A method for detecting anomalies of a UAV based on signal strength and a graph neural network
By constructing an undirected radius graph and graph neural network, combined with a multi-head attention model and a multi-layer pooling network, the dynamic adaptability and robustness of anomaly detection in multi-UAV systems are solved, achieving efficient and low-cost anomaly detection suitable for variable environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV OF TECH
- Filing Date
- 2025-08-22
- Publication Date
- 2026-05-15
AI Technical Summary
Existing anomaly detection methods for multi-UAV systems are difficult to adapt to dynamic topology and mission heterogeneity, have high computational overhead and poor robustness, and lack dynamic modeling and discrimination mechanisms for RSSI signals in multi-UAV cooperative missions.
Based on signal strength and graph neural networks, this method generates reward maps, Gaussian target maps, and obstacle maps by constructing an undirected radius graph. It uses a multi-head attention model and a multi-layer pooling graph neural network for feature embedding and anomaly scoring, and combines a loss function for training to achieve real-time detection of UAV anomalies.
It improves the accuracy and robustness of identifying abnormal drones in multi-drone systems, reduces deployment costs, and can adapt to dynamic topology changes, making it suitable for urban airspace surveillance, intelligent patrols, and complex drone swarm collaboration scenarios.
Smart Images
Figure CN121117874B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent safety monitoring technology for multi-UAV systems, specifically relating to a method for detecting UAV anomalies based on signal strength and graph neural networks. Background Technology
[0002] With the widespread deployment of multi-Drone Systems (MDS) in highly dynamic missions such as military patrols, disaster relief, and smart logistics, the security of these systems is becoming increasingly serious. If unauthorized or malfunctioning drones infiltrate a normal swarm, it can lead to mission disruptions, resource waste, or even severe security incidents.
[0003] Traditional anomaly detection schemes rely on centralized processing, fixed rule-based judgment, or global communication mechanisms, which are difficult to adapt to dynamic topologies and task heterogeneity, and suffer from problems such as high computational overhead, poor robustness, and insufficient scalability.
[0004] Received Signal Strength Indicator (RSSI), as an easily accessible and low-cost wireless communication feature, is widely used in network ranging and spatial perception. However, existing methods lack dynamic modeling and discrimination mechanisms for its abnormal patterns in multi-UAV cooperative tasks. Meanwhile, Graph Neural Networks (GNNs), especially Graph Attention Networks (GATs), have shown powerful capabilities in processing structured data. However, most current models rely on static graph structures and fixed neighborhood perception, making it difficult to effectively address the challenges posed by communication noise, spatial dynamics, and behavioral heterogeneity in multi-UAV systems. Therefore, there is an urgent need for an anomaly detection mechanism that is oriented towards multi-source perception and possesses online adaptability. Summary of the Invention
[0005] The purpose of this invention is to solve the problems of the prior art and to propose an anomaly detection method for unmanned aerial vehicles based on signal strength and graph neural networks.
[0006] To achieve the above objectives, the technical solution provided by this invention is as follows:
[0007] A method for detecting unmanned aerial vehicle (UAV) anomalies based on signal strength and graph neural networks includes:
[0008] Constructing an undirected radius graph in real time using drones as nodes;
[0009] Real-time construction of reward graph, Gaussian target graph, obstacle graph and spatial preference graph for each node;
[0010] The reward map, Gaussian target map, obstacle map, and spatial preference map are encoded to obtain the fused feature embedding for each node;
[0011] Construct a multi-head attention model, embed the fused features of each node into the multi-head attention model, and obtain the final node embedding of each node;
[0012] Construct a multi-level pooled graph neural network model, generate a node feature matrix based on the final node embedding of each node, and input the adjacency matrix and node feature matrix of the undirected radius graph into the multi-level pooled graph neural network model to obtain the global pooled aggregate embedding.
[0013] The loss function is calculated based on global pooling aggregation embedding and then backpropagated.
[0014] Calculate the anomaly score for each node. When the anomaly score is greater than a preset threshold, the current node is judged as abnormal. When the anomaly score is less than or equal to the preset threshold, the current node is judged as normal.
[0015] Furthermore, the construction of the reward graph, Gaussian target graph, obstacle graph, and spatial preference graph for each node includes:
[0016] Construct the reward graph for each node, expressed by the formula:
[0017]
[0018] in, For the first Each node travels from its current point to points within the sensing area. Path cost, For the first The speed of each node Indicates the first Each node is from the current point to the point... distance, It is a constant;
[0019] Construct the Gaussian target map for each node, expressed by the formula:
[0020]
[0021] in, This represents a point within the observable region of node 𝑖. At this point, for sensing neighbor nodes Spatial perception response intensity, Represents a node Perceiving neighbor nodes, Represents the sensing neighbor node The set, Represents a node Points within the observable area, express The set, Represents a node At point Location and sensing neighbor nodes The distance between them Point With sensing neighbor nodes Euclidean distance component in the x-axis direction, Point With sensing neighbor nodes Euclidean distance component in the y-axis direction Represents a node The standard deviation of the sensing range in the x-axis direction Represents a node The standard deviation of the sensing range in the y-axis direction;
[0022] The point cloud information collected by each drone is acquired and discretized into a two-dimensional raster image, which is the obstacle map.
[0023] Multiplying the reward map and the Gaussian target map yields the spatial preference map.
[0024] Furthermore, the encoding of the reward map, Gaussian target map, obstacle map, and spatial preference map to obtain the fused feature embedding for each node includes:
[0025] The reward map, Gaussian target map, obstacle map, and spatial preference map are respectively input into a cascaded structure including a multilayer perceptron, max pooling, and convolution modules for encoding, and the graph embeddings of the reward map, Gaussian target map, obstacle map, and spatial preference map are output respectively.
[0026] The various graph embeddings are fused to obtain the fused feature embedding for each node.
[0027] Furthermore, the step of embedding the fused features of each node into the multi-head attention model to obtain the final node embedding for each node includes:
[0028] The fused features of each node are embedded into the input multi-head attention model, and the attention weights assigned by each node to each neighbor node in each attention head are calculated, as expressed by the following formula:
[0029]
[0030] in, Indicates the first Nodes in each attention head For neighboring nodes The attention weight assigned This represents the LeakyReLU activation function. It is the first Feature transformation matrix of each attention head Indicates the first The attention vector to be learned from each attention head. Indicates the first The transpose of the attention vector to be learned for each attention head. Represents a node The fusion feature embedding, Representing neighboring nodes The fusion feature embedding, Representing neighboring nodes The fusion feature embedding, This represents a vector concatenation operation. Represents a node The neighborhood group, Represents a node With nodes The difference in signal strength, Represents a node With nodes The difference in signal strength;
[0031] All nodes output by the attention head are embedded and concatenated, as shown by the following formula:
[0032]
[0033] in, Represents the spliced nodes Node embedding, Indicates the number of attention heads;
[0034] The node embeddings of each spliced node are input to the residual connection layer and the graph normalization layer, and the final node embeddings of each node are output.
[0035] Furthermore, the step of inputting the adjacency matrix and node feature matrix of the undirected radius graph into a multi-layer pooling graph neural network model to obtain a global pooling aggregation embedding includes:
[0036] The adjacency matrix and node feature matrix of the undirected radius graph are input into a multi-layer pooled graph neural network model to obtain the assignment probability matrix.
[0037] The formula for calculating the distribution probability matrix is as follows:
[0038]
[0039] in, Indicates the first The assignment probability matrix of a layered pooling graph neural network. Indicates the first Layered pooling graph neural networks, Indicates the first The node feature matrix as input to a layered pooling graph neural network. Indicates the first The adjacency matrix of the input to the layer pooling graph neural network. This represents the softmax classifier;
[0040] The adjacency matrix and node feature matrix are iteratively updated based on the allocation probability matrix. The node feature matrix of the final layer is then subjected to global average pooling to obtain a global pooling aggregate embedding.
[0041] Furthermore, the iterative update of the adjacency matrix and node feature matrix based on the allocation probability matrix, and the final layer's node feature matrix undergoing global average pooling to obtain a globally pooled aggregated embedding, includes:
[0042] The adjacency matrix and node feature matrix are iteratively updated based on the allocation probability matrix, as expressed by the formula:
[0043]
[0044]
[0045] in, Indicates the first The node feature matrix as input to a layered pooling graph neural network. Indicates the first The adjacency matrix of the input to the layer pooling graph neural network. Indicates transpose;
[0046] Multiplying the node feature matrix input to the final layer with the transpose of the assignment probability matrix obtained from the final layer yields the global pooling aggregate embedding.
[0047] Furthermore, a joint loss function is used to train a multi-layer pooled graph neural network model, which is expressed by the formula:
[0048]
[0049] in, Denotes the joint loss function. This represents the loss for classifying outlier nodes. This represents the cluster consistency loss. This represents the allocation entropy regularization loss. and This represents the regularization weight hyperparameter;
[0050]
[0051] in, This represents the number of nodes in an undirected radius graph. Represents a node The true value is 1 for abnormal and 0 for normal. Indicates the mean deviation of signal strength The predicted value calculated using the Sigmoid activation function. This represents the Sigmoid activation function;
[0052]
[0053] in, This represents the original node feature matrix. This represents the allocation probability matrix obtained in the final layer. This represents the node feature matrix of the final layer input. For global pooling aggregation embedding, It is the Frobenius norm;
[0054]
[0055] in, Indicates the number of cluster center nodes. Represents the nodes in the allocation probability matrix Assigned to the The probability of each cluster center.
[0056] Furthermore, the formula for calculating the anomaly score for each node is as follows:
[0057]
[0058]
[0059]
[0060] in, This represents the mean signal strength of all nodes. This represents the number of nodes in an undirected radius graph. Represents a node signal strength, Represents a node The mean deviation of signal strength Represents a node Abnormal scores, Represents the nodes in each attention head For neighboring nodes Attention weights assigned The nodes obtained after average aggregation For neighboring nodes The final attention weight assigned, Represents a node The set of neighbors.
[0061] This invention introduces a graph attention network (GAT) to compute attention weights between nodes, effectively enhancing the ability to express and model RSSI signal deviation features; it constructs a hierarchical graph representation by introducing a differentiable graph pooling mechanism, realizing multi-scale structure modeling and feature aggregation; and it combines attention weights and RSSI mean deviation for anomaly scoring, improving the accuracy and robustness of identifying abnormal drones.
[0062] Furthermore, this invention possesses technical advantages such as low deployment cost and adaptability to dynamic topologies: Firstly, the RSSI feature relied upon is a signal strength index that can be directly obtained from the communication protocol layer, and the entire detection process does not require additional sensor deployment or high computing resource support, significantly reducing deployment costs; Secondly, this method constructs an undirected graph structure constrained by the communication radius in real time, enabling the graph neural network to dynamically adjust the graph structure when the topology changes with the task scenario (such as node addition or communication interruption), adapting to the heterogeneity and dynamic changes in the communication topology of the UAV system, thereby improving the generalization ability of the detection system in a variable environment.
[0063] In summary, this invention significantly improves the ability to identify communication anomalies in multi-UAV systems while maintaining low deployment costs, and is applicable to typical scenarios such as urban airspace supervision, intelligent patrol, security missions, and complex UAV swarm collaboration. Attached Figure Description
[0064] Figure 1 This is a flowchart of an anomaly detection method for unmanned aerial vehicles (UAVs) based on signal strength and graph neural networks according to the present invention. Detailed Implementation
[0065] 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.
[0066] like Figure 1 As shown, the present invention provides a method for detecting anomalies in unmanned aerial vehicles (UAVs) based on signal strength and graph neural networks, comprising the following steps:
[0067] S1: Construct a radius graph structure with drones as nodes, and determine edge connections based on communication distance.
[0068] Model the drone swarm as an undirected radius graph. ,in, Represents a set of drone nodes. This represents the set of communication edges between any two nodes.
[0069]
[0070] If any two nodes The Euclidean distance satisfies Then construct the edge ,in, This is the communication radius threshold.
[0071] S2: Generate a four-channel feature map through a heterogeneous sensing mechanism and extract a high-dimensional embedding representation.
[0072] S2.1: Constructing a Reward Map: This represents the path cost from the drone's current location to each point in the perception area, calculated as follows:
[0073]
[0074] in, For the first Each node is located from its current point to points within the observable region. Path cost, For the first The speed of each node Indicates the first Each node is from the current point to the point... distance, To prevent division by zero of small constants.
[0075] S2.2: Constructing a Gaussian Map: This map represents the spatial heat of the target region and describes the spatial heat distribution of each target within the current task area. The calculation formula is:
[0076]
[0077] in, This represents a point within the observable region of node 𝑖. At this point, for sensing neighbor nodes Spatial perception response intensity, Represents a node Perceiving neighbor nodes, express The set, Represents a node Points within the observable area, express The set, Represents a node At point With sensing neighbor nodes The distance between them Point With sensing neighbor nodes Euclidean distance component in the x-axis direction, Point With sensing neighbor nodes Euclidean distance component in the y-axis direction Represents a node The standard deviation of the sensing range in the x-axis direction Represents a node The standard deviation of the sensing range in the y-axis direction.
[0078] S2.3: Constructing the Obstacle Map: Discretize the point cloud information collected by the sensing devices on the UAV into a two-dimensional grid map, where 1 represents the presence of obstacles and 0 represents the absence of obstacles. This Boolean matrix is the obstacle map, indicating the distribution of obstacles within the area.
[0079] S2.4: Construct a spatial preference map: This is obtained by multiplying the reward map and the Gaussian target map, which strengthens the preference for the target's neighborhood.
[0080] S2.5: Input the reward map, Gaussian target map, obstacle map and spatial preference map into a cascaded structure including multilayer perceptron, max pooling and convolution modules for encoding, and output the graph embeddings of the reward map, Gaussian target map, obstacle map and spatial preference map respectively; fuse the graph embeddings to obtain the fused feature embedding of each node.
[0081] The aforementioned multi-channel image data (reward map, Gaussian target map, obstacle map, and spatial preference map) are encoded into feature embedding vectors using a multilayer perceptron (MLP), max pooling, and convolutional modules. ,Right now:
[0082]
[0083] in, Indicates multi-channel features, This refers to the maneuverability attributes of the drone (flight speed, acceleration, current pitch, roll, yaw angle, etc.). This represents the perceptual feature extraction function, which consists of a multilayer perceptron (MLP), max pooling, convolutional modules, etc.
[0084] S3: Introduce the graph attention mechanism (GAT) to perform weighted modeling of neighboring nodes, highlighting signal deviation characteristics.
[0085] S3.1: Attention Coefficient Calculation. Let the nodes in the radius graph be... with neighboring nodes The input features are and Suppose that a total of H attention heads are used, the first one... The calculation of each attention point is as follows:
[0086]
[0087] in, Indicates the first Nodes in each attention head For neighboring nodes The attention weight assigned This represents the LeakyReLU activation function. It is the first Feature transformation matrix of each attention head Indicates the first The attention vector to be learned from each attention head. Indicates the first The transpose of the attention vector to be learned for each attention head. Represents a node The fusion feature embedding, Representing neighboring nodes The fusion feature embedding, Representing neighboring nodes The fusion feature embedding, This represents a vector concatenation operation. Represents nodes in the radius graph The set of neighbors; Represents a node With nodes The difference in signal strength, Represents a node With nodes The difference in signal strength.
[0088] S3.2: Multi-head attention aggregation. The results from multiple heads are fused by splicing.
[0089]
[0090] in, Represents the spliced nodes Node embedding, This indicates the number of attention heads.
[0091] S3.3: Residual connection and normalization mechanism.
[0092] To enhance gradient stability and multi-scale feature fusion capabilities, residual connectivity and layer normalization operations are introduced:
[0093]
[0094] in, Represents the merged node The final node embedding is used as the input vector for the multi-layer pooled graph neural network model. This indicates a normalization operation.
[0095] S4: Apply a learnable graph pooling mechanism to improve graph modeling quality and detection accuracy.
[0096] S4.1: Pooling Mechanism Construction. In an undirected radius graph, Represents the adjacency matrix. Represents the node feature matrix, This represents the number of nodes in an undirected radius graph. A node assignment probability matrix is generated through a learnable clustering mapping mechanism. ,Will The original nodes are mapped as follows: Cluster center nodes. The allocation probability matrix is generated using a pooling graph neural network in a multi-layer pooling graph neural network model:
[0097]
[0098] in, Indicates the first The assignment probability matrix of a layered pooling graph neural network. Indicates the first Layered pooling graph neural networks are used to generate node-cluster assignment relationships (which can be single-layer or multi-layer GNNs, such as GCN, GraphSAGE, GAT, etc.). Indicates the first The node feature matrix as input to a layered pooling graph neural network. Indicates the first The adjacency matrix of the input to the layer pooling graph neural network. This represents a row-normalized softmax classifier used to map the original assigned scores to a probability distribution. When When it is 1, , .
[0099] S4.2: Graph Embedding and Structure Update.
[0100] The node feature matrix and adjacency matrix after pooling are updated as follows:
[0101]
[0102]
[0103] in, Indicates the first The node feature matrix as input to a layered pooling graph neural network. Indicates the first The adjacency matrix of the input to the layer pooling graph neural network.
[0104] S4.3 Iteration and Global Pooling.
[0105] Through multi-level iterative updates ( , The number of nodes in the graph is reduced layer by layer while retaining key structural information until a highly compressed graph representation is obtained. The node feature matrix input to the final layer is multiplied by the transpose of the assignment probability matrix obtained from the final layer to obtain the global pooling aggregate embedding.
[0106] Multi-level pooled graph neural network models serve to perform hierarchical feature extraction and structural compression, making node embeddings more discriminative before entering anomaly scoring. The loss function is calculated based on global pooling aggregated embeddings and backpropagated, enhancing the performance of S3.3. It is more representative and enhances the global structural expressiveness and robustness of node embedding.
[0107] S4.4: Optimize the multi-head attention model and the multi-layer pooling graph neural network model.
[0108] To improve training performance, a joint loss function is introduced:
[0109]
[0110] in, Denotes the joint loss function. This represents the loss for classifying outlier nodes; This represents the clustering consistency loss, used to enhance the semantic consistency of nodes within the same cluster; This represents the allocation entropy regularization loss, used to prevent the allocation matrix from degenerating into a high-entropy random distribution; and This represents the regularization weight hyperparameter.
[0111] Anomaly Node Classification Loss To supervise the identification of abnormal nodes, a binary classification cross-entropy loss is used, in the following form:
[0112]
[0113] in, Represents a node The true label (1 for abnormal, 0 for normal); Indicates the mean deviation of signal strength The probability of a prediction being an anomaly, calculated using the Sigmoid activation function. This represents the Sigmoid activation function.
[0114] Cluster Consistency Loss The node representation used after constraint graph pooling should retain the original structural features to prevent feature drift, and its definition is as follows:
[0115]
[0116] in, This represents the original node feature matrix. This represents the allocation probability matrix obtained in the final layer. This represents the node feature matrix of the final layer input. For global pooling aggregation embedding, It is the Frobenius norm.
[0117] Allocation entropy regularization loss Used to prevent the allocation matrix from degenerating into a high-entropy uniform distribution or a completely random distribution, thereby improving the expressiveness of the cluster structure, it is defined as:
[0118]
[0119] in, Indicates the number of cluster center nodes. Represents the nodes in the allocation probability matrix Assigned to the The probability of each cluster center. The allocation entropy regularization loss encourages the allocation matrix to have sparsity and distinctiveness, avoiding the invalid structure where all nodes are evenly distributed among all clusters.
[0120] S5: Calculate anomaly scores based on attention weights and signal strength deviations to achieve efficient identification.
[0121] Each node of The mean deviation is defined as follows:
[0122]
[0123] in, Represents a node The mean deviation of signal strength Represents a node signal strength, The average signal strength of all nodes in the network is given by the following formula:
[0124]
[0125] in, Representative set The number of nodes within the system, i.e., the number of drones.
[0126] For each node Define anomaly scoring The sum of the attention weights of its neighboring nodes and the signal strength deviation:
[0127]
[0128] in, Represents a node Abnormal scores, express The nodes obtained after average aggregation For neighboring nodes The final attention weight assigned.
[0129] like If the drone is identified as an abnormal target, then the drone is considered an abnormal target. This is the threshold for abnormal scoring in the preliminary review.
[0130] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.
Claims
1. A method for detecting anomalies in unmanned aerial vehicles (UAVs) based on signal strength and graph neural networks, characterized in that, The UAV anomaly detection method based on signal strength and graph neural networks includes: Constructing an undirected radius graph in real time using drones as nodes; Real-time construction of reward graph, Gaussian target graph, obstacle graph and spatial preference graph for each node; The reward map, Gaussian target map, obstacle map, and spatial preference map are encoded to obtain the fused feature embedding for each node; Construct a multi-head attention model, embed the fused features of each node into the multi-head attention model, and obtain the final node embedding of each node; Construct a multi-level pooled graph neural network model, generate a node feature matrix based on the final node embedding of each node, and input the adjacency matrix and node feature matrix of the undirected radius graph into the multi-level pooled graph neural network model to obtain the global pooled aggregate embedding. The loss function is calculated based on global pooling aggregation embedding and then backpropagated. Calculate the anomaly score for each node. When the anomaly score is greater than a preset threshold, the current node is judged as abnormal. When the anomaly score is less than or equal to the preset threshold, the current node is judged as normal. in, The construction of the reward graph, Gaussian target graph, obstacle graph, and spatial preference graph for each node includes: Construct the reward graph for each node, expressed by the formula: in, For the first Each node travels from its current point to points within the sensing area. Path cost, For the first The speed of each node Indicates the first Each node is from the current point to the point... distance, It is a constant; Construct the Gaussian target map for each node, expressed by the formula: in, This represents a point within the observable region of node 𝑖. At this point, for sensing neighbor nodes Spatial perception response intensity, Represents a node Perceiving neighbor nodes, Represents the sensing neighbor node The set, Represents a node Points within the observable area, express The set, Represents a node At point Location and sensing neighbor nodes The distance between them Point With sensing neighbor nodes Euclidean distance component in the x-axis direction, Point With sensing neighbor nodes Euclidean distance component in the y-axis direction Represents a node The standard deviation of the sensing range in the x-axis direction Represents a node The standard deviation of the sensing range in the y-axis direction; The point cloud information collected by each drone is acquired and discretized into a two-dimensional raster image, which is the obstacle map. Multiplying the reward map and the Gaussian target map yields the spatial preference map.
2. The UAV anomaly detection method based on signal strength and graph neural network according to claim 1, characterized in that, The encoding of the reward map, Gaussian target map, obstacle map, and spatial preference map to obtain the fused feature embedding for each node includes: The reward map, Gaussian target map, obstacle map, and spatial preference map are respectively input into a cascaded structure including a multilayer perceptron, max pooling, and convolution modules for encoding, and the graph embeddings of the reward map, Gaussian target map, obstacle map, and spatial preference map are output respectively. The various graph embeddings are fused to obtain the fused feature embedding for each node.
3. The UAV anomaly detection method based on signal strength and graph neural network according to claim 1, characterized in that, The step of embedding the fused features of each node into the multi-head attention model to obtain the final node embedding for each node includes: The fused features of each node are embedded into the input multi-head attention model, and the attention weights assigned by each node to each neighbor node in each attention head are calculated, as expressed by the following formula: in, Indicates the first Nodes in each attention head For neighboring nodes The attention weight assigned This represents the LeakyReLU activation function. It is the first Feature transformation matrix of each attention head Indicates the first The attention vector to be learned from each attention head. Indicates the first The transpose of the attention vector to be learned for each attention head. Represents a node The fusion feature embedding, Representing neighboring nodes The fusion feature embedding, Representing neighboring nodes The fusion feature embedding, This represents a vector concatenation operation. Represents a node The neighborhood group, Represents a node With nodes The difference in signal strength, Represents a node With nodes The difference in signal strength; All nodes output by the attention head are embedded and concatenated, as shown by the following formula: in, Represents the spliced nodes Node embedding, Indicates the number of attention heads; The node embeddings of each spliced node are input to the residual connection layer and the graph normalization layer, and the final node embeddings of each node are output.
4. The UAV anomaly detection method based on signal strength and graph neural network according to claim 1, characterized in that, The step of inputting the adjacency matrix and node feature matrix of the undirected radius graph into a multi-layer pooled graph neural network model to obtain a global pooled aggregate embedding includes: The adjacency matrix and node feature matrix of the undirected radius graph are input into a multi-layer pooled graph neural network model to obtain the assignment probability matrix. The formula for calculating the distribution probability matrix is as follows: in, Indicates the first The assignment probability matrix of a layered pooling graph neural network. Indicates the first Layered pooling graph neural networks, Indicates the first The node feature matrix as input to a layered pooling graph neural network. Indicates the first The adjacency matrix of the input to the layer pooling graph neural network. This represents the softmax classifier; The adjacency matrix and node feature matrix are iteratively updated based on the allocation probability matrix. The node feature matrix of the final layer is then subjected to global average pooling to obtain a global pooling aggregate embedding.
5. The UAV anomaly detection method based on signal strength and graph neural network according to claim 4, characterized in that, The process of iteratively updating the adjacency matrix and node feature matrix based on the allocation probability matrix, and finally obtaining the global pooled aggregate embedding by global average pooling of the node feature matrix, includes: The adjacency matrix and node feature matrix are iteratively updated based on the allocation probability matrix, as expressed by the formula: in, Indicates the first The node feature matrix as input to a layered pooling graph neural network. Indicates the first The adjacency matrix of the input to the layer pooling graph neural network. Indicates transpose; Multiplying the node feature matrix input to the final layer with the transpose of the assignment probability matrix obtained from the final layer yields the global pooling aggregate embedding.
6. The UAV anomaly detection method based on signal strength and graph neural network according to claim 1, characterized in that, A multi-layer pooled graph neural network model is trained using a joint loss function, which is expressed by the formula: in, Denotes the joint loss function. This represents the loss for classifying outlier nodes. This represents the cluster consistency loss. This represents the allocation entropy regularization loss. and This represents the regularization weight hyperparameter; in, This represents the number of nodes in an undirected radius graph. Represents a node The true value is 1 for abnormal and 0 for normal. Indicates the mean deviation of signal strength The predicted value calculated using the Sigmoid activation function. This represents the Sigmoid activation function; in, This represents the original node feature matrix. This represents the allocation probability matrix obtained in the final layer. This represents the node feature matrix of the final layer input. For global pooling aggregation embedding, It is the Frobenius norm; in, Indicates the number of cluster center nodes. Represents the nodes in the allocation probability matrix Assigned to the The probability of each cluster center.
7. The UAV anomaly detection method based on signal strength and graph neural network according to claim 1, characterized in that, The formula for calculating the anomaly score for each node is as follows: in, This represents the mean signal strength of all nodes. This represents the number of nodes in an undirected radius graph. Represents a node signal strength, Represents a node The mean deviation of signal strength Represents a node Abnormal scores, Represents the nodes in each attention head For neighboring nodes Attention weights assigned The nodes obtained after average aggregation For neighboring nodes The final attention weight assigned, Represents a node The set of neighbors.