Method for identifying key vehicles in complex traffic network for intelligent driving car decision planning and vehicle device

Abstract

The application discloses a key vehicle identification method for intelligent driving automobile decision planning in a complex traffic network and a vehicle device, utilizes a complex network theory to model an automatic driving vehicle and a surrounding complex environment of the automatic driving vehicle, forms a complex network of the vehicle, then sorts the importance of the relationship between surrounding vehicles and the vehicle through a graph neural network (GNN) based on an attention mechanism, i.e., a graph attention network (GAT), identifies a key node of the vehicle, and converts the key vehicle identification problem which has an important influence on the safety of the automatic driving automobile into the importance identification problem of the key node in the complex network. Furthermore, the graph attention network is used to identify an abnormal node of the complex network formed by the automatic driving automobile, i.e., a node which suddenly becomes a great threat to the vehicle, so as to prevent the abnormal node from influencing the dynamic complex network of the vehicle and protect the safety of the vehicle.

Description

Technical Field

[0001] This invention belongs to the field of intelligent driving, specifically a key vehicle identification method and vehicle equipment for intelligent vehicles that have a significant impact on safety by integrating complex networks and graph neural networks based on attention mechanisms. Background Technology

[0002] With the development, popularization, and deployment of autonomous vehicles, they will coexist with conventional vehicles on the road for a considerable period. Furthermore, as the market penetration rate of autonomous vehicles continues to increase, it will inevitably have a significant impact on existing road traffic flow. Compared to manually driven vehicles, autonomous vehicles have a smaller impact on mixed traffic flow. In other words, with less right-of-way, making intelligent decision-making in complex mixed traffic environments becomes crucial. Applying human driving behavior decisions to autonomous driving decision-making, identifying the vehicle's key interaction targets—that is, finding critical vehicles—is essential and represents the first step in making effective decisions.

[0003] Complex networks are highly complex networks, abstractions of complex systems, and generally possess properties such as self-organization, self-similarity, small-world characteristics, and scale-free operation. Autonomous vehicles, on the other hand, are integrated systems combining perception, decision-making, and control. Autonomous vehicles operate within a complex system, their movement influenced by the complex environment and surrounding individuals. Therefore, for autonomous vehicles, the interaction of the surrounding environment and individual influences constitutes the complex network of autonomous driving. Treating vehicles as nodes and roads as edges and constraints, the critical vehicle identification problem, which significantly impacts the safety of autonomous vehicles, can be transformed into the importance identification problem of key nodes within a complex network.

[0004] With the development of graph neural networks (GNNs), GNNs have become a representation learning method for depicting the relationships between nodes in a graph. Attention mechanisms are now widely used in sequence-based tasks, as they amplify the influence of the most important parts of the data. This invention proposes how attention mechanisms can be used in network graph structures. By aggregating neighboring nodes through self-attention, the accuracy of the model is improved, and node weights are assigned to complete the task of identifying key nodes in complex networks. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and vehicle-mounted equipment for identifying key vehicle nodes that significantly impact the safety of autonomous vehicles, which integrates complex networks and attention-based graph neural networks.

[0006] According to the method provided by the present invention, the solution is as follows:

[0007] The vehicle key node identification method provided by this invention mainly includes the following aspects: modeling autonomous vehicles and their surrounding complex environment using complex network theory to form a complex network of the vehicle; then ranking the importance of the relationship between the vehicle and surrounding vehicles using a graph neural network (GNN) based on an attention mechanism, namely a graph attention network (GAT), to identify key nodes of the vehicle; and further identifying abnormal nodes, i.e., nodes whose threat to the vehicle suddenly increases, through the graph attention network to the complex network formed by autonomous vehicles, to prevent abnormal nodes from affecting the dynamic complex network of the vehicle and to protect vehicle safety.

[0008] The method for identifying the importance of key vehicle nodes includes the following steps:

[0009] Step 1: Using complex network theory, dynamically model the driving environment of the autonomous vehicle as it moves, with itself as the master node, including:

[0010] Step 1.1: First, based on complex network theory, the vehicle and other vehicles on the road are considered as nodes in a complex network, with the road as the edge, forming a complex network model that includes surrounding neighboring vehicle nodes, thus generating a dynamic complex network for the vehicle:

[0011] G = (V, B, W, P)

[0012] Where G is a time-varying complex dynamic network, V is the set of nodes in the time-varying complex dynamic network G, B is the set of edges in the time-varying complex dynamic network G, representing the connections between nodes, W is the weight of the nodes in the time-varying complex dynamic network G, and P is the function of the edges in the complex dynamic network G, representing the relationship between nodes.

[0013] Step 1.2: Based on the established model, the vehicle in the complex network is considered as the master node. The master node is connected to its neighboring nodes within three orders of magnitude. Driving feature information of the vehicle's neighboring nodes is collected, primarily including the path length, speed, and acceleration between the two vehicles. A corresponding three-degree-of-freedom driving feature vector is generated from these three parameters, forming the input layer of the graph neural network. The graph neural network then processes the feature vectors of the complex network nodes. The constructed three-degree-of-freedom driving feature vector is as follows:

[0014] v j =(dij, a, a) + )

[0015] Where d ij This represents the path length between two nodes, where 'a' is the node speed. + For node acceleration.

[0016] Step 2: Aggregate the feature vectors of the input layer using a graph neural network based on an attention mechanism, parameterize the driving feature information, and assign weights through a multi-head attention mechanism to obtain attention parameters and determine the importance of nodes. This includes:

[0017] Step 2.1: Use a graph neural network to aggregate the collected feature information of complex network nodes, forming a set of node features, then we have:

[0018] v = {v1, v2, ..., v} n}, v j ∈B F

[0019] Where n is the number of nodes, F is the number of features for each node, and v is the number of features of the vehicle's neighboring nodes. i The set consisting of (i = 1, 2, ..., n).

[0020] And by regenerating a new set of node features (with potentially different cardinalities F′) at the input layer, we have:

[0021] v′={v′1, v′2,..., v′ n}, v′ j ∈B F′

[0022] Where v′ is the set of node output features.

[0023] Step 2.2: Map the vehicle master node and its neighbor nodes using a graph attention network, defined by a weight matrix W∈B. F The features of each vehicle node are parameterized. To give the nodes sufficient expressive power, the input features are transformed into higher-level features, requiring a learnable linear transformation. Therefore, this invention utilizes a weight matrix W∈B... F A parameterized shared linear transformation is applied to each node. Then, self-attention—a shared attention mechanism—is performed on each vehicle node: R F′ ×R F′ →R and calculate the attention coefficient, then we have:

[0024] x ij =(Wv i Wv j ), j∈V

[0025] Where x ij V represents the importance of node j to node i, W is the weight matrix of driving feature parameters, and v i v j This refers to the driving characteristic information of this vehicle node.

[0026] Furthermore, to facilitate comparison of attention coefficients across different vehicle nodes, the importance of nodes is normalized, resulting in:

[0027]

[0028] Where a ij x is the normalized vehicle attention coefficient. ik Let exp represent the importance of each node to the master node, and let softmax be the logistic regression function.

[0029] Because nodes have a shared attention mechanism, to obtain the weight of each vehicle node, an activation function is used to concatenate the result vectors of nodes i and j. This prevents the attention value from being calculated so that nodes only relate to their neighboring nodes and ignore the main node. Therefore, adding an activation function ensures that the main node cannot be ignored, thus yielding:

[0030]

[0031] LeakyReLu is an activation function in neural networks used to handle nonlinear problems, while This maps concatenated vectors to real numbers for a feedforward neural network.

[0032] The formula for normalization coefficients has also been updated to:

[0033]

[0034] Where || represents the connection relationship, which is a concatenation operation.

[0035] After obtaining the normalized coefficients of the vehicles, the linear combination of the features of the corresponding vehicle nodes is calculated as the final output features of the vehicle nodes. An activation function is then applied to transform the nonlinear problem into a linear one, resulting in the following equation:

[0036]

[0037] v′ j This is the final output feature representing the importance of node j to node i.

[0038] Step 2.3: Since there are three driving feature parameters in this invention, a multi-head attention mechanism (K=3) is used. By evaluating the importance of individual node features, three independent attention mechanisms are established for these three node features. The feature information of the three nodes is parameterized, and after aggregation, a new node importance coefficient is obtained. To stabilize the self-attention learning process, the learned features are concatenated, resulting in the following formula:

[0039]

[0040] k represents the number of attention mechanism heads, which is 3 in this invention. The normalized attention coefficient W is calculated for the 3-times attention mechanism. k This is the weight matrix for the corresponding input linear transformation.

[0041] In graph attention networks, a multi-head attention mechanism is implemented. Since the weights of features differ, simply connecting them is not ideal. This invention averages the features based on their respective proportions, as shown in the following equation:

[0042]

[0043] Softmax is used to apply the nonlinear process of multi-head attention, while v′ j Then update the feature values ​​to the final output.

[0044] Step 3: Identify abnormal nodes in the complex network of autonomous vehicles using a graph attention network. This involves introducing a graph attention mechanism through graph embedding, capturing the communication relationships between vehicles to analyze abnormal behavior, and abstracting this into a complex network structure for graph-based anomaly detection. Specifically, the graph attention mechanism is used to detect abnormal vehicle behavior, analyzing the impact of changes in vehicle node driving feature information on vehicle safety. If an anomaly is detected, the node is updated and its importance is reassessed. This includes the following steps:

[0045] Step 3.1: During vehicle operation, the vehicle dynamic network is in a dynamic equilibrium state. However, this equilibrium is disrupted when the vehicle makes decisions, causing changes in the dynamic network formed by the autonomous vehicle as the main node. Vehicle behavior within the network is random, and the structural information of entities connected by random, anomalous links often lacks similarity compared to normal behavior. Based on this assumption, the behavioral features of each node in the network can be learned and updated through aggregation. Therefore, the anomalous features of node feature vectors, i.e., the anomalous driving feature information, are utilized.

[0046] Step 3.2: For the nodes of the vehicle network. Suppose that there exists an edge e between node pair (i, j). ij The anomaly probability can be expressed as P((i,j)|B(i), B(j)∈[0,1]), where B(i) and B(j) represent the sets of neighboring nodes of i and j, respectively. By applying hierarchical attention judgment to the feature vectors of nodes, the distance, speed, and acceleration of vehicle nodes are hierarchically divided to determine the anomaly of vehicle nodes. The discretized state set L = {non-existent, normal, abnormal} can be converted into a numerical representation, thereby obtaining the anomaly probability of the edges between nodes.

[0047] Based on the eigenvectors, the initialization matrix H is obtained. (0) =B F Knowing the current state of the vehicle node and its relationship with the master node, we process it using a graph attention mechanism, using the following formula:

[0048] Current t =GAT(H (o) )

[0049] Current t GAT represents the dynamic update mechanism for node embedding in the graph attention layer, which captures and acquires information representing short-term vehicle nodes.

[0050] Then, using a function, the anomaly probability of that node is calculated:

[0051]

[0052] Where a and b are the contrast parameters of the attention layer, p(i,j) is the anomaly probability of the vehicle node, σ is the sigmoid function, and β is the hyperparameter. The control probability is within [0, 1].

[0053] Step 3.3: After obtaining the abnormal probability of the node, determine whether the abnormal node poses a threat to itself, i.e., p(i,j)>0.5. Update the driving feature vector of the abnormal node, use the graph attention network layer to update the behavioral features of the vehicle node, thereby obtaining a new feature vector, and then calculate the importance of the node to the master node through learning.

[0054] If p(i,j)<0.5, the original graph attention network will still be used for computation, and no network point embedding will be performed.

[0055] Step 4: After the graph attention network learns and processes the nodes and identifies abnormal nodes, the vehicle node feature v′ is output after normalization. j By using the connection layers of a graph neural network, the differences between nodes in the network are obtained, and the final feature value v′ is used as the basis for this analysis. j Arrange the nodes in ascending order of importance, and the maximum value is the value of the most critical node. This sorts the nodes by importance and ultimately identifies the critical vehicle nodes. This allows us to determine the vehicles that have the greatest impact on the safety of our own vehicle, and then decide how to interact with the vehicles to make the right decisions.

[0056] The present invention also proposes a vehicle-mounted device capable of performing the aforementioned key vehicle identification method.

[0057] The beneficial effects of this invention are:

[0058] 1. This invention models roads and autonomous vehicles and other vehicles on roads based on the cognition of complex network theory. It uses complex network theory to determine the key nodes of vehicles, transforming the problem of identifying key vehicles that have an important impact on the safety of autonomous vehicles into the problem of identifying the importance of key nodes in complex networks.

[0059] 2. This invention utilizes a graph neural network method based on an attention mechanism for key node identification in complex networks. By using a graph attention network to transform the feature information of vehicle nodes into more easily understood attention coefficients, the importance of vehicle nodes to the master node (i.e., the self-vehicle) can be determined, thereby reducing the complexity of the environment during the movement of autonomous vehicles and improving the cognitive ability of autonomous vehicles.

[0060] 3. This invention addresses abnormal nodes in vehicle networks by using graph embedding. It utilizes graph attention mechanisms to capture communication relationships between vehicles to analyze abnormal behavior and abstracts them into a complex network structure. In graph-based anomaly detection, the discretized state set is transformed into the anomaly probability of vehicle nodes using vehicle node feature information. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method of the present invention.

[0062] Figure 2 Flowchart for determining key nodes of a vehicle using a graph attention network

[0063] Figure 3 Example diagram of anomaly detection at key vehicle nodes Detailed Implementation

[0064] The invention will now be further described with reference to the accompanying drawings.

[0065] The technical solution of the present invention will now be described in detail, clearly and completely with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0066] This invention provides a vehicle recognition method suitable for autonomous vehicles to accurately perceive complex environments. It integrates complex network theory and attention-based graph neural networks, transforming the critical vehicle recognition problem, which significantly impacts the safety of autonomous vehicles, into the problem of identifying the importance of key nodes in a complex network. The attention mechanism is used in the network structure. By aggregating neighboring nodes through self-attention, the model's accuracy is improved, and node weights are assigned to complete the key node identification task. Furthermore, the graph attention network identifies abnormal nodes—those whose threat to the vehicle suddenly increases—within the complex network formed by autonomous vehicles, preventing abnormal nodes from affecting the dynamic and complex network and protecting vehicle safety. Figure 1As shown, the specific steps include:

[0067] (1) Using complex network theory, a dynamic model of the driving environment is formed by the autonomous vehicle itself as the main node during its movement. Based on complex network theory, the vehicle and other vehicles on the road are regarded as complex network nodes, and the road is used as the edge to form a complex network model including the surrounding neighboring vehicle nodes, thus generating a dynamic complex network of the vehicle:

[0068] G = (V, B, W, P)

[0069] Where G is a time-varying complex dynamic network, V is the set of nodes in the time-varying complex dynamic network G, B is the set of edges in the time-varying complex dynamic network G, representing the connections between nodes, W is the weight of the nodes in the time-varying complex dynamic network G, and P is the function of the edges in the complex dynamic network G, representing the relationship between nodes.

[0070] In this invention, based on the established model, the vehicle in the complex network is considered as the master node. The master node is connected to its neighboring nodes within three orders of magnitude. Driving feature information of the vehicle's neighboring nodes is collected, primarily including the path length, speed, and acceleration between the two vehicles. These three parameters generate corresponding three-degree-of-freedom feature vectors, which form the input layer of the graph neural network. The graph neural network then processes the feature vectors of the complex network nodes. The constructed driving feature information feature vector is as follows:

[0071] v j =(d ij ,a,a + )

[0072] Where d ij This represents the path length between two nodes, where 'a' is the node speed. + For node acceleration.

[0073] (2) After modeling a complex network, a graph neural network is used to evaluate important nodes and rank them to obtain information for interaction with the target vehicle, such as... Figure 2 As shown, it includes the following steps:

[0074] First, a graph neural network is used to aggregate the collected feature information of complex network nodes, forming a set of node features, then:

[0075] v = {v1, v2, ..., v} n}, v j ∈B F

[0076] Where n is the number of nodes, F is the number of features for each node, and v is the feature set of the vehicle's neighboring nodes.

[0077] And by regenerating a new set of node features (with potentially different cardinalities F′) at the input layer, we have:

[0078] v′={v′1, v′2,..., v′ n}, v′ j ∈B F′

[0079] Where v′ is the set of node output features.

[0080] Then, a graph attention network is used to map the vehicle master node and its neighbor nodes, using a weight matrix W∈B. F The features of each vehicle node are parameterized. To give the nodes sufficient expressive power, the input features are transformed into higher-level features, requiring a learnable linear transformation. Therefore, this invention utilizes a weight matrix W∈B... F A parameterized shared linear transformation is applied to each node. Then, self-attention—a shared attention mechanism—is performed on each vehicle node: R F′ ×R F′ →R and calculate the attention coefficient, then we have:

[0081] x ij =(Wv i Wv j ), j∈V

[0082] Where x ij V represents the importance of node j to node i, W is the weight matrix of driving feature parameters, and v i v j This refers to the driving characteristic information of this vehicle node.

[0083] Furthermore, to facilitate comparison of attention coefficients across different vehicle nodes, the importance of nodes is normalized, resulting in:

[0084]

[0085] Where a ij x is the normalized vehicle attention coefficient. ik Let exp represent the importance of each node to the master node, and let softmax be the logistic regression function.

[0086] Because nodes have a shared attention mechanism, to obtain the weight of each vehicle node, an activation function is used to concatenate the result vectors of nodes i and j. This prevents the attention value from being calculated so that nodes only relate to their neighboring nodes and ignore the main node. Therefore, adding an activation function ensures that the main node cannot be ignored, thus yielding:

[0087]

[0088] LeakyReLu is the activation function in a neural network used to handle nonlinear problems, while θ T This maps concatenated vectors to real numbers for a feedforward neural network.

[0089] The formula for normalization coefficients has also been updated to:

[0090]

[0091] Where || represents the connection relationship, which is a concatenation operation.

[0092] After obtaining the normalized coefficients of the vehicles, the linear combination of the features of the corresponding vehicle nodes is calculated as the final output features of the vehicle nodes. An activation function is then applied to transform the nonlinear problem into a linear one, resulting in the following equation:

[0093]

[0094] v′ j This is the final output feature representing the importance of node j to node i.

[0095] Based on the above importance assessment for a single node feature, since this invention has three driving feature parameters, a multi-head attention mechanism (K=3) is used to establish three independent attention mechanisms for these three node features. The feature information of the three nodes is parameterized, and after aggregation, a new node importance coefficient is obtained. In order to stabilize the self-attention learning process, the learned features are concatenated, resulting in the following formula:

[0096]

[0097] k represents the number of attention mechanism heads, which is 3 in this invention. The normalized attention coefficient W is calculated for the 3-times attention mechanism. k This is the weight matrix for the corresponding input linear transformation.

[0098] In graph attention networks, a multi-head attention mechanism is implemented. Since the weights of features differ, simply connecting them is not ideal. This invention averages the features based on their respective proportions, as shown in the following equation:

[0099]

[0100] Softmax is used to apply the nonlinear process of multi-head attention, while v′ j Then update the feature values ​​to the final output.

[0101] (3) In the process of aggregating learning information, some unexpected situations will inevitably occur. During the driving process, a dynamic balance will be formed between vehicles. When the vehicle makes a decision, this balance will be broken. Therefore, how to detect the anomaly of the network graph is a very important thing. It can capture the importance of the relationship in the graph.

[0102] Therefore, this invention proposes using graph embedding and graph attention mechanisms to analyze abnormal behavior by capturing the communication relationships between vehicles, and abstracting this into a complex network structure for graph-based anomaly detection. In other words, it utilizes graph attention mechanisms to detect abnormal vehicle behavior. Figure 3 As shown, it mainly includes the following aspects:

[0103] Because the network is a dynamic network with autonomous vehicles as the main nodes, changes occur. Vehicle behavior within the network is random, and the structural information of entities connected by random, anomalous links often lacks similarity to normal behavior. Based on this assumption, the behavioral features of each node in the network can be learned and updated through aggregation. Therefore, by utilizing the anomalies in the node's feature vectors—specifically, the anomalies in driving feature information—the behavioral features of that node can be updated, thereby deriving a new node feature vector.

[0104] By employing graph embedding and introducing a graph attention mechanism, abnormal behavior is analyzed by capturing the communication relationships between vehicles, and this is abstracted into a complex network structure for graph-based anomaly detection. Specifically, graph attention is used to detect abnormal vehicle behavior. This includes:

[0105] For the nodes in the vehicle network, suppose there exists an edge e between node pair (i, j). ij The anomaly probability can be expressed as P((i,j)|B(i), B(j)∈[0,1]), where B(i) and B(j) represent the sets of neighboring nodes of i and j, respectively. By applying hierarchical attention judgment to the feature vectors of nodes, the distance, speed, and acceleration of vehicle nodes are hierarchically divided to determine the anomaly of vehicle nodes. The discretized state set L = {non-existent, normal, abnormal} can be converted into a numerical representation, thereby obtaining the anomaly probability of the edges between nodes.

[0106] Based on the eigenvectors, the initialization matrix H is obtained. (0) =B F Knowing the current state of the vehicle node and its relationship with the master node, we process it using a graph attention mechanism, using the following formula:

[0107] Current t =GAT(H (O) )

[0108] Current tGAT represents the dynamic update mechanism for node embedding in the graph attention layer, which captures and acquires information representing short-term vehicle nodes.

[0109] Then, the abnormal probability of this node is calculated using a function:

[0110]

[0111] Where a and b are the contrast parameters of the attention layer, p(i,j) is the anomaly probability of the vehicle node, σ is the sigmoid function, and β is the hyperparameter. The result is controlled to be within [0, 1].

[0112] Finally, after obtaining the abnormal probability of the node, it is determined whether the abnormal node poses a threat to itself, i.e., p(i,j)>0.5. The driving feature vector of the abnormal node is updated, and the behavioral features of the vehicle node are updated using the graph attention network layer to obtain a new feature vector. Then, the importance of the node to the master node is calculated through learning.

[0113] If p(i,j)<0.5, the original graph attention network will still be used for computation, and no network point embedding will be performed.

[0114] (4) Figure 2 As shown, after the graph attention network learns and processes the nodes, identifies abnormal nodes, and normalizes the data, it outputs the vehicle node feature v′. j By using the connection layers of a graph neural network, the differences between nodes in the network are obtained, and the final feature value v′ is used as the basis for this analysis. j Arrange the nodes in ascending order of importance, and the maximum value is the value of the most critical node. This sorts the nodes by importance and ultimately identifies the critical vehicle nodes. This allows us to determine the vehicles that have the greatest impact on the safety of our own vehicle, and then decide how to interact with the vehicles to make the right decisions.

[0115] This invention also proposes a vehicle-mounted device capable of performing the aforementioned key vehicle identification method.

[0116] In summary, this invention, based on the cognition of complex network theory, uses complex network theory to determine key vehicle nodes, transforming the critical vehicle identification problem, which has a significant impact on the safety of autonomous vehicles, into the problem of identifying the importance of key nodes in complex networks. Furthermore, it utilizes a graph neural network method based on an attention mechanism for key node identification in complex networks. By using a graph attention network, the feature information of vehicle nodes is transformed into more easily understood attention coefficients, thus deriving the importance of vehicle nodes to the master node, i.e., the self-vehicle.

[0117] Based on the above method, this invention also utilizes a graph attention mechanism to capture the communication relationships between vehicles to analyze abnormal behavior, and abstracts it into a complex network structure. In graph-based anomaly detection, the discretized state set is transformed into the anomaly probability of vehicle nodes using the feature information of vehicle nodes.

[0118] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. All equivalent methods or modifications that do not depart from the technology of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying key vehicles in a complex traffic network for intelligent vehicle decision planning, characterized in that, The application relates to a method for identifying abnormal nodes in a complex network formed by an autonomous vehicle, and belongs to the field of autonomous driving. The method comprises the following steps: Step 1: dynamically modeling a driving environment by taking the autonomous vehicle as a main node in a complex network theory; Step 2: aggregating feature vectors in an input layer by using a graph neural network based on an attention mechanism, parameterizing driving feature information, and obtaining an attention parameter and node importance by using a multi-head attention mechanism; Step 4: After the learning process of the graph attention network for the nodes and the determination of the abnormal nodes, the vehicle node features are output after normalization The difference of the nodes in the network is obtained through the connection layer of the graph neural network, so as to sort the importance of the nodes, and finally obtain the key vehicle nodes; Step 3: identifying abnormal nodes in a complex network formed by the autonomous vehicle by using a graph attention network; introducing a graph attention mechanism by using a graph embedding method; analyzing abnormal behaviors by capturing communication relationships between vehicles; detecting abnormal behaviors of the vehicles by using the graph attention mechanism; analyzing influences on safety of the autonomous vehicle by using changes in driving feature information of the vehicle nodes; and updating and learning the nodes and confirming the importance of the nodes again when abnormal behaviors occur. The specific implementation of step 1 comprises the following steps: Step 1.1: regarding the autonomous vehicle and other vehicles on a road as complex network nodes based on the complex network theory, regarding the road as an edge to form a complex network model comprising surrounding neighbor vehicle nodes, and generating a vehicle dynamic complex network: Wherein G is a time-varying complex dynamic network, V is a node set in the time-varying complex dynamic network G, B is a set of edges in the time-varying complex dynamic network G, represents a connection between nodes, W is a weight of a node in the time-varying complex dynamic network G, and P is a function of an edge in the complex dynamic network G, representing a relationship between nodes; The driving feature information mainly includes path length between the two vehicles , speed and acceleration ​ Step 1.2: regarding the autonomous vehicle in the complex network as a main node, connecting the main node and neighbor nodes within three orders, collecting driving feature information of the neighbor nodes of the autonomous vehicle, constructing a corresponding three-degree driving feature information feature vector from the driving feature information, and constituting an input layer of a graph neural network to process the feature vector of the complex network node by using the graph neural network; wherein denotes the path length between two nodes, is the node velocity, is the node acceleration; The constructed driving feature information feature vector is as follows: The specific implementation of step 2 comprises the following steps: where n is the number of nodes and F is the number of features per node, is a set of features for the vehicle neighbor nodes; Step 2.2: Map the vehicle master node and neighbor nodes using a graph attention network, which is a weighted matrix Parametrize the features of each vehicle node. To give the nodes sufficient expressive power, the input features are converted into higher-level features, which requires a learnable linear transformation using a weight matrix Parametrize the shared linear transformation and apply it to each node. Then, perform self-attention on each vehicle node, which is a shared attention mechanism: and calculate the attention coefficient, which is: wherein representing a node for a node of importance, is a weight matrix for driving characteristic parameters, is driving characteristic information for the vehicle node; Step 2.1: collecting complex network node feature information by using a graph neural network to form a group of node features, and then calculating a linear combination of the features of the corresponding vehicle nodes as a final output feature of the vehicle nodes; is the final derived node for the node importance of the output feature, is the normalized vehicle attention coefficient; Then, an activation function is applied to convert a nonlinear problem into a linear problem, so that the following formula is obtained: k is the number of attention mechanism heads, k = 3, is the normalized attention coefficient for the 3rd attention mechanism calculation, is the weight matrix of the corresponding input linear transformation; Step 2.3: using a multi-head attention mechanism to judge the importance of a single node feature, establishing three independent attention mechanisms for the three node features, parameterizing the feature information of the three nodes, and obtaining a new node importance coefficient after aggregation, so as to stabilize the self-attention learning process and obtain the following formula by splicing the features to be learned: wherein a non-linear process to apply multi-head attention, and are updated as the final output feature values; Step 2.1 further comprises regenerating a new set of node features at the input layer, with potentially different cardinality Then, there is: wherein is a set of node output features; The multi-head attention mechanism is executed in the graph attention network, and since the weights of the features are different, only connection is not ideal, and the features are averaged according to the proportion of the parameters, as shown in the following formula: wherein is the normalized vehicle attention coefficient, is the importance of other nodes to the main node, is the language index function with is the base, is the logistic regression function; Since the node has a shared attention mechanism, to get the weight of each vehicle node, the node and the result vector of the node are spliced, which prevents the attention value from causing the node to only have a relationship with the neighbor node and ignores the main node. After adding the activation function, the main node cannot be ignored, so the result is obtained: wherein is an activation function in a neural network for handling non-linear problems, and is a feedforward neural network mapping the concatenated vector onto the real numbers; Step 2.2 further comprises normalizing the importance of the nodes, and then the following formula is obtained: wherein representing a connection relationship, for splicing operation; The normalization coefficient formula is also updated as follows: The specific implementation of step 3 comprises the following steps: Step 3.1: During the driving process of the vehicle, the dynamic network of the vehicle is in a state of dynamic balance, but when the vehicle makes a decision, the balance is broken, and the dynamic network formed by the autonomous vehicle as the main node will change, and the behavior of the vehicle in the network is randomly generated. Compared with normal behavior, the structure information of entities connected by random abnormal links often lacks similarity. Based on this assumption, the behavior characteristics of each node in the network can be learned and updated through aggregation, and all the abnormalities of the node feature vector, that is, the abnormalities of the driving feature information, are used to judge the abnormal nodes; Step 3.2: For the nodes of the vehicle network, let the pair of nodes be There is an edge , whose anomaly probability can be expressed as , where , , respectively, represent the neighbor node sets of and ; by using hierarchical attention judgment on the feature vectors of the nodes, the distance, speed and acceleration of the vehicle nodes are hierarchically divided to judge the abnormal situation of the vehicle nodes, the discrete state set L = {non-existent, normal, abnormal} can be converted into a numerical representation, thereby obtaining the anomaly probability of the edge between the nodes; According to the eigenvector, an initialization matrix is derived Knowing the state of the current vehicle node and the relationship state with the main node, the graph attention mechanism is used for processing, as follows: wherein information capture acquisition representing short-term vehicle nodes, dynamic update mechanism of node embedding for graph attention layers; The abnormal probability of the node is calculated by using the function: wherein with is a contrast parameter for the attention layer, is an anomaly probability for the vehicle node, is a function, is a hyperparameter, the control probability is within [0, 1]; Step 3.3: After obtaining the abnormal probability of the node, it is judged whether the abnormal node is a threat to itself. If there is a threat, the driving feature vector of the abnormal node is updated, the behavior characteristics of the vehicle node are updated by using the graph attention network layer, and the new feature vector is obtained. Then, the importance of the node to the main node is calculated by learning and calculation; If there is no threat, the original graph attention network is still calculated, and the network point is not embedded.

2. The method of claim 1, wherein, The method for judging whether the abnormal node is a threat is as follows: when the abnormal node is a threat, when the abnormal node is not a threat.

3. An apparatus for a vehicle, characterized by comprising: The vehicle equipment can perform the key vehicle identification method of any one of claims 1-2.

Images