time dimension features on the current event decision according to the event feature, the embedding representation of the sub-region, and the time dimension feature, so as to weight the embedding representation of the sub-region and the time dimension feature, and obtain a weighted sub-region characteristic embedding and a weighted time characteristic embedding;

[0018] a decision output module configured to output an early warning level and a treatment scheme according to the event feature, the weighted sub-region characteristic embedding, and the weighted time characteristic embedding.

[0019] In summary, the slow traffic emergency early warning method and system based on artificial intelligence provided in the present application comprehensively depict the complexity of the slow traffic environment by constructing a knowledge graph of the slow traffic region, taking the static characteristics, dynamic characteristics of each sub-region, and the spatial correlation and functional correlation between sub-regions as node and edge attributes. On this basis, the graph neural network is used to encode the characteristics of each sub-region to obtain its embedding representation, effectively capturing the deep correlation and features between regions. Further, the event feature and the time dimension feature of the emergency are combined, and the attention network is introduced to calculate the regional attention and time attention weight, and the sub-region embedding representation and the time dimension feature are weighted, realizing the deep fusion of multi-source heterogeneous information and the highlighting of key information. Finally, the early warning level and the treatment scheme are output according to the fused features. Through the above technical solutions, the present application effectively solves the problems of lack of scene adaptability of early warning decision and failure to fully consider the differences in regional and time resilience in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flowchart of the slow traffic emergency early warning method based on artificial intelligence provided in the present application.

[0021] Figure 2 A schematic diagram of the slow traffic emergency early warning system based on artificial intelligence provided in the present application.

[0022] Figure 3 A schematic diagram of the graph neural network.

[0023] In the figure: 1, graph acquisition module; 2, encoding module; 3, event information acquisition module; 4, attention calculation module; 5, decision output module. DETAILED DESCRIPTION

[0024] The technical solutions in the present application will be described clearly and completely in the present application in combination with the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. The components of the present application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0025] It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Meanwhile, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0026] Reference Figure 1 The present application provides an artificial intelligence-based slow traffic emergency early warning method, the steps of the method comprising:

[0027] A1. Obtain the knowledge graph of the slow traffic area; the knowledge graph takes the static and dynamic characteristics of each sub-area of the slow traffic area as node attributes, and takes the spatial and functional correlations between sub-areas as edge attributes;

[0028] A2. According to the knowledge graph, encode the characteristics of each sub-area using a graph neural network to obtain the embedding representation of each sub-area;

[0029] A3. Obtain the event characteristics and time dimension characteristics of the emergency;

[0030] A4. According to the event characteristics, the embedding representation of the sub-area, and the time dimension characteristics, use an attention network to calculate the regional attention weight of each sub-area on the current event decision and the time attention weight of different time dimension characteristics on the current event decision, to weight the embedding representation of the sub-area and the time dimension characteristics, to obtain the weighted sub-area characteristic embedding and the weighted time characteristic embedding;

[0031] A5. According to the event characteristics, the weighted sub-area characteristic embedding, and the weighted time characteristic embedding, output the early warning level and the disposal scheme.

[0032] The slow traffic emergency early warning method based on artificial intelligence proposed in the present application can improve the intelligent level and decision efficiency of the urban slow traffic system in response to emergencies. The core of the method is to build and use the knowledge graph of the slow traffic area, combined with the deep learning model, to realize accurate early warning of emergencies and intelligent recommendation of disposal schemes.

[0033] Specifically, the "slow traffic area" refers to an area in the city where walking and non-motorized vehicles are the main means of transportation, usually including multiple sub-areas of different types. These areas are characterized by dense pedestrian flow, diverse and complex traffic modes, and the occurrence of emergencies may have a significant impact on pedestrian and non-motorized vehicle safety. "Knowledge graph" is a structured knowledge representation form that stores and manages entities in the real world and their relationships in the form of a graph. In the present application, the knowledge graph is used to integrate the static characteristics (such as geographic information, historical pedestrian density, traffic infrastructure density, emergency resource distribution, historical event frequency and type) and dynamic characteristics (such as real-time pedestrian density, real-time weather conditions, real-time traffic flow) of the slow traffic area as node attributes, and the spatial correlation (such as the connection relationship between adjacent sub-areas) and functional correlation (such as the functional dependence relationship between sub-areas with commuting or resting correlation) between sub-areas as edge attributes. Through the knowledge graph, the complex environment of the slow traffic area can be fully and multi-dimensionally described.

[0034] "Graph neural network" is a neural network model specifically designed to handle graph-structured data, which can effectively capture the dependency relationships between nodes (i.e., sub-areas) and the topological structure information of the graph. In the present application, the graph neural network is used to encode the characteristics of each sub-area in the knowledge graph, converting it into a low-dimensional, dense "embedding representation". This embedding representation can capture the deep semantic information and structural information of the sub-area, providing high-quality input for subsequent decision-making. In the field of artificial intelligence, embedding representation refers to a technique that maps high-dimensional, sparse data (such as text, images, nodes or edges in graph-structured data) into a low-dimensional, dense vector space; this mapping aims to capture the semantic information, structural information or contextual relationships in the original data, so that data points with similar semantics or associations are close to each other in the new vector space.

[0035] "Event features" refer to information describing the attributes of the incident itself, such as event type (e.g., traffic accident, person falling, facility failure, etc.), event severity (e.g., minor, moderate, severe), event location (e.g., specific latitude and longitude or sub-area identifier), and event impact range (e.g., impact radius or list of affected sub-areas). "Temporal dimension features" refer to information related to the time of the incident, such as the time period of the incident (e.g., morning rush hour, evening rush hour, off-peak, night), day type (e.g., weekday, weekend), and holiday information (e.g., statutory holidays, special event days). These features are crucial for assessing the urgency and potential impact of the incident.

[0036] "Attention network" is a deep learning model that can assign different weights to input data based on their importance, highlighting key information. In this application, attention network is used to calculate the "regional attention weight" of each sub-area on the current event decision and the "temporal attention weight" of different temporal dimension features on the current event decision. These weights reflect the importance of different regions and temporal dimension features in the current event context, and through the weighting mechanism, the model can pay more attention to the most relevant regions and time information for the current event, resulting in "weighted sub-area feature embedding" and "weighted temporal feature embedding".

[0037] "Warning level" refers to the warning level determined based on the nature, severity, potential impact, and other factors of the incident, such as level 1 warning (highest), level 2 warning, level 3 warning, etc. "Disposal scheme" refers to the recommended response measures for a specific warning level and incident situation, such as dispatching emergency vehicles, evacuating crowds, issuing traffic control information, activating emergency plans, etc.

[0038] The core steps of the slow traffic incident warning method based on artificial intelligence proposed in this application include:

[0039] First, the knowledge graph of the slow traffic area is acquired. The construction of the knowledge graph is the basis of the present application, which takes the static and dynamic characteristics of each sub-area of the slow traffic area as the node attribute, and takes the spatial correlation and functional correlation between the sub-areas as the edge attribute. For example, the static geographic information of each sub-area such as geographic boundary, area, road type, etc. can be collected manually or imported from the existing geographic information system (GIS). At the same time, the historical flow density of each sub-area in different time periods, the density of traffic infrastructure (such as bicycle lane, sidewalk), the distribution of emergency resources (such as hospital, fire station, police station), and the frequency and type of past events can be obtained from the historical database. For dynamic characteristics, the data stream of the sensor network (such as video monitoring, geomagnetic sensor, people flow counter) can be accessed in real time to obtain the real-time flow density of each sub-area, the real-time weather conditions (such as temperature, humidity, rainfall), and the real-time traffic flow (such as bicycle flow, pedestrian flow). When constructing the edge attribute, the channel connection relationship between adjacent sub-areas can be identified according to the geographical location relationship of the sub-areas, and it is taken as a spatial correlation. For example, if two sub-areas are directly connected by a road or a bridge, a spatial correlation edge can be established. In addition, the functional dependency relationship between sub-areas with commuting correlation (such as residential area and business district) or rest correlation (such as park and surrounding catering area) can also be identified according to the city planning, functional zoning or historical data analysis, and it is taken as a functional correlation. For example, if a large number of residents commute from a residential area to a business district, a functional correlation edge can be established between the two sub-areas. In this way, a knowledge graph that fully reflects the complex characteristics of the slow traffic area can be constructed.

[0040] Second, according to the knowledge graph, the characteristics of each sub-region are encoded using a graph neural network to obtain the embedding representation of each sub-region. After obtaining the knowledge graph, it is necessary to convert the graph structure data into numerical vectors that can be processed by machine learning models. For example, a graph convolutional network (GCN) can be used to aggregate the features of nodes (sub-regions) in the knowledge graph. Specifically, the initial feature vector of each sub-region can be composed of its static characteristics and dynamic characteristics. The graph convolutional network aggregates the feature information of neighboring nodes and combines its own features to generate more representative sub-region features. For example, the embedding representation of a sub-region not only contains its own real-time crowd density and traffic infrastructure density, but also integrates the crowd density and traffic infrastructure density information of its neighboring sub-regions, thereby capturing local spatial dependencies. In addition, a graph attention network (GAT) can also be used to calculate the attention weights between each sub-region and its neighbor nodes (also known as neighboring sub-regions), and then perform weighted summation on the neighbor node features according to these weights. For example, for a particular sub-region, its neighbor sub-regions with similar crowd density may be assigned higher attention weights, which means that the features of these neighbor sub-regions have a greater impact on its embedding representation. In this way, the graph neural network can learn the structural and semantic information of each sub-region in the knowledge graph and generate an embedding representation that reflects its multi-dimensional characteristics.

[0041] Third, obtain the event characteristics and time dimension characteristics of the emergency event. When an emergency event occurs or is identified, it is necessary to obtain the key information related to the event in a timely manner. For example, event characteristics can be extracted from event reports through manual input, sensor data analysis, or natural language processing techniques. Event characteristics can include event type (such as traffic accident, person falling, facility damage, crowd gathering, etc.), event severity (such as minor, moderate, severe, urgent, etc.), event location (such as specific latitude and longitude coordinates or sub-region identifier), and event impact range (such as affected road segment, area, or estimated number of affected people). At the same time, the time dimension characteristics of the event occurrence need to be obtained. For example, the timestamp of the event occurrence can be automatically recorded, and the time period (such as morning peak, evening peak, flat peak, night), date type (such as weekday, weekend), and whether it is a holiday, etc. information can be extracted from it. These features will be important inputs for subsequent attention networks and decision outputs.

[0042] Then, according to the event feature, the embedding representation of the sub-region, and the time dimension feature, the attention network is used to calculate the regional attention weight of each sub-region to the current event decision and the time attention weight of different time dimension features to the current event decision, so as to weight the embedding representation of the sub-region and the time dimension feature, and obtain the weighted sub-region characteristic embedding and the weighted time characteristic embedding. For example, a multi-head attention network can be used to calculate multiple attention weights in parallel. Specifically, the event feature and the embedding representation of each sub-region can be fused to form an event-region feature fusion vector. Then, the fusion vector is input into the attention network, so that the attention network learns and outputs the regional attention weight of each sub-region to the current event decision. For example, for a pedestrian falling event occurring in a park, the attention network can give higher regional attention weight to the park sub-region and its surrounding rest area, and give lower weight to the commercial area sub-region far away from the park. Similarly, the event feature and the time dimension feature can be fused to form an event-time feature fusion vector, which is input into the attention network to calculate the time attention weight of different time dimension features to the current event decision. For example, for a traffic accident occurring at night, the attention network can give higher weight to the time dimension feature of "night" because the visibility is low and the emergency response is difficult at night, which requires more urgent attention. After obtaining the regional attention weight, the embedding representation of each sub-region can be weighted and summed to obtain the weighted sub-region characteristic embedding, so as to highlight the most relevant regional information of the current event. Similarly, according to the time attention weight, the time dimension features are weighted and summed to obtain the weighted time characteristic embedding, so as to emphasize the most relevant time information of the current event.

[0043] Finally, according to the event feature, the weighted sub-region characteristic embedding, and the weighted time characteristic embedding, the warning level and the disposal scheme are output. After obtaining the sub-region characteristic embedding and the time characteristic embedding weighted by the attention mechanism, they need to be fused with the original event feature to form a comprehensive event situation feature vector. For example, the three types of feature vectors can be spliced or fused through a multi-layer perception machine. Then, the fused event situation feature is input into a classification or regression model, which can be another neural network layer or a decision model, for predicting the warning level of the emergency event. For example, the model can output "first-level warning", "second-level warning", or "third-level warning", etc. At the same time, the model can also intelligently recommend the most suitable disposal scheme according to the event situation feature and in combination with a preset disposal scheme library. For example, for a traffic accident occurring in a commercial area with a first-level warning, the system can recommend disposal schemes such as "immediately dispatching traffic police and ambulances, blocking the affected road section, and issuing a detour notice". In this way, the present application can provide accurate warning and feasible disposal suggestions according to the real-time situation of the event.

[0044] The slow traffic emergency early warning method based on artificial intelligence proposed in the present application significantly improves the accuracy and adaptability of early warning decision-making by introducing a knowledge graph and an attention mechanism. Traditional early warning systems often rely on preset rules or simple decision tree models, making it difficult to cope with complex and variable slow traffic environments, resulting in excessive response of emergency resources or unnecessary waste. For example, a pedestrian fall event occurring on a park trail may have a much lower emergency level than a traffic accident on a city main road, but the traditional system may not be able to distinguish such differences in scenarios. The present application builds a knowledge graph containing static characteristics, dynamic characteristics, spatial correlations and functional correlations, and uses graph neural networks to encode the characteristics of each sub-region, which can comprehensively capture the complex environmental information of slow traffic areas. This enables the system to more carefully understand the characteristics of different sub-regions and their mutual influences, thereby avoiding a "one-size-fits-all" early warning strategy. In addition, existing systems often struggle to make optimal decisions in the first instance when information is uncertain or events are dynamically evolving, which may delay response due to insufficient information or issue unnecessary warnings due to excessive caution. The present application introduces an attention network to dynamically calculate regional attention weights and time attention weights based on event features, sub-region embedding representations and time dimension features, and weights the sub-region characteristic embeddings and time dimension features. This mechanism enables the model to focus on the most relevant regional and temporal information for the current event, effectively addressing information uncertainty and adapting to the dynamic development of events. For example, under night conditions, even relatively minor events can significantly increase the difficulty of handling and potential risks, and the attention mechanism can give these time dimension features higher weights, thereby improving the sensitivity and accuracy of early warning.

[0045] In summary, the present application models slow traffic areas in multiple dimensions using a knowledge graph and encodes features using a graph neural network, and dynamically adjusts the importance of regional and temporal dimensions using an attention mechanism, ultimately outputting accurate early warning levels and handling schemes. Compared with existing technologies, the present application can more comprehensively consider the complex characteristics of slow traffic areas and the dynamic evolution of events, thereby generating more accurate and adaptive early warning levels and handling schemes, effectively overcoming the limitations of early warning decision-making in existing technologies, and improving the intelligent level and decision-making efficiency of urban slow traffic systems in response to emergencies.

[0046] In some embodiments, step A1 comprises:

[0047] A101. Obtain the static characteristics of each sub-region of the slow traffic area; the static characteristics include geographic information, historical crowd density, traffic infrastructure density, emergency resource distribution, and historical event occurrence frequency and type;

[0048] A102. Obtain the dynamic characteristics of each sub-region of the slow traffic area; the dynamic characteristics include real-time crowd density, real-time weather condition, and real-time traffic flow;

[0049] A103. Obtain the channel connection relationship between adjacent sub-regions in the slow traffic area as the spatial correlation between the sub-regions;

[0050] A104. Obtain the functional dependency relationship between sub-regions with commuting correlation or resting correlation in the slow traffic area as the functional correlation between the sub-regions;

[0051] A105. Construct the knowledge graph according to the static characteristics, the dynamic characteristics, the spatial correlation, and the functional correlation.

[0052] Specifically, in step A101, the static characteristics refer to the relatively stable and slowly changing attribute information of each sub-region in the slow traffic area. The geographic information can include the area, shape, terrain, road network layout, etc. of the sub-region, which provides the basic spatial background of the region; the historical crowd density can refer to the average distribution density or peak density of personnel in the sub-region in the past period of time, which reflects the normal crowd characteristics of the region; the traffic infrastructure density can include the distribution density of facilities such as bicycle lanes, sidewalks, and bus stops, which evaluates the traffic carrying capacity of the region; the emergency resource distribution can refer to the location and quantity of emergency facilities such as medical points, fire stations, and shelters, which evaluates the emergency response capability of the region; the historical event occurrence frequency and type can refer to the types and occurrence times of sudden events that have occurred in the sub-region in the past period of time, which aims to reveal the potential risk patterns of the region. These static characteristics as the basic attributes of the nodes in the knowledge graph provide stable background information for subsequent analysis.

[0053] In step A102, the dynamic characteristics refer to the real-time changing and time-sensitive attribute information of each sub-region in the slow traffic area. The real-time crowd density can refer to the personnel density of the sub-region at the current time obtained through sensors, mobile signaling data, etc., which reflects the instantaneous activity of the region; the real-time weather condition can include temperature, humidity, rainfall, wind speed, etc., which evaluates the impact of weather on slow traffic; the real-time traffic flow can refer to the number of slow traffic tools such as pedestrians and bicycles passing through the sub-region at the current time, which reflects the instantaneous traffic pressure of the region. These dynamic characteristics as the real-time updated attributes of the nodes in the knowledge graph can capture the instantaneous state changes of the region.

[0054] In step A103, the spatial correlation refers to the correlation between different sub-regions in the slow traffic area based on geographical location proximity or physical connectivity. Specifically, the channel connection relationship can refer to whether there is a physical connection such as a road, a bridge, a tunnel, etc. between the sub-regions, so that personnel or vehicles can directly move from one sub-region to another. This correlation is an edge attribute in the knowledge graph, and its purpose is to describe the spatial topology between sub-regions and reflect the propagation path of the emergency in space.

[0055] In step A104, the functional correlation refers to the correlation between different sub-regions in the slow traffic area based on their functional attributes or personnel activity patterns. Specifically, the commuting correlation can refer to a large number of daily commuting between two sub-regions, such as the relationship between residential areas and work areas; the rest correlation can refer to two sub-regions serving leisure and entertainment activities together, such as the relationship between parks and commercial streets. This functional dependence relationship is an edge attribute in the knowledge graph, and its purpose is to describe the functional coupling between sub-regions and reflect the functional impact range of the emergency.

[0056] In step A105, according to the static characteristics, dynamic characteristics, spatial correlations and functional correlations obtained above, the knowledge graph can be constructed. Specifically, each sub-region can be represented as a node in the knowledge graph, and its static characteristics and dynamic characteristics are attributes of the node; and the spatial correlation and functional correlation between the sub-regions are represented as edges connecting these nodes as attributes of the edges. Thus, the constructed knowledge graph can comprehensively and structurally represent the complex information of the slow traffic area, providing high-quality input for subsequent graph neural network coding.

[0057] Through the above technical solutions, the present application can construct a more comprehensive, fine and time-effective slow traffic area knowledge graph. The knowledge graph not only contains the static basic information of the sub-region, but also integrates the real-time changing dynamic data, and clearly distinguishes the spatial correlation and the functional correlation, so that the understanding of the slow traffic area is more in-depth and stereoscopic. Thus, high-quality, multi-dimensional input is provided for subsequent graph neural network coding, which significantly improves the accuracy and representativeness of the embedding representation of each sub-region, and further provides a more solid data foundation and situational awareness capability for the early warning decision of the emergency, effectively improving the accuracy and reliability of the early warning.

[0058] In some embodiments, with reference to Figure 3 , the graph neural network includes a graph convolution network, a graph attention network, and a multi-layer perception;

[0059] Step A2 includes:

[0060] A201. Extract static and dynamic characteristics of each sub-region, as well as spatial and functional correlations between sub-regions from the knowledge graph;

[0061] A202. According to the static characteristics, dynamic characteristics, spatial correlations and functional correlations, use graph convolution network to aggregate features of each sub-region, and obtain aggregated sub-region features;

[0062] A203. According to the aggregated sub-region features, use graph attention network to calculate attention weights between each sub-region and its neighbor nodes, and weight sum neighbor node features according to the attention weights, to determine weighted sub-region features;

[0063] A204. Use multilayer perceptron to perform nonlinear transformation on the weighted sub-region features of each sub-region, to obtain embedding representation of each sub-region.

[0064] Specifically, in step A201, the knowledge graph is constructed according to the above step A1, and the node attributes contain the static and dynamic characteristics of each sub-region, and the edge attributes represent the spatial and functional correlations between sub-regions. This step aims to provide structured input data for subsequent graph neural network processing.

[0065] Further, in step A202, graph convolution network is used to locally aggregate features of each sub-region. This means that the features of each sub-region not only contain its own static and dynamic characteristics, but also fuse the features of its directly connected neighbor sub-regions, as well as the connection information embodied through spatial and functional correlations. In this way, local structural information and neighborhood context information of sub-regions can be effectively captured.

[0066] In addition, in step A203, graph attention network is introduced to solve the problem that graph convolution network may treat all neighbor nodes equally during aggregation. By calculating attention weights, the system can identify neighbor nodes that contribute more to the embedding representation of the current sub-region and give them higher weights. In this way, adaptive weighted aggregation of neighbor node features can be realized, so that the weighted sub-region features can more accurately reflect their importance and relevance in complex networks.

[0067] Finally, in step A204, multilayer perceptron is applied to perform nonlinear mapping on the weighted sub-region features processed by graph convolution network and graph attention network. This nonlinear transformation can enhance the expression ability of the model, so that it can learn more complex and high-dimensional feature representations, thereby converting the original, aggregated and weighted sub-region features into compact and rich semantic information embedding representation, providing high-quality input for subsequent event warning decision-making.

[0068] The scheme of the present application systematically solves the limitations in the above encoding process by refining the graph neural network into a combination of graph convolution network, graph attention network and multilayer perceptron. First, the graph convolution network aggregates the features of the neighboring nodes, so that the feature representation of each sub-region can be integrated into its local context information, thereby capturing the structural features implied by the spatial and functional correlations between sub-regions. Second, the graph attention network further optimizes the feature aggregation process. It can adaptively assign different attention weights according to the correlation between the sub-region and its neighboring nodes, thereby highlighting the neighbor information that is more important to the current sub-region embedding representation, and effectively handling heterogeneous correlation relationships. Finally, the multilayer perceptron performs nonlinear transformation on the aggregated and weighted features, greatly enhancing the nonlinear representation ability of the model, so that the final sub-region embedding representation can more comprehensively and deeply reflect the complex characteristics of the sub-region and its role in the slow traffic network. It is precisely due to this hierarchical and refined processing mechanism that the characteristics of each sub-region can be efficiently and accurately encoded.

[0069] Through the above technical scheme, the present application can obtain a sub-region embedding representation with more robustness and representation ability. Compared with the basic scheme, this combined graph neural network structure can more effectively integrate the static characteristics, dynamic characteristics and complex spatial and functional correlations of each sub-region within the slow traffic region, overcoming the problem of insufficient feature capture that may exist in a single model. As a result, the obtained embedding representation can more accurately reflect the real state and mutual influence of each sub-region, providing higher quality input for subsequent emergency warning decision-making, thereby significantly improving the accuracy and reliability of the warning.

[0070] Preferably, step A202 can include:

[0071] According to the static characteristics and the dynamic characteristics, the feature similarity between each sub-region and its neighboring nodes is calculated;

[0072] According to the feature similarity, the connection weights between each sub-region and its neighboring nodes represented by the spatial correlation and the functional correlation are adjusted;

[0073] According to the adjusted connection weights, the graph convolution network is used to aggregate the features of each sub-region, to obtain the aggregated sub-region features.

[0074] Specifically, before performing feature aggregation, it is necessary to first calculate the feature similarity between each sub-region and its neighbor nodes. The feature similarity can be understood as a measure of the closeness of two sub-regions in static and dynamic characteristics. For example, the cosine similarity, the inverse of the Euclidean distance, or a small neural network can be used to learn the similarity score between the static and dynamic characteristic vectors of two sub-regions. The purpose is to quantify the inherent correlation strength between sub-regions, rather than just their physical or functional connections.

[0075] Further, according to the calculated feature similarity, the connection weights between each sub-region and its neighbor nodes represented by the spatial correlation and the functional correlation are adjusted. This means that the weights of the edges representing the spatial correlation and the functional correlation between sub-regions in the original knowledge graph are no longer fixed values, but are dynamically adjusted according to their feature similarity. For example, if the feature similarity of two sub-regions is high, the connection weight between them can be enhanced; conversely, if the similarity is low, the connection weight can be weakened. This adjustment can be done in various ways, such as multiplying the original connection weight by the similarity score, or converting the similarity to a weight gain through a mapping function. The purpose is to enable the graph convolution network to pay more attention to neighbor nodes with higher feature similarity to the current sub-region when aggregating information, thereby achieving more accurate information transmission.

[0076] Thus, when using a graph convolution network to aggregate features of each sub-region, these adjusted connection weights will be used. The graph convolution network will weight the features of neighbor nodes according to these new weights, and update them in combination with its own features, to finally obtain the aggregated sub-region features.

[0077] The scheme of the present application dynamically adjusts the connection weights in the graph convolution network by introducing feature similarity, solving the problem that traditional graph convolution networks ignore the inherent similarity differences between nodes when aggregating features. Specifically, by calculating the similarity between the static and dynamic characteristics of each sub-region, sub-regions that are more similar in actual properties can be identified. It is precisely because this similarity is used to adjust the connection weights that the graph convolution network can assign higher weights to neighbor nodes that are more similar in characteristics to the target sub-region when performing feature aggregation operations. This ensures that in the process of information propagation, information from more relevant neighbors can be more effectively integrated, so that the aggregated sub-region features can more accurately reflect the comprehensive information of their environment.

[0078] By the technical solution, the accuracy and effectiveness of the sub-region feature aggregation can be improved. By dynamically adjusting the connection weight, the graph convolution network can adaptively capture deeper and feature-based correlations between sub-regions, rather than relying only on the preset topology. This helps to generate more representative and discriminative sub-region embedding representations, providing a more solid foundation for subsequent attention calculation and early warning decision-making, thereby improving the overall accuracy and reliability of slow traffic emergency warning.

[0079] In some preferred embodiments, the following is described by a specific example. Assume that there are two adjacent sub-regions, sub-region A and sub-region B. The static characteristics of sub-region A can include that it is a business district, the historical pedestrian flow density is high, and the dynamic characteristics are real-time pedestrian flow density and real-time traffic flow. The static characteristics of sub-region B can include that it is a residential area, the historical pedestrian flow density is moderate, and the dynamic characteristics are real-time pedestrian flow density and real-time traffic flow.

[0080] First, the cosine similarity between the static characteristic vectors and the dynamic characteristic vectors of sub-region A and sub-region B can be calculated. If the calculation result shows that although they are adjacent in space, due to one being mainly a business area and the other being mainly a residential area, there are significant differences in their pedestrian flow patterns and traffic flow patterns, resulting in a relatively low feature similarity.

[0081] Second, according to this low feature similarity, the spatial correlation connection weight between sub-region A and sub-region B is adjusted. For example, if the original spatial connection weight is 1, it can be multiplied by a similarity-based attenuation factor (for example, if the similarity is 0.3, the attenuation factor can be 0.5), so that the adjusted connection weight becomes 0.5. This means that when the graph convolution network performs feature aggregation, the weight of sub-region A obtaining information from sub-region B is appropriately reduced, and vice versa.

[0082] Thus, when the graph convolution network performs feature aggregation on sub-region A, it will obtain information from sub-region B according to the adjusted connection weight. Since the weight of sub-region B is reduced, the aggregated features of sub-region A will be more influenced by other neighbors (e.g., another business district sub-region) with more similar characteristics, so that the aggregated features can more accurately reflect the characteristics of sub-region A as a business district. This dynamic adjustment mechanism ensures that even if there are connected sub-regions in terms of physics or function, the strength of information transmission will be optimized according to the similarity of their inherent characteristics, avoiding excessive aggregation of irrelevant information and improving the accuracy of feature representation.

[0083] Preferably, step A203 can include:

[0084] According to the aggregated sub-region features, feature similarities between each sub-region and different types of neighbor nodes are calculated;

[0085] According to the feature similarities, spatial attention weights between each sub-region and spatially-related neighbor nodes, and functional attention weights between each sub-region and functionally-related neighbor nodes are calculated respectively;

[0086] According to the spatial attention weights and the functional attention weights, weighted neighbor node features that fuse spatial and functional information are obtained by weighted sum of the aggregated sub-region features of the spatially-related neighbor nodes and the functionally-related neighbor nodes;

[0087] According to the weighted neighbor node features, combined with the aggregated sub-region features of the sub-region itself, weighted sub-region features are obtained.

[0088] Among them, the feature similarities between each sub-region and different types of neighbor nodes can be realized in multiple ways such as cosine similarity, dot product similarity or neural network layer, the purpose of which is to quantify the proximity of the center sub-region and different types of neighbor nodes in the feature space. For example, for spatially-related neighbor nodes, their feature similarities reflect more the proximity in geographical information or physical connection; while for functionally-related neighbor nodes, they reflect the functional coordination or dependency relationship. Further, according to the feature similarities, spatial attention weights and functional attention weights are calculated respectively. This means that the system will independently evaluate the importance of spatially-related and functionally-related relationships to the center sub-region. For example, independent attention mechanisms (such as independent linear transformation layers and softmax functions) can be used to process the neighbor node features of spatially-related and functionally-related relationships respectively, so as to obtain their respective attention weights. Thus, according to the calculated spatial attention weights and functional attention weights, the weighted sum of the aggregated sub-region features of spatially-related neighbor nodes and functionally-related neighbor nodes is performed. This process realizes the fusion of information of different types of neighbor nodes, and the size of the weight reflects the importance of this type of neighbor node to the current sub-region feature representation. In this way, weighted neighbor node features that fuse spatial and functional information can be obtained, which more comprehensively reflect the complex environment in which the sub-region is located. Finally, the weighted neighbor node features that fuse spatial and functional information are combined with the aggregated sub-region features of the sub-region itself. This combination can be a simple concatenation, summation or fusion through an additional neural network layer, the purpose of which is to effectively integrate the inherent characteristics of the sub-region itself with the contextual information from its different types of neighbor nodes, so as to obtain the final weighted sub-region features, which will be used for subsequent nonlinear transformation to generate embedding representation.

[0089] The scheme of the present application introduces differentiated processing of different types of neighbor nodes (spatially related neighbor nodes and functionally related neighbor nodes, spatially related neighbor nodes are neighbor nodes having spatial correlation with the central sub-region, and functionally related neighbor nodes are neighbor nodes having functional correlation with the central sub-region) in the graph attention network calculation process, and calculates the attention weights thereof respectively, thereby solving the problem that the traditional graph attention network may not finely capture the differentiated influence of different types of correlation on the sub-region feature representation. Specifically, by calculating the feature similarity between each sub-region and different types of neighbor nodes, and based thereon, calculating the spatial attention weight and the functional attention weight respectively, the model can adaptively adjust the attention degree of the neighbor node features according to the importance of the actual correlation type. Thus, the weighted sum of the aggregated sub-region features of the spatially related neighbor nodes and the functionally related neighbor nodes can more accurately fuse the context information from different types of correlation to form a more representative weighted neighbor node feature. Finally, the weighted neighbor node feature is combined with the aggregated sub-region feature of the sub-region itself, ensuring that the final sub-region feature not only contains its own information, but also fully absorbs important context information from different types of neighbors after fine weighting, thereby improving the accuracy and richness of the sub-region embedding representation.

[0090] Through the above technical scheme, the present application can more finely capture the complex relationship between each sub-region and its neighbor nodes in the slow traffic area, especially in distinguishing spatial correlation and functional correlation. This detailed attention mechanism enables the model to more accurately understand and utilize the information contained in different types of correlation when encoding the sub-region characteristics, avoiding feature confusion or information loss that may be caused by simple aggregation. Thus, the obtained sub-region embedding representation will be more accurate and expressive, and can more effectively reflect the real state and potential risks of the sub-region in the complex traffic environment, thereby providing more reliable input for subsequent emergency warning decision-making, significantly improving the accuracy and robustness of the warning.

[0091] In some embodiments, the event features include event type, event severity, event occurrence location, and event impact range;

[0092] The time dimension features include the time period of the occurrence of the emergency event, the date type, and the holiday information;

[0093] Step A4 includes:

[0094] A401. According to the event features, the embedding representation of the sub-region, and the time dimension features, using a multi-head attention network, respectively calculating the regional attention weight of each sub-region on the current event decision, and the time attention weight of different time dimension features on the current event decision;

[0095] A402. According to the regional attention weights, the embedding representations of the sub-regions are weighted and summed to obtain weighted sub-region characteristic embeddings;

[0096] A403. According to the time attention weights, the time-dimension features are weighted and summed to obtain weighted time characteristic embeddings.

[0097] Among them, the event features aim to comprehensively describe the essential attributes and potential impacts of the emergency. Specifically, the event type can include traffic accidents, natural disasters (such as heavy rain, snow), public safety incidents, etc.; the event severity can be quantified into different levels such as slight, moderate, severe, or evaluated through specific indicators such as casualty number, property loss; the event occurrence location can be accurate to specific geographic coordinates or sub-region number; the event influence range defines the size of the area affected by the event or the number of affected people. The acquisition of these features can be achieved through various ways such as sensor data, manual input, news report analysis, etc., and the purpose is to provide comprehensive event background information for subsequent warning decision-making.

[0098] The time-dimension features are used to capture the temporal context information of the event occurrence, which is crucial for understanding the dynamic evolution of the event and predicting its impact. The time period of the event occurrence can be refined to hours, minutes, such as distinguishing peak hours, flat peak hours or night; the date type can distinguish weekdays, weekends or specific dates; the holiday information explicitly indicates whether the event occurs on a statutory holiday or a special holiday, because holidays are usually accompanied by significant changes in human flow and traffic patterns. The introduction of these time-dimension features helps the model identify the potential risks and impact patterns of the event under different time backgrounds.

[0099] Specifically, in step A401, multi-head attention networks are used to calculate multiple attention weights in parallel, thereby capturing the complex relationships between event, region and time features from different "representation subspaces". For example, one attention head may focus on the association between event type and regional passenger flow density, while another attention head may focus on the association between event severity and time period. Through this multi-head mechanism, the regional attention weights of each sub-region on the current event decision and the time attention weights of different time-dimension features on the current event decision can be more comprehensively and robustly captured.

[0100] In step A402, the calculated regional attention weights are applied to the embedding representations of the sub-regions. This means that for the sub-regions more relevant to the current event context, their embedding representations will obtain higher weights in the weighted sum process, thereby playing a greater role in subsequent decision-making. The weighted sub-region characteristic embeddings can more accurately reflect the importance of each sub-region under the current event background.

[0101] Similarly, in step A403, temporal attention weights are applied to the temporal dimension features. This allows the temporal dimension features that are more relevant to the current event context (e.g., if the event occurs during rush hour, the weight of the "rush hour" temporal feature will be higher) to be more strongly expressed after weighted summation. The weighted temporal feature embedding can more finely capture the impact of the time context of the event occurrence on the decision.

[0102] The scheme of the present application can more comprehensively and deeply understand the inherent properties of the emergency and the time context of its occurrence by defining the event features and the temporal dimension features in detail and introducing a multi-head attention network for weight calculation. Specifically, the detailed definition of the event features allows the model to grasp the essence of the event from multiple dimensions (type, severity, location, impact range) and provide rich semantic information for subsequent attention calculation. The introduction of the temporal dimension features allows the model to perceive the dynamic context of the event occurrence, for example, the same type of event may have completely different impacts and warning needs in different time periods (such as weekday rush hour and weekend night).

[0103] The application of the multi-head attention network is the core of the present scheme. Its working principle is that by running multiple attention mechanisms in parallel, each mechanism learns different feature association patterns. For example, one attention head may focus on the association between the event occurrence location and a specific sub-region, while another may focus on the association between the event type and the historical event occurrence frequency. This multi-angle feature fusion and weight distribution mechanism allows the model to extract the most critical regional and temporal information from the complex event context and assign appropriate weights to them. Thus, the weighted sub-region feature embedding and the weighted temporal feature embedding can more accurately reflect which regions and time factors are most critical to the warning decision under the current event context, thereby providing more accurate input for subsequent warning level determination and disposition scheme formulation.

[0104] By the technical solution, the accuracy and the fine level of the slow traffic emergency early warning can be improved. Specifically, the detailed definition of the event features and the time dimension features enables the model to capture the deep semantic and dynamic context information of the event, avoiding the early warning deviation caused by information loss or coarse-grained description. The introduction of the multi-head attention network enables the model to evaluate the complex interaction among the event, the region and the time from multiple dimensions and a more comprehensive perspective, so as to calculate more representative and discriminative region attention weights and time attention weights. Thus, the weighted sub-region characteristic embedding and the weighted time characteristic embedding can more accurately reflect the importance of the sub-regions and the time dimension features to the early warning decision under the current event situation, thereby improving the accuracy of the early warning level determination and the pertinence of the disposal scheme, and effectively improving the emergency response capability and the resilience of the slow traffic system.

[0105] Preferably, step A401 can include:

[0106] According to the event features, the embedding representation of the sub-region and the time dimension features, an event-region feature fusion vector and an event-time feature fusion vector are respectively constructed;

[0107] The event-region feature fusion vector is input into a first attention head group of the multi-head attention network, a plurality of region attention weights are calculated in parallel by using the first attention head group, and the region attention weights are aggregated to obtain the region attention weights of the sub-regions to the current event decision;

[0108] The event-time feature fusion vector is input into a second attention head group of the multi-head attention network, a plurality of time attention weights are calculated in parallel by using the second attention head group, and the time attention weights are aggregated to obtain the time attention weights of the different time dimension features to the current event decision.

[0109] Specifically, the event features include the event type, the event severity, the event occurrence position and the event influence range. The embedding representation of the sub-region is a vector representation obtained by encoding the characteristics of each sub-region by using a graph neural network, which contains the static characteristics and the dynamic characteristics of each sub-region. The time dimension features include the time period of the emergency event, the date type and the holiday information.

[0110] where the constructing event-region feature fusion vector refers to fusing the event feature of the current emergency event with the embedding representation of each sub-region in the slow traffic region. This fusion can be realized in various ways, for example, the event feature vector and the sub-region embedding representation vector can be spliced, element-wise added, or nonlinearly transformed through a fully connected layer, etc. to form a comprehensive vector that can reflect both the event context and the region characteristics. The purpose is to provide a unified input containing event context information for the subsequent calculation of region attention weights.

[0111] Similarly, the constructing event-time feature fusion vector refers to fusing the event feature of the current emergency event with the time dimension feature. This can also be realized in various ways such as splicing, element-wise addition, or nonlinear transformation, etc. to generate a comprehensive vector that can reflect both the event context and the time characteristics. The purpose is to provide a unified input containing event context information for the subsequent calculation of time attention weights.

[0112] In practical applications, the multi-head attention network is designed to contain multiple independent attention heads, each of which can learn different attention patterns. The present application divides these attention heads into different groups. Specifically, the first attention head group is specifically used to process the event-region feature fusion vector, and each attention head inside it calculates different region attention weights in parallel. These parallel calculated weights are then aggregated, for example by averaging, weighted summation or another fully connected layer, to obtain the final region attention weights of each sub-region on the current event decision. This design enables the model to capture the relevance between events and regions from multiple perspectives.

[0113] At the same time, the second attention head group is specifically used to process the event-time feature fusion vector, and each attention head inside it calculates different time attention weights in parallel. These parallel calculated weights are also aggregated to obtain the final time attention weights of different time dimension features on the current event decision. This design enables the model to capture the relevance between events and time dimensions from multiple perspectives.

[0114] The scheme of the present application effectively decouples the complex interaction between events and regions, events and time dimensions by fusing event features with regional characteristics and time characteristics respectively and processing them using different attention head groups in the multi-head attention network. Specifically, the construction of the event-region feature fusion vector enables the calculation of regional attention weights to focus more on the influence of events at different spatial locations, while the construction of the event-time feature fusion vector enables the calculation of time attention weights to focus more on the influence of events in different time periods. By inputting these fusion vectors into the first and second attention head groups of the multi-head attention network respectively, different types of attention weights can be learned and optimized independently and in parallel in their respective context environments. This separation processing mechanism avoids interference between different dimensional features, enabling the model to more accurately capture the unique contributions of regional and time dimensions in event situations, thereby providing more accurate attention information for subsequent weighting and decision-making.

[0115] Through the above technical solution, the present application can realize more refined and targeted calculation of regional attention weights and time attention weights. This separated and parallel attention mechanism significantly improves the model's ability to identify the importance of different dimensional features in complex situations, avoiding confusion caused by feature interaction, thereby making the final warning level and disposal scheme decision more accurate and reliable. In addition, this design also enhances the model's interpretability, enabling analysts to more clearly understand which regional factors and time factors play a key role in a particular emergency.

[0116] In some preferred embodiments, the following is described through a specific example. Suppose in a slow traffic area, a "runaway electric bicycle hits a person" emergency event occurs.

[0117] First, the system obtains the event features of the event, such as event type "traffic accident", severity "moderate", location "a certain intersection in A sub-region", and influence range "the intersection and surrounding 50 meters". At the same time, the current time dimension features are obtained, such as time period "work peak period (17:00-18:00)", date type "workday", and holiday information "non-holiday". In addition, the system has obtained the embedding representation of each sub-region through the graph neural network, such as the embedding representation of A sub-region being vector VA, the embedding representation of B sub-region being vector VB, etc.

[0118] Next, to compute the region attention weights, the system constructs event-region feature fusion vectors. For example, the event feature vector is concatenated with the embedding representation VAof the A sub-region, forming a fusion vector F_EA. Similarly, F_EBis formed by concatenating with the embedding representation VBof the B sub-region, and so on. These fusion vectors F_EA, F_EB... are input to the first attention head group of the multi-head attention network. Multiple attention heads within this head group compute in parallel, for example, one head might focus on the association between event type and region traffic infrastructure density, another head might focus on the association between event severity and region historical people flow density. The region attention weights (e.g., 0.7 for the A sub-region, 0.2 for the B sub-region) from these parallel computations are then aggregated to obtain the final region attention weights.

[0119] Meanwhile, to compute the time attention weights, the system constructs event-time feature fusion vectors. For example, the event feature vector is concatenated with the time dimension feature vector, forming a fusion vector F_ET. This fusion vector F_ETis input to the second attention head group of the multi-head attention network. Multiple attention heads within this head group compute in parallel, for example, one head might focus on the association between event type and time period (peak hours), another head might focus on the association between event severity and date type (weekdays). The time attention weights (e.g., 0.8 for peak hours, 0.6 for weekdays) from these parallel computations are then aggregated to obtain the final time attention weights.

[0120] In this way, the system can accurately identify which sub-regions (e.g., the A sub-region gets a high weight due to its high people flow density and traffic infrastructure density) and which time dimensions (e.g., peak hours get a high weight due to its high traffic flow) are more relevant and important to the decision-making of the “electric bicycle out of control hitting people” event.

[0121] In some embodiments, step A5 comprises:

[0122] A501. Fusing the event feature, the weighted sub-region characteristic embedding, and the weighted time characteristic embedding to obtain a fused event context feature;

[0123] A502. Determining the warning level and the handling scheme according to the event context feature.

[0124] Specifically, in step A501, the event features, the weighted sub-region characteristic embeddings, and the weighted time characteristic embeddings are fused to generate a unified and comprehensive event context feature. This fusion can be achieved in various ways, for example, the feature vectors can be spliced to form a longer feature vector, or a fully connected neural network layer can be used to perform nonlinear transformation and combination of the features to extract higher-level abstract features. The purpose of fusion is to integrate information from different dimensions (events themselves, spatial regions, time) into a unified representation, so that subsequent decision-making processes can fully consider all relevant factors. Thus, the fused event context feature can more completely describe the comprehensive information of the background, affected area, and occurrence period of the current emergency.

[0125] In step A502, the warning level and the disposal scheme are determined according to the fused event context feature. This determination process can utilize a pre-trained classification model or regression model, for example, it can be a multi-layer perceptron, support vector machine, or decision tree model. The model is trained to map the input event context feature to a predefined warning level (e.g., level one warning, level two warning, level three warning, etc.) and the corresponding disposal scheme. The determination of the warning level aims to assess the potential harm and urgency of the event, while the determination of the disposal scheme aims to provide targeted countermeasures, such as traffic control recommendations, emergency resource allocation instructions, personnel evacuation paths, etc.

[0126] The scheme of the present application can combine the essential attributes of the emergency, its potential impact in a specific region, and the particularity of the occurrence time period by first fusing the event features, the weighted sub-region characteristic embeddings, and the weighted time characteristic embeddings. It is precisely due to this deep fusion of multi-dimensional information that the system can form a more comprehensive and accurate event context feature. On this basis, by analyzing the fused event context feature, the system can more accurately assess the severity, impact range, and urgency of the event, and thus determine the most appropriate warning level and disposal scheme. This mechanism of first fusion and then decision-making ensures the scientificity and rationality of the decision, avoiding one-sided judgments that may be caused by single-dimensional information.

[0127] By the technical solution, the event feature of the emergency, the sub-region feature influenced by the emergency, and the time dimension feature of the emergency are effectively integrated to form a high-dimensional and comprehensive event situation feature. The comprehensive feature representation enables the subsequent warning level determination and disposal scheme formulation process to fully consider the complexity and variability of the emergency, thereby significantly improving the accuracy of the warning and the pertinence and practicality of the disposal scheme. Thus, more reliable decision support can be provided for the slow traffic management department, the response efficiency and disposal capability of the emergency are effectively improved, and the negative influence of the emergency on the slow traffic system is minimized.

[0128] In some embodiments of the present application, the warning level and the disposal scheme are determined according to the fused event situation feature, which can be determined by looking up a table or referring to the coping scheme (including the determined warning level and disposal scheme) that achieves the best effect under similar event situation feature conditions in historical processing data. However, in actual application, relying only on the event situation feature may not be able to fully evaluate the deep influence of the emergency on the urban slow traffic system, resulting in that the determination of the warning level and the disposal scheme lacks consideration of the urban resilience, thereby affecting the accuracy of the warning and the effectiveness of the disposal. In this regard, the present application further proposes a step of determining the warning level and the disposal scheme, which introduces the urban resilience index to more comprehensively evaluate the potential influence of the emergency, so as to generate a more targeted and effective warning and disposal scheme.

[0129] Preferably, the step A502 can include:

[0130] obtaining the emergency resource reserve, the traffic network redundancy, and the historical disaster recovery data of the slow traffic region as the urban resilience index;

[0131] evaluating the potential influence of the emergency on the resilience of the urban slow traffic system according to the event situation feature and the urban resilience index to obtain an event resilience influence evaluation result;

[0132] determining the warning level according to the event resilience influence evaluation result and a preset warning level determination rule;

[0133] determining the disposal scheme according to the warning level, the event resilience influence evaluation result, and a preset disposal scheme library.

[0134] The urban resilience indicator refers to the quantitative measurement of the ability of the urban slow traffic system to resist, adapt, and recover from the impact of a sudden event. Specifically, emergency resource reserves can include the distribution and quantity of information on medical sites, shelters, rescue teams, and material reserve points in geographic space, as well as their schedulability; the traffic network redundancy can refer to the ability of the slow traffic network to have backup paths or alternative solutions, such as the existence of secondary roads or trails available for pedestrians or non-motor vehicles to pass when the main road is blocked; historical disaster recovery data can include historical records of the time required for the urban slow traffic system to recover to normal operation, resources invested, and effective measures taken after similar events in the past. These indicators can be obtained through various means such as city big data platforms, geographic information systems (GIS), historical statistical data, and emergency management department reports.

[0135] Further, the event resilience impact assessment result is obtained by comprehensively analyzing the interaction between the event context characteristics of the sudden event and the urban resilience indicators. For example, an evaluation model can be constructed that can quantify the impact degree of the urban slow traffic system and its recovery ability under a specific event context. The evaluation result aims to provide a comprehensive perspective to understand the potential impact of the event on the long-term or short-term stability of the urban slow traffic system.

[0136] Therefore, the determination of the warning level is no longer based solely on the characteristics of the event itself, but is combined with the urban response and recovery capabilities. The preset warning level determination rule can be a multi-dimensional decision matrix or a machine learning model, with the event resilience impact assessment result as input and the specific warning level as output. The determination of the disposal scheme is based on the warning level and the event resilience impact assessment result, and the most suitable response measures are selected or generated from the preset disposal scheme library. The disposal scheme library can include standardized or customized plans for different event types, different warning levels, and different resilience impact situations, such as traffic control schemes, personnel evacuation paths, emergency material allocation plans, and information release strategies.

[0137] The scheme of the present application effectively solves the problem that the warning and disposal scheme may not be accurate and effective only by relying on the event situation characteristics by introducing the urban resilience indicators when determining the warning level and disposal scheme. Specifically, when a sudden event occurs, the event situation characteristics are first obtained, which provides the nature and influence range information of the event itself. On this basis, by obtaining the urban resilience indicators such as emergency resource reserves of slow traffic area, traffic network redundancy, and historical disaster recovery data, the carrying capacity and recovery potential of the city when facing such events can be comprehensively understood. Then, the event situation characteristics and the urban resilience indicators are comprehensively analyzed to evaluate the potential impact of the sudden event on the resilience of the urban slow traffic system. This evaluation process can reveal the deep damage that the event may cause and the weak link of the urban system in response and recovery. For example, even if the same type of event occurs in an area with sufficient emergency resources and high traffic network redundancy, its resilience impact may be much smaller than in an area with scarce resources and fragile traffic network. In this way, a more insightful event resilience impact evaluation result can be obtained. Finally, based on this evaluation result, combined with the preset warning level determination rules, a more accurate and realistic warning level can be determined. At the same time, according to the warning level, the event resilience impact evaluation result, and the preset disposal scheme library, a more targeted and effective disposal scheme can be selected or generated. For example, for the area with weak system recovery capacity according to the resilience impact evaluation result, even if the event is generally severe, a higher level of warning and more comprehensive disposal measures may be triggered to deploy resources in advance, minimize losses, and speed up recovery.

[0138] Through the above technical scheme, the present application can overcome the limitation of the traditional method that the urban system resilience is not considered in the sudden event warning and disposal. By introducing the urban resilience indicators and conducting event resilience impact evaluation, the determination of the warning level is more scientific and accurate, which can more truly reflect the potential impact of the event on the urban slow traffic system and the recovery difficulty. At the same time, the disposal scheme is also more targeted and forward-looking, which can make full use of the existing resources of the city and optimize the emergency response strategy, thereby significantly improving the resilience and recovery capacity of the urban slow traffic system in response to sudden events, effectively protecting the safety of pedestrians and non-motor vehicle users, and minimizing the impact of events on the operation of the city.

[0139] Reference Figure 2 The present application provides a slow traffic sudden event warning system based on artificial intelligence, which comprises:

[0140] The atlas acquisition module 1 is configured to acquire a knowledge graph of the slow traffic area, wherein the knowledge graph takes static characteristics and dynamic characteristics of each sub-area of the slow traffic area as node attributes, and takes spatial correlation and functional correlation between the sub-areas as edge attributes (for details, refer to step A1 in the foregoing description) ;

[0141] The encoding module 2 is configured to encode characteristics of each sub-area by using a graph neural network according to the knowledge graph, to obtain an embedding representation of each sub-area (for details, refer to step A2 in the foregoing description) ;

[0142] The event information acquisition module 3 is configured to acquire event features and time dimension features of the sudden event (for details, refer to step A31 in the foregoing description) ;

[0143] The attention calculation module 4 is configured to calculate, by using an attention network, a regional attention weight of each sub-area to a current event decision and a time attention weight of different time dimension features to the current event decision, according to the event features, the embedding representation of the sub-area, and the time dimension features, to weight the embedding representation of the sub-area and the time dimension features, to obtain a weighted sub-area characteristic embedding and a weighted time characteristic embedding (for details, refer to step A4 in the foregoing description) ;

[0144] The decision output module 5 is configured to output an early warning level and a disposal scheme according to the event features, the weighted sub-area characteristic embedding, and the weighted time characteristic embedding (for details, refer to step A5 in the foregoing description).

[0145] The above description is only an embodiment of the present application and is not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for early warning of sudden incidents in slow-moving traffic based on artificial intelligence, characterized in that, The steps of this method include: A1. Obtain a knowledge graph of the slow traffic area; the knowledge graph uses the static and dynamic characteristics of each sub-region of the slow traffic area as node attributes, and the spatial and functional associations between sub-regions as edge attributes; A2. Based on the knowledge graph, the characteristics of each sub-region are encoded using a graph neural network to obtain the embedding representation of each sub-region; A3. Obtain the event characteristics and time-related characteristics of the emergency; A4. Based on the event features, the embedding representation of the sub-region, and the time dimension features, use an attention network to calculate the regional attention weight of each sub-region for the current event decision and the time attention weight of different time dimension features for the current event decision, so as to weight the embedding representation of the sub-region and the time dimension features to obtain the weighted sub-region feature embedding and the weighted time feature embedding. A5. Based on the event characteristics, the weighted sub-region characteristic embedding, and the weighted time characteristic embedding, output the warning level and handling plan; The event characteristics include event type, event severity, event location, and event impact scope; The time dimension features include the time period, date type, and holiday information of the emergency; Step A4 includes: A401. Based on the event features, the embedded representation of the sub-regions, and the time dimension features, a multi-head attention network is used to calculate the regional attention weight of each sub-region for the current event decision and the time attention weight of different time dimension features for the current event decision. A402. Based on the region attention weight, perform a weighted summation on the embedding representation of the sub-region to obtain the weighted sub-region feature embedding; A403. Based on the time attention weight, the time dimension features are weighted and summed to obtain the weighted time feature embedding; Step A401 includes: Based on the event features, the embedded representation of the sub-region, and the time dimension features, an event-region feature fusion vector and an event-time feature fusion vector are constructed respectively. The event-region feature fusion vector is input into the first attention head group of the multi-head attention network. The first attention head group is used to calculate multiple region attention weights in parallel, and the region attention weights are aggregated to obtain the region attention weights of each sub-region for the current event decision. The event-time feature fusion vector is input into the second attention head group of the multi-head attention network. Multiple time attention weights are calculated in parallel using the second attention head group, and the time attention weights are aggregated to obtain the time attention weights of the different time dimension features for the current event decision.

2. The method for early warning of sudden incidents in slow-moving traffic based on artificial intelligence according to claim 1, characterized in that, Step A1 includes: A101. Obtain the static characteristics of each sub-region of the slow-traffic area; the static characteristics include geographic information, historical pedestrian density, traffic infrastructure density, emergency resource distribution, and the frequency and type of historical events; A102. Obtain the dynamic characteristics of each sub-area of ​​the slow-traffic zone; the dynamic characteristics include real-time pedestrian density, real-time weather conditions, and real-time traffic flow. A103. Obtain the channel connection relationship between adjacent sub-regions in the slow traffic area as the spatial association between the sub-regions; A104. Obtain the functional dependencies between sub-regions in the slow traffic area that have commuting or rest associations, as the functional associations between the sub-regions; A105. Construct the knowledge graph based on the static characteristics, the dynamic characteristics, the spatial associations, and the functional associations.

3. The method for early warning of sudden incidents in slow-moving traffic based on artificial intelligence according to claim 1, characterized in that, The graph neural network includes graph convolutional networks, graph attention networks, and multilayer perceptrons; Step A2 includes: A201. Extract the static and dynamic characteristics of each sub-region, as well as the spatial and functional relationships between the sub-regions, from the knowledge graph; A202. Based on the static characteristics, dynamic characteristics, spatial correlation, and functional correlation, a graph convolutional network is used to aggregate the features of each sub-region to obtain the aggregated sub-region features; A203. Based on the aggregated sub-region features, use a graph attention network to calculate the attention weights between each sub-region and its neighboring nodes, and perform a weighted summation of the neighboring node features based on the attention weights to determine the weighted sub-region features. A204. Using a multilayer perceptron, a nonlinear transformation is performed on the weighted features of each sub-region to obtain the embedded representation of each sub-region.

4. The method for early warning of sudden incidents in slow-moving traffic based on artificial intelligence according to claim 3, characterized in that, Step A202 includes: Based on the static and dynamic characteristics, calculate the feature similarity between each sub-region and its neighboring nodes; Based on the feature similarity, the connection weights between each sub-region represented by the spatial association and the functional association and its neighboring nodes are adjusted; Based on the adjusted connection weights, a graph convolutional network is used to aggregate features of each sub-region to obtain the aggregated sub-region features.

5. The method for early warning of sudden incidents in slow-moving traffic based on artificial intelligence according to claim 3, characterized in that, Step A203 includes: Based on the aggregated sub-region features, calculate the feature similarity between each sub-region and different types of neighboring nodes; Based on the aforementioned feature similarity, the spatial attention weight between each sub-region and its spatially associated neighboring nodes, as well as the functional attention weight between each sub-region and its functionally associated neighboring nodes, are calculated respectively. Based on the spatial attention weight and the functional attention weight, the aggregated sub-region features of the spatially associated neighbor nodes and the functionally associated neighbor nodes are weighted and summed to obtain the weighted neighbor node features that fuse spatial and functional information. Based on the weighted neighbor node features, combined with the aggregated sub-region features of the sub-region itself, the weighted sub-region features are obtained.

6. The method for early warning of sudden incidents in slow-moving traffic based on artificial intelligence according to claim 1, characterized in that, Step A5 includes: A501. The event features, the weighted sub-region feature embedding, and the weighted time feature embedding are fused to obtain the fused event context features; A502. Determine the warning level and the response plan based on the characteristics of the event context.

7. An artificial intelligence-based early warning system for sudden incidents in slow-moving traffic, characterized in that, The system includes: The knowledge graph acquisition module is used to acquire the knowledge graph of the slow traffic area; the knowledge graph uses the static and dynamic characteristics of each sub-region of the slow traffic area as node attributes, and the spatial and functional associations between sub-regions as edge attributes. The encoding module is used to encode the characteristics of each sub-region based on the knowledge graph using a graph neural network to obtain the embedded representation of each sub-region; The event information acquisition module is used to acquire the event characteristics and time dimension characteristics of sudden events; The attention calculation module is used to calculate the regional attention weight of each sub-region for the current event decision and the temporal attention weight of different temporal features for the current event decision using an attention network, based on the event features, the embedding representation of the sub-region, and the temporal dimension features. This is used to weight the embedding representation of the sub-region and the temporal dimension features to obtain the weighted sub-region feature embedding and the weighted temporal feature embedding. The decision output module is used to output the warning level and the handling plan based on the event characteristics, the weighted sub-region characteristic embedding, and the weighted time characteristic embedding. The event characteristics include event type, event severity, event location, and event impact scope; The time dimension features include the time period, date type, and holiday information of the emergency; When the attention calculation module calculates the regional attention weights of each sub-region for the current event decision and the temporal attention weights of different temporal features for the current event decision using an attention network, based on the event features, the embedding representations of the sub-regions, and the temporal dimension features, and then weights the embedding representations of the sub-regions and the temporal dimension features to obtain the weighted sub-region feature embeddings and the weighted temporal feature embeddings, the following steps are performed: A401. Based on the event features, the embedded representation of the sub-regions, and the time dimension features, a multi-head attention network is used to calculate the regional attention weight of each sub-region for the current event decision and the time attention weight of different time dimension features for the current event decision. A402. Based on the region attention weight, perform a weighted summation on the embedding representation of the sub-region to obtain the weighted sub-region feature embedding; A403. Based on the time attention weight, the time dimension features are weighted and summed to obtain the weighted time feature embedding; Step A401 includes: Based on the event features, the embedded representation of the sub-region, and the time dimension features, an event-region feature fusion vector and an event-time feature fusion vector are constructed respectively. The event-region feature fusion vector is input into the first attention head group of the multi-head attention network. The first attention head group is used to calculate multiple region attention weights in parallel, and the region attention weights are aggregated to obtain the region attention weights of each sub-region for the current event decision. The event-time feature fusion vector is input into the second attention head group of the multi-head attention network. Multiple time attention weights are calculated in parallel using the second attention head group, and the time attention weights are aggregated to obtain the time attention weights of the different time dimension features for the current event decision.

Images