Method for constructing small sample vehicle abnormal behavior detection model based on meta learning and application
Patent Information
- Application Number
- CN202410603857.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-15
- Publication Date
- 2026-09-25
- Estimated Expiration
- 2044-05-15
AI Technical Summary
对于许多现实问题，如车辆行驶时的异常行为检测，在较短时间段内，难以获取大量标注数据，或者人工标注的成本较高，由此，产生了小样本问题，小样本问题的数据集中仅包含少数几个标注的样本
[0040](1)本发明中的基于元学习的小样本车辆异常行为检测模型构建方法，通过将不同时间段对应的车辆行驶数据构建为对应的图结构数据，每个节点代表车辆，该图结构数据不仅包含丰富的属性信息，还包含节点之间复杂的结构信息，基于该图结构数据进行车辆异常行为检测，能够提升检测的准确度，同时，通过设计异常评分模块对车辆行为异常进行评分，在元学习训练过程中，将表征车辆行为异常的车辆行为异常评分与异常节点标签或正常节点标签进行损失计算，如此，将异常行为检测抽象为二分类问题，实现了将元学习迁移至小样本车辆异常行为检测上，能够充分利用元学习的优势有效利用小样本数据，进而提升小样本车辆异常行为检测的准确度。
Smart Images
Figure CN118468180B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of small sample vehicle abnormal behavior detection technology, and more specifically, relates to a method for constructing and applying a small sample vehicle abnormal behavior detection model based on meta-learning. Background Technology
[0002] In recent years, various deep learning methods have achieved excellent results in various machine learning tasks, and have also promoted the rapid deployment of deep learning applications for autonomous driving, such as intelligent transportation and vehicle modeling. However, the high performance of various deep learning methods often depends on the support of massive amounts of labeled data. For many real-world problems, such as the detection of abnormal behavior of vehicles, it is difficult to obtain a large amount of labeled data in a short period of time, or the cost of manual labeling is too high. This has led to the few-sample problem, in which the dataset contains only a few labeled samples.
[0003] Existing technologies typically perform anomaly detection by reconstructing the attributes and structural information of vehicle driving data. However, since the dataset for small sample problems contains only a few labeled samples, the accuracy of this detection method is not high. For training samples with few labeled data, unsupervised learning methods can also be used for model training. However, unsupervised learning methods are prone to misidentifying noise as anomalies, leading to low accuracy in the identification results.
[0004] Meta-learning enables models to quickly learn new tasks with a small number of samples, and it also has advantages such as high learning efficiency and strong generalization ability. However, meta-learning is mostly used in problems such as image classification and regression. How to transfer meta-learning to the detection of abnormal vehicle behavior with a small number of samples to improve the accuracy of abnormal vehicle behavior detection with a small number of samples is a current challenge. Summary of the Invention
[0005] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and application for constructing a small sample vehicle abnormal behavior detection model based on meta-learning, the purpose of which is to improve the accuracy of small sample vehicle abnormal behavior detection.
[0006] To achieve the above objectives, according to a first aspect of the present invention, a method for constructing a few-sample vehicle abnormal behavior detection model based on meta-learning is provided, comprising:
[0007] Vehicle driving data from different time periods are constructed into a corresponding graph structure data G = (N, A, X, L); where node N represents a vehicle, adjacency matrix A is used to characterize the structural information between nodes, node feature X is the attribute information of the corresponding node, and anomaly label L is used to characterize whether the current node is an abnormal node.
[0008] The graph structure data is divided into a set of non-overlapping meta-training tasks and a set of targets;
[0009] A small-sample vehicle abnormal behavior detection network is trained using the aforementioned meta-training task set and target set in a meta-learning manner to obtain a small-sample vehicle abnormal behavior detection model; wherein, the small-sample vehicle abnormal behavior detection network includes:
[0010] The feature extraction module is used to obtain the features of each sample node in the meta-training task set or target set by fusing attribute information and spatial information;
[0011] An anomaly scoring module is used to obtain an anomaly score for vehicle behavior based on the features output by the feature extraction module.
[0012] During the meta-learning training process, the abnormal vehicle behavior score is compared with the abnormal node label or normal node label to calculate the loss, and the network parameters are updated in reverse. When the preset training rounds are reached, a well-trained small sample vehicle abnormal behavior detection model is obtained.
[0013] Furthermore, the small-sample vehicle abnormal behavior detection network is trained using the aforementioned meta-training task set and target set in a meta-learning manner, including:
[0014] The meta-training task set is input into the small sample vehicle abnormal behavior detection network, and the vehicle behavior abnormal score is output. The vehicle behavior abnormal score is compared with the abnormal node label or normal node label to calculate the loss, and the network parameters are updated in reverse. After reaching the preset training rounds, a preliminarily trained small sample vehicle abnormal behavior detection model is obtained.
[0015] The target set is input into the pre-trained model to output anomaly vehicle behavior scores. The anomaly vehicle behavior scores are compared with the abnormal node labels or normal node labels to calculate the loss, and the network parameters are updated in reverse. After a preset number of training rounds, a trained small-sample vehicle anomaly behavior detection model is obtained.
[0016] Furthermore, the feature extraction module includes:
[0017] A graph encoder is used to obtain the feature representation of each node in graph structure data, wherein the feature representation includes the node's attribute information and structural information;
[0018] The self-attention submodule is used to perform self-attention perception on the feature representation to focus on the correlation between distant nodes and obtain self-attention features;
[0019] The fusion submodule is used to fuse the feature representation output by the graph encoder with the self-attention feature to obtain the fused self-attention feature. The fused self-attention feature includes the attribute information and spatial information of each node in the graph structure data after fusion.
[0020] Furthermore, the self-attention submodule includes:
[0021] Three parallel convolutional layers are used to extract query feature vectors, key-value feature vectors, and candidate feature vectors from the feature representations obtained from the graph encoder, respectively.
[0022] The self-attention computation layer is used to perform multimodal multiplication on the query feature vector and the key-value feature vector, divide the result by the number of elements in the candidate feature vector, and then pass the softmax function to obtain the self-attention weight vector.
[0023] The self-attention feature extraction layer is used to perform tensor multiplication on the self-attention weight vector and the candidate feature vector to obtain the self-attention features.
[0024] Furthermore, the graph encoder is a graph encoder with multiple GNN layers connected in series.
[0025] Furthermore, the loss function during training. for:
[0026]
[0027]
[0028] Where, dev(v) i ) represents the node v output by the anomaly scoring module. i abnormal scores s i and reference score μ i The deviation between them, reference score μ i Let σ be the average of multiple sampled fractions that satisfy a Gaussian prior distribution. r y is the standard deviation of the sample scores; i For node v i The corresponding label's true value, if node v i If y is a marked abnormal node, then i =1, otherwise, y i =0; m represents the confidence interval.
[0029] According to a second aspect of the present invention, a method for detecting anomalous vehicle behavior based on meta-learning is provided, comprising:
[0030] Construct vehicle driving data into a corresponding graph structure data;
[0031] The graph structure data is input into a trained small sample vehicle abnormal behavior detection model to obtain a vehicle behavior abnormality score; wherein the trained small sample vehicle abnormal behavior detection model is constructed by the small sample vehicle abnormal behavior detection model construction method described in any one of the first aspects.
[0032] According to a third aspect of the present invention, a system for constructing a small-sample vehicle abnormal behavior detection model based on meta-learning is provided, including a computer-readable storage medium and a processor;
[0033] The computer-readable storage medium is used to store executable instructions;
[0034] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the method for constructing a small sample vehicle abnormal behavior detection model as described in any of the first aspects.
[0035] According to a fourth aspect of the present invention, a small-sample vehicle abnormal behavior detection system based on meta-learning is provided, comprising a computer-readable storage medium and a processor;
[0036] The computer-readable storage medium is used to store executable instructions;
[0037] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the small sample vehicle abnormal behavior detection method described in the second aspect.
[0038] According to a fifth aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method for constructing a small sample vehicle abnormal behavior detection model as described in any of the first aspects, or / and the method for detecting small sample vehicle abnormal behavior as described in the second aspect.
[0039] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:
[0040] (1) The method for constructing a small-sample vehicle abnormal behavior detection model based on meta-learning in this invention constructs a graph structure data corresponding to vehicle driving data in different time periods. Each node represents a vehicle. This graph structure data not only contains rich attribute information, but also contains complex structural information between nodes. Based on this graph structure data, the accuracy of vehicle abnormal behavior detection can be improved. At the same time, by designing an abnormal scoring module to score vehicle behavior abnormalities, during the meta-learning training process, the vehicle behavior abnormality score representing vehicle behavior abnormality is compared with the abnormal node label or normal node label for loss calculation. In this way, abnormal behavior detection is abstracted into a binary classification problem, realizing the transfer of meta-learning to small-sample vehicle abnormal behavior detection. It can make full use of the advantages of meta-learning to effectively utilize small-sample data, thereby improving the accuracy of small-sample vehicle abnormal behavior detection.
[0041] (2) Further, by performing preliminary training on the small sample vehicle abnormal behavior detection network through the meta-training task set, the initial parameters of the small sample vehicle abnormal behavior detection model with better network parameters can be obtained. Based on the initial model with better parameters, the initial model is trained again using a target set that does not overlap with the meta-training task set. At this time, less labeled data and less training time can be used to complete the training of the initial model, and thus the trained small sample vehicle abnormal behavior detection model is obtained.
[0042] (3) As a preferred embodiment, considering the overfitting problem caused by using multi-layer graph convolutional neural networks to capture long-distance node information in graph data structures, this invention utilizes the self-attention mechanism to enable elements at different positions in the graph structure data to interact better based on the structural and attribute information extracted by the graph neural network, thereby achieving direct capture of long-distance dependencies between any two positions, thus maintaining more information and further improving the detection accuracy of small sample models.
[0043] (4) Preferably, the graph encoder is constructed using multiple graph neural network layers. These layers encode each node as a low-dimensional latent representation. The graph neural network follows the neighborhood message passing mechanism and aggregates features from the local neighborhood in an iterative manner to calculate the node representation. This can capture the node dependencies in the graph data structure and maintain the structural information of the graph structure data, thereby further improving the accuracy of the small sample anomaly detection model.
[0044] (5) As a preferred option, the design of the self-attention submodule fully considers every attribute, location and spatial information of the sample data, and can extract effective features more effectively than traditional graph neural networks.
[0045] (6) As a preferred option, the loss function uses a reference score as a reference for quantifying the degree of deviation between the abnormal score and the normal node score, so that the model assigns a larger abnormal score to those nodes whose features deviate significantly from the normal nodes, so that the abnormal nodes can obtain higher scores and facilitate the detection of abnormal vehicle behavior. Attached Figure Description
[0046] Figure 1 This is a schematic diagram of a small sample graph anomaly node detection framework that integrates meta-learning in an embodiment of the present invention.
[0047] Figure 2 This is a schematic diagram of the framework that integrates the self-attention mechanism and the graph encoder in an embodiment of the present invention. Detailed Implementation
[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0049] In this invention, the terms "first," "second," etc., used in the invention and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.
[0050] Example 1
[0051] The method for constructing a few-sample vehicle abnormal behavior detection model based on meta-learning provided in this embodiment of the invention includes:
[0052] Vehicle driving data for different time periods are constructed into a graph structure data G = (N, A, X, L), where node N represents a vehicle, and the adjacency matrix A is used to characterize the traffic network formed by the vehicle driving routes, reflecting the complex structural information between nodes. Each element in the adjacency matrix A indicates whether intersections, important buildings, etc. on the vehicle driving route are adjacent; node features X are the attribute information of the corresponding node, obtained through the vehicle driving data corresponding to the node; anomaly labels L are used to characterize whether the current node is an abnormal node, and only a small number of labeled abnormal nodes are included.
[0053] The graph-structured data is divided into a set of non-overlapping meta-training tasks and a set of objectives.
[0054] A small sample vehicle abnormal behavior detection network is trained using a meta-training task set and target set in a meta-learning manner to obtain a small sample vehicle abnormal behavior detection model; the small sample vehicle abnormal behavior detection network includes a feature extraction module and an anomaly scoring module.
[0055] The feature extraction module is used to obtain the features of each sample node in the meta-training task set or target set by fusing attribute information and spatial information;
[0056] The anomaly scoring module is used to obtain anomaly scores for vehicle behavior based on the node attribute information and spatial information output by the feature extraction module.
[0057] During the meta-learning training process, the abnormal vehicle behavior score output by the abnormal scoring module is compared with the abnormal node label or normal node label to calculate the loss and update the network parameters in reverse. When the preset training rounds are reached, a well-trained small sample vehicle abnormal behavior detection model is obtained.
[0058] like Figure 1 The diagram shown is a schematic of a few-sample graph anomaly node detection framework based on fusion meta-learning provided in an embodiment of the present invention. Figure 1 In this context, fθ represents the feature extraction function of the feature extraction module, and gθ represents the function of the anomaly scoring module.
[0059] In this embodiment of the invention, vehicle driving data corresponding to historical time periods is acquired and constructed into corresponding graph structure data. The graph structure data is then divided into non-overlapping meta-training task sets and target sets based on different time periods. A sample in either the meta-training task set or the target set represents the graph structure data corresponding to a specific time period. Graph structure data corresponding to different time periods collectively constitute the meta-training task set and the target set. The meta-training task set contains a small number of labeled anomalous node samples, while the target set contains even fewer labeled anomalous node samples compared to the meta-training task set. In this embodiment of the invention, each task set and target set is divided into a training set, a test set, and a validation set, with no overlap between the target task set and the meta-training task set.
[0060] Specifically, a small-sample vehicle abnormal behavior detection network is trained using a meta-training task set and target set in a meta-learning manner, including:
[0061] First, the meta-training task set is input into the feature extraction module to obtain the features of each sample node in the meta-training task set by fusing attribute information and spatial information. Then, the obtained fused node attribute information and spatial information are input into the anomaly scoring module to obtain the vehicle behavior anomaly score, and the loss is calculated. The network parameters are then updated in reverse to obtain the pre-trained small sample vehicle abnormal behavior detection model.
[0062] Then, the pre-trained model is fine-tuned based on the target set. Specifically, the target set is input into the pre-trained model, and the pre-trained model is trained again. Similarly, during the training process, the network parameters are updated in reverse through loss calculation to reach the preset training rounds and obtain the final trained small sample vehicle abnormal behavior detection model.
[0063] Thus, by initially training the small-sample vehicle anomaly detection network using a meta-training task set, initial parameters for a small-sample vehicle anomaly detection model with optimized network parameters can be obtained. Based on this optimized initial model, the initial model is then trained again using a target set that does not overlap with the meta-training task set. This allows for training the initial model with less labeled data and less training time, resulting in a fully trained small-sample vehicle anomaly detection model. In this invention, the introduction of meta-learning technology can better utilize graph data knowledge from the same source domain. During the training phase, meta-knowledge of real anomalies is extracted from different small-sample graph anomaly node detection tasks, and further fine-tuning is performed for new tasks, enabling the model to quickly and effectively adapt to target data.
[0064] As a further design of the present invention, the feature extraction module includes a graph encoder, a self-attention submodule, and a fusion submodule;
[0065] Graph encoders are used to obtain feature representations of each node in graph structure data, which include the node's attribute information and structural information;
[0066] The self-attention submodule is used to perform self-attention perception on the feature representation of each node, so as to focus on the correlation between distant nodes in the entire input data and obtain self-attention features;
[0067] The fusion submodule is used to fuse the feature representation of each node output by the graph encoder with the self-attention feature to obtain the fused self-attention feature. The fused self-attention feature contains the features of the fused attribute information and spatial information of each node in the graph structure data.
[0068] Considering the overfitting problem caused by using multi-layer graph convolutional neural networks to capture long-distance node information in graph data structures, in this embodiment of the invention, based on the structural and attribute information of the graph structure data extracted simultaneously by the graph neural network, a self-attention mechanism is used to enable elements at different positions in the graph structure data to interact better, thereby achieving direct capture of long-distance dependencies between any two positions. This can maintain more information and improve the detection accuracy of the small sample model.
[0069] Preferably, in this embodiment of the invention, a graph encoder consisting of L concatenated GNN layers is used to learn the feature representations of the nodes. The graph encoder calculation formula is as follows:
[0070] H 1 =GNN 1 (A, X),
[0071] F = GNN L (A, H) L-1 )
[0072] Where A represents the adjacency matrix; X represents the node features; H 1 F represents the output of the first GNN layer; F is the node feature representation output from the graph encoder based on L concatenated GNN layers. The graph encoder is compatible with architectures based on arbitrary GNNs; in this embodiment of the invention, it is implemented using simple graph convolution.
[0073] In this embodiment of the invention, the graph encoder described above is preferably constructed from multiple graph neural network layers. These layers encode each node as a low-dimensional latent representation. The graph neural network follows a neighborhood message passing mechanism and iteratively aggregates features from the local neighborhood to compute the node representation. To capture node dependencies in the graph data structure, multiple graph neural network layers are stacked in the graph encoder, which can maintain the structural information of the graph data and thus improve the accuracy of the few-shot anomaly detection model.
[0074] Preferably, in this embodiment of the invention, the self-attention submodule includes:
[0075] Three parallel convolutional layers are used to extract query feature vectors, key-value feature vectors, and candidate feature vectors from the features obtained from the graph encoder, respectively.
[0076] The self-attention computation layer is used to perform multimodal multiplication on the query feature vector and the key-value feature vector, divide the result by the number of elements in the candidate feature vector, and then pass it through the softmax function to obtain the self-attention weight vector.
[0077] The self-attention feature extraction layer performs tensor multiplication on the self-attention weight vector and the candidate feature vector to obtain tensor self-attention features. Preferably, the self-attention feature calculation formula is as follows:
[0078]
[0079] Among them, F i * F represents the self-attention feature. i Let Q represent the feature representation of the i-th node extracted by the graph encoder, Q represent the query feature vector, K represent the key-value feature vector, and d represent the tensor (Q). i *K i The number of elements in ), where V represents the candidate feature vector.
[0080] The present invention preferably uses the above-mentioned self-attention calculation formula, which, because it fully considers every attribute, location and spatial information of the sample data, can extract effective features more effectively than traditional graph neural networks.
[0081] like Figure 2 The diagram shown is a schematic representation of the framework integrating a self-attention mechanism and a graph encoder in an embodiment of the present invention. Figure 2 p ini w represents the feature of node i. i The node features after self-attention, λ i This represents the element value in the adjacency matrix A.
[0082] Preferably, the loss calculation method during the training phase of the small-sample vehicle abnormal behavior detection model is as follows:
[0083]
[0084]
[0085] in, Represents the global loss, dev(v) i ) represents node v in the graph structure data output by the anomaly scoring module. i abnormal scores s i and reference score μ i Deviation between; reference score μ i Let σ be the average of k sample scores that satisfy a Gaussian prior distribution, where k is chosen empirically. r y represents the standard deviation of the sample scores. i For node v i The corresponding true value of the label, in this embodiment of the invention, if node v i For the marked abnormal nodes, y i The value of y is 1 in all other cases. i The value is 0; m represents the confidence interval and defines the radius of deviation.
[0086] In this embodiment of the invention, the reference score μ i for:
[0087]
[0088] The sampled scores that constitute the reference scores satisfy a Gaussian prior distribution:
[0089]
[0090] Where, μ, σ 2 These are the mean and variance of the Gaussian prior distribution, respectively.
[0091] The preferred deviation loss calculation formula of this invention uses a reference score as a reference to quantify the degree of deviation between the abnormal score and the normal node score, so that the model assigns a larger abnormal score to nodes whose features deviate significantly from those of normal nodes, so that abnormal nodes can obtain higher scores and facilitate the detection of abnormal vehicle behavior.
[0092] The method for constructing a few-sample vehicle abnormal behavior detection model based on meta-learning in this invention constructs a graph structure data based on vehicle driving data corresponding to different time periods. Each node represents a vehicle. This graph structure data not only contains rich attribute information but also complex structural information between nodes. Vehicle abnormal behavior detection based on this graph structure data can improve the accuracy of detection. At the same time, by designing an anomaly scoring module, a vehicle behavior anomaly score is obtained to characterize vehicle behavior abnormalities. During the meta-learning training process, the vehicle behavior anomaly score is compared with the abnormal node label or normal node label for loss calculation. In this way, abnormal behavior detection is abstracted into a binary classification problem, realizing the transfer of meta-learning to few-sample vehicle abnormal behavior detection. It can fully utilize the advantages of meta-learning to effectively utilize few-sample data, solve the problem of difficulty in detecting small-sample anomalies in human-machine-object systems, and improve the accuracy of few-sample vehicle abnormal behavior detection.
[0093] Furthermore, by introducing a feature extraction module that integrates a self-attention mechanism and a graph neural network, this invention better represents the attribute and structural information of graph structure data, thereby further improving the accuracy of vehicle abnormal behavior detection.
[0094] Example 2
[0095] This invention provides a few-sample vehicle abnormal behavior detection method based on meta-learning, including:
[0096] Vehicle driving data is constructed into a corresponding graph structure and input into a pre-trained small-sample vehicle abnormal behavior detection model for fine-tuning, resulting in a vehicle behavior anomaly score. The pre-trained small-sample vehicle abnormal behavior detection model is constructed using the model construction method described in Example 1 above.
[0097] Example 3
[0098] This invention provides a system for constructing a few-sample vehicle abnormal behavior detection model based on meta-learning, including a computer-readable storage medium and a processor;
[0099] Computer-readable storage media are used to store executable instructions;
[0100] The processor is used to read executable instructions stored in a computer-readable storage medium and execute the method for constructing a small sample vehicle abnormal behavior detection model in Example 1.
[0101] Example 4
[0102] This invention provides a few-sample vehicle abnormal behavior detection system based on meta-learning, including a computer-readable storage medium and a processor;
[0103] Computer-readable storage media are used to store executable instructions;
[0104] The processor is used to read executable instructions stored in a computer-readable storage medium and execute the small sample vehicle abnormal behavior detection method in Embodiment 2.
[0105] Example 5
[0106] This invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for constructing a small sample vehicle abnormal behavior detection model as in Embodiment 1, or / and the method for detecting small sample vehicle abnormal behavior as in Embodiment 2.
[0107] This invention integrates self-attention mechanisms and traditional graph neural networks to maintain the attribute and structural features of graph-structured data; it uses a bias loss network to train the model to assign higher anomaly scores to abnormal nodes; and it integrates meta-learning techniques to extract meta-knowledge of real anomalies from different auxiliary tasks during the meta-training phase, and further fine-tunes the model for new tasks, enabling the model to quickly and effectively adapt to target data and improve the accuracy of detecting abnormal vehicle behavior in small samples.
[0108] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for constructing a few-sample vehicle abnormal behavior detection model based on meta-learning, characterized in that, include: Vehicle driving data from different time periods are constructed into a corresponding graph structure data G=(N,A,X,L); where node N represents a vehicle, adjacency matrix A is used to characterize the structural information between nodes, node feature X is the attribute information of the corresponding node, and anomaly label L is used to characterize whether the current node is an abnormal node. The graph structure data is divided into a set of non-overlapping meta-training tasks and a set of targets; A small-sample vehicle abnormal behavior detection network is trained using the aforementioned meta-training task set and target set in a meta-learning manner to obtain a small-sample vehicle abnormal behavior detection model; wherein, the small-sample vehicle abnormal behavior detection network includes: The feature extraction module is used to obtain the features of each sample node in the meta-training task set or target set by fusing attribute information and spatial information; An anomaly scoring module is used to obtain an anomaly score for vehicle behavior based on the features output by the feature extraction module. During the meta-learning training process, the vehicle behavior anomaly score is compared with the anomaly node label or normal node label to calculate the loss, and the network parameters are updated in reverse; when the preset training rounds are reached, a well-trained small sample vehicle abnormal behavior detection model is obtained. The feature extraction module includes: A graph encoder is used to obtain the feature representation of each node in graph structure data, wherein the feature representation includes the node's attribute information and structural information; The self-attention submodule is used to perform self-attention perception on the feature representation to focus on the correlation between distant nodes and obtain self-attention features; The fusion submodule is used to fuse the feature representation output by the graph encoder with the self-attention feature to obtain the fused self-attention feature, wherein the fused self-attention feature includes the attribute information and spatial information features of each node in the graph structure data after fusion. Loss function during training for: in, Nodes output by the anomaly scoring module abnormal scores and reference scores The deviation between them, reference score Let be the average of multiple sample scores that satisfy a Gaussian prior distribution. The standard deviation of the sample scores; For nodes The corresponding label's actual value, if the node If it is a marked abnormal node, then =1, otherwise, =0; This represents the confidence interval.
2. The method for constructing a small-sample vehicle abnormal behavior detection model according to claim 1, characterized in that, The small-sample vehicle abnormal behavior detection network is trained using the aforementioned meta-training task set and target set in a meta-learning manner, including: The meta-training task set is input into the small sample vehicle abnormal behavior detection network, and the vehicle behavior abnormal score is output. The vehicle behavior abnormal score is compared with the abnormal node label or normal node label to calculate the loss, and the network parameters are updated in reverse. After reaching the preset training rounds, a preliminarily trained small sample vehicle abnormal behavior detection model is obtained. The target set is input into the pre-trained model to output anomaly vehicle behavior scores. The anomaly vehicle behavior scores are compared with the abnormal node labels or normal node labels to calculate the loss, and the network parameters are updated in reverse. After a preset number of training rounds, a trained small-sample vehicle anomaly behavior detection model is obtained.
3. The method for constructing a small-sample vehicle abnormal behavior detection model according to claim 1, characterized in that, The self-attention submodule includes: Three parallel convolutional layers are used to extract query feature vectors, key-value feature vectors, and candidate feature vectors from the feature representations obtained from the graph encoder, respectively. The self-attention computation layer is used to perform multimodal multiplication on the query feature vector and the key-value feature vector, divide the result by the number of elements in the candidate feature vector, and then pass the softmax function to obtain the self-attention weight vector. The self-attention feature extraction layer is used to perform tensor multiplication between the self-attention weight vector and the candidate feature vector to obtain the self-attention features.
4. The method for constructing a small-sample vehicle abnormal behavior detection model according to claim 1, characterized in that, The graph encoder is a graph encoder that cascades multiple GNN layers.
5. A method for detecting abnormal vehicle behavior in small samples based on meta-learning, characterized in that, include: Construct vehicle driving data into a corresponding graph structure data; The graph structure data is input into the trained small sample vehicle abnormal behavior detection model to obtain a vehicle behavior abnormality score; wherein the trained small sample vehicle abnormal behavior detection model is constructed by the small sample vehicle abnormal behavior detection model construction method according to any one of claims 1-4.
6. A system for constructing a few-sample vehicle abnormal behavior detection model based on meta-learning, characterized in that, Includes computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the method for constructing a small sample vehicle abnormal behavior detection model according to any one of claims 1-4.
7. A few-sample vehicle abnormal behavior detection system based on meta-learning, characterized in that, Includes computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the small sample vehicle abnormal behavior detection method according to claim 5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for constructing a small sample vehicle abnormal behavior detection model as described in any one of claims 1-4, or / and the method for detecting small sample vehicle abnormal behavior as described in claim 5.