Urban traffic event semantic recognition method based on knowledge graph

Abstract

The invention discloses an urban traffic event semantic recognition method based on a knowledge graph, and relates to the technical field of intelligent traffic and artificial intelligence, and the method comprises the steps: obtaining the multi-modal traffic data of urban traffic, and constructing a knowledge graph model; preprocessing and feature extraction are carried out on the multi-modal traffic data, the extracted multi-modal features are mapped to entity nodes of a knowledge graph model, a fusion feature vector is generated, and semantic embedding coding is carried out on the fusion feature vector through a graph neural network; constructing an event inference rule base based on a semantic embedding coding result, and performing multi-layer inference calculation on the fusion feature vector by using a graph convolutional neural network to obtain a matching strength score of the candidate traffic event and a standard event mode in the knowledge graph; and in combination with the event space-time constraint condition and the historical event mode, outputting a traffic event recognition result, confidence evaluation and disposal suggestions. According to the invention, the accuracy and practicability of urban traffic event identification are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent transportation and artificial intelligence, and particularly relates to a city traffic event semantic recognition method based on a knowledge graph. BACKGROUND

[0002] With the acceleration of urbanization and the continuous growth of traffic demand, the city traffic system generates a large amount of diversified traffic data in the operation process, including speed, flow and density information collected by traffic sensors, image data obtained by video monitoring, and text and event records from social media and traffic management platforms. These multi-modal data can reflect the overall situation of traffic operation, but due to the heterogeneous data sources, complex semantic expression and high correlation between events, the traditional traffic event recognition methods based on single data source or static rules have obvious deficiencies in accuracy and real-time performance. In recent years, as a structured semantic representation method, the knowledge graph has been widely used in the field of intelligent transportation. By organizing traffic events, participants, time and space information in the form of entities and relationships, it can provide semantic support for traffic event reasoning and analysis. However, existing methods mostly focus on static knowledge representation or single modal data mapping, lacking deep fusion of multi-modal features and semantic reasoning ability, resulting in limited accuracy and generalization ability of event recognition results in complex traffic environment. In addition, the application of existing knowledge graphs in traffic event recognition tasks mostly stays at the level of data visualization and statistical analysis, and has not formed a high-precision event recognition mechanism based on deep semantic embedding and event reasoning rules, which is difficult to meet the needs of real-time and accurate event recognition and response of intelligent traffic management.

[0003] CN117312577A is a traffic event knowledge graph construction method based on multi-layer semantic graph convolutional neural network, which proposes to define an event ontology hierarchical system from top to bottom, and constructs an event extraction model based on the OneRel model, uses a multi-layer semantic graph convolutional neural network to capture deep entity relationship semantic information, constructs a college traffic safety event knowledge graph, and realizes the fusion of global semantic dependency analysis and syntax analysis. This method can improve the recognition ability of long-distance entities and generate a relatively complete knowledge graph. However, its application scope is concentrated in the college traffic safety scene, lacks comprehensive fusion processing of multi-modal traffic data (such as video, sensor, text, etc.), and has not established an event reasoning rule library to realize dynamic event recognition based on semantic embedding, which is difficult to directly promote to complex urban traffic environment;

[0004] CN115658916A A method for constructing a knowledge graph based on traffic accidents and an application method, proposes to construct a knowledge graph from structured traffic accident data, and use it for multi-level, multi-dimensional accident analysis, which broadens the accident prevention ideas. The advantages of this method are that it provides semantic data support for accident analysis, but the type of data set constructed is relatively single, mainly for historical accident data, lacks processing capability for real-time multi-modal traffic data, and only relies on knowledge semantic relationship network in event identification, lacks reasoning mechanism combined with spatio-temporal constraint conditions and historical event mode, which limits its applicability in real-time identification of dynamic events. SUMMARY

[0005] In view of the problems of the existing urban traffic event identification technology, such as insufficient fusion of multi-modal traffic data, limited deep semantic reasoning capability, and lack of dynamic event identification combined with spatio-temporal constraint conditions, the present application is proposed.

[0006] Therefore, the problem to be solved by the present application is how to construct a method that can fuse multi-modal traffic data, has deep semantic embedding and multi-layer reasoning capability, and combines spatio-temporal constraint conditions and historical event mode to realize high-precision, real-time urban traffic event identification.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] In a first aspect, the present application embodiment provides a method for semantic identification of urban traffic events based on a knowledge graph, which comprises,

[0009] Obtaining multi-modal traffic data of urban traffic, constructing a knowledge graph model;

[0010] Preprocessing and feature extraction are performed on the multi-modal traffic data, the extracted multi-modal features are mapped to the entity nodes of the knowledge graph model, a fusion feature vector is generated, and semantic embedding coding is performed on the fusion feature vector through a graph neural network;

[0011] Based on the semantic embedding coding result, an event reasoning rule base is constructed, a graph convolutional neural network is used to perform multi-layer reasoning calculation on the fusion feature vector, and the matching strength score of the candidate traffic event and the standard event mode in the knowledge graph is obtained;

[0012] Based on the matching strength score, combined with the event spatio-temporal constraint condition and the historical event mode, the traffic event identification result, the confidence evaluation and the disposal suggestion are output.

[0013] As a preferred scheme of the city traffic event semantic recognition method based on the knowledge graph, wherein: based on the matching strength score, combined with the event space-time constraint condition and the historical event mode, the traffic event recognition result, the confidence evaluation and the disposal suggestion are output, including:

[0014] The matching strength score is processed by multi-level threshold classification, a four-level recognition confidence classification system is established according to the numerical range of the matching strength score, and a high-confidence candidate event is obtained;

[0015] The space-time attributes of the high-confidence candidate event are extracted from the event instance layer of the knowledge graph model, and the space-time overlap analysis is performed on the real-time road network topology data;

[0016] For the high-confidence candidate event verified by the space-time constraint condition, the historical event mode matching module is called from the dynamic rule layer of the event reasoning rule library, and the historical matching coefficient is calculated;

[0017] The matching strength score and the historical matching coefficient are comprehensively evaluated by using a weighted fusion algorithm to generate a confidence score;

[0018] According to the numerical interval of the confidence score, the high-confidence candidate event is classified into a confirmed event, a to-be-checked event and a false positive event;

[0019] For the confirmed event, the associated disposal scheme is extracted from the semantic relationship layer of the knowledge graph model, and the multi-level disposal suggestion is generated combined with the real-time road network state data;

[0020] The traffic event recognition result, the confidence score and the disposal suggestion are packaged as a structured output message, and are pushed to the traffic management platform through the event instance layer interface of the knowledge graph model.

[0021] As a preferred scheme of the city traffic event semantic recognition method based on the knowledge graph, wherein: the matching strength score is obtained by,

[0022] The standard event mode is extracted from the knowledge graph model, the feature vector coding processing is performed, and the deep graph convolution reasoning network dedicated to event reasoning is established;

[0023] The attribute information of the standard event mode is converted into a 128-dimensional feature vector with the same dimension as the semantic embedding coding result by using the graph neural network;

[0024] The fusion feature vector is input into the deep graph convolution reasoning network for forward reasoning calculation, and the event reasoning feature representation is generated by four-layer convolution operation;

[0025] The sliding window detection algorithm is used to identify the traffic event mode in the event reasoning feature representation, and a candidate traffic event set is generated;

[0026] The candidate traffic event is preliminarily screened, and a multi-dimensional similarity calculation method is used to evaluate the matching strength of the candidate traffic event and the standard event mode, and a matching strength score is generated;

[0027] The candidate traffic event is verified by a causal relationship reasoning rule based on an event reasoning rule base;

[0028] The candidate traffic event verified by the causal relationship reasoning is verified by the spatio-temporal constraint, and a traffic event prediction result is generated.

[0029] As a preferred scheme of the city traffic event semantic recognition method based on the knowledge graph, wherein: the matching strength score includes semantic similarity and structural similarity; the semantic similarity calculates the similarity degree of the feature vector of the candidate traffic event and the feature vector of the standard event mode through the cosine similarity; the structural similarity calculates the topological structure matching degree of the event node based on the adjacency relationship matrix in the knowledge graph model.

[0030] As a preferred scheme of the city traffic event semantic recognition method based on the knowledge graph, wherein: the event reasoning rule base includes a basic rule layer, a composite rule layer and a dynamic rule layer; the basic rule layer is a trigger condition rule for storing a single traffic event; the composite rule layer is a composite reasoning rule defining a joint trigger of multiple events, and a rule combination matrix is established by the event severity level, the influence range, the duration and the disposal urgency; the dynamic rule layer is a weight coefficient of the reasoning rule dynamically adjusted according to the historical traffic event statistical data and the real-time road network state.

[0031] As a preferred scheme of the city traffic event semantic recognition method based on the knowledge graph, wherein: the knowledge graph model includes a traffic event concept layer CTG={C1,C2,...,C n}, an event instance layer ITG={I1,I2,...,I m} and a semantic relationship layer RTG={R1,R2,...,R k}, and the three-layer structure is associated by a semantic mapping function CTG*ITG→RTG.

[0032] As a preferred scheme of the city traffic event semantic recognition method based on the knowledge graph, wherein: the fusion feature vector is fused by feature vector splicing operation of visual feature vector, numerical feature vector and text feature vector mapped to the same entity node, including:

[0033] According to the type attribute of the entity node in the knowledge graph model, the visual feature vector, the numerical feature vector and the text feature vector are respectively mapped to the vehicle entity node, the road entity node and the event entity node, and a corresponding relationship between the feature vector and the entity node is established;

[0034] A graph structure of the adjacency relationship of the entity node in the knowledge graph model is constructed, the fusion feature vector is taken as an initial feature of a graph node, input to a graph neural network for forward propagation, and a node embedding representation with global perception ability is formed through two-layer graph convolution operation;

[0035] A multi-layer graph attention network is used for semantic embedding coding of the node embedding representation, an attention weight learning mechanism is used for adjusting the contribution degree of different neighbor nodes, and a semantic embedding coding result is generated.

[0036] As a preferred scheme of the city traffic event semantic recognition method based on the knowledge graph, wherein: the extraction method of the visual feature vector, the numerical feature vector and the text feature vector is,

[0037] Interframe difference and background modeling preprocessing are performed on the monitoring video stream, a moving target area is extracted and scale normalization is performed;

[0038] A convolutional neural network is used to extract features from the normalized target area, obtain visual features, and perform dimension reduction compression through principal component analysis to generate a visual feature vector;

[0039] The sensor structured data is time window segmented and outlier detected for preprocessing, a sliding window statistical method is used to extract traffic flow parameters to generate numerical features, and the numerical features are subjected to Z-score standardization processing to generate a standardized numerical feature vector;

[0040] The text structured data is segmented and part-of-speech tagged for preprocessing, a bidirectional long short-term memory network based on an attention mechanism is used to extract semantic information of traffic event description to generate text features capturing event type and severity;

[0041] The text features are converted into vector representation through a word embedding model to generate a text feature vector.

[0042] In a second aspect, an embodiment of the present application provides a computer device comprising a memory and a processor, the memory storing a computer program, wherein the computer program instructions are executed by the processor to implement the steps of the city traffic event semantic recognition method based on the knowledge graph according to the first aspect of the present application.

[0043] In a third aspect, an embodiment of the present application provides a computer readable storage medium having stored thereon a computer program, wherein the computer program instructs a processor to implement the steps of the method for identifying a city traffic event based on a knowledge graph according to the first aspect of the present application.

[0044] Compared with the prior art, the present application has the following beneficial effects: by acquiring multi-modal traffic data of city traffic and constructing a knowledge graph model covering an event concept layer, an event instance layer and a semantic relationship layer, unified semantic representation of heterogeneous traffic information is realized, and data islands are effectively broken; by specific preprocessing and feature extraction of different modal data, visual, numerical and text features are mapped to knowledge graph entity nodes and fused, semantic embedding representation with global perception ability is generated by using a graph neural network and a multi-layer attention mechanism, and deep fusion of cross-modal features and improvement of context correlation are realized; based on the semantic embedding encoding result, an event reasoning rule library containing basic rules, compound rules and dynamic rules is constructed, and combined with a deep graph convolution reasoning network, multi-dimensional similarity calculation, causal relationship verification and space-time constraint analysis, a matching strength score of a candidate event and a standard mode is obtained, so that events with high correlation and high credibility are effectively screened; combined with the matching strength score, the space-time constraint condition and the historical event mode, a confidence score is generated by a weighted fusion algorithm, and events are classified according to the score result, multi-level disposal suggestions are generated by extracting disposal schemes and real-time road network states, and the accuracy and practicability of city traffic event identification are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. Among them:

[0046] Figure 1 The flowchart of the method for identifying a city traffic event based on a knowledge graph. DETAILED DESCRIPTION

[0047] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0048] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from the description, and those skilled in the art can make similar generalizations without departing from the concept of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0049] Second, the "one embodiment" or "an embodiment" referred to herein can include a particular feature, structure, or characteristic. The various embodiments appearing at different places in this specification are not necessarily all cumulative or mutually exclusive of each other.

[0050] As described in the above background, the existing methods focus on static knowledge representation or mapping of single modal data, lack of deep fusion of multi-modal features and semantic reasoning ability, resulting in limited accuracy and generalization ability of event recognition results in complex traffic environment. In addition, the application of existing knowledge graph in traffic event recognition task mostly stays at the level of data visualization and statistical analysis, and has not formed a high-precision event recognition mechanism based on deep semantic embedding and event reasoning rules, which is difficult to meet the demand of intelligent traffic management for real-time and accurate event recognition and response. In view of the above problems, the present application provides a kind of city traffic event semantic recognition method based on knowledge graph.

[0051] Figure 1 The flow chart of a kind of city traffic event semantic recognition method based on knowledge graph according to an embodiment of the present application. As shown in Figure 1 In a kind of city traffic event semantic recognition method based on knowledge graph, it includes:

[0052] S1: acquiring multi-modal traffic data of city traffic, and constructing a knowledge graph model;

[0053] Specifically, multi-modal traffic data is collected through a city traffic management platform;Frame extraction and target detection preprocessing are performed on the monitoring video stream to obtain video structured data containing vehicle position, type and motion trajectory;At the same time, the traffic sensor data is denoised and standardized, and sensor structured data containing traffic flow, speed and occupancy rate is obtained;Text cleaning algorithm based on keyword filtering and geographic location label is used to preprocess social media text data, and text structured data related to traffic events is obtained.

[0054] Preferably, a multi-modal data timestamp alignment mechanism is established, and a nearest neighbor time matching algorithm is used to synchronize and align data from different sources according to time dimension, and the time window is set to δt=30 seconds, so that the multi-modal data in the same time window can be effectively fused.

[0055] It should be noted that the multi-modal traffic data includes monitoring video stream distributed in the main intersection of the city, magnetic induction traffic sensor data buried under the road and social media text data of traffic APP.

[0056] Specifically, the association weight information between entities in the semantic relationship layer is calculated, and the semantic association strength of vehicle entities, road entities, time entities and event entities is quantified by analyzing the co-occurrence frequency of entities in the structured data of video, sensor and text.

[0057] The semantic relationship layer is fused with the traffic event concept layer and the event instance layer to complete the construction of the knowledge graph model.

[0058] Preferably, the semantic relationship layer includes causal relationship, temporal relationship, spatial relationship and influence relationship, and the association degree between different event types is represented by a relationship strength matrix.

[0059] In an optional embodiment, the knowledge graph model supports semantic representation and reasoning query of multi-modal data; the knowledge graph model includes a traffic event concept layer CTG={C1, C2,..., C n}, an event instance layer ITG={I1, I2,..., I m} and a semantic relationship layer RTG={R1, R2,..., R k}, and the three-layer structure establishes an association relationship through a semantic mapping function CTGxITG→RTG.

[0060] It should be noted that the traffic event concept layer defines the basic attribute framework of vehicle entities, road entities, time entities and event entities; the vehicle entity includes vehicle type, license plate and driving direction attributes; the road entity includes road segment name, lane number and speed limit attributes; the event instance defines the triggering conditions and determination rules of typical traffic events such as traffic congestion, traffic accident and road construction based on traffic management specifications and historical event statistical data.

[0061] In an optional embodiment, the knowledge graph model is constructed based on urban traffic scenarios, and includes four main event categories: road traffic events, public traffic events, special vehicle events and sudden traffic conditions; each event category is set with four core attribute dimensions: event severity level {first level, second level, third level, fourth level}, impact range {local, regional, city-wide}, duration {short-term, medium-term, long-term} and disposal emergency degree {low, medium, high, very high}.

[0062] S2: Preprocess and feature extract the multi-modal traffic data, map the extracted multi-modal features to the entity nodes of the knowledge graph model, generate a fusion feature vector, and perform semantic embedding coding on the fusion feature vector through a graph neural network;

[0063] Specifically, inter-frame difference and background modeling preprocessing are performed on the monitored video stream, the moving target region is extracted and normalized in scale; a convolutional neural network is used to extract features from the normalized target region to obtain visual features, and principal component analysis is used for dimension reduction and compression to generate a visual feature vector; the sensor structured data is preprocessed by time window segmentation and outlier detection, the traffic flow parameters are extracted by a sliding window statistical method to generate numerical features, and the numerical features are standardized by Z-score to generate a standardized numerical feature vector; the text structured data is preprocessed by word segmentation and part-of-speech tagging, and a bidirectional long short-term memory network based on an attention mechanism is used to extract semantic information of traffic event description to generate text features capturing event type and severity; the text features are converted into vector representation by a word embedding model to generate a text feature vector.

[0064] Preferably, the establishment of the word embedding model comprises: constructing a traffic domain corpus, wherein the traffic domain corpus contains text data of traffic event reports, traffic control texts and traffic regulation documents; preprocessing the traffic domain corpus, wherein the preprocessing includes: text cleaning, removing special characters and irrelevant symbols; word segmentation processing, enhanced word segmentation using a traffic domain professional dictionary; stop word filtering, retaining keywords with traffic semantic features; based on the preprocessed corpus, training an initial word vector using Word2Vec or GloVe algorithm, wherein the dimension of the initial word vector is set to 200-300 dimensions; performing semantic enhancement on the initial word vector through a traffic domain knowledge graph: using entity relationships in the knowledge graph as constraint conditions; using a graph neural network to optimize the spatial distribution of the word vector; using an attention mechanism to differentiate vectors for polysemous words, generating word vector representations with context differentiation ability for the semantic differences of the same word in different traffic contexts; optimizing the word vector space through negative sampling technology to make word vectors with similar traffic semantics close in the vector space.

[0065] It should be noted that the word embedding model adopts a dynamic learning rate adjustment strategy during training, with an initial learning rate of 0.025, which linearly decays to 0.001 with training rounds.

[0066] It should be noted that the multi-modal features include visual features, numerical features and text features; the visual features include vehicle appearance, color and shape information; the visual feature vector retains vehicle type identification and motion state judgment; the traffic flow parameters include vehicle flow rate of change, average vehicle speed and lane occupancy rate; the text feature vector encodes the category and impact range information of the traffic event.

[0067] Further, the fusion feature vector is fused by a feature vector splicing operation of the visual feature vector, the numerical feature vector and the text feature vector mapped to the same entity node, including: according to the type attribute of the entity node in the knowledge graph model, the visual feature vector, the numerical feature vector and the text feature vector are respectively mapped to the vehicle entity node, the road entity node and the event entity node, and the corresponding relationship between the feature vector and the entity node is established; a graph structure of the adjacency relationship of the entity node in the knowledge graph model is constructed, the fusion feature vector is taken as the initial feature of the graph node, input to the graph neural network for forward propagation, and through two-layer graph convolution operation, the node embedding representation with global perception ability is formed; the multi-layer graph attention network is used for semantic embedding coding of the node embedding representation, the contribution degree of different neighbor nodes is adjusted through the attention weight learning mechanism, and the semantic embedding coding result is generated.

[0068] In an optional implementation, the graph neural network includes a first layer graph convolution and a second layer graph convolution; the first layer graph convolution aggregates the feature information of the neighbor nodes through a message passing mechanism, and a specific calculation formula is as follows:

[0069]

[0070] wherein, is the output feature vector of node i at the first layer, tanh (*) is a hyperbolic tangent activation function, N(i) is a neighbor node set of node i, and ij is the attention weight of node i and neighbor node j, F vi is the visual feature vector of node i, F nj is the numerical feature vector of neighbor node j, ||*||2 is an L2 norm, and F tj is the text feature vector of neighbor node j, K is a normalization parameter quantity, and γ k is the kth normalization coefficient, W (1) is the first layer weight matrix.

[0071] It should be noted that the output value range of the first layer graph convolution is in the interval [-1, 1], wherein close to 1 indicates that the node has high importance and strong neighborhood correlation, close to -1 indicates that the node and the neighborhood have a negative correlation, and close to 0 indicates that the node is in a neutral state.

[0072] In an optional implementation, the second layer graph convolution continues to perform feature aggregation and nonlinear transformation, and a specific calculation formula is as follows:

[0073]

[0074] wherein, is the output feature vector of node i at the second layer, σ (*) is a Sigmoid activation function, and μ ijis the attention weight of the second layer, is the output feature of the neighbor node j in the first layer, R ij is the semantic relation matrix of node i and node j, ||*||1 is the L1 norm, is the cosine function, δ is the regularization bias term, W (2) is the second layer weight matrix.

[0075] It should be noted that the second layer graph convolution output value range is in the interval [0, 1], wherein close to 1 indicates that the node has a very high traffic event prediction value after deep semantic embedding, close to 0 indicates that the node semantic information is weak, and the intermediate value indicates that the node has certain semantic expression ability and needs to be combined with other information to judge.

[0076] S3: Based on the semantic embedding coding result, an event reasoning rule library is constructed, and a multi-layer reasoning calculation is performed on the fusion feature vector by using a graph convolutional neural network to obtain a matching strength score of the candidate traffic event and the standard event mode in the knowledge graph;

[0077] Specifically, the event reasoning rule library includes a basic rule layer, a composite rule layer and a dynamic rule layer; the basic rule layer stores the triggering condition rules of single traffic events, including the determination rules of four basic events of road traffic events, public traffic events, special vehicle events and sudden traffic conditions; the composite rule layer defines the composite reasoning rules triggered by multiple events, and establishes a rule combination matrix through the event severity level {first level, second level, third level, fourth level}, the influence range {local, regional, city-wide}, the duration {short-term, medium-term, long-term} and the disposal emergency degree {low, medium, high, very high}; the dynamic rule layer dynamically adjusts the weight coefficient of the reasoning rule according to the historical traffic event statistical data and the real-time road network state.

[0078] Further, the standard event mode is extracted from the knowledge graph model, the feature vector is encoded, and a deep graph convolution reasoning network dedicated to event reasoning is established.

[0079] It should be noted that the standard event mode includes road traffic events, public traffic events, special vehicle events and sudden traffic events; the deep graph convolution reasoning network increases a third layer reasoning convolution layer and a fourth layer decision convolution layer based on the second layer graph convolution output of the graph neural network.

[0080] In an optional implementation, the third layer reasoning convolution layer adopts a gating mechanism to selectively propagate the semantic embedding coding result, and the specific formula is as follows:

[0081]

[0082] wherein, is the output feature vector of node i in the third layer, is an element-level multiplication operation, and Gi is a gating value vector for node i, is the output feature of neighbor node j at the second layer, W (3) is a third layer weight matrix, β ij is a gating weight coefficient, is the output feature of neighbor node j at the third layer, b g is a gating bias vector.

[0083] It should be noted that the third layer reasoning convolution output value domain is in the interval [0, 1], wherein close to 1 indicates that the node has high reasoning value after being filtered by the gating mechanism, close to 0 indicates that the reasoning information importance of the node is low, and the intermediate value indicates that the node has certain reasoning ability.

[0084] In an optional implementation, the fourth layer decision convolution layer aggregates global context information through a self-attention mechanism to generate an event reasoning feature representation, and the specific formula is as follows:

[0085]

[0086]

[0087] wherein, is an output feature vector of node i at the fourth layer, ReLU(*) is a rectified linear unit activation function, V is a set of all nodes in the graph, A ij is a self-attention weight of node i to node j, W (4) is a fourth layer weight matrix, d is a feature dimension, is a transposition operation on the output feature vector of node j at the third layer .

[0088] It should be noted that the fourth layer decision convolution output value domain is in the interval [0, +∞), wherein +∞ indicates that the node has a very high contribution to the traffic event decision, close to 0 indicates that the node has a smaller impact on the event reasoning decision, and the intermediate value indicates that the node plays an auxiliary role in the decision process.

[0089] Further, the attribute information of the standard event mode is converted into a 128-dimensional feature vector with the same dimension as the semantic embedding encoding result by using the graph neural network.

[0090] Specifically, the fusion feature vector is input into the deep graph convolution reasoning network for forward reasoning calculation, and an event reasoning feature representation is generated through four convolution operations; a sliding window detection algorithm is used to identify the traffic event mode in the event reasoning feature representation to generate a candidate traffic event set; the candidate traffic events are preliminarily screened, and a multi-dimensional similarity calculation method is used to evaluate the matching strength of the candidate traffic events and the standard event mode to generate a matching strength score.

[0091] In optional embodiments, a preliminary screening is performed on the candidate traffic incident, a feature activation threshold is set, when the matching intensity score is greater than or equal to the activation threshold, the corresponding candidate traffic incident is marked as a high-confidence incident, and a real-time response mechanism is automatically triggered, including pushing an alarm to a traffic management platform, associating an incident instance node in a knowledge graph to update dynamic attributes (such as severity level and impact range), and activating a multi-incident joint disposal strategy in a composite rule layer; when the matching intensity score is less than the activation threshold, the candidate incident is determined to be a low-confidence or noise interference, the candidate incident is filtered and stored in a historical database for offline analysis, a low-confidence log is generated and manual review is prompted, for example, a road section anomaly is detected but does not reach an incident threshold.

[0092] It should be noted that the activation threshold is based on statistical analysis of historical traffic incident data and field expert experience (such as incident severity standards in traffic management specifications), and a critical value is selected by ROC curve or precision-recall curve to achieve the optimal balance between false positive rate and false negative rate.

[0093] Preferably, the matching intensity score includes semantic similarity and structural similarity; the semantic similarity calculates the similarity degree of the feature vector of the candidate traffic incident and the feature vector of the standard incident mode through cosine similarity; the structural similarity calculates the topological structure matching degree of the event node based on the adjacency relationship matrix in the knowledge graph model.

[0094] It should be noted that the related formula of the matching intensity score is as follows:

[0095] Score = w1 · S se + w2 · S st

[0096]

[0097] wherein Score is the matching intensity score, w1 is the semantic similarity weight, w2 is the structural similarity weight, w3 is the spatio-temporal similarity weight, S se is the semantic similarity score, V c is the feature vector of the candidate traffic incident, V s is the feature vector of the standard incident mode, S st is the structural similarity score, M is the upper bound of the range of the degree of all nodes in the graph, D c (a) is the number of nodes with degree a in the candidate incident graph, D s (a) is the number of nodes with degree a in the standard incident graph, E c is the number of edges of the candidate incident graph, E s is the number of edges of the standard incident graph.

[0098] Further, based on the causality rules of the event reasoning rule base, the causality reasoning verification is performed on the candidate traffic events with high matching strength score; the spatio-temporal constraint verification is performed on the candidate traffic events verified by the causality reasoning verification, and the traffic event prediction result is generated.

[0099] In an optional embodiment, a forward chain reasoning algorithm is adopted to trace back possible cause events and predict possible result events from the trigger condition of the candidate traffic event; the event causality chain path is constructed through the causality, time sequence, spatial relationship and influence relationship of the semantic relationship layer in the knowledge graph model; a causality confidence threshold is set to filter out the candidate traffic events with complete causality chain and higher confidence than the threshold; meanwhile, a reverse causality verification mechanism is established to trace back the cause events from the result events, and the rationality of the causality is verified bidirectionally.

[0100] In an optional embodiment, the spatio-temporal constraint verification includes time constraint check and spatial constraint check; the time constraint check verifies whether the occurrence time of the candidate traffic event conforms to the historical statistical rule and the traffic management time window limit; the spatial constraint check verifies whether the occurrence position of the candidate traffic event conforms to the road network topology and the geographical space constraint condition; an event conflict detection algorithm is adopted to identify mutually exclusive events in the same spatio-temporal region, and the conflicting candidate traffic events are eliminated through priority sorting and confidence comparison; an event consistency scoring mechanism is established to generate the traffic event prediction result by comprehensively considering the time consistency, spatial consistency and logical consistency.

[0101] It should be noted that the comprehensive confidence is marked for the traffic event prediction result, and the detailed path information of the event reasoning is output, including the triggered reasoning rule, the participated knowledge graph entity node and the key decision points in the reasoning process.

[0102] S4: Based on the matching strength score, the spatio-temporal constraint condition and the historical event mode, the traffic event identification result, the confidence evaluation and the disposal suggestion are output.

[0103] Specifically, the matching strength score is processed by multi-level threshold grading, and a four-level identification confidence classification system is established according to the numerical range of the matching strength score; the spatio-temporal attributes of the high-confidence candidate event are extracted from the event instance layer of the knowledge graph model, and the spatio-temporal overlap analysis is performed with the real-time road network topology data.

[0104] In an optional embodiment, when the matching strength score is greater than or equal to 0.85, the high-confidence candidate event is marked; when 0.55 is less than the matching strength score and the matching strength score is less than 0.55, the low-confidence candidate event is marked.

[0105] It should be noted that the space-time attribute includes event occurrence timestamp and geographic coordinate range; the space-time overlap analysis adopts a GeoHash encoding matching algorithm to verify whether the high-confidence candidate event set meets the preset space-time constraint condition; the verification result is fed back to the knowledge graph model and the event reasoning rule library to update the entity node attribute and the rule weight coefficient.

[0106] Further, for the high-confidence candidate event verified by the space-time constraint condition, a historical event pattern matching module is called from the dynamic rule layer of the event reasoning rule library to calculate a historical matching coefficient; a weighted fusion algorithm is used to comprehensively evaluate the matching strength score and the historical matching coefficient to generate a confidence score; according to the numerical interval of the confidence score, the high-confidence candidate event is classified into a confirmed event, a to-be-checked event and a false positive event.

[0107] It should be noted that the historical event pattern matching module is based on a sliding time window to count the frequency and distribution law of the high-confidence candidate event occurring in a historical period T; the weight parameter of the weighted fusion algorithm is dynamically adjusted according to the domain knowledge defined in the event concept layer of the knowledge graph model;

[0108] In an optional embodiment, the event classification method comprises: if the confidence score is greater than or equal to a first preset confidence threshold, the high-confidence candidate event is classified as a confirmed event; if the confidence score is less than the first preset confidence threshold and greater than or equal to a second preset confidence threshold, the high-confidence candidate event is classified as a to-be-checked event; and if the confidence score is less than the second preset confidence threshold, the high-confidence candidate event is classified as a false positive event.

[0109] It should be noted that the first preset confidence threshold is greater than the second preset confidence threshold; the first preset confidence threshold is determined based on the minimum confidence value of a typical confirmed event in historical traffic event data; and the second preset confidence threshold is determined based on the maximum confidence value of a typical false positive event in historical traffic event data.

[0110] Further, for the confirmed event, an associated disposal scheme is extracted from the semantic relationship layer of the knowledge graph model to generate a multi-level disposal suggestion in combination with real-time road network state data; the traffic event recognition result, the confidence score and the disposal suggestion are packaged as a structured output message to be pushed to a traffic management platform through an event instance layer interface of the knowledge graph model.

[0111] Preferably, the multi-level disposal suggestion adopts a case-based reasoning method: according to the identified event type and severity, a similar case is retrieved from a historical disposal case library to generate a disposal suggestion scheme in combination with the current traffic condition and available resources, including a resource scheduling strategy, a diversion route planning and an emergency response timing arrangement.

[0112] In summary, the application realizes unified semantic representation of heterogeneous traffic information and effectively breaks data silos by obtaining multi-modal traffic data of urban traffic and constructing a knowledge graph model covering an event concept layer, an event instance layer and a semantic relationship layer; through specific preprocessing and feature extraction of different modal data, visual, numerical and text features are mapped to knowledge graph entity nodes and fused, semantic embedding representation with global perception ability is generated by using graph neural network and multi-layer attention mechanism, and deep fusion of cross-modal features and improvement of context correlation are realized; an event reasoning rule library containing basic rules, composite rules and dynamic rules is constructed based on semantic embedding encoding results, and the matching strength score of candidate events and standard modes is obtained by combining deep graph convolution reasoning network, multi-dimensional similarity calculation, causal relationship verification and spatio-temporal constraint analysis, so that events with high correlation and high credibility are effectively screened; the confidence score is generated by a weighted fusion algorithm in combination with the matching strength score, spatio-temporal constraint conditions and historical event modes, and events are classified according to the score results, multi-level treatment suggestions are generated by extracting treatment schemes and real-time road network states, and the accuracy and practicality of urban traffic event identification are significantly improved.

[0113] The embodiment also provides a computer device suitable for the case of the knowledge graph-based urban traffic event semantic identification method, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the knowledge graph-based urban traffic event semantic identification method proposed in the above embodiment.

[0114] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device. In addition, the input device can be an external keyboard, touchpad or mouse, etc.

[0115] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A knowledge graph-based urban traffic event semantic identification method, characterized in that: The application relates to a method for identifying a traffic event based on a knowledge graph model. The method comprises the following steps: acquiring multi-modal traffic data of urban traffic, and constructing a knowledge graph model; The method comprises the following steps: preprocessing and extracting features of the multi-modal traffic data, mapping the extracted multi-modal features to entity nodes of the knowledge graph model, generating a fusion feature vector, and performing semantic embedding coding on the fusion feature vector through a graph neural network; Based on the semantic embedding coding result, an event reasoning rule library is constructed, and a multi-layer reasoning calculation is performed on the fusion feature vector through a graph convolutional neural network to obtain a matching strength score of a candidate traffic event and a standard event mode in the knowledge graph; Based on the matching strength score, an event spatio-temporal constraint condition and a historical event mode are combined to output a traffic event identification result, a confidence evaluation and a disposal suggestion. 2.The knowledge graph-based urban traffic event semantic identification method of claim 1, wherein: Based on the matching strength score, an event spatio-temporal constraint condition and a historical event mode are combined to output a traffic event identification result, a confidence evaluation and a disposal suggestion, which comprises the following steps: The matching strength score is subjected to multi-level threshold grading processing, a four-level identification confidence classification system is established according to the numerical range of the matching strength score, and a high-confidence candidate event is obtained; The spatio-temporal attributes of the high-confidence candidate event are extracted from the event instance layer of the knowledge graph model, and spatio-temporal overlap analysis is performed on the real-time road network topology data; For the high-confidence candidate event verified through the spatio-temporal constraint condition, a historical event mode matching module is called from the dynamic rule layer of the event reasoning rule library, and a historical matching coefficient is calculated; The matching strength score and the historical matching coefficient are comprehensively evaluated through a weighted fusion algorithm to generate a confidence score; According to the numerical interval of the confidence score, the high-confidence candidate event is classified into a confirmed event, a to-be-checked event and a false positive event; For the confirmed event, an associated disposal scheme is extracted from the semantic relationship layer of the knowledge graph model, and a multi-level disposal suggestion is generated in combination with real-time road network state data; The traffic event identification result, the confidence score and the disposal suggestion are encapsulated as a structured output message, and are pushed to a traffic management platform through an event instance layer interface of the knowledge graph model. 3.The knowledge graph-based urban traffic event semantic identification method of claim 2, wherein: The method for obtaining the matching strength score comprises the following steps: A standard event mode is extracted from the knowledge graph model, feature vector coding processing is performed, and a deep graph convolution reasoning network dedicated to event reasoning is established; The attribute information of the standard event mode is converted into a 128-dimensional feature vector with the same dimension as the semantic embedding coding result through a graph neural network; The fusion feature vector is input into the deep graph convolution reasoning network for forward reasoning calculation, and an event reasoning feature representation is generated through four-layer convolution operation; A sliding window detection algorithm is used to identify a traffic event mode in the event reasoning feature representation to generate a candidate traffic event set; The candidate traffic events are subjected to preliminary screening, and a multi-dimensional similarity calculation method is used to evaluate the matching strength of the candidate traffic events and the standard event mode to generate a matching strength score; The candidate traffic events are subjected to causal relationship reasoning verification based on the causal relationship rules of the event reasoning rule library; The candidate traffic events verified through the causal relationship reasoning verification are subjected to spatio-temporal constraint verification to generate a traffic event prediction result. 4.The knowledge graph-based urban traffic event semantic identification method of claim 3, wherein: The matching strength score includes semantic similarity and structural similarity; the semantic similarity calculates the similarity degree of the feature vector of the candidate traffic event and the feature vector of the standard event mode through cosine similarity; and the structural similarity calculates the topological structure matching degree of the event node based on the adjacency relationship matrix in the knowledge graph model. 5.The knowledge graph-based urban traffic event semantic identification method of claim 3, wherein: The event reasoning rule library includes a basic rule layer, a composite rule layer and a dynamic rule layer; the basic rule layer stores the trigger condition rules of a single traffic event; the composite rule layer defines the composite reasoning rules of joint triggering of multiple events, and establishes a rule combination matrix through the event severity level, influence range, duration and disposal urgency; and the dynamic rule layer dynamically adjusts the weight coefficients of the reasoning rules according to historical traffic event statistical data and real-time road network state. 6.The knowledge graph-based urban traffic event semantic identification method of claim 4, wherein: The knowledge graph model comprises a traffic event concept layer CTG={C1, C2,..., C n}, an event instance layer ITG={I1, I2,..., I m}, and a semantic relationship layer RTG={R1, R2,..., R k}, and the three layers are associated by a semantic mapping function CTGxITG→RTG.

7. The knowledge graph-based urban traffic event semantic identification method of claim 6, wherein: The fusion feature vector is obtained by fusing the visual feature vector, the numerical feature vector and the text feature vector mapped to the same entity node through a feature vector splicing operation, and includes: According to the type attribute of the entity node in the knowledge graph model, the visual feature vector, the numerical feature vector and the text feature vector are respectively mapped to the vehicle entity node, the road entity node and the event entity node, and the corresponding relationship between the feature vector and the entity node is established; A graph structure of the adjacency relationship of the entity node in the knowledge graph model is constructed, the fusion feature vector is taken as the initial feature of the graph node, input to the graph neural network for forward propagation, and through two-layer graph convolution operation, a node embedding representation with global perception ability is formed; A multi-layer graph attention network is used to perform semantic embedding coding on the node embedding representation, the contribution degrees of different neighbor nodes are adjusted through an attention weight learning mechanism, and a semantic embedding coding result is generated. 8.The knowledge graph-based urban traffic event semantic identification method of claim 7, wherein: The extraction method of the visual feature vector, the numerical feature vector and the text feature vector is, Inter-frame difference and background modeling preprocessing are performed on the monitoring video stream, the moving target region is extracted and scale normalization is performed; A convolutional neural network is used to extract features from the normalized target region, obtain visual features, and perform dimension reduction compression through principal component analysis to generate a visual feature vector; Time window segmentation and outlier detection preprocessing are performed on the sensor structured data, a sliding window statistical method is used to extract traffic flow parameters to generate numerical features, and the numerical features are subjected to Z-score standardization processing to generate a standardized numerical feature vector; Text structured data is subjected to word segmentation and part-of-speech tagging preprocessing, a bidirectional long short-term memory network based on an attention mechanism is used to extract semantic information of traffic event description to generate text features capturing event type and severity; A word embedding model is used to convert the text features into vector representation to generate a text feature vector. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the knowledge graph-based urban traffic event semantic recognition method of any one of claims 1-8.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the knowledge graph-based urban traffic event semantic recognition method of any one of claims 1-8.

Images