An Abnormal Prediction Method, Device and Storage Medium for Intelligent Transportation System

Abstract

The present invention relates to an abnormal prediction method, device and storage medium for an intelligent transportation system based on a contrast graph neural network. The method includes: obtaining original data, where the original data includes a traffic accident data set and a road data set; extracting information from the road data set to obtain a traffic topology graph, spatial features, spatio-temporal features and additional features of the road, and combining with the traffic accident data set to obtain a graph data set; establishing an abnormal prediction model for the intelligent transportation system based on the contrast graph neural network, where the abnormal prediction model for the intelligent transportation system uses a contrastive learning loss function for model pre-training; performing data augmentation on the graph data set and pre-training the model based on the augmented graph data set; taking the graph data as input and predicting the abnormal situation of each node in the graph based on the abnormal prediction model for the intelligent transportation system. Compared with the prior art, the present invention has the advantages of solving the problem of abnormal data imbalance and high prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of traffic anomaly prediction, and in particular to an intelligent traffic system anomaly prediction method, device and storage medium based on a contrast graph neural network. Background Art

[0002] The Internet of Things-based intelligent traffic system (ITS) is expected to improve the safety and reliability of traffic networks without affecting service quality. Participants in the intelligent traffic system have a large number of sensors and cameras to collect a large amount of data. Artificial intelligence has been embedded in a large number of fields that were previously labor-intensive, such as traffic prediction, traffic light control, and traffic anomaly detection. Most neural networks require large-scale datasets and computing power, which are called data-hungry models, and the performance of the models highly depends on the datasets.

[0003] The success of artificial intelligence is partly attributed to its ability to extract features in Euclidean data. A graph is a common and natural data structure used to represent real-world data. More and more data is represented by graphs, such as in chemistry, e-commerce, and citation networks. Driven by the excellent feature capture ability of graph neural networks (GNNs), GNNs have been applied in fields that require the use of non-Euclidean data represented by graphs. Graph neural networks have the characteristic of satisfying additional information input. A graph contains the inherent relationships between nodes, edges, and global features. GNNs are a group of connectivity-driven models that address the need for deep geometric learning. Some studies use GNNs as encoders to transform the input graph, including the original node features. The encoder compresses the node representation by leveraging a large amount of potential node connectivity.

[0004] Abnormal situations in cities can pose a great danger to public safety. The relationships between some intelligent devices are dynamic. However, most of the data generated by participants is normal. The lack of abnormal events will reduce the performance of the model. Some studies improve performance by manually injecting abnormal events, but this ignores the potential representativeness of nodes in the graph neural network.

[0005] Therefore, it is urgent to solve the problem of the imbalance between normal data and abnormal data distribution in the intelligent traffic system, and then improve the intelligence and reliability of the traffic network without affecting normal traffic. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent traffic system anomaly prediction method, device and storage medium based on a contrast graph neural network, which uses a graph data augmentation method to expand abnormal data, improve the model prediction accuracy, and realize the prediction of abnormal states in the intelligent traffic system.

[0007] The object of the present invention can be achieved by the following technical solutions:

[0008] An abnormal prediction method for an intelligent transportation system based on a contrast graph neural network, comprising the following steps:

[0009] Step 1) Obtain original data, where the original data includes a traffic accident data set and a road data set;

[0010] Step 2) Extract information from the road data set to obtain the traffic topology graph, spatial features, spatio-temporal features, and additional features of the road, and combine with the traffic accident data set to obtain a graph data set;

[0011] Step 3) Establish an abnormal prediction model for the intelligent transportation system based on a contrast graph neural network, and the abnormal prediction model for the intelligent transportation system uses a contrast learning loss function for model pre-training;

[0012] Step 4) Perform data augmentation on the graph data set, and pre-train the model based on the augmented graph data set;

[0013] Step 5) Use the graph data as input, and predict the abnormal conditions of each node in the graph based on the abnormal prediction model of the intelligent transportation system.

[0014] The road data set includes road topology data, road average speed data, POI data, date, and weather data.

[0015] The said step 2) includes the following steps:

[0016] Step 21) Generate a traffic topology graph based on the road topology data;

[0017] Step 22) Based on the road topology data and POI data, use a graph-based spatial convolution layer to extract spatial features;

[0018] Step 23) Based on the road average speed data, use a graph-based spatio-temporal convolution layer to extract spatio-temporal features;

[0019] Step 24) Based on the date and weather data, use an additional feature extraction layer to extract additional features;

[0020] Step 25) Concatenate the spatial features, spatio-temporal features, and additional features to obtain a feature representation;

[0021] Step 26) Combine the traffic accident data set and the feature representation to obtain a graph data set.

[0022] The graph-based spatial convolution layer includes several graph convolutional neural network layers stacked and embedded with a residual structure, where the output of the l-th layer is:

[0023]

[0024] Among them, N(j) represents the combination of the neighbor nodes of point j, and α ij represents the product of the degrees of node i and node j, and W l , b l represents the parameter of the l-th layer of convolution, and σ is the ReLU activation function.

[0025] The Graph-based spatio-temporal convolution layer includes the superposition of several graph convolutional neural network layers and traditional convolutional network layers, and embeds a residual structure. The spatio-temporal features are obtained by extracting temporal features and fusing spatial features. Among them, the method for extracting temporal features is as follows:

[0026]

[0027] Among them, H K represents the extracted temporal features, represents the input data of the (k - 1)-th layer, σ is the ReLU activation function, F is the number of convolutional kernels of this layer, is the coefficient matrix and bias matrix of the (k - 1)-th layer of convolutional layer.

[0028] The additional feature extraction layer includes the superposition of several fully connected layers and activation functions. Each feature is encoded in One-Hot form to extract additional features.

[0029] The data augmentation for the graph dataset includes topological augmentation based on the traffic topology graph and feature augmentation based on node and edge masks. Among them,

[0030] The topological augmentation based on the traffic topology graph includes the following steps: Sampling K-hop in the nearest neighbor manner with the node where a traffic accident occurs as the center, and performing random walks on the current node on this basis, so as to obtain the

[0031] The feature augmentation based on node and edge masks includes the following steps: For the current randomly delete or add a pre-configured percentage of edges that are more than a preset distance from the central node, and partially perturb its features to obtain the feature augmentation.

[0032] The contrastive learning loss function is:

[0033]

[0034] Among them, Y represents the label of the sample, taking 0 or 1; represents the Euclidean distance of the i-th sample, and m represents the hyperparameter of the hinge loss function.

[0035] An intelligent transportation system anomaly prediction device based on a contrast graph neural network, including a memory, a processor, and a program stored in the memory, characterized in that when the processor executes the program, the method described above is implemented.

[0036] A storage medium, on which a program is stored, and when the program is executed, the method described above is implemented.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] (1) The present invention focuses on the problem of the imbalance between normal data and abnormal data distribution in the intelligent transportation system, and uses the method of graph data augmentation to enable the model to learn the essential features of the data, improving the prediction accuracy of anomalies in the intelligent transportation system. Managers can better perceive the abnormal state of the current intelligent transportation system and make early deployments for the abnormal state.

[0039] (2) The present invention introduces a feature extraction algorithm of deep learning. This algorithm uses the method of graph neural network to learn the features of the data, and uses the contrast loss function to train the model, enabling the model to learn the essential features of the data, and thus improving the intelligence and reliability of the transportation network without affecting normal traffic.

[0040] (3) The data augmentation method proposed by the present invention increases the model's observation perspective of the data, enabling the model to learn the essential features of the data, and further improving the generalization ability and prediction accuracy of the model.

[0041] (4) The data augmentation method of the present invention can be deployed within any intelligent transportation anomaly prediction framework based on graph neural network, and has good compatibility.

[0042] (5) The present invention adds additional features to participate in model learning, considers more factors affecting traffic anomalies, and further improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a structural diagram of the intelligent transportation system anomaly prediction model based on the contrast graph neural network of the present invention;

[0044] Figure 2 It is a schematic diagram of pre-training the model based on the contrast learning loss function of the present invention;

[0045] Figure 3 It is a schematic diagram of the index results during the training process of the present invention;

[0046] Figure 4 It is a comparison graph of the prediction results between the present invention and other baseline models. Specific Embodiment

[0047] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and provides detailed implementation manners and specific operation processes. However, the protection scope of the present invention is not limited to the following embodiments.

[0048] This embodiment provides an abnormal prediction method for an intelligent transportation system based on a contrastive graph neural network, including the following steps:

[0049] Step 1) Obtain original data, where the original data includes a traffic accident dataset and a road dataset.

[0050] The road dataset includes road topology data, road average speed data, POI (Position of Interest) data, date, and weather data. POI data refers to the position of a specific point that a person may find useful or interesting, representing the density of people on this road, which is a fine-grained feature affecting traffic accidents and plays an important role in traffic accident prediction. Date and weather data include weather, temperature, humidity, pressure, and whether it is a holiday, etc.

[0051] Step 2) Extract the information of the road dataset to obtain the traffic topology graph Spatial features, spatio-temporal features, and additional features, and combine them with the traffic accident dataset to obtain a graph dataset.

[0052] Step 21) Generate a traffic topology graph based on the road topology data.

[0053] Step 22) Based on the road topology data and POI data, use a Graph-based spatial convolution layer to extract spatial features.

[0054] The spatial convolution layer is defined as convolving a graph using the adjacency matrix of the graph, and is used to extract spatial features, including road length (a category data in the road topology data) and POI distribution, which are fine-grained data directly affecting traffic anomalies. The main challenge of the spatial convolution layer is how to convolve the neighborhood of nodes and maintain local invariance. Given a traffic topology graph where V represents the set of road nodes, and ε represents the edges in the graph, represents The adjacency matrix. The spatial convolution layer of the present invention includes a stack of several graph convolutional neural network layers and embeds a residual structure to prevent gradient vanishing. The spatial convolution layer extracts the features of each node in the graph and propagates the node features to its neighbors according to the above definition. However, when the network is very deep, after the weight update of each mini-batch, the feature distribution may be distorted. To stabilize the learning process and significantly reduce the number of training times of the deep network, the present invention applies batch normalization in the spatial convolution layer.

[0055] The output of the l-th layer of the spatial convolution layer is:

[0056]

[0057] where, N(j) represents the combination of the neighbor nodes of point j, and α ij represents the product of the degrees of node i and node j, W l , b l represents the parameters of the l-th layer convolution, and σ is the ReLU activation function.

[0058] Step 23) Based on the road average speed data, use the Graph-based spatio-temporal convolution layer to extract spatio-temporal features.

[0059] Although RNN-based networks are widely used in the field of time series, the RNN used for traffic prediction cannot quickly adapt to dynamic changes. The spatial convolution layer applies graph convolution to capture spatial features from neighbors, and the spatio-temporal convolution layer is used to extract features related to time X t , representing the risk of traffic anomalies in the time domain.

[0060] The Graph-based spatio-temporal convolution layer includes a stack of several graph convolutional neural network layers and traditional convolutional network layers and embeds a residual structure, and obtains spatio-temporal features by extracting time features and fusing spatial features, wherein the method for extracting time features is:

[0061]

[0062] where, H K represents the extracted time features, represents the input data of the (k - 1)-th layer, σ is the ReLU activation function, F is the number of convolutional kernels of this layer, is the coefficient matrix and bias matrix of the (k - 1)-th layer convolutional layer.

[0063] After extracting the time features, the present invention jointly applies spatial graph convolution and standard convolution to merge the features of consecutive time. The consecutive time features refer to the average speed of each road, which reflects the risk of accidents in the time dimension, and the average speed of motor vehicles on this road is closely related to the probability of accidents.

[0064] Step 24) Based on the date and weather data, use an additional feature extraction layer to extract additional features.

[0065] The additional feature extraction layer consists of a stack of several fully connected layers (FC layers) and activation functions. Each feature is encoded in One-Hot form to obtain the additional features. The essence of the FC layer is to transform one feature space into another. Each FC layer slightly reduces the dimension.

[0066] The additional feature extraction layer is responsible for extracting data that is independent of space and time in the road traffic network, including meteorological factors and holiday features. Meteorological conditions are an important factor in traffic anomalies. Adverse weather affects driving safety. For example, haze reduces visibility, and rain reduces road friction. In addition, on holidays or weekends, the probability of traffic congestion or accidents is higher. Therefore, the additional feature extraction layer also extracts this feature.

[0067] Step 25) Concatenate the spatial features, spatio-temporal features, and additional features to obtain a feature representation.

[0068] Step 26) Combine the traffic accident dataset and the feature representation to obtain a graph dataset.

[0069] Step 3) Establish an intelligent transportation system anomaly prediction model based on a contrastive graph neural network, as Figure 1 shown. The intelligent transportation system anomaly prediction model uses a contrastive learning loss function for model pre-training. The last two layers of the model are classification layers.

[0070] Step 4) Perform data augmentation on the graph dataset and pre-train the model based on the augmented graph dataset.

[0071] Step 41) Perform data augmentation on the graph dataset, specifically including topological augmentation based on the traffic topology graph and feature augmentation based on node and edge masks.

[0072] Step 411) Topological augmentation based on the traffic topology graph : Sample K-hop in the nearest neighbor manner with the node where a traffic accident occurs as the center, and perform random walks on the current node on this basis to obtain the

[0073] Step 412) Feature augmentation based on node and edge masks: For the current randomly delete or add ρ% of the edges that are more than a preset distance from the central node, and partially perturb its features to obtain the feature augmentation of the current .

[0074] The graph augmentation strategy of the present invention satisfies: where It is an enhanced distribution conditional on the original Graph. Subgraphs are generated by sampling K-hop numbers centered on abnormal nodes, and the selection of hyperparameters has a great impact on the results.

[0075] In this embodiment, ρ is taken as 5. The edges that are more than a preset distance from the central node are defined as the edges that are more than 2 hops from the central node.

[0076] Step 42) Pre-train the model based on the augmented graph dataset, as Figure 2 shown.

[0077] The present invention adopts a parameter pre-training method based on graph contrast learning. Standard convolution in Euclidean space is not suitable for data based on the Graph structure. The main idea of spatial convolution is to rearrange the vertices into a specific grid form. Similar to CNN on images, the features of each node are extracted from the average of its neighbors by applying spatial graph convolution. However, the connections of the graph are represented by the adjacency matrix , which is different from the default assumption of their neighbors in images. According to the convolution theorem, the spectral convolution is defined as:

[0078]

[0079] where represents the Fourier transform.

[0080] The goal of the model of the present invention is to learn the feature representation for the input data, rather than the specific label. Specifically, taking x i ∈R d as the input and outputting the representation Z i of the road network, the network structure can be expressed as a function f θ :R d →R o . The present invention adopts a contrast learning loss function to pre-train the network parameters, and sends the augmented positive and negative samples into the network, while pulling the positive samples closer and making the distance between the positive samples and the negative samples farther, so as to obtain a better parameter initialization space for subsequent supervised learning training.

[0081] The contrast learning loss function is:

[0082]

[0083] where Y represents the label of the sample, taking 0 or 1; represents the Euclidean distance of the i-th sample, and m represents the hyperparameter of the hinge loss function.

[0084] Step 5) Take the graph data as the input and predict the abnormality of each node in the graph based on the intelligent transportation system abnormality prediction model.

[0085] When the above functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0086] This embodiment uses K-fold cross-validation to train the model, and the various indicators during the training process are as Figure 3 shown.

[0087] This experiment is compared with the current mainstream model frameworks, and the comparison result graph is as Figure 4 shown. The results show that the present invention has varying degrees of improvement in indicators such as accuracy, F1 score, recall rate, and AUC compared with the baseline model. With the support of the present invention, the abnormal prediction accuracy of the intelligent transportation system can reach 88%, effectively improving the representation of unbalanced data.

[0088] In summary, for the abnormal prediction method of the intelligent transportation system based on the contrast graph neural network proposed by the present invention, in the established experimental environment, through data augmentation of graph data and pre-training of model parameters, the abnormal state prediction at the road level with hourly granularity is achieved. For the trained model, the subgraph topology with features to be tested is fed into the network to obtain the abnormal prediction of nodes. From the experimental results, the expected effect is achieved.

[0089] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. An abnormal prediction method for an intelligent transportation system based on a contrastive graph neural network, characterized in that It includes the following steps: Step 1) Obtain the original data, where the original data includes a traffic accident dataset and a road dataset; Step 2) Extract the information of the road dataset to obtain the traffic topology map, spatial features, spatio-temporal features, and additional features of the road, and combine them with the traffic accident dataset to obtain a graph dataset; Step 3) Establish an intelligent transportation system anomaly prediction model based on a contrastive graph neural network. The intelligent transportation system anomaly prediction model uses a contrastive learning loss function for model pre-training; Step 4) Perform data augmentation on the graph dataset and pre-train the model based on the augmented graph dataset; Step 5) Use the graph data as input and predict the anomaly situation of each node in the graph based on the intelligent transportation system anomaly prediction model; Step 2) includes the following steps: Step 21) Generate a traffic topology map based on the road topology data; Step 22) Based on the road topology data and POI data, use a graph-based spatial convolutional layer to extract spatial features; Step 23) Based on the road average speed data, use a graph-based spatio-temporal convolutional layer to extract spatio-temporal features; Step 24) Based on the date and weather data, use an additional feature extraction layer to extract additional features; Step 25) Concatenate the spatial features, spatio-temporal features, and additional features to obtain a feature representation; Step 26) Combine the traffic accident dataset and the feature representation to obtain a graph dataset; The data augmentation for the graph dataset includes topology augmentation based on the traffic topology graph and feature augmentation based on node and edge masks. Among them, The above-mentioned traffic topology graph-based topological augmentation includes the following steps: sampling K-hop in the nearest neighbor manner with the node where a traffic accident occurs as the center, and on this basis, performing random walks on the current node to obtain ; The feature augmentation based on node and edge masks includes the following steps: For the current , randomly delete or add a preconfigured percentage of edges that are more than a preset distance away from the central node, and partially perturb their features to obtain the feature augmentation of the current ; The contrastive learning loss function is: in, Indicates the label of the sample, which can be 0 or 1; Indicates i The Euclidean distance of samples, m Represents the hyperparameters of the hinge loss function.

2. The abnormal prediction method for an intelligent transportation system based on a contrast graph neural network according to claim 1, wherein The road dataset includes road topology data, road average speed data, POI data, date, and weather data.

3. The abnormal prediction method for an intelligent transportation system based on a contrastive graph neural network according to claim 1, wherein The Graph-based spatial convolution layer includes several stacked graph convolutional neural network layers and embeds a residual structure. Among them, the output of the l -th layer is: Among them, represents the combination of neighbor nodes of the representative point represents the product of the degrees of node and node represents the parameter of the th layer convolution is the ReLU activation function.​ 4. The abnormal prediction method of an intelligent transportation system based on a contrast graph neural network according to claim 1, wherein The graph-based spatio-temporal convolutional layer includes a superposition of several graph convolutional neural network layers and traditional convolutional network layers, and embeds a residual structure. Spatio-temporal features are obtained by extracting time features and fusing spatial features. The method for extracting time features is: Among them, represents the extracted time feature, denotes the input data of the th layer, is the ReLU activation function, F is the number of convolutional kernels of this layer, , is the coefficient matrix and bias matrix of the (k - 1)th convolutional layer.

5. The abnormal prediction method of an intelligent transportation system based on a contrast graph neural network according to claim 1, characterized in that The additional feature extraction layer includes a superposition of several fully connected layers and activation functions. Each feature is encoded in One-Hot form to extract additional features.

6. An abnormal prediction device for an intelligent transportation system based on a contrastive graph neural network, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1-5.

7. A storage medium, on which a program is stored, characterized in that, When the program is executed, it implements the method according to any one of claims 1-5.

Images