CN117894168A - A traffic flow anomaly detection method based on graph contrastive learning network - Google Patents

Abstract

The invention discloses a traffic flow anomaly detection method based on a graph comparison learning network, and belongs to the technical field of traffic flow monitoring. The invention takes the occupancy rate, speed and flow information of the urban road network structure and traffic flow as road section characteristics, designs an area clustering algorithm to form urban area representation, and constructs an urban road section map structure and an urban area map structure; based on two urban traffic map structures, a space-time encoder-decoder network with shared parameters is designed, the network is composed of a space-time attention module consisting of a map coiler layer and a time self-attention layer and a designed flow change extraction and comparison learning layer, and traffic jams and abnormal road sections are efficiently detected under different time and dynamic environments. The invention can efficiently model traffic conditions in different time and dynamic environments, and is helpful for better understanding the space-time characteristics of urban traffic; the real-time and dynamic changes of urban traffic can be processed, and the effectiveness of traffic management is improved.

Description

A traffic flow anomaly detection method based on graph contrastive learning network

Technical Field

The present invention belongs to the technical field of traffic flow monitoring, and in particular relates to a traffic flow anomaly detection method based on a graph contrast learning network.

Background technique

As the urban population continues to grow, urban traffic congestion has become a serious challenge, which not only affects the daily lives of urban residents, but also has a profound impact on the economic, environmental and social sustainability of cities. However, current related work still has shortcomings in the process of detecting congestion.

First, the urban traffic network is a dynamic and complex system. Traffic congestion changes at different times and locations. Traditional congestion detection methods mainly focus on static road segment information, such as road topology and historical traffic data, and it is difficult to obtain dynamic spatiotemporal traffic characteristics and associations. Therefore, more accurate methods are needed to capture spatiotemporal changes and anomalies.

Second, defining traffic congestion anomalies is complex because anomalies can take many forms, such as temporary road closures, large-scale traffic accidents, or traffic congestion caused by meteorological factors. Clear anomaly definitions and calibration methods are needed to facilitate detection.

Summary of the invention

In order to solve the problems in the prior art, the present invention proposes a traffic flow anomaly detection method based on a graph contrast learning network.

The technical solution of the present invention is as follows:

The present invention provides a traffic flow anomaly detection method based on a graph contrast learning network, which comprises the following steps:

S1: The urban traffic network is regarded as a graph structure. Each node in the graph structure represents a road section in the city. The characteristics of the road section are embedded into the graph structure. At the same time, an adjacency matrix is constructed to indicate whether there is a connection between the road sections. Finally, the urban road section graph structure is obtained.

S2: Set an allocation vector for each node of the urban road segment graph structure obtained in S1, and the allocation vector is used to indicate which region the road segment belongs to; all allocation vectors constitute an allocation matrix Z; use a clustering method to optimize the allocation matrix so that the road segments in each region have highly similar features and the feature differences between different regions are maximized, thereby obtaining an urban region graph structure;

S3: construct a spatiotemporal encoder-decoder network, which includes two spatiotemporal encoders with the same structure and one spatiotemporal decoder. The first spatiotemporal encoder and the second spatiotemporal encoder take the information in the urban road segment graph structure and the urban area graph structure as input respectively, and extract the features of the urban road segment graph structure and the urban area graph structure respectively; the output results of the two spatiotemporal encoders are spliced and input into the spatiotemporal decoder, and the spatiotemporal decoder captures the spatiotemporal relationship in the input features and generates the category probability of the traffic flow at the current moment to distinguish between normal and abnormal traffic flow data;

S4: Use historical traffic flow data to train the spatiotemporal encoder-decoder network; use the trained spatiotemporal encoder and spatiotemporal decoder to detect traffic flow anomalies.

According to a preferred solution of the present invention, S1 specifically comprises:

The urban traffic network is regarded as a graph structure G = (V, E), where each node represents a road section in the city; V represents a node set and E represents an edge set; for each node _vi , in each time segment t, the following road section characteristics are defined: traffic speed, road section occupancy rate, and road section flow rate. The road section characteristics are embedded in the graph structure. The edges in the graph structure are the traffic topology structure. At the same time, an adjacency matrix A is constructed, where A[i][j] indicates whether there is a connection between road section i and road section j, A[i][j] = 1 indicates that there is a connection between road section i and road section j, and A[i][j] = 0 indicates that there is no connection; finally, the urban road section graph structure is represented as _Gs = ( _Vs , _Es , A).

According to a preferred embodiment of the present invention, S2 is specifically:

Generate an assignment vector z _i for each node in the urban road segment graph structure obtained by S1, indicating which region the road segment belongs to; these vectors constitute the assignment matrix Z = (z ₁ , ..., z _n ) ∈ ^RN*M , where N represents the number of road segments and M represents the number of regions; use the K-means clustering method to optimize the assignment matrix so that the road segments in each region have highly similar features and the feature differences between different regions are maximized. The optimization objective function is defined as Wherein, z _m,n represents the probability of road segment n being assigned to region m; an adjacency matrix A _r is constructed according to the distances between regions, and finally, the urban region graph structure is obtained, which is represented as _Gr = (V _r , _Er , A _r ).

According to a preferred embodiment of the present invention, in S3, the two spatiotemporal encoders each include a first graph convolution layer, a temporal self-attention layer, and a second graph convolution layer;

The first graph convolution layer of the first spatiotemporal encoder performs graph convolution operations on the node-level spatial features of urban road sections; its temporal self-attention layer integrates the node-level temporal features to obtain the temporal dependency of historical traffic flow information; its second graph convolution layer further improves the abstract representation of features;

The first graph convolution layer of the second spatiotemporal encoder performs graph convolution operations on the regional-level spatial features of urban road sections; its temporal self-attention layer integrates the regional-level temporal features to obtain the temporal dependency of historical traffic flow information; and its second graph convolution layer further improves the abstract representation of features.

According to a preferred embodiment of the present invention, the spatiotemporal encoder-decoder network further comprises a contrastive learning layer, which obtains the outputs of the two spatiotemporal encoders and uses a self-supervised learning method to distinguish positive samples from negative samples to improve the detection accuracy of traffic flow anomalies;

The contrastive loss function of the contrastive learning layer is defined as follows:

Among them, _Ui is the central node feature in each cluster at the regional level after using the K-means clustering algorithm, _Pi is the in-cluster sample, _Ni is the out-cluster sample, and sim is the similarity measure between embedded vectors.

According to a preferred solution of the present invention, in S4, the data used to train the spatiotemporal encoder-decoder network are: traffic network G and historical data X, where X∈R ^T*N*V represents the traffic dynamics of all nodes V in T time slices, and the training goal is to enable the spatiotemporal encoder-decoder network to identify which nodes are abnormal at time T+1;

The method of using the trained spatiotemporal encoder and spatiotemporal decoder to detect traffic flow anomalies is as follows: inputting information on the urban road segment graph structure and the urban area graph structure at the current moment, and using the trained spatiotemporal encoder and spatiotemporal decoder to predict the traffic flow anomaly at the next moment.

The beneficial effects of the present invention include:

The present invention designs a two-level graph representation, including a regional level and a node level, for efficiently acquiring global and regional traffic rules.

The present invention takes the urban road network structure and the occupancy rate, speed and flow information of traffic flow as the road section characteristics, designs a regional clustering algorithm to form the urban area representation, and constructs the urban road section graph structure and the urban area graph structure; based on the two urban traffic graph structures, a parameter-sharing spatiotemporal encoder-decoder network is designed. The network consists of a spatiotemporal attention module composed of a graph convolution layer and a temporal self-attention layer, and a designed flow change extraction contrast learning layer, which can efficiently detect traffic congestion and abnormal sections under different time and dynamic environments.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for detecting anomalies in traffic flow based on a graph contrast learning network according to the present invention.

Detailed ways

The present invention is further described and illustrated below in conjunction with specific embodiments. The embodiments are merely exemplary of the present disclosure and do not define the scope of limitation. The technical features of each embodiment of the present invention may be combined accordingly without conflicting with each other.

As shown in FIG1 , the traffic flow anomaly detection method based on graph contrast learning network of the present invention comprises the following steps:

S1: The urban traffic network is regarded as a graph structure. Each node in the graph structure represents a road section in the city. The characteristics of the road section are embedded into the graph structure. At the same time, an adjacency matrix is constructed to indicate whether there is a connection between the road sections. Finally, the urban road section graph structure is obtained.

S2: Set an allocation vector for each node of the urban road segment graph structure obtained in S1, and the allocation vector is used to indicate which region the road segment belongs to; all allocation vectors constitute an allocation matrix Z; use a clustering method to optimize the allocation matrix so that the road segments in each region have highly similar features and the feature differences between different regions are maximized, thereby obtaining an urban region graph structure;

S3: construct a spatiotemporal encoder-decoder network, which includes two spatiotemporal encoders with the same structure and one spatiotemporal decoder. The two spatiotemporal encoders take the information in the urban road segment graph structure and the urban area graph structure as input respectively, and extract the features of the urban road segment graph structure and the urban area graph structure respectively; the output results of the two spatiotemporal encoders are spliced and input into the spatiotemporal decoder. The spatiotemporal decoder captures the spatiotemporal relationship in the input features and generates the category probability of the traffic flow at the current moment to distinguish between normal and abnormal traffic flow data;

S4: Use historical traffic flow data to train the spatiotemporal encoder-decoder network; use the trained spatiotemporal encoder and spatiotemporal decoder to detect traffic flow anomalies.

Specifically, in the present invention, a two-level graph structure representation method is adopted to process the regional level and node level features of urban road sections respectively. This two-level representation method not only helps to extract high-dimensional traffic data features, but also can fully capture the laws of urban traffic at different spatial scales.

In a specific embodiment of the present invention, S1 is: the present invention regards the urban traffic network as a graph structure G = (V, E), and each node represents a road section in the city. Wherein, V represents a node set, and E represents a set of edges. For each node _vi , in each time segment t, the following features are defined: traffic speed (sp), road section occupancy (op), road section flow (fl), and the features of the road section are embedded in the graph structure. The edges in the graph are traffic topological structures, and an adjacency matrix (AdjacencyMatrix) is constructed at the same time, which is represented as an adjacency matrix A, where A[i][j] indicates whether there is a connection between road section i and road section j. This is usually a binary matrix, where A[i][j] = 1 indicates that there is a connection between road section i and road section j, and A[i][j] = 0 indicates that there is no connection. Finally, the urban road section graph structure is represented as _Gs = ( _Vs , _Es , A).

In a specific embodiment of the present invention, S2 is: Since the road sections in the city are different, such as main roads and branch roads, combined with the different functions of different sections in the city, functional areas will be formed, and the traffic flow and speed are greatly affected by the road function and the city function. Therefore, the present invention generates an allocation vector z _i for each road section (node), indicating which area the road section belongs to. These vectors constitute the allocation matrix Z = (z ₁ , ..., z _n ) ∈ ^RN*M , where N represents the number of road sections and M represents the number of areas. In order to better represent the urban area and determine the number of areas M, the present invention uses the optimal cluster allocation matrix Z to ensure that the road sections in each area have the highest similarity within the class and the maximum difference between different areas as the principle, and generates the optimal cluster allocation matrix. The optimization objective function can be defined as a combination of maximizing the similarity within the class and minimizing the difference between classes. The present invention is achieved by iteratively updating the elements of the cluster allocation matrix Z. Among them, the objective function is defined as:

Where zn, k represents the elements of the cluster assignment matrix Z. The adjacency matrix A _r is constructed according to the distances between regions. Finally, the urban region graph structure is represented as G _r =(V _r , E _r , A _r ).

In cities, there are many traffic flow patterns, such as morning and evening rush hours, weekend travel, holiday changes, etc. The feature information needs to be fused in time and space to efficiently detect traffic congestion and abnormal sections in different time and dynamic environments. In order to achieve time and space fusion and more accurate anomaly detection, the present invention designs a time and space encoder-decoder network suitable for detecting traffic flow anomalies.

The spatiotemporal encoder of the present invention is composed of a sandwich structure of two layers of graph convolution (GCN) and a temporal self-attention layer, and the encoder structure for the urban area graph structure and the urban road section graph structure is the same. First, the first step of the encoder is to perform graph convolution operations on the spatial features of the area level and node level of the urban road section respectively.

This is achieved by applying GCN.

Among them, ^Hl is the feature representation after the first graph convolutional layer operation, σ(·) is the activation function, A is the adjacency matrix of the corresponding graph structure of the input, D is the degree matrix, ^W1 is the weight matrix of the first graph convolutional layer, the subscript i represents the i-th time step, and X is the node feature information of the corresponding graph structure of the input.

Next is a temporal self-attention layer: The key component of the spatiotemporal encoder is the temporal self-attention layer. This layer is responsible for integrating the temporal features at the regional and node levels respectively, and obtaining the temporal dependency of historical traffic flow information. By introducing the temporal attention mechanism, this module is able to capture the temporal correlation of different road sections after obtaining spatial information, considering the changes of urban traffic in different time and dynamic environments. The temporal self-attention layer is an important component for traffic flow anomaly detection to capture the dependencies between different time steps. Here, this embodiment will describe the temporal self-attention layer in detail, with special attention to the temporal mutation part. The representation of the temporal self-attention layer is as follows:

The calculation formulas for query (Q), key (K), and value (V) are as follows:

Q＝H ^l W _Q ，K＝H ^l W _k ，V＝H ^l W _v

H ^l represents the output of the first graph convolution layer and the input of the temporal self-attention layer. W _Q , W _K and W _V are the linear transformation weight matrices of query, key and value, respectively.

Use the attention weight Attention _ij (Q, K, V) to perform weighted summation on the value vectors of neighboring nodes to obtain the aggregated features

Combined aggregation features And the characteristics of the node itself, update the characteristics of the node through a nonlinear transformation:

Among them, W ^l is the learned weight matrix; is the updated feature, i.e., the output of the temporal self-attention layer.

After the temporal attention layer, the present invention performs a second layer of graph convolution operation to further improve the abstract representation of features.

The second graph convolutional layer operates as follows:

in, represents the output of the temporal self-attention layer multiplied by the adjacency matrix, and W ² is the weight matrix of the second graph convolutional layer. The output of this layer/> Includes more advanced regional level features.

Furthermore, the present invention also obtains additional attention to the mutations in time through the mutation mask of the temporal self-attention mechanism. Afterwards, the present invention designs a contrastive learning layer to improve the detection accuracy of traffic flow anomalies. Contrastive learning is a powerful self-supervised learning method. The main idea is to learn how to distinguish positive samples (similar data points) from negative samples (dissimilar data points). In traffic flow anomaly detection, the anomaly should be different from the surrounding environment from the perspective of node attributes. From a spatial perspective, the data points with high similarity between the node and the surrounding are positive samples. Data points far away from the time step, or data points from different areas are selected as negative samples for learning, because these data points are more likely to be dissimilar. The goal of contrastive learning is to allow the model to distinguish positive samples from negative samples. The present invention is achieved by constructing a contrast loss function, which is defined as follows:

Among them, _Ui is the central node feature in each cluster at the regional level after using the K-means clustering algorithm, _Pi is the in-cluster sample, _Nj is the out-cluster sample, and sim is the similarity measure between embedded vectors.

The goal of the present invention is to restore or generate the original traffic flow data. In order to achieve this goal, the present invention splices the high-dimensional representation of the two-level urban traffic graph network, and then uses a spatiotemporal decoder structure of a layer of GCN (graph convolutional network) and a layer of MLP (multi-layer perceptron) to reconstruct the original traffic flow in a regular traffic environment. GCN is used to capture spatiotemporal relationships, while MLP is used for further processing and generating outputs. In traffic flow anomaly detection, the goal of the present invention is to detect the category probability of traffic flow at the current moment to distinguish between normal and abnormal traffic flow data. Here, the present invention adopts cross entropy loss, in which the output of the model is a probability distribution, and the traffic flow anomaly detection task is regarded as a binary classification problem (normal and abnormal), and the loss function is defined as follows:

Where N represents the number of urban road segments, _yi is the actual category label (0 and 1, normal and abnormal), is the probability of abnormal traffic flow on this road section predicted by the model.

The data used to train the spatiotemporal encoder-decoder network are: traffic network G and historical data X, where X∈R ^{T *N*V} represents the traffic dynamics of all nodes V in T time slices, and the training goal is to enable the spatiotemporal encoder-decoder network to identify which nodes are abnormal at time T+l. After the spatiotemporal encoder-decoder network is trained, the information of the urban road segment graph structure and the urban area graph structure at the current moment is input, and the trained spatiotemporal encoder and spatiotemporal decoder are used to predict the traffic flow abnormality at the next moment, thereby realizing traffic flow abnormality detection. By considering the urban road network structure and various traffic flow characteristics, the present invention can efficiently model traffic conditions in different time and dynamic environments, which helps to better understand the spatiotemporal characteristics of urban traffic; it can handle the real-time and dynamic changes of urban traffic, which helps to identify congestion and abnormal conditions in real time under different scenarios, and improves the effectiveness of traffic management.

The above-mentioned embodiments only express several implementation methods of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. For ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims (10)

1. The traffic flow anomaly detection method based on the graph comparison learning network is characterized by comprising the following steps of:

s1: taking the urban traffic network as a drawing structure, wherein each node in the drawing structure represents one road section of a city, embedding the characteristics of the road section into the drawing structure, and constructing an adjacent matrix to represent whether the road sections are connected or not, so as to finally obtain the drawing structure of the urban road section;

s2: setting an allocation vector for each node of the urban road section graph structure obtained in the step S1, wherein the allocation vector is used for indicating which region the road section belongs to; all allocation vectors form an allocation matrix Z; optimizing the distribution matrix by adopting a clustering method to ensure that road sections in each region have highly similar characteristics and the characteristic difference between different regions is maximized, so as to obtain a city region diagram structure;

s3: constructing a space-time encoder-decoder network, wherein the space-time encoder-decoder network comprises two space-time encoders with the same structure and a space-time decoder, and the first space-time encoder and the second space-time encoder respectively take information in a city road map structure and a city area map structure as input and respectively extract the characteristics of the city road map structure and the city area map structure; the output results of the two space-time encoders are spliced and then input into a space-time decoder, the space-time decoder captures the space-time relationship in the input characteristics, and the class probability of the traffic flow at the current moment is generated to distinguish normal traffic flow data from abnormal traffic flow data;

s4: training a space-time encoder-decoder network using historical traffic stream data; and detecting abnormal traffic flow by using the trained space-time encoder and the trained space-time decoder.

2. The traffic flow anomaly detection method based on graph comparison learning network as claimed in claim 1, wherein the step S1 specifically comprises:

taking the urban traffic network as a mapping structure G= (V, E), wherein each node represents one road section of a city; wherein V represents a node set and E represents a set of edges; for each node v _i In each time segment t, the following road segment characteristics are defined: the traffic flow speed, the road section occupancy and the road section flow are embedded into a graph structure, the edges in the graph structure are traffic topological structures, and an adjacency matrix A is constructed at the same time, wherein A [ i ]][j]Indicating whether there is a connection between road section i and road section j, ai][j]=1 indicates that there is a connection between the road section i and the road section j, a [ i ]][j]=0 indicates no connection; finally, the city road map structure is denoted as G _s ＝(V _s ,E _s ,A)。

3. The traffic flow anomaly detection method based on graph comparison learning network as claimed in claim 1, wherein the step S2 is specifically:

generating an allocation vector z for each node in the city road graph structure obtained in S1 _i Indicating which region the road section belongs to; these vectors form an allocation matrix z= (Z) ₁ ,...,z _n )∈R ^N*M Where N represents the number of road segments and M represents the number of areas; optimizing distribution by adopting K-means clustering methodThe matrix maximizes the feature differences between the different regions with highly similar features for road segments within each region, and the optimized objective function is defined asWherein z is _m,n Representing the probability that road segment n is assigned to region m; constructing an adjacency matrix A based on the distance between regions _r Finally, the resulting urban area map structure is denoted as G _r ＝(V _r ,E _r ,A _r )。

4. The traffic flow anomaly detection method based on graph contrast learning network of claim 1, wherein in S3, both the space-time encoders comprise a first graph convolutional layer, a temporal self-attention layer, and a second graph convolutional layer;

a first graph convolution layer of a first space-time encoder performs graph convolution operation on the spatial characteristics of the node level of the urban road section; the time self-attention layer integrates the time characteristics of the node level, and the time dependence of the historical traffic flow information is obtained; the second graph convolution layer further improves abstract representation of the features;

the first graph convolution layer of the second space-time encoder performs graph convolution operation on the spatial characteristics of the regional level of the urban road section; the time self-attention layer integrates the time characteristics of the regional level, and the time dependence of the historical traffic flow information is obtained; its second graph convolution layer further improves the abstract representation of the feature.

5. The traffic flow anomaly detection method based on graph comparison learning network of claim 4, wherein in S3, the first graph convolution layer and the second graph convolution layer are implemented by GCN;

the first graph convolution layer operates to:

wherein H is ^l Is the firstThe feature representation after the operation of a graph convolution layer, σ (·) is the activation function, a is the adjacency matrix of the corresponding graph structure of the input, D is the degree matrix, W ¹ Is the weight matrix of the first graph convolution layer, the subscript i represents the ith time step, and X is the node characteristic information of the corresponding graph structure.

6. The traffic flow anomaly detection method based on graph comparison learning network of claim 4, wherein in S3, the temporal self-attention layer is represented as follows:

wherein d _k Is the dimension of the query and key; the calculation formula for query Q, key K and value V is as follows:

Q＝H ^l W _Q ,K＝H ^l W _k ,V＝H ^l W _v

H ^l representing the output of the first coiler layer, being the input of the temporal self-attention layer, W _Q 、W _K And W is _V A linear transformation weight matrix of query, key and value, respectively;

attention weighting _ij (Q, K, V) weighting and summing the value vectors of the neighbor nodes to obtain an aggregate characteristic

Incorporating polymeric featuresAnd the characteristics of the node itself, updating the characteristics of the node by a nonlinear transformation:

wherein W is ^l Is a learned weight matrix;is an updated feature, i.e., the output of the immediate self-attention layer.

7. The traffic flow anomaly detection method based on graph-contrast learning network of claim 4, wherein the second graph convolution layer operates to:

wherein,the output representing the temporal self-attention layer is multiplied by the adjacency matrix, W ² Is the weight matrix of the second graph convolutional layer.

8. The traffic flow anomaly detection method based on the graph contrast learning network of claim 4, wherein the space-time encoder-decoder network further comprises a contrast learning layer, the contrast learning layer obtains the outputs of the two space-time encoders, and a self-supervision learning method is adopted to distinguish positive samples from negative samples so as to improve the detection accuracy of traffic flow anomalies;

the contrast loss function of the contrast learning layer is defined as follows:

wherein U is _i To use the central node feature in each cluster in the region level after the K-means clustering algorithm, P _i Is an intra-cluster sample, N _j Is an out-of-cluster sample and sim is a similarity measure between embedded vectors.

9. The traffic flow anomaly detection method based on graph contrast learning network of claim 1, wherein the space-time decoder in S3 comprises a layer of graph convolution network and a layer of multi-layer perceptron; the graph convolution network is used for capturing the space-time relationship, and the multi-layer perceptron is used for further processing and generating output;

the output of the space-time decoder is a probability distribution, the traffic flow anomaly detection task is regarded as a classification problem that distinguishes between normal and anomaly, and the loss function of the space-time decoder is defined as follows:

wherein N represents the number of city segments, y _i Is the actual class label, i.e., 0 and 1, normal and abnormal,is the probability of the model predicted traffic flow anomaly for that road segment.

10. The traffic flow anomaly detection method based on graph-contrast learning network of claim 1, wherein in S4, the training space-time encoder-decoder network adopts the following data: traffic network G and history data X, where X ε R ^T*N*V Representing traffic dynamics of all nodes V over T time slices, the goal of training is to have the space-time encoder-decoder network identify which nodes are abnormal at time t+1;

the traffic flow anomaly detection by using the trained space-time encoder and the trained space-time decoder is as follows: and inputting the information of the urban road map structure and the urban area map structure at the current moment, and predicting the abnormal condition of the traffic flow at the next moment by using the trained space-time encoder and the trained space-time decoder.