Unmanned aerial vehicle cluster communication network full-end reliability prediction method based on gnn and network signature

Abstract

The application relates to a kind of GNN and network signature-based UAV cluster communication network full end reliability prediction method, belong to communication network technical field, adopt two-stage prediction: first stage, determine the topological structure characteristic variable of UAV cluster communication network, based on simulation generation training data set, design, train and optimize graph neural network model, realize the prediction of network signature vector, signature only relies on network structure characteristics, does not rely on link reliability parameter;Second stage, based on signature vector, combined with link reliability parameter, use the full end reliability certainty mixed expression based on signature, calculate the full end reliability of network under the link failure scene.The application reduces the sensitivity of model to the change of reliability parameter, improves the generalization ability and applicability of prediction method under different network scale and reliability conditions.

Description

Technical Field

[0001] This invention relates to a method for predicting the reliability of all ends of a UAV swarm communication network based on GNN and network signature, belonging to the field of communication network technology. Background Technology

[0002] Drone swarms have been widely used in key areas such as agricultural and forestry plant protection, emergency rescue, and energy inspection, becoming an important technical means to ensure mission efficiency and operational safety. Especially in complex scenarios such as post-disaster communication relay and large-scale inspections, drones, with their advantages of mobility, flexibility, and rapid deployment, can effectively compensate for the problems of damaged or insufficient coverage of traditional communication infrastructure, making their application value particularly prominent.

[0003] Modern UAV swarm systems typically employ a mesh-based self-organizing network communication architecture, consisting of a ground control station and multiple aerial UAV nodes forming a dynamic communication network. In this architecture, some UAV nodes can directly establish communication links with the ground control station, serving as information access nodes between the swarm and the ground station. The remaining UAV nodes achieve collaborative communication within the swarm and between the swarm and the ground station through multi-hop relay transmission with access nodes or other UAV nodes, thus simultaneously meeting the needs of long-distance communication and close-range collaboration. For such complex and scalable communication networks, end-to-end reliability (i.e., the probability that all nodes in the network remain connected) is a core indicator for measuring the system's communication capabilities and operational security, and also an important basis for network design, deployment, and optimization.

[0004] In the end-to-end reliability analysis of UAV swarm communication networks, the causes of network failures are diverse, including node-level failures such as node hardware malfunctions, as well as link-level failures such as communication link interruptions. Engineering practice shows that in UAV swarm systems with a certain degree of redundant node configuration, node failures can be mitigated to some extent through backup or replacement mechanisms. However, communication links are susceptible to various factors such as electromagnetic interference, changes in flight distance, and environmental obstruction, and their failure probability is usually significantly higher than that of node failures, becoming the main factor restricting the overall reliability of the network.

[0005] However, the analysis and evaluation of end-to-end reliability in complex communication networks dominated by link failures still faces many challenges. First, accurate calculation of end-to-end reliability is an NP-hard problem; the computational complexity increases exponentially with network size, making it difficult to apply directly to practical engineering scenarios. Second, while commonly used methods such as Monte Carlo simulation can obtain high-precision reliability assessment results, ensuring statistical accuracy usually requires extensive random sampling, resulting in long computation times and high resource consumption, making it difficult to meet the real-time analysis needs of UAV swarm communication networks in dynamic link environments.

[0006] In recent years, some studies have attempted to introduce deep learning methods to predict network reliability, directly outputting reliability assessment results by learning network topology features and related parameters. However, such methods still have certain limitations in engineering applications: on the one hand, some methods are only trained for specific link reliability parameters, requiring model retraining when link reliability conditions change, resulting in high application costs; on the other hand, while using link reliability parameters and network structure features as model input can achieve good prediction results under specific conditions, the model is highly dependent on the distribution of link reliability parameters, limiting its generalization ability across different network sizes or link conditions. Furthermore, existing deep learning methods are mostly based on traditional deep neural network frameworks, and their adaptability and expressive power for network systems with complex topological structures still have room for improvement.

[0007] In summary, existing computational methods and prediction techniques are insufficient in terms of accuracy, efficiency, and generalization ability to simultaneously meet the requirements of full-end reliability assessment of UAV swarm communication networks under complex topology and dynamic link conditions. There is an urgent need for a full-end reliability prediction method that can balance computational efficiency, prediction accuracy, and engineering applicability. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention discloses a method for predicting the overall reliability of UAV swarm communication networks based on GNN and network signatures. The specific technical solution is as follows:

[0009] A method for predicting the end-to-end reliability of UAV swarm communication networks based on GNN and network signatures includes the following steps:

[0010] Step 1, First Stage: Predict the signature vector based on GNN, as follows:

[0011] Step 1.1: Abstract the UAV swarm communication network as an undirected graph G(V,E) and construct the UAV swarm network system;

[0012] Step 1.2: Determine the network structure characteristic variables. Select node characteristics and network topology as the structural characteristic variables of the UAV swarm network system. The node characteristics include node degree and node clustering coefficient, and the network topology is selected as the adjacency matrix.

[0013] Step 1.3: Generate a sample network dataset for training the GNN model using simulation methods;

[0014] Step 1.4: Construct GNN. The goal of constructing GNN is to design a deep learning proxy model that takes network node features and adjacency matrix as input and outputs network signature vector.

[0015] Step 1.5: Train and optimize the GNN model;

[0016] Step 1.6: Predict the signature vector S of the trained GNN network;

[0017] Step 2, Second Stage: After predicting the network signature vector S in the first stage, substitute the signature vector S into the hybrid expression formula for network end-to-end reliability based on signatures to calculate the reliability of any link. Network end-to-end reliability The calculation formula is as follows:

[0018] ;

[0019] in, The signature vector obtained from the first stage GNN prediction is the first... One portion, , This is the value for link reliability. The dimension of the signature vector is the total number of edges in the network. j acts as a traversal index in the formula, representing the actual physical meaning of the number of edges in the network that are working normally.

[0020] Furthermore, the specific implementation process of step 1.1 is as follows:

[0021] Each vertex in the vertex set V of the undirected graph G(V,E) corresponds to a drone or a ground base station in the drone swarm network system described in step 1.1. An undirected edge is established in the undirected graph G(V,E) between two vertices if and only if there is an actual communication connection between the drones or ground base stations corresponding to the two vertices that enables data transmission. The undirected edge carries a weight value, which specifically represents the probability of successfully establishing a connection for the corresponding communication link, i.e., the link reliability value described in step 2. , In this way, the undirected graph G(V,E) corresponding to the UAV swarm network system is constructed.

[0022] Furthermore, the specific implementation process of step 1.3 is as follows:

[0023] The number of nodes generated using the Python library is [number]. The number of links is arrive For any integer fully connected network, N sets of fully connected networks with different topological parameters are generated, where N is the preset sample size, serving as the sample network dataset for training the GNN model. For each sample network, the structural feature variables described in step 1.2 are extracted and integrated into the feature data of the sample network. Considering the actual fault scenarios of the UAV swarm communication network, link failure is set as the only cause of network malfunction, that is, assuming that all nodes are completely reliable, only the on / off state of the communication link affects the network connectivity, and each link failure event is independent and identically distributed with equal failure probability. The Monte Carlo simulation algorithm is used to calculate the signature vector of each sample network. For each sample network, the random simulation of the on / off state of the link is repeated M times, where M is the preset number of simulations. The connectivity characteristics of the network under each simulation are statistically analyzed, and the network signature vector corresponding to the sample network is obtained based on the statistical results. This signature vector is used as the label of the sample network in the sample dataset.

[0024] Furthermore, the GNN full-process architecture design in step 1.4 is as follows:

[0025] Design 1), Input Layer, including:

[0026] Node feature matrix X: The node feature matrix X has a dimension of [number of nodes, 2], where each node contains a 2-dimensional feature of node degree and node clustering coefficient;

[0027] Edge index: The edge index is obtained by transforming the adjacency matrix and is used to represent the topological connection relationship of the graph. Its format is [2, number of links];

[0028] Design 2), Graph Convolutional Layer GCNConv, is used for core feature extraction, including:

[0029] The first convolutional layer GCNConv1 in the first figure performs neighborhood feature aggregation and linear transformation on the node feature matrix X of the input layer. The input → output dimension is: 2 → hidden layer 1 dimension.

[0030] The second image shows the convolutional layer GCNConv2, which further extracts deep features based on the output features of GCNConv1. The input → output dimension is: hidden layer 1 dimension → hidden layer 2 dimension.

[0031] The mathematical expression for the graph convolutional layer GCNConv is as follows:

[0032] First convolutional layer: ;

[0033] Second convolutional layer: ;

[0034] in, Let A be an adjacency matrix with self-loops. W1 is the identity matrix; W2 is the trainable learning matrix of the first convolutional layer and W1 is the trainable learning matrix of the second convolutional layer. The degree matrix is ​​a diagonal matrix with self-loops: the diagonal elements are the degree of each node in the network, containing self-loops, and the remaining elements are 0; X is the node feature matrix, in the format [number of nodes × number of node features], H(1) represents the output feature matrix of the first convolutional layer GCNConv1, and H(2) represents the output feature matrix of the second convolutional layer GCNConv2. GCNConv, representing graph convolutional layers, introduces non-linear transformation activation functions, enabling the model to fit complex graph structure features rather than simply performing linear aggregation of neighborhood features.

[0035] Design 3), graph pooling layer, the graph pooling layer is used to compress global information, and adopts global average pooling operation: the feature vector output by the graph convolutional layer H(2) of all nodes in the same network is calculated along the node dimension to extract global graph-level features from local node features, so as to obtain global graph-level feature vectors as input to the next layer;

[0036] Design 4), output layer, including:

[0037] Fully connected layer fc: The input dimension is the global graph-level feature dimension output by the graph pooling layer, and the output dimension is the dimension of the network signature vector S;

[0038] Softmax normalization: Apply the Softmax function to the output of the fully connected layer fc to obtain a probability vector, which is the network signature vector S.

[0039] The expression for the Softmax function is as follows:

[0040] ;

[0041] in, Indicates the first Input values The output after performing the Softmax transformation, where j is the index parameter used to iterate through the elements at each position in the signature vector, and e... xi This indicates that each element of the signature vector is exponentialized;

[0042] To enable GNNs to be adapted for training on sample networks with different topologies, for A network with 10 nodes sets the input and final output dimensions of the GNN sample labels to a uniform value. ,Right now The maximum number of links in a node network, if the actual signature dimension of the sample network is less than... If the missing dimension is filled with 0, then the signature tag value will be filled with 0.

[0043] Furthermore, the GNN end-to-end architecture also includes the following design:

[0044] Design a): After the graph pooling layer and before the output layer, add a Dropout layer; the Dropout layer randomly shuts down some neurons at a dropout_rate, where the dropout_rate ranges from 0.1 to 0.5, to suppress model overfitting and improve the generalization ability of GNN to UAV swarm communication networks with different topologies.

[0045] Design b): Activation function selection: After each layer operation of the first convolutional layer GCNConv1 and the second convolutional layer GCNConv2, a ReLU activation function is introduced to achieve layer-by-layer nonlinear transformation, avoid gradient vanishing, and enhance the feature representation capability of the model; the expression of the ReLU function is:

[0046] In the formula, x is the output feature value of the graph convolutional layer.

[0047] Design c) Improvement of the Softmax function: Based on the characteristics of the network signature vector, a decision layer is added between the fully connected layer fc of the output layer and the Softmax normalization layer. The input of the decision layer is the output vector of the fully connected layer fc. For neurons whose index positions exceed the interval [0, NL] in the vector, their output values ​​are forcibly assigned to a preset minimum value, which is -1e9 or -∞, where NL is the total number of edges in the sample network, i.e., the actual number of links. After processing by the decision layer, the result is input into the Softmax normalization layer, so that the Softmax output result corresponding to the neurons exceeding the interval [0, NL] is 0, so as to adapt to sample networks with different numbers of links and ensure that the signature vector outputs a valid probability value only for the dimension corresponding to the actual link.

[0048] Furthermore, the specific implementation process of step 1.5 is as follows:

[0049] During the training of GNN, KL divergence is used as the core loss function, and root mean square error (RMSE) is used as an auxiliary performance monitoring indicator to form a multi-dimensional training performance monitoring system. KL divergence guides the optimization direction of GNN parameters by quantifying the information difference between the predicted probability distribution and the true distribution. During the calculation process, a logarithmic transformation log() is applied to the output of GNN. At the same time, the batchmean reduction strategy is used to take the mean of the KL divergence calculation results of each batch of training data to achieve gradient balancing backpropagation of batch data.

[0050] The expressions for KL divergence and RMSE are as follows:

[0051] ;

[0052] ;

[0053] in, This represents the actual signature vector, i.e., the result obtained through Monte Carlo simulation. This represents the signature vector predicted by the GNN, where i is the position index, indicating the i-th component of the signature vector. The dimension representing the signature vector;

[0054] During training, the Adam optimization algorithm is used to dynamically adjust the trainable parameters of the GNN. Gradient clipping is combined to suppress gradient explosion, and L2 regularization is achieved by integrating weight decay terms to avoid model overfitting and improve training stability.

[0055] Furthermore, during GNN training, a Bayesian optimization method is introduced to automate the configuration of key hyperparameters such as hidden layer dimension, learning rate, and Dropout rate in the entire GNN architecture.

[0056] Furthermore, during GNN training, a dynamic learning rate adjustment mechanism is designed to automatically reduce the learning rate when the validation loss stagnates, in order to balance convergence speed and GNN performance.

[0057] Furthermore, the specific implementation process of step 1.6 is as follows:

[0058] Based on the optimal hyperparameter combination of the GNN determined by the Bayesian optimization method, the GNN model is finally trained until the loss is verified to be stable and convergent. After training, the structural feature variables corresponding to the UAV swarm communication network to be evaluated are input into the trained GNN model. The result output by the model after calculation is the network signature vector S corresponding to the network to be evaluated.

[0059] Working principle of the invention:

[0060] Based on the structural characteristics and reliability computing principles of UAV swarm communication networks, this invention constructs a two-stage end-to-end reliability prediction mechanism, the working principle of which is as follows.

[0061] First, in the initial stage, the UAV swarm communication network is abstracted as a graph structure model, where nodes represent UAV communication nodes and edges represent communication links between nodes. Based on this graph structure, structural feature variables characterizing the network topology are extracted, including node degree, clustering coefficient, and adjacency matrix. Based on these network structural features, a corresponding training sample dataset is generated using simulation methods. A graph neural network model is constructed and trained using the network structural features as input and the system signature vector as output. Through multi-layer propagation and aggregation of network structural information, the graph neural network learns the mapping relationship between the network topology and the system signature, thereby achieving prediction of the UAV swarm communication network system signature without relying on link reliability parameters.

[0062] Secondly, in the second stage, based on the system signature prediction results obtained in the first stage and combined with the reliability parameters of the communication link, a hybrid expression of system signature and link reliability parameters is used to perform deterministic calculation of the overall reliability of the UAV swarm communication network under edge failure scenarios. Since the system signature can characterize the structural reliability contribution of the network under different failure sequences, this calculation process does not require the introduction of a learning model again to obtain the overall reliability assessment results of the network.

[0063] By employing the two-stage processing method described above, this invention decouples the structural prediction process of system signatures from the numerical calculation process related to reliability parameters, ensuring that link reliability parameters do not participate in deep learning model training. This effectively reduces the model's dependence on the distribution of specific reliability parameters. This working principle not only improves the computational efficiency of end-to-end reliability prediction but also significantly enhances the generalization ability and stability of the prediction method under different network sizes and link reliability conditions, making it particularly suitable for reliability assessment of large-scale UAV swarm communication networks.

[0064] Compared with the prior art, the present invention has at least the following beneficial effects:

[0065] 1. Achieve effective decoupling between structural reliability and parameter reliability, thereby improving the generalization ability of the prediction model.

[0066] This invention introduces system signatures as intermediate representations to separate the structural reliability characteristics and link reliability parameters of UAV swarm communication networks. In the prediction stage, the link reliability parameters are no longer introduced into the deep learning model training process, thereby effectively reducing the model's dependence on specific link reliability distributions. This eliminates the need to retrain the model when the link reliability parameters change, significantly improving the generalization ability and applicability of the method under different network sizes and link conditions.

[0067] 2. Avoid directly learning full-end reliability to improve prediction stability and interpretability.

[0068] Compared to existing prediction methods that directly use network topology and link reliability parameters as input and end-to-end reliability as output, this invention first predicts network signatures and then calculates end-to-end reliability based on a deterministic reliability hybrid expression. This gives the prediction results clear physical and mathematical meaning, effectively reduces the fitting pressure of deep learning models on random parameters, and improves the stability and interpretability of the prediction results.

[0069] 3. Significantly reduces computational complexity and improves the evaluation efficiency of large-scale networks.

[0070] This invention utilizes graph neural networks to learn network structure features, avoiding the exponentially increasing computational complexity in traditional full-end reliability accuracy calculations. Compared to Monte Carlo simulation methods, it can significantly reduce computation time and resource consumption while ensuring prediction accuracy, making it particularly suitable for UAV swarm communication network scenarios with a large number of nodes and complex topologies.

[0071] 4. Suitable for dynamic link environments, meeting the real-time evaluation needs of engineering projects.

[0072] When link reliability parameters change, this invention does not require retraining the graph neural network model. It only needs to perform deterministic calculations based on the predicted network signature and the updated link reliability parameters to quickly obtain new end-to-end reliability assessment results, thereby effectively supporting the real-time reliability analysis and assurance needs of UAV swarm communication networks in dynamic environments.

[0073] 5. Fully leverage the advantages of graph neural networks in modeling network structure data to improve prediction accuracy.

[0074] This invention uses graph neural networks to model UAV swarm communication networks, which can fully characterize the network topology, node relationships and local structural characteristics. Compared with traditional deep neural network-based methods, it is more suitable for dealing with the reliability prediction problem of network systems and effectively improves the accuracy and robustness of system signature prediction. Attached Figure Description

[0075] Figure 1 This is a flowchart of the algorithm of the present invention;

[0076] Figure 2 This is a schematic diagram of a drone swarm communication network according to an embodiment of the present invention, taking a network with 11 nodes and 21 links as an example;

[0077] Figure 3 This is a comparison chart of the signature prediction accuracy and overall reliability prediction accuracy of the method of this invention with the traditional Monte Carlo simulation method. Detailed Implementation

[0078] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention. It should be noted that, due to limitations in hardware resources, the embodiments of the present invention are only illustrated using an 11-node UAV swarm communication network as an example; the method of the present invention can be applied to UAV swarm communication networks with larger node scales in both theory and engineering practice, and the larger the network node scale, the more significant the technical advantages of the method of the present invention compared with traditional methods.

[0079] I. Overview of the Implementation Environment

[0080] This embodiment focuses on the application scenario of post-disaster communication relay, taking a cluster communication network of 9 drones and 2 ground base stations as the research object, and realizes real-time and accurate prediction of the reliability of the entire network.

[0081] The core hardware and software configuration of this embodiment is as follows:

[0082] 1. Hardware environment: 12th Gen Intel® Core™ i5-12400F 2.50 GHz, NVIDIA GeForce 3060.

[0083] 2. Software environment: The operating system is Windows 11, the programming language is Python 3.9, and the core dependency libraries include: PyTorch 2.0, PyTorch Geometric 2.3.1, NetworkX 2.8.8, NumPy, and Pandas.

[0084] II. Implementation Steps

[0085] Combined with appendix Figure 1 The technical solution of this invention is divided into two stages: GNN prediction of network signature vector and calculation of end-to-end reliability based on signature hybrid formula. The specific implementation of each stage and step is as follows:

[0086] Step 1, First Stage: Predicting the Signature Vector Based on GNN:

[0087] Step 1.1: Construct the undirected graph G(V,E) corresponding to the UAV swarm communication network, see [link / reference]. Figure 2 The specific process is as follows:

[0088] 1. Vertex set V: consists of 11 vertices, each corresponding to one of the 9 drones (corresponding to...). Figure 2 (Nodes numbered 1 to 9) and 2 ground base stations (corresponding to) Figure 2 (Nodes numbered 10 to 11 in the middle).

[0089] 2. Edge set E: An undirected edge is established if and only if there is an actual communication link between the entities corresponding to two vertices (drone-drone, drone-ground base station); in this embodiment, the finally constructed undirected graph includes 21 edges, and the specific connection relationship of the edges is in the form of an integer pair of "source node - target node" (such as 1-2, 4-5, 8-11, etc.), and is stored after being sorted in lexicographical order;

[0090] 3. Edge weight: Each edge carries a weight value p (0 < p < 1), representing the probability of successful establishment of the corresponding communication link (i.e., link reliability).

[0091] Step 1.2: Determine the network structure feature variables, and the specific process is as follows:

[0092] 1. Node features: Select the node degree and the node clustering coefficient. The node degree is the number of edges directly connected to this node (for example, Figure 2 if there are 3 edges directly connected to node 1, then the node degree of this node is 3); the clustering coefficient is calculated using the standard definition. For nodes with a degree less than 2, their clustering coefficients are directly set to 0, and all feature values are retained with 6 decimal places of precision;

[0093] 2. Network topology structure: It is represented by an adjacency matrix. The adjacency matrix is an 11×11 binary matrix. The element A(i,j)=1 indicates that there is an undirected edge between node i and node j, and A(i,j)=0 indicates that there is no undirected edge;

[0094] 3. Feature integration: Organize the degrees and clustering coefficients of 11 nodes in the order of node numbers to form a node feature matrix X with a dimension of [11,2] (each row corresponds to the 2-dimensional features of a node).

[0095] Step 1.3: Generate a sample network dataset for training the GNN model through simulation, and the specific process is as follows:

[0096] 1. Sample network generation: Generate a fully connected network with a fixed number of nodes of 11 and the number of links L randomly taking values within the range of [10, 55] (the maximum number of links in an 11-node fully connected network is 55). Adopt the "spanning tree + random edge addition" strategy: First generate a spanning tree with 10 edges as the basic connected structure, and then randomly select L - 10 edges from the remaining 45 possible edges for supplementation; Repeat this process until 20,000 groups of sample networks are generated, and each sample is stored as an independent text file in the standardized edge list format;

[0097] 2. Sample feature extraction: For each sample network, extract the node degree and the node clustering coefficient, organize them into a node feature matrix of [11,2], and then store the feature data of all samples in the format of "sample ID + 22 feature values" (11 node degrees + 11 clustering coefficients) as a CSV file;

[0098] 3. Sample Label Generation: Link failure is set as the sole cause of network failure (nodes are completely reliable), and each link failure is independent and identically distributed. A Monte Carlo simulation algorithm is used to calculate the signature vector of each sample network as the sample label. Specific parameters: number of iterations C = 20000; number of independent experiments for each sample M = 10000. In each experiment, the edge order is randomly shuffled and then deleted sequentially. The edge deletion sequence number when the network first becomes disconnected is recorded, and the frequency of each sequence number is counted (frequency replaces probability), resulting in a 55-dimensional signature probability distribution vector (adapted to the maximum number of links in an 11-node network). If the actual number of links L in the sample network is less than 55, the label value of the missing dimension is filled with 0.

[0099] 4. Dataset splitting: The 20,000 samples were randomly divided into a training set (16,000 sets) and a test set (4,000 sets) in an 8:2 ratio.

[0100] Step 1.4: Construct the GNN model. The specific process is as follows:

[0101] We construct a complete GNN model consisting of an input layer, a graph convolutional layer, a graph pooling layer, and an output layer. The core objective is to achieve the mapping from input (node ​​feature matrix X + adjacency matrix) to output (network signature vector). The specific design is as follows:

[0102] 1. Input Layer: Includes three core inputs: ① Node feature matrix X (dimension [11,2]); ② Edge index (derived from the adjacency matrix, format [2,21], recording the source and target node indices of all edges); ③ Batch processing index (used for multi-graph parallel training).

[0103] 2. Graph Convolutional Layer (GCNConv): 2 layers in total, used for core feature extraction;

[0104] ① The first convolutional layer GCNConv1: Input dimension 2 → Output dimension is the hidden layer dimension (32-256) within the Bayesian optimization search range, which performs neighborhood aggregation and linear transformation on node features;

[0105] ② The second convolutional layer GCNConv2: Input dimension (the hidden layer dimension output by the first layer) → Output dimension is the next hidden layer dimension (32-256) within the Bayesian optimization search range, further extracting deep topological features;

[0106] ③ Mathematical expression:

[0107] First convolutional layer: ;

[0108] Second convolutional layer: ;

[0109] in, Let A be the adjacency matrix with self-loops (A is the adjacency matrix and W is the identity matrix), W1 is the trainable learning matrix of the first graph convolutional layer, and W2 is the trainable learning matrix of the second graph convolutional layer. It is a degree matrix with self-loops (the diagonal elements are the node degree + 1, and the other elements are 0);

[0110] 3. Graph Pooling Layer: Global average pooling is used to calculate the average value of the output features of the GCNConv2 hidden layer along the node dimension, resulting in a global graph-level feature vector with the same dimension as the hidden layer.

[0111] 4. Dropout layer: Set after the graph pooling layer, dropout_rate is a parameter within the Bayesian optimization search range (0.1~0.5). By randomly turning off a corresponding proportion of neurons, it suppresses model overfitting and improves generalization ability;

[0112] 5. Output layer: Consists of a fully connected layer and an improved Softmax layer.

[0113] ① Fully connected layer fc: Input dimension (i.e., the hidden layer dimension output by the graph pooling layer) → Output dimension 55 (the maximum number of links to adapt to an 11-node network);

[0114] ② Decision layer: Add a decision layer between the fully connected layer and the Softmax layer. If the index position exceeds [0,21], force the value to be -1e9 (approximately -∞).

[0115] ③ Softmax layer: Normalizes the output of the decision layer to obtain a 55-dimensional probability vector (i.e., network signature vector).

[0116] 6. Activation function: After each GCNConv layer, a ReLU activation function is applied, with the expression Relu(x) = max(0,x), to achieve non-linear transformation and avoid gradient vanishing;

[0117] 7. Parameter initialization: The weights of all linear layers and GCNConv layers are initialized using the Xavier normal distribution, and the bias term is uniformly set to 0.

[0118] Step 1.5: Train and optimize the GNN model:

[0119] The training strategy employs a full-process approach, encompassing "loss function + optimization algorithm + hyperparameter optimization + dynamic learning rate adjustment." The specific process is as follows:

[0120] 1. Performance monitoring system: Using KL divergence as the core loss function, combined with root mean square error (RMSE) as an auxiliary indicator.

[0121] ①KL divergence: Quantizes the distribution difference between the predicted signature vector Spred and the true signature vector Strue obtained from Monte Carlo simulation, guiding parameter optimization. During calculation, a log() transformation is applied to Spred, and a batchmean reduction strategy (averaging the loss for each batch) is used to achieve gradient equalization backpropagation. The expression is:

[0122] ;

[0123] ②RMSE: Monitors the numerical error between Spred and Strue, expressed as:

[0124] ;

[0125] 2. Optimization Algorithm: The Adam optimizer is adopted with the following parameters: first moment estimation decay rate β1=0.9, second moment estimation decay rate β2=0.999, and numerical stability coefficient ε=1e-8. Gradient clipping is implemented to suppress gradient explosion, and L2 regularization is achieved by integrating weight decay terms.

[0126] 3. Hyperparameter Optimization: A Bayesian optimization method is introduced to achieve automated configuration, constructing a search space including five key hyperparameters: ① GCNCV hidden layer dimension (32–256); ② Adam learning rate (1e-4–1e-3); ③ Dropout rate (0.1–0.5); ④ batch size (32–128); ⑤ epochs (100–300 training rounds, with model performance validated every 10 rounds). Optimization process: First, five sets of hyperparameters are randomly selected for initial exploration, then the objective function is optimized (i.e., minimizing KL divergence + RMSE) through 15 iterations, finally outputting the optimal hyperparameter combination.

[0127] 4. Dynamic adjustment of learning rate: The ReduceLROnPlateau scheduler is used. When the validation loss decreases by less than 1% for three consecutive training rounds, the current learning rate is multiplied by a decay factor of 0.1. The number of adjustments is limited to no more than three times, and the adjusted learning rate is no less than 1e-5.

[0128] Step 1.6: Predict the signature vector based on the trained GNN. The specific process is as follows:

[0129] After Bayesian optimization, the optimal hyperparameter combination is determined and the convergence verification of the GNN model is completed based on this combination (ensuring that the loss fluctuation range of 5 consecutive verification rounds is <1%). The network features to be evaluated extracted in step 1.2 (the node feature matrix X with dimensions [11,2] and the edge index with dimensions [2,21]) are input into the trained optimal GNN model. After the model forward operation, a 55-dimensional network signature vector is directly output. In this embodiment, the calculation time of this step is only 0.01s.

[0130] Step 2, Second Stage: Calculating the overall reliability of the network based on the signature hybrid formula

[0131] Substitute the signature vector obtained in step 1.6 into the hybrid formula for network reliability based on signatures to calculate the end-to-end reliability R(p) for any link reliability p, as shown in the following formula:

[0132] ;

[0133] Where Sk is the k-th component of the signature vector obtained in step 1.6, and p is the link reliability (in this embodiment, a step size of 0.1 is used to show the network-wide reliability when p takes values ​​of 0-1); after substituting into the calculation, the network-wide reliability R(p) of the UAV swarm communication network in this embodiment is obtained, as shown below:

[0134] Network reliability results under different p-values:

[0135] p=0.0: R(p)=0.000000

[0136] p=0.1: R(p)=0.000710

[0137] p=0.2: R(p)=0.006006

[0138] p=0.3: R(p)=0.033208

[0139] P=0.4: R(P)=0.122881

[0140] P=0.5: R(P)=0.306019

[0141] P=0.6: R(P)=0.550077

[0142] P=0.7: R(P)=0.772155

[0143] P=0.8: R(P)=0.913282

[0144] P=0.9: R(P)=0.976963

[0145] P=1.0: R(P)=1.000000

[0146] III. Implementation Results Verification

[0147] To systematically verify the effectiveness and superiority of the present invention, a control experiment was designed to compare the performance of the method of the present invention with that of the traditional Monte Carlo simulation method. The experimental conditions were kept strictly consistent: a unified hardware environment and the same network to be evaluated (11 nodes and 21 edges) to eliminate interference from irrelevant variables.

[0148] The core performance indicators of the experiment include signature computation efficiency and prediction accuracy. Specific data are as follows: Traditional Monte Carlo simulation takes 15 seconds to compute a signature, while the method of this invention only takes 0.01 seconds. In terms of accuracy, the RMSE error of the signature vector predicted by this invention is 0.0169 compared with the signature vector obtained by Monte Carlo simulation. The overall reliability result calculated based on the GNN-predicted signature vector has an RMSE error of 0.0194 compared with the reliability result obtained based on the Monte Carlo simulation signature.

[0149] Experimental results show: See Figure 3 The method of this invention is close to the traditional Monte Carlo simulation method in terms of both signature prediction accuracy and overall reliability prediction accuracy. However, the signature calculation efficiency is improved by 1500 times, which can efficiently respond to the real-time evaluation requirements of the dynamic link topology of UAV clusters, and significantly improve the engineering applicability and promotion value of the technical solution of this invention.

[0150] IV. Engineering Application Advantages of the Invention

[0151] 1. Reducing Dependence on Measured Link Reliability p: In engineering practice of UAV swarm communication networks, the reliability p-value of a single communication link is often affected by factors such as airspace interference, dynamic node topology, and hardware performance fluctuations, resulting in high measurement costs, poor real-time performance, and difficulty in guaranteeing data accuracy. The network signature fast prediction model and two-stage evaluation paradigm based on GNN proposed in this invention can, to some extent, eliminate the dependence on the measured value of link reliability p, and directly complete the efficient evaluation of overall reliability through predicted network signatures.

[0152] 2. Flexible and adaptable reliability classification criteria: This invention does not specify the classification criteria for network reliability, which is highly flexible and can be set by users according to specific engineering scenario requirements (such as task priority, survivability requirements, deployment environment, etc. of drone swarms).

[0153] V. Supplementary Explanation

[0154] 1. Terminology Definitions: The meanings of all terms (including technical and scientific terms) in this embodiment are consistent with the common understanding of those skilled in the art. For terms already defined in general dictionaries, they should be understood as their conventional meaning in the existing technical context, and no idealized or overly formal special interpretations are given unless otherwise explicitly defined in this embodiment.

[0155] 2. Parameter flexibility: The specific parameters in this embodiment (such as the number of samples, training rounds, hyperparameter range, etc.) are all exemplary settings. Those skilled in the art can adjust them independently according to the actual engineering scenario without deviating from the core technical idea of ​​the present invention.

[0156] 3. Alternative implementation: This embodiment uses GCNConv as the graph convolutional layer. Those skilled in the art can also replace it with other graph convolutional structures such as GATConv. Only the parameter initialization and selection method of the corresponding layer need to be adjusted, and the other steps remain unchanged, so that the same technical effect can still be achieved.

[0157] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for predicting the overall reliability of a UAV swarm communication network based on GNN and network signature, characterized in that, Includes the following steps: Step 1, First Stage: Predict the signature vector based on GNN, as follows: Step 1.1: Abstract the UAV swarm communication network as an undirected graph. To build a drone swarm network system; Step 1.2: Determine the network structure characteristic variables. Select node characteristics and network topology as the structural characteristic variables of the UAV swarm network system. The node characteristics include node degree and node clustering coefficient, and the network topology is selected as the adjacency matrix. Step 1.3: Generate a sample network dataset for training the GNN model using simulation methods. The specific implementation process is as follows: The number of nodes generated using the Python library is [number]. The number of links is arrive For any integer fully connected network, N sets of fully connected networks with different topological parameters are generated, where N is the preset sample size, serving as the sample network dataset for training the GNN model. For each sample network, the structural feature variables described in step 1.2 are extracted and integrated into the feature data of the sample network. In conjunction with the actual fault scenarios of the UAV swarm communication network, link failure is set as the only cause of network functional abnormality. That is, it is assumed that all nodes are completely reliable, and only the on / off state of the communication link affects the network connectivity. Moreover, each link failure event is independent and identically distributed with equal failure probability. The signature vector of each sample network is calculated using the Monte Carlo simulation algorithm. For each sample network, the random simulation of the on / off state of the link is repeated M times, where M is the preset number of simulations. The connectivity characteristics of the network under each simulation are statistically analyzed. Based on the statistical results, the network signature vector corresponding to the sample network is obtained, and the signature vector is used as the label of the sample network in the sample dataset. Step 1.4: Construct GNN. The goal of constructing GNN is to design a deep learning proxy model that takes network node features and adjacency matrix as input and outputs network signature vector. Step 1.5: Train and optimize the GNN model; Step 1.6: Predict the signature vector S of the trained GNN network; Step 2, Second Stage: After predicting the network signature vector S in the first stage, substitute the signature vector S into the hybrid expression formula for network end-to-end reliability based on signatures to calculate the reliability of any link. Network end-to-end reliability The calculation formula is as follows: ， in, The signature vector obtained from the first stage GNN prediction is the first... One portion, , This is the value for link reliability. The dimension of the signature vector is the total number of edges in the network. j acts as a traversal index in the formula, representing the actual physical meaning of the number of edges in the network that are working normally.

2. The method for predicting the overall reliability of UAV swarm communication networks based on GNN and network signatures according to claim 1, characterized in that, The specific implementation process of step 1.1 is as follows: Each vertex in the vertex set V of the undirected graph G(V,E) corresponds to a drone or a ground base station in the drone swarm network system described in step 1.

1. An undirected edge is established in the undirected graph G(V,E) between two vertices if and only if there is an actual communication connection between the drones or ground base stations corresponding to the two vertices that enables data transmission. The undirected edge carries a weight value, which specifically represents the probability of successfully establishing a connection for the corresponding communication link, i.e., the link reliability value described in step 2. , In this way, the undirected graph G(V,E) corresponding to the UAV swarm network system is constructed.

3. The method for predicting the overall reliability of UAV swarm communication networks based on GNN and network signatures according to claim 1, characterized in that, The complete GNN architecture design in step 1.4 is as follows: Design 1), Input Layer, including: Node feature matrix X: The node feature matrix X has a dimension of [number of nodes, 2], where each node contains a 2-dimensional feature of node degree and node clustering coefficient; Edge index: The edge index is obtained by transforming the adjacency matrix and is used to represent the topological connection relationship of the graph. Its format is [2, number of links]; Design 2), Graph Convolutional Layer GCNConv, is used for core feature extraction, including: The first convolutional layer GCNConv1 in the first figure performs neighborhood feature aggregation and linear transformation on the node feature matrix X of the input layer. The input → output dimension is: 2 → hidden layer 1 dimension. The second image shows the convolutional layer GCNConv2, which further extracts deep features based on the output features of GCNConv1. The input → output dimension is: hidden layer 1 dimension → hidden layer 2 dimension. The mathematical expression for the graph convolutional layer GCNConv is as follows: First convolutional layer: , Second convolutional layer: , in, Let A be the adjacency matrix with self-loops, and I be the identity matrix. W1 is the trainable learning matrix of the first layer of graph convolution, and W2 is the trainable learning matrix of the second layer of graph convolution. It is a degree matrix with self-loops, specifically a diagonal matrix: the diagonal elements are the degree of each node in the network, containing self-loops, and the remaining elements are 0; The node feature matrix is ​​in the format [number of nodes × number of node features]. This represents the output feature matrix of the first convolutional layer GCNConv1. This represents the output feature matrix of the second convolutional layer, GCNConv2. GCNConv, representing graph convolutional layers, introduces non-linear transformation activation functions, enabling the model to fit complex graph structure features rather than simply performing linear aggregation of neighborhood features. Design 3), graph pooling layer, the graph pooling layer is used to compress global information, and adopts global average pooling operation: the feature vector output by the graph convolutional layer H(2) of all nodes in the same network is calculated along the node dimension to extract global graph-level features from local node features, so as to obtain global graph-level feature vectors as input to the next layer; Design 4), output layer, including: Fully connected layer fc: The input dimension is the global graph-level feature dimension output by the graph pooling layer, and the output dimension is the dimension of the network signature vector S; Softmax normalization: Apply the Softmax function to the output of the fully connected layer fc to obtain a probability vector, which is the network signature vector S. The expression for the Softmax function is as follows: ， in, Indicates the first Input values The output after performing the Softmax transformation, where j is the index parameter used to iterate through the elements at each position in the signature vector, and e... xi This indicates that each element of the signature vector is exponentialized; To enable GNNs to be adapted for training on sample networks with different topologies, for A network with 10 nodes, where the input and final output dimensions of the GNN sample labels are uniformly set to 1. ,Right now The maximum number of links in a network of nodes, if the actual signature dimension of the sample network is less than... If the missing dimension is filled with 0, then the signature tag value will be filled with 0.

4. The method for predicting the overall reliability of UAV swarm communication networks based on GNN and network signatures according to claim 3, characterized in that, The full-process architecture of GNN also includes the following designs: Design a): After the graph pooling layer and before the output layer, add a Dropout layer; the Dropout layer randomly shuts down some neurons at a dropout_rate, where the dropout_rate ranges from 0.1 to 0.5, to suppress model overfitting and improve the generalization ability of GNN to UAV swarm communication networks with different topologies. Design b) Activation function selection: After each layer of the graph convolutional layers GCNConv1 and GCNConv2, a ReLU activation function is introduced to achieve layer-by-layer nonlinear transformation, avoid gradient vanishing, and enhance the feature representation capability of the model; the expression of the ReLU activation function is: In the formula, x is the output feature value of the graph convolutional layer; Design c) Improvement of Softmax function: Based on the characteristics of the network signature vector, a decision layer is added between the fully connected layer fc of the output layer and the Softmax normalization layer; the input of the decision layer is the output vector of the fully connected layer fc, and for neurons whose index positions in the vector exceed the interval [0, NL], their output values ​​are forcibly assigned to a preset minimum value, which is -1e9 or -∞, where NL is the total number of edges in the sample network, i.e., the actual number of links; After processing by the decision layer, the result is then input into the Softmax normalization layer, so that the Softmax output result corresponding to neurons outside the [0, NL] interval is 0, in order to adapt to sample networks with different numbers of links and ensure that the signature vector outputs valid probability values ​​only for the dimensions corresponding to the actual links.

5. The method for predicting the overall reliability of UAV swarm communication networks based on GNN and network signatures according to claim 1, characterized in that, The specific implementation process of step 1.5 is as follows: During the training of GNN, KL divergence is used as the core loss function, and root mean square error (RMSE) is used as an auxiliary performance monitoring indicator to form a multi-dimensional training performance monitoring system. KL divergence guides the optimization direction of GNN parameters by quantifying the information difference between the predicted probability distribution and the true distribution. During the calculation process, a logarithmic transformation log() is applied to the output of GNN. At the same time, the batchmean reduction strategy is used to take the mean of the KL divergence calculation results of each batch of training data to achieve gradient balancing backpropagation of batch data. The expressions for KL divergence and RMSE are as follows: ， ， in, This represents the actual signature vector, i.e., the result obtained through Monte Carlo simulation. This represents the signature vector predicted by the GNN, where i is the position index, indicating the i-th component of the signature vector. The dimension representing the signature vector; During training, the Adam optimization algorithm is used to dynamically adjust the trainable parameters of the GNN. Gradient clipping is combined to suppress gradient explosion, and L2 regularization is achieved by integrating weight decay terms to avoid model overfitting and improve training stability.

6. The method for predicting the overall reliability of a UAV swarm communication network based on GNN and network signature as described in claim 5, characterized in that, During GNN training, Bayesian optimization is introduced to automate the configuration of key hyperparameters such as hidden layer dimension, learning rate, and Dropout rate in the entire GNN architecture.

7. The method for predicting the overall reliability of a UAV swarm communication network based on GNN and network signature as described in claim 5, characterized in that, During GNN training, a dynamic learning rate adjustment mechanism is designed to automatically reduce the learning rate when the validation loss stagnates, in order to balance convergence speed and GNN performance.

8. The method for predicting the overall reliability of a UAV swarm communication network based on GNN and network signature as described in claim 6, characterized in that, The specific implementation process of step 1.6 is as follows: Based on the optimal hyperparameter combination of the GNN determined by the Bayesian optimization method, the GNN model is finally trained until the loss is verified to be stable and convergent. After training, the structural feature variables corresponding to the UAV swarm communication network to be evaluated are input into the trained GNN model. The result output by the model after calculation is the network signature vector S corresponding to the network to be evaluated.

Images