An electromagnetic-network situation integrated perception method based on reciprocal learning

Abstract

This invention discloses an integrated electromagnetic-network situational awareness method based on reciprocal learning, comprising: dividing the target network into nodes, establishing edges and edge weights to form a graph; generating electromagnetic features and network features to constitute initial node features; inputting the initial node features in parallel into branches A and B, and constructing electromagnetic feature matrices and network feature matrices based on the electromagnetic features and network features of all nodes, respectively; obtaining an enhanced network feature matrix, a network quality representation, and an electromagnetic quality representation in branch A; obtaining an electromagnetic quality representation and a network quality representation in branch B; constructing a total diversity loss; weighted fusion of the electromagnetic quality representations and network quality representations from branches A and B to obtain a joint quality representation; constructing a total training loss for training to complete the perception of electromagnetic features and network state features. This invention achieves unified and accurate perception of spectrum sensing, link connectivity, and QoS status in complex environments.

Description

Technical Field

[0001] This invention relates to the field of information engineering, specifically to an integrated electromagnetic-network situational awareness method based on reciprocal learning. Background Technology

[0002] In current electromagnetic and network quality sensing systems, electromagnetic and network states are assessed separately, resulting in "island phenomena" in the sensing models across different dimensions. This leads to gaps and misjudgments in the understanding of complex interference. On the electromagnetic side, traditional sensing models can capture local electromagnetic features such as spectrum occupancy and signal strength, but they are completely detached from higher-level information such as network topology and link quality, making it difficult to assess the actual threat of interference to communication links. Conversely, on the network side, while sensing models can monitor node connectivity and link load, they are insufficiently aware of the channel signal-to-noise ratio degradation and sudden increase in bit error rate caused by strong electromagnetic interference at the lower levels, leading to misjudgments of the upper-level state.

[0003] In the development of electromagnetic spectrum situational awareness, spectrum sensing can be divided into single-user spectrum sensing and multi-user collaborative spectrum sensing based on the number of participating users. Single-user spectrum sensing methods can be further divided into traditional model-driven methods and modern data-driven methods, with modern data-driven methods including machine learning and deep learning methods. Multi-user collaborative sensing methods can be categorized according to their cooperation methods: centralized collaborative sensing, distributed collaborative sensing, relay collaborative sensing, and clustered collaborative sensing.

[0004] Single-user spectrum sensing refers to the process where a sensing user independently samples local radio frequency signals to determine the presence of a primary user. Traditional model-driven methods primarily rely on the differences between the primary user's communication signals and noise statistical characteristics to determine the channel state, such as energy detection, matched filtering detection, cyclostationary detection, and eigenvalue detection. With the development of deep learning technology, after offline training on large amounts of data, sensing performance has seen a significant improvement compared to machine learning algorithms.

[0005] Among the four cooperative modes of multi-user collaborative spectrum sensing, centralized cooperative sensing has a simple structure, good real-time performance, and is widely used. Distributed cooperative sensing has high reliability and accuracy, but the system operation is complex. Relay-type cooperative spectrum sensing is mainly used when some sensing users have poor communication environments and cannot effectively transmit information; it forwards information through sensing users with better channel conditions. Clustered cooperative sensing, based on centralized cooperative sensing, divides sensing users into multiple clusters, with each cluster independently completing sensing and then reporting the results to the fusion center.

[0006] Current network state perception approaches the issue from different dimensions such as signals, protocols, data, and models to construct state perception of network connectivity and performance. These methods can be categorized into perception methods based on signal processing and pattern recognition, perception methods based on network protocols and active probing, perception methods based on data-driven approaches and machine learning, and perception methods based on graph theory and network modeling.

[0007] Currently, spectrum sensing technology has shifted from traditional model-driven methods such as energy detection and cyclostationary detection to modern data-driven deep learning methods, significantly improving feature extraction capabilities in complex background noise under a collaborative sensing architecture. Simultaneously, network state sensing has formed a multi-dimensional sensing system based on signal processing, network protocols (such as the OLSR protocol), data-driven approaches, and graph theory modeling, enabling topology maintenance and anomaly detection. However, the "island phenomenon" between the electromagnetic and network dimensions remains prominent in existing sensing systems: electromagnetic sensing models are often detached from the upper-layer network topology, making it difficult to assess the actual threat of interference to communication links; while network-side models lack sufficient perception of underlying causes such as signal-to-noise ratio (SNR) degradation and increased bit error rate due to strong electromagnetic interference at the physical layer, leading to gaps and misjudgments in the understanding of composite interference. Therefore, there is an urgent need for an integrated situational awareness technology that can transcend the physical layer electromagnetic features and network layer state, achieving deep fusion of dual-domain features and collaborative reasoning. Summary of the Invention

[0008] The purpose of this invention is to provide an integrated electromagnetic-network situational awareness method based on reciprocal learning, which effectively eliminates the "island phenomenon" of awareness and achieves unified and accurate awareness of spectrum perception, link connectivity and QoS status in complex environments.

[0009] To achieve the above functions, this invention designs an integrated electromagnetic-network situational awareness method based on reciprocal learning. For a target network containing multiple receivers and multiple communication links, the following steps S1-S4 are executed to complete the perception of electromagnetic quality and network quality:

[0010] Step S1: Based on the receiver location and the nodes of the communication link, divide the target network into nodes, establish the edges between each node, and assign edge weights to form a graph; for each node in the graph, generate electromagnetic features and network features respectively to further constitute the initial features of the node.

[0011] Step S2: Input the initial features of the nodes in parallel into branches A and B. Based on the electromagnetic features and network features of all nodes, construct the electromagnetic feature matrix and the network feature matrix, respectively. In branch A, obtain the enhanced network feature matrix through the attention mechanism; obtain the network quality representation through the enhanced network feature matrix; obtain the electromagnetic quality representation through graph convolutional propagation. In branch B, obtain the electromagnetic quality representation through the reverse attention mechanism and graph convolutional propagation; obtain the network quality representation through the gating mechanism.

[0012] Step S3: Construct the total diversity loss based on the electromagnetic domain diversity loss and the network domain diversity loss; weight and fuse the electromagnetic quality representation and network quality representation obtained from branch A and branch B respectively; concatenate the fused electromagnetic quality representation and network quality representation to obtain the joint quality representation;

[0013] Step S4: Based on the joint quality representation, construct a shared feature encoding, and construct spectrum occupancy loss, link connectivity classification loss, and network QoS regression loss to form the total training loss; train based on the total training loss, and output spectrum occupancy, link connectivity, and network QoS results to complete the perception of electromagnetic features and network state features.

[0014] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0015] To address the problems of traditional situation generation methods in cross-domain fusion, such as independent modeling of the electromagnetic and network domains, fragmented feature representation spaces, and neglect of cross-domain causal relationships, which prevent electromagnetic sensing results from providing effective priors for network state inference and from correcting spectral situation estimates in the reverse direction of network state changes, ultimately resulting in low overall sensing accuracy and weak generalization ability in complex adversarial environments, this invention constructs a reciprocal learning-based electromagnetic-network situational awareness algorithm (Reciprocal Learning Network, RLNet). This algorithm utilizes cross-domain attention mechanisms and quality-aware graph convolution to achieve deep interweaving and causal inference between electromagnetic features and network states. Through two-branch reciprocal interaction, it extracts the correlation information between the two domains, prompting branch A to use network features to assist electromagnetic representation, and branch B to use electromagnetic quality to guide network feature reshaping. This effectively eliminates the "island phenomenon" in sensing, achieving unified and accurate sensing of spectrum perception, link connectivity, and QoS status in complex environments. It overcomes the technical challenges of cross-domain information loss, difficulty in quantitatively tracing the causes of complex interference, and misjudgment of upper-level states in traditional sensing systems, laying a solid foundation for the robust operation and precise resource management of communication networks in complex dynamic electromagnetic environments. Attached Figure Description

[0016] Figure 1 This is a flowchart of an integrated electromagnetic-network situational awareness method based on reciprocal learning, according to an embodiment of the present invention.

[0017] Figure 2 It is a graph constructed after adaptively partitioning the target network into nodes and adding edges, according to an embodiment of the present invention.

[0018] Figure 3 This is a flowchart of reciprocal learning provided according to an embodiment of the present invention;

[0019] Figure 4 This is a flowchart of constructing the total training loss according to an embodiment of the present invention;

[0020] Figure 5 This is a comparison chart of spectrum sensing detection probabilities under different signal-to-noise ratios provided in the embodiments of the present invention;

[0021] Figure 6 This is a comparison chart of false alarm probabilities for spectrum sensing under different signal-to-noise ratios provided in an embodiment of the present invention;

[0022] Figure 7 This is a comparison chart of link delay errors at different signal-to-noise ratios provided in the embodiments of the present invention;

[0023] Figure 8 This is a comparison chart of packet loss rate errors at different signal-to-noise ratios according to embodiments of the present invention;

[0024] Figure 9 This is a comparison chart of effective rate error of the downlink under different signal-to-noise ratios according to embodiments of the present invention;

[0025] Figure 10 This is a comparison chart of link reliability F1 scores under different signal-to-noise ratios provided by embodiments of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0027] This invention provides an integrated electromagnetic-network situational awareness method based on reciprocal learning, targeting a network comprising multiple receivers and multiple communication links, referring to... Figure 1 Perform the following steps S1-S4 to complete the sensing of electromagnetic quality and network quality:

[0028] Step S1: Based on the receiver location and the nodes of the communication link, divide the target network into nodes, establish the edges between each node, and assign edge weights to form a graph; for each node in the graph, generate electromagnetic features and network features respectively to further constitute the initial features of the node.

[0029] The specific steps of step S1 are as follows:

[0030] Step S1.1: Based on the receiver location and the nodes of the communication link, the nodes of the target network are adaptively partitioned using a Voronoi diagram. Edges are supplemented using the Delaunay triangulation method, and edge weights are assigned to each edge to construct the graph. The constructed graph is referenced... Figure 2 ;

[0031] The specific steps of step S1.1 are as follows:

[0032] Step S1.1.1: Extract key points, including the location of each receiver and the source and destination node locations of all communication links. ,in, , These are the source and destination node positions for the nth communication link, respectively. The total number of communication links; using each key point as a seed point, the target network is divided into several units using a Voronoi diagram. The distance from any node in each unit to the corresponding seed point is less than the distance to other seed points, ensuring complete spatial coverage without overlap.

[0033] Step S1.1.2: For the preset high-density traffic area, divide it into four equal-area sub-regions and add virtual nodes to improve sensing accuracy; the final generated node set is denoted as... ,in, Represents each node in the target network. The total number of nodes. Adaptively determined based on the distribution of communication links;

[0034] Step S1.1.3: If the i-th node in the target network The j-th node The Euclidean distance between them is less than the connecting radius. Then establish an edge To ensure graph connectivity, the Delaunay triangulation method is used, with a distance threshold set. If the Euclidean distance between two nodes exceeds the threshold, the nodes are considered disconnected, and an edge is added between them. Based on the Gaussian radial basis function (RBF), each edge is assigned a weight, as shown in the following formula:

[0035] ;

[0036] in, For the i-th node The j-th node Edge weights between them For nodes and nodes The square of the distance, This represents an exponential function with the natural constant e as its base. The scale parameter is the average distance between nodes to adaptively match the spatial scale of the graph. The physical meaning of edge weights is that spatially adjacent nodes are more likely to have similar electromagnetic environments and network states, and their information interaction should be stronger. The greater the distance, the smaller the weight, and the faster the information decays.

[0037] Step S1.1.4: Construct the adjacency matrix based on edge weights Calculate the degree matrix degree matrix diagonal elements ,in, For nodes and nodes The connection strength is obtained by adding a self-loop. ,in, To enhance the adjacency matrix, For the identity matrix, and The corresponding degree matrix Complete the construction of the graph.

[0038] Step S1.2: Based on the IQ data packets collected by each receiver, electromagnetic features are generated for all nodes in the graph after filtering, feature extraction, dimensionality reduction, and spatial estimation. ;

[0039] Set threshold Valid subbands were selected, and features were extracted using the Fast Fourier Transform (FFT) method. Dimensionality reduction was performed using PCA, and electromagnetic features were obtained through inverse distance interpolation using a two-ray model. .

[0040] Step S1.3: Extract network features for each node. , with electromagnetic characteristics Concatenating with positional codes to form the initial features of the nodes. .

[0041] Network characteristics This includes sender-side features, receiver-side features, and latency features. The sender-side features include the number of packets sent. Number of bytes sent Transmission throughput Packet sending rate The characteristics of the receiving side correspond to the characteristics of the transmitting side.

[0042] Step S2: Input the initial features of the nodes in parallel into branches A and B. Based on the electromagnetic features and network features of all nodes, construct the electromagnetic feature matrix and the network feature matrix, respectively. In branch A, obtain the enhanced network feature matrix through the attention mechanism; obtain the network quality representation through the enhanced network feature matrix; obtain the electromagnetic quality representation through graph convolutional propagation. In branch B, obtain the electromagnetic quality representation through the reverse attention mechanism and graph convolutional propagation; obtain the network quality representation through the gating mechanism.

[0043] Reference Figure 3 Based on reciprocal learning between branches A and B, the specific steps performed by branch A in step S2 are as follows:

[0044] Step S2.1.1: Construct an electromagnetic characteristic matrix from the electromagnetic characteristics of all nodes. After two layers of graph convolution, the electromagnetic mass representation of branch A is obtained. ,in This represents a two-layer graph convolution process; each node has incorporated the electromagnetic information of its multi-hop neighbors, forming a high-level representation of the local electromagnetic field.

[0045] Step S2.1.2: Construct a network feature matrix from the network features of all nodes. The query matrix of branch A is obtained by linear projection. ;in, The query matrix for branch A The weight matrix represents the electromagnetic mass of branch A. The key matrix of branch A is obtained by projecting the matrix separately. Sum matrix ;

[0046] Attention score is calculated as follows:

[0047] ;

[0048] ;

[0049] in, Let the original attention score be the sum of the attention from the i-th node to the j-th node in branch A. For the i-th node of branch A, the first node is the first node. The original attention scores of each node. Let be the query vector for the i-th node in branch A. Let be the key vector of the j-th node in branch A. This is a scaling factor used to prevent the dot product from becoming too large as the dimension increases, which could lead to gradient vanishing. Let be the attention score of the i-th node in branch A to the j-th node. Its physical meaning is the degree to which the network state of the i-th node depends on the electromagnetic environment of the j-th node. If the electromagnetic interference at a certain node is strong and the attention score is large, it means that the interference source has a significant impact on the network performance of this node. The total number of nodes; This represents an exponential function with the natural constant e as its base; in the example, To avoid the softmax saturation caused by an excessively large dot product;

[0050] Based on network feature matrix The enhanced network feature matrix of branch A is calculated as follows:

[0051] ;

[0052] in, This represents the augmented network feature matrix for branch A. The electromagnetic information aggregated in branch A represents the electromagnetic information obtained from the network perspective of the nodes, weighted and aggregated according to the importance of the electromagnetic environment of each node. This represents the electromagnetic information of each node in branch A. The total number of nodes. The weight matrix for the electromagnetic information aggregated in branch A;

[0053] Step S2.1.3: Enhanced network feature matrix of branch A After another layer of graph convolution propagation, the network quality representation of output branch A is obtained. It encodes "the state that the network should exhibit under a given electromagnetic environment".

[0054] Step S2.1.4: Define the network dynamic similarity between nodes as follows, using the trend of change rather than the absolute value to measure the similarity between nodes:

[0055] ;

[0056] in, Let i be the network dynamic similarity between the i-th node and the j-th node. Let i be the change in the network feature of the i-th node. Let j be the change in the network feature of the j-th node. Gaussian kernel bandwidth;

[0057] To avoid distorting the contrastive learning process by abnormal nodes due to mutations, anomaly detection is performed on all nodes based on the Z-score. If the i-th node satisfies... ,in, For branch A The network standard score of each node measures the degree to which the variation in network characteristics of that node deviates from the global mean. This represents the average of the changes in network characteristics across all nodes. The standard deviation of the changes in network features across all nodes is used to label the i-th node as an outlier, and its comparison weight is set according to the following formula:

[0058] ;

[0059] in, Let be the comparison weight of the i-th node in branch A. Let i be the anomaly label for the i-th node of branch A;

[0060] The cosine similarity score between the hidden representation of the i-th node in branch A and the hidden representations of all nodes is calculated as follows:

[0061] ;

[0062] ;

[0063] in, For temperature hyperparameters, in the examples, ; Let be the network quality representation of the i-th node in branch A; Let be the normalized network quality representation latent vector of the i-th node in branch A. Let be the normalized network quality representation latent vector of the j-th node in branch A. Let be the cosine similarity between the latent vectors of the i-th node and the j-th node in branch A;

[0064] The soft contrast loss for constructing branch A is as follows:

[0065] ;

[0066] in, This represents the soft contrast loss of branch A. For the i-th node in branch A and the i-th node... Cosine similarity of the latent vectors of each node; The total number of nodes;

[0067] The soft contrast loss formula states that if the i-th node and the j-th node have similar trends (i.e., high dynamic similarity), then their latent representations should be similar (i.e., high cosine similarity). The numerator is the weighted positive samples, and the denominator is all samples. By minimizing this loss, the model allows dynamically similar nodes to cluster in the latent space.

[0068] Step S2.1.5: Calculate the mass distance between adjacent nodes. ,in Let be the network quality representation of the i-th node in branch A. Let j be the network quality representation of the j-th node in branch A;

[0069] Based on the quality distance between adjacent nodes, the original edge weights are modulated to obtain the modulated edge weights as follows:

[0070] ;

[0071] in, This represents the modulated edge weight between the i-th node and the j-th node in branch A; that is, adjacent nodes with similar network quality may also have similar electromagnetic environments (because electromagnetic influences the network), so strong connections should be maintained. When the network quality difference is large, the edge weight is attenuated to reduce noise propagation. It is the quality distance modulation intensity, which controls the attenuation rate of edge weights due to differences in network quality.

[0072] Based on the modulated edge weights, a modulated graph is obtained. Two layers of graph convolution are then performed on the modulated graph, and network feature representations are injected as cross-domain biases.

[0073] ;

[0074] in, Indicates the branch A's first The electromagnetic quality representation of the layer, This represents the network quality representation of the output of branch A. The adjacency matrix after branch A modulation. for The corresponding degree matrix, For normalized graph convolution kernels, For branch A The electromagnetic quality representation of the layer, For branch A The learnable weight matrix representing the layer electromagnetic mass. For branch A Learnable weight matrix for layer cross-domain projection;

[0075] The electromagnetic mass representation of the final output branch A It encodes "the rational distribution of the electromagnetic environment under network quality constraints".

[0076] In step S2, the specific steps performed by branch B are as follows:

[0077] Step S2.2.1: In branch B, the statistical features of the IQ data packets are first encoded into a temporal representation. Then, through a reverse cross-domain attention mechanism, information is extracted from the electromagnetic features of the IQ data packets to obtain the enhanced electromagnetic feature matrix. The electromagnetic mass representation of branch B is obtained through graph convolution propagation. ;

[0078] Step S2.2.2: Construct the power vector of the i-th node Calculate the power change at the i-th node. ;in, Let represent the power vector of the i-th node at time t. Let represent the power vector of the i-th node at time t-1;

[0079] The dynamic similarity of the electromagnetic domain should not be based on abstract PCA characteristic changes, but rather on the time-varying power of each subband of the physical quantity. The electromagnetic dynamic similarity between nodes is defined as follows:

[0080] ;

[0081] in, This represents the electromagnetic dynamic similarity between the i-th node and the j-th node. This represents the power change at the j-th node; The kernel scale parameter for electromagnetic dynamic similarity controls the sensitivity of power variation differences to similarity.

[0082] For all nodes, perform anomaly detection. If the i-th node satisfies... ,in, Let the electromagnetic standard score of the i-th node in branch B be the value of the i-th node. Then, mark the i-th node as an anomalous node. The average of the power changes at all nodes. Let be the standard deviation of the power changes at all nodes; set the comparison weight of the i-th node according to the following formula:

[0083] ;

[0084] in, Let be the comparison weight for the i-th node in branch B. Let i be the anomaly label for the i-th node of branch B;

[0085] The soft contrast loss for constructing branch B is as follows:

[0086] ;

[0087]

[0088] in, This represents the soft contrast loss of branch B. Let represent the cosine similarity between the i-th node and the j-th node in branch B in the electromagnetic feature latent space. This indicates that the i-th node in branch B is related to the i-th node. Cosine similarity between nodes in the electromagnetic feature latent space This represents the normalized electromagnetic latent vector of the j-th node in branch B. Let represent the normalized electromagnetic latent vector of the i-th node in branch B. This refers to temperature hyperparameters. The total number of nodes;

[0089] Step S2.2.3: In branch B, a gating mechanism is used to calculate the gating vector for each node based on the electromagnetic mass representation of each node:

[0090] ;

[0091] in, This represents the gate vector of the i-th node in branch B. This represents the Sigmoid function. For the gated weight matrix, This is the gated bias vector; This represents the electromagnetic mass of the i-th node in branch B.

[0092] If a node experiences strong electromagnetic interference, the gating vector will approach 1 (allowing passage) in the "packet loss rate" and "latency" dimensions, and approach 0 (suppressing) in the "throughput" dimension, reflecting the performance pattern caused by the interference. In the graph convolution of the gating mechanism, the network quality representation is first multiplied element-wise by the gating vector before propagation, while simultaneously injecting the electromagnetic quality representation:

[0093] ;

[0094] in, Indicates branch B. Layer network quality representation, For activation function, Indicates branch B. Layer network quality representation, This represents the gate vector in branch B. For the first The layer network quality representation is a self-propagating learnable weight matrix, and a linear transformation is performed on the gated network quality representation. For branch B Learnable weight matrix for cross-domain projection of layers. It represents the propagation of network quality after gating. It is element-wise multiplication; It is a cross-domain injection of electromagnetic mass representation. Normalized graph convolution kernel;

[0095] Step S2.2.4: Network quality representation of the final output branch B .

[0096] Step S3: Construct the total diversity loss based on the electromagnetic domain diversity loss and the network domain diversity loss; weight and fuse the electromagnetic quality representation and network quality representation obtained from branch A and branch B respectively; concatenate the fused electromagnetic quality representation and network quality representation to obtain the joint quality representation;

[0097] The specific steps of step S3 are as follows:

[0098] Step S3.1: By using negative cosine similarity constraints to ensure that the features learned by the two branches are complementary rather than redundant, the electromagnetic domain diversity loss is constructed as follows:

[0099] ;

[0100] in, This indicates the loss of diversity in the electromagnetic domain. Let represent the electromagnetic mass of the i-th node in branch A. Let represent the electromagnetic mass of the i-th node in branch B; Let be the total number of nodes; minimizing the electromagnetic domain diversity loss is equivalent to maximizing the directional difference between the features of the two branches. When the cosine similarity approaches 0 (orthogonal), the two branches encode completely different information dimensions, and their complementarity is maximized.

[0101] The loss of network domain diversity is as follows:

[0102] ;

[0103] in, This indicates a loss of network domain diversity. Let be the network quality representation of the i-th node in branch A. Let represent the network quality of the i-th node in branch B; The total number of nodes;

[0104] The total diversity loss is constructed as follows:

[0105] ;

[0106] in, Indicates the total diversity loss, This indicates the loss of diversity in the electromagnetic domain;

[0107] Step S3.2: Concatenate the electromagnetic mass representations of the i-th node learned from branch A and branch B respectively to obtain the concatenated electromagnetic mass representation. The weights of the electromagnetic mass representation of the i-th node in branches A and B are obtained according to the following formula:

[0108] ;

[0109] in, , These are the weights representing the electromagnetic mass of the i-th node in branches A and B, respectively. This represents the splicing electromagnetic mass of the i-th node. This indicates softmax normalization. This indicates MLP (Multilayer Perceptron) learning;

[0110] The fusion strategy learned by the multilayer perceptron has node adaptability: nodes in areas with dense interference may rely more on branch B, while nodes far from the interference source may rely more on branch A.

[0111] The weighted and fused electromagnetic quality is represented as follows:

[0112] ;

[0113] in, This represents the electromagnetic quality of the fused i-th node.

[0114] The same method is used to obtain the network quality representation after fusion of the i-th node. ;

[0115] Step S3.3: Concatenate the fused electromagnetic quality representation and network quality representation of the i-th node along the dimension:

[0116] ;

[0117] in, This represents the joint quality representation of the i-th node. The network quality representation after fusion at the i-th node. The joint quality representation contains information from both electromagnetic and network dimensions, and captures the causal relationship between them through two-branch reciprocal learning, serving as the unified input for all subsequent perception tasks.

[0118] Step S4: Based on the joint quality representation, construct a shared feature encoding, and construct spectrum occupancy loss, link connectivity classification loss, and network QoS (Quality of Service) regression loss to form the total training loss; output the spectrum occupancy, link connectivity, and network QoS results, and train based on the total training loss to complete the perception of electromagnetic features and network state features.

[0119] Reference Figure 4 The specific steps of step S4 are as follows:

[0120] Step S4.1: For the joint quality representation of all nodes, share a single encoding layer to compress the 128-dimensional joint features into a 64-dimensional general representation; as shown in the following equation:

[0121] ;

[0122] in, For activation function, For the joint quality representation of all nodes, To share the weight matrix, For shared bias vectors; Encode the shared features of all nodes;

[0123] Step S4.2: Perform binary classification to determine whether each sub-band is occupied independently. Input the spliced ​​shared feature code and electromagnetic quality representation, and output... A probability vector of dimension 1, where each dimension corresponds to the spectral occupancy probability of a subband; subbands with probabilities ≤ 0.5 are defined as spectral holes, and a spectral occupancy loss is constructed:

[0124] ;

[0125] in, For spectrum occupancy loss, For the i-th node The real label on the figure For the i-th node Predicted spectrum occupancy probability of each sub-band; The total number of nodes. This represents the total number of sub-bands.

[0126] Step S4.3: Use a binary classification head to determine whether the effective rate of a node is less than 30% of the maximum rate. The structure of the binary classification head is as follows:

[0127] ;

[0128] in, This is a prediction of the link connectivity degradation at the i-th node. For learnable weight matrix, Encode the shared features of the i-th node. For bias terms;

[0129] Constructing a link connectivity classification loss:

[0130] ;

[0131] in, The loss is the classification loss for the link connectivity of the i-th node. The true connectivity label for the i-th node;

[0132] Step S4.4: Predict the latency of the i-th node Packet loss rate Effective rate Three continuous indicators, with different dimensions, are used to construct the network QoS regression loss using weighted MSE:

[0133] ;

[0134] in, For network QoS regression loss, , and Let be the latency, packet loss rate, and effective rate of the i-th node, respectively. , and These are the weights of the metrics for latency, packet loss rate, and effective rate of the i-th node, respectively. ; ; , , and These are the statistical variances of latency, packet loss rate, and effective rate on the training set, respectively, used to measure the numerical dispersion of each indicator;

[0135] Step S4.5: Introduce learnable parameters for spectrum occupancy loss respectively. Learnable parameters of link connectivity classification loss Learnable parameters of network QoS regression loss The weights for the spectrum occupancy loss are obtained after softmax normalization. Weights of link connectivity classification loss Weights of network QoS regression loss ;

[0136] ;

[0137] parameter Automatic optimization through backpropagation and gradient descent is used to achieve dynamic multi-task balancing.

[0138] Step S4.6: Construct the total training loss as follows:

[0139] ;

[0140] in, This represents the total training loss. Tasks with high losses in the early stages of training will automatically have their weights increased. As the loss of a task decreases, its weight will be transferred to other tasks, achieving dynamic multi-task balancing.

[0141] Figure 5 This is a comparison of the detection probabilities of the proposed method under different signal-to-noise ratios (SNRs). The comparison curves of detection probabilities (PDs) show that the proposed method exhibits optimal performance under all SNR conditions. In extremely low SNR environments (-15dB), the detection probabilities of traditional LSTM, GAT, and Transformer models are only between 83.45% and 85.64%, while the proposed method reaches 89.81%, indicating that it has stronger signal feature extraction capabilities under complex background noise. As the SNR increases, the detection probability of the proposed method rapidly increases and remains stable, reaching a near-ideal detection level of 99.75% at 0dB, significantly outperforming the comparison models and verifying the high efficiency of this model in network state recognition tasks.

[0142] Figure 6This is a comparison chart of the false alarm probability (PF) of the spectrum sensing method under different signal-to-noise ratios (SNRs) according to the present invention. In the experimental analysis of the PF, the present invention also demonstrates a significant competitive advantage. At the worst channel quality (-15dB), the PF of models such as LSTM and GAT all exceed 21%, leading to numerous false alarms and resource waste. The present invention successfully suppresses the PF to a low level of 1.46%. As the SNR environment improves, the PF of the present invention remains below 2%, consistently lower than the compared algorithms. This low PF characteristic reflects the model's excellent ability to identify interference noise, accurately distinguishing real signals from random noise, and ensuring the accuracy of subsequent decisions.

[0143] Figure 7 This is a comparison chart of link delay errors under different signal-to-noise ratios (SNRs) using the method of this invention. Regarding communication delay performance, the trends of different algorithms with SNR are basically consistent, i.e., as signal quality improves, processing and feedback delays gradually decrease. However, this invention maintains the lowest delay response throughout the entire observation interval. This is mainly due to its optimized model structure, which reduces redundant computational overhead, thereby shortening the closed-loop response time from state awareness to resource scheduling. This low-latency characteristic is of great significance for communication scenarios with high real-time requirements, proving that this invention can effectively ensure the rapid issuance and execution of communication commands while maintaining recognition accuracy.

[0144] Figure 8 This is a comparison chart of the packet loss rate error of the method of this invention under different signal-to-noise ratios. It can be seen that in low signal-to-noise ratio environments with severe interference, the packet loss rate of this invention is significantly lower than that of the DRL-ST and DDPG algorithms, demonstrating its strong anti-interference capability in harsh environments. As the signal-to-noise ratio increases, the packet loss rate of this invention rapidly decreases to near zero, and its rate of decrease is significantly faster than that of the comparative algorithms. This extremely low packet loss performance means that, with the optimization of this algorithm, the retransmission probability during data transmission is greatly reduced, and the system maintains high throughput while ensuring an extremely high data transmission success rate.

[0145] Figure 9This is a comparison chart of the effective rate error of the link under different signal-to-noise ratios (SNRs) using the method of this invention. Regarding the effective rate performance of the communication system, comparing this invention with two reinforcement learning algorithms, DRL-ST and DDPG, reveals that the effective rate curve of this invention consistently ranks at the top across the entire SNR range. Especially in the low SNR region, the advantages of this invention are more pronounced; at -15dB, its normalized effective rate reaches 0.682, significantly higher than DDPG's 0.635 and DRL-ST's 0.612. With improvements in the channel environment, this invention can more quickly approach the maximum transmission rate and stabilize at around 0.975 at 15dB. This indicates that through more accurate network state awareness, this invention can provide better decision support for transmission strategies, thereby maximizing the utilization of spectrum resources.

[0146] Figure 10 This is a comparison chart of the link reliability F1 score under different signal-to-noise ratios using the method of this invention. Experimental data shows that the F1 score of this invention remains above 97.5% across the entire signal-to-noise ratio range. In contrast, the F1 score curves of DRL-ST and DDPG fluctuate significantly at low signal-to-noise ratios and remain at a generally low level. This excellent balance between high precision and high recall achieved by the present invention ensures that the system can still provide highly reliable sensing results when facing dynamically changing link environments.

[0147] Therefore, this invention demonstrates outstanding performance in the network state awareness stage, with its high detection rate, low false alarm rate, and stable F1 score laying a solid foundation for perception. This enhanced perception and recognition capability directly translates into performance gains at the communication layer, achieving a significant breakthrough not only in effective data rate but also exhibiting clear advantages in reducing system latency and packet loss rate. Experimental data comprehensively proves that this invention can achieve efficient, reliable, and real-time communication goals in complex and ever-changing electromagnetic environments through deep collaboration between low-level perception and high-level control, possessing extremely high academic value and promising practical application prospects.

[0148] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for integrated electromagnetic-network situational awareness based on reciprocal learning, characterized in that, For a target network containing multiple receivers and multiple communication links, perform the following steps S1-S4 to complete the sensing of electromagnetic quality and network quality: Step S1: Based on the receiver location and the nodes of the communication link, divide the target network into nodes, establish the edges between each node, and assign edge weights to form a graph; for each node in the graph, generate electromagnetic features and network features respectively to further constitute the initial features of the node. Step S2: Input the initial features of the nodes in parallel into branches A and B. Based on the electromagnetic features and network features of all nodes, construct the electromagnetic feature matrix and the network feature matrix, respectively. In branch A, obtain the enhanced network feature matrix through the attention mechanism; obtain the network quality representation through the enhanced network feature matrix; obtain the electromagnetic quality representation through graph convolutional propagation. In branch B, obtain the electromagnetic quality representation through the reverse attention mechanism and graph convolutional propagation; obtain the network quality representation through the gating mechanism. Step S3: Construct the total diversity loss based on the electromagnetic domain diversity loss and the network domain diversity loss; weight and fuse the electromagnetic quality representation and network quality representation obtained from branch A and branch B respectively; concatenate the fused electromagnetic quality representation and network quality representation to obtain the joint quality representation; The specific steps of step S3 are as follows: Step S3.1: Construct the electromagnetic domain diversity loss as follows: ； in, This indicates the loss of diversity in the electromagnetic domain. Let represent the electromagnetic mass of the i-th node in branch A. Let represent the electromagnetic mass of the i-th node in branch B; The total number of nodes; The loss of network domain diversity is as follows: ； in, This indicates a loss of network domain diversity. Let be the network quality representation of the i-th node in branch A. Let represent the network quality of the i-th node in branch B; The total number of nodes; The total diversity loss is constructed as follows: ； in, Indicates the total diversity loss, This indicates the loss of diversity in the electromagnetic domain; Step S3.2: Concatenate the electromagnetic mass representations of the i-th node learned from branch A and branch B respectively to obtain the concatenated electromagnetic mass representation. The weights of the electromagnetic mass representation of the i-th node in branches A and B are obtained according to the following formula: ； in, , Let be the weights representing the electromagnetic mass of the i-th node in branches A and B, respectively. This represents the splicing electromagnetic mass of the i-th node. This indicates softmax normalization. This represents multilayer perceptron learning; The weighted and fused electromagnetic quality is represented as follows: ； in, This represents the electromagnetic quality of the fused i-th node. The same method is used to obtain the network quality representation after fusion of the i-th node. ; Step S3.3: Concatenate the fused electromagnetic quality representation and network quality representation of the i-th node along the dimension: ； in, This represents the joint quality representation of the i-th node. Let be the network quality representation after the fusion of the i-th node; Step S4: Based on the joint quality representation, construct a shared feature encoding, and construct spectrum occupancy loss, link connectivity classification loss, and network QoS regression loss to form the total training loss; train based on the total training loss, and output spectrum occupancy, link connectivity, and network QoS results to complete the perception of electromagnetic features and network state features; The specific steps of step S4 are as follows: Step S4.1: For the joint quality representation of all nodes, a single coding layer is shared, as follows: ； in, For activation function, For the joint quality representation of all nodes, To share the weight matrix, For shared bias vectors; Encode the shared features of all nodes; Step S4.2: Perform binary classification to determine whether each subband is occupied independently. Input the concatenated shared feature encoding and electromagnetic quality representation, and output a probability vector, with each dimension corresponding to the spectrum occupancy probability of a subband. Subbands with a probability ≤ 0.5 are defined as spectrum holes, and a spectrum occupancy loss is constructed. ； in, For spectrum occupancy loss, For the i-th node The real label on the figure For the i-th node Predicted spectrum occupancy probability of each sub-band; The total number of nodes. This represents the total number of sub-bands. Step S4.3: Use a binary classification head to determine whether the effective rate of a node is less than 30% of the maximum rate. The structure of the binary classification head is as follows: ； in, This is a prediction of the link connectivity degradation at the i-th node. For learnable weight matrix, Encode the shared features of the i-th node. For bias terms; Constructing a link connectivity classification loss: ； in, The loss is the classification loss for the link connectivity of the i-th node. The true connectivity label for the i-th node; Step S4.4: Predict the latency of the i-th node Packet loss rate Effective rate The network QoS regression loss is constructed using three continuous metrics as follows: ； in, For network QoS regression loss, , and Let be the latency, packet loss rate, and effective rate of the i-th node, respectively. , and These are the weights of the metrics for latency, packet loss rate, and effective rate of the i-th node, respectively. ; ; , , and These are the statistical variances of latency, packet loss rate, and effective rate on the training set, respectively. Step S4.5: Introduce learnable parameters for spectrum occupancy loss respectively. Learnable parameters of link connectivity classification loss Learnable parameters of network QoS regression loss The weights for the spectrum occupancy loss are obtained after softmax normalization. Weights of link connectivity classification loss Weights of network QoS regression loss ; Step S4.6: Construct the total training loss as follows: ； in, This represents the total training loss.

2. The electromagnetic-network situational awareness method based on reciprocal learning according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: Based on the receiver location and the nodes of the communication link, the nodes of the target network are adaptively partitioned using a Voronoi graph, the edges are supplemented using the Delaunay triangulation method, and edge weights are assigned to each edge to construct the graph. Step S1.2: Based on the IQ data packets collected by each receiver, electromagnetic features are generated for all nodes in the graph after filtering, feature extraction, dimensionality reduction, and spatial estimation. ; Step S1.3: Extract network features for each node. , with electromagnetic characteristics Concatenating with positional codes to form the initial features of the nodes. .

3. The electromagnetic-network situational awareness method based on reciprocal learning according to claim 2, characterized in that, The specific steps of step S1.1 are as follows: Step S1.1.1: Extract key points, including the location of each receiver and the source and destination node locations of all communication links. ,in, , These are the source and destination node positions for the nth communication link, respectively. The total number of communication links; using each key point as a seed point, the target network is divided into several units using a Voronoi diagram, where the distance from any node in each unit to the corresponding seed point is less than the distance to other seed points; Step S1.1.2: For the preset high-density traffic area, divide it into four equal-area sub-regions and add virtual nodes. The final node set is denoted as . ,in, Represents each node in the target network. The total number of nodes; Step S1.1.3: If the i-th node in the target network The j-th node The Euclidean distance between them is less than the connecting radius. Then establish an edge The Delaunay triangulation method is used, and a distance threshold is set. If the Euclidean distance between two nodes is greater than the distance threshold, the two nodes are considered not connected, and an edge is added between them. Based on the Gaussian radial basis function, each edge is assigned a weight, as shown in the following formula: ； in, For the i-th node The j-th node Edge weights between them For nodes and nodes The square of the distance, This represents an exponential function with the natural constant e as its base. The average distance between nodes is used as the scale parameter to adaptively match the spatial scale of the graph. Step S1.1.4: Construct the adjacency matrix based on edge weights Calculate the degree matrix degree matrix diagonal elements ,in, For nodes and nodes The connection strength is calculated, and self-loops are added to obtain an enhanced adjacency matrix with self-loops. ,in, To enhance the adjacency matrix, For the identity matrix, and The corresponding degree matrix Complete the construction of the graph.

4. The electromagnetic-network situational awareness method based on reciprocal learning according to claim 3, characterized in that, In step S1.2, a threshold is set. Valid subbands were selected, and features were extracted using the Fast Fourier Transform (FFT) method. Dimensionality reduction was performed using PCA, and electromagnetic features were obtained through inverse distance interpolation using a two-ray model. .

5. The electromagnetic-network situational awareness method based on reciprocal learning according to claim 4, characterized in that, In step S1.3, network features This includes sender-side features, receiver-side features, and latency features. The sender-side features include the number of packets sent. Number of bytes sent Transmission throughput Packet sending rate The characteristics of the receiving side correspond to the characteristics of the transmitting side.

6. The electromagnetic-network situational awareness method based on reciprocal learning according to claim 5, characterized in that, In step S2, the specific steps performed by branch A are as follows: Step S2.1.1: Construct an electromagnetic characteristic matrix from the electromagnetic characteristics of all nodes. After two layers of graph convolution, the electromagnetic mass representation of branch A is obtained. ,in This indicates a two-layer graph convolution process; Step S2.1.2: Construct a network feature matrix from the network features of all nodes. The query matrix of branch A is obtained by linear projection. ;in, The query matrix for branch A The weight matrix represents the electromagnetic mass of branch A. The key matrix of branch A is obtained by projecting the matrix separately. Sum matrix ; Attention score is calculated as follows: ； ； in, Let be the original attention score of the i-th node to the j-th node in branch A. For the i-th node of branch A, the first node is the first node. The original attention scores of each node. Let be the query vector for the i-th node in branch A. Let be the key vector of the j-th node in branch A. This is the scaling factor; Let the attention score between the i-th node and the j-th node in branch A be denoted as . The total number of nodes; This represents an exponential function with the natural constant e as its base. Based on network feature matrix The enhanced network feature matrix of branch A is calculated as follows: ； in, This represents the augmented network feature matrix for branch A. This represents the electromagnetic information aggregated in branch A. This represents the electromagnetic information of each node in branch A. The total number of nodes. The weight matrix for the electromagnetic information aggregated in branch A; Step S2.1.3: Enhanced network feature matrix of branch A After another layer of graph convolution propagation, the network quality representation of output branch A is obtained. ; Step S2.1.4: Define the network dynamic similarity between nodes as follows: ； in, Let i be the network dynamic similarity between the i-th node and the j-th node. Let i be the change in the network feature of the i-th node. Let j be the change in the network feature of the j-th node. Gaussian kernel bandwidth; For all nodes, anomaly detection is performed based on Z-score. If the i-th node satisfies... ,in, Let i be the network standard score of the i-th node in branch A. This represents the average of the changes in network characteristics across all nodes. The standard deviation of the changes in network features across all nodes is used to label the i-th node as an outlier, and its comparison weight is set according to the following formula: ； in, Let be the comparison weight of the i-th node in branch A. Let i be the anomaly label for the i-th node of branch A; The cosine similarity score between the hidden representation of the i-th node in branch A and the hidden representations of all nodes is calculated as follows: ； ； in, This refers to temperature hyperparameters. Let i be the network quality representation of the i-th node in branch A; Let be the normalized network quality representation latent vector of the i-th node in branch A. Let be the normalized network quality representation latent vector of the j-th node in branch A. Let be the cosine similarity between the latent vectors of the i-th node and the j-th node; The soft contrast loss for constructing branch A is as follows: ； in, This represents the soft contrast loss of branch A. For the i-th node in branch A and the i-th node... Cosine similarity of the latent vectors of each node The total number of nodes; Step S2.1.5: Calculate the mass distance between adjacent nodes. ,in Let be the network quality representation of the i-th node in branch A. Let j be the network quality representation of the j-th node in branch A; Based on the quality distance between adjacent nodes, the original edge weights are modulated to obtain the modulated edge weights as follows: ； in, This represents the modulated edge weight between the i-th node and the j-th node in branch A. It is the mass distance modulation intensity; Based on the modulated edge weights, a modulated graph is obtained. Two layers of graph convolution are then performed on the modulated graph, and network feature representations are injected as cross-domain biases. ； in, Indicates the branch A's first The electromagnetic quality representation of the layer, For activation function, This represents the network quality representation of the output of branch A. The adjacency matrix after branch A modulation. for The corresponding degree matrix, For normalized graph convolution kernels, For branch A The learnable weight matrix representing the layer electromagnetic mass. For branch A Learnable weight matrix for layer cross-domain projection; Electromagnetic mass representation of the final output branch A .

7. The electromagnetic-network situational awareness method based on reciprocal learning according to claim 6, characterized in that, In step S2, the specific steps performed by branch B are as follows: Step S2.2.1: In branch B, the statistical features of the IQ data packets are first encoded into a temporal representation. Then, through a reverse cross-domain attention mechanism, information is extracted from the electromagnetic features of the IQ data packets to obtain the enhanced electromagnetic feature matrix. The electromagnetic mass representation of branch B is obtained through graph convolution propagation. ; Step S2.2.2: Construct the power vector of the i-th node Calculate the power change at the i-th node. ;in, Let represent the power vector of the i-th node at time t. Let represent the power vector of the i-th node at time t-1; The electromagnetic dynamic similarity between nodes is defined as follows: ； in, This represents the electromagnetic dynamic similarity between the i-th node and the j-th node. This represents the power change at the j-th node. The kernel scale parameter for electromagnetic dynamic similarity; For all nodes, perform anomaly detection. If the i-th node satisfies... ,in, Let the electromagnetic standard score of the i-th node in branch B be the value of the i-th node. Then, mark the i-th node as an anomalous node. The average of the power changes at all nodes. Let be the standard deviation of the power changes at all nodes; set the comparison weight of the i-th node according to the following formula: ； in, Let be the comparison weight for the i-th node in branch B. Let i be the anomaly label for the i-th node of branch B; The soft contrast loss for constructing branch B is as follows: ； ； in, This represents the soft contrast loss of branch B. Let represent the cosine similarity between the i-th node and the j-th node in branch B in the electromagnetic feature latent space. This indicates that the i-th node in branch B is related to the i-th node. Cosine similarity between nodes in the electromagnetic feature latent space This represents the normalized electromagnetic latent vector of the j-th node in branch B. Let represent the normalized electromagnetic latent vector of the i-th node in branch B. This refers to temperature hyperparameters. Step S2.2.3: In branch B, a gating mechanism is used to calculate the gating vector for each node based on the electromagnetic mass representation of each node: ； in, This represents the gate vector of the i-th node in branch B. This represents the Sigmoid function. For the gated weight matrix, This is the gated bias vector; This represents the electromagnetic mass of the i-th node in branch B. In the gated graph convolution, the network quality representation is first multiplied element-wise by the gate vector and then propagated, while simultaneously injecting the electromagnetic quality representation: ； in, Indicates branch B. Layer network quality representation, For activation function, Indicates branch B. Layer network quality representation, This represents the gate vector in branch B. For the first The layer network quality representation is a self-propagating learnable weight matrix, and a linear transformation is performed on the gated network quality representation. For branch B Learnable weight matrix for cross-domain projection of layers. It represents the propagation of network quality after gating. It is element-wise multiplication; It is a cross-domain injection of electromagnetic mass representation. Normalized graph convolution kernel; Step S2.2.4: Network quality representation of the final output branch B .

Images