A method for reconstructing frequency usage patterns based on spectral knowledge graphs

Abstract

This application belongs to the field of satellite communication technology and discloses a frequency usage situation reconstruction method based on a spectrum knowledge graph. The method includes: Step 1, deploying a low-Earth orbit satellite constellation and collecting spatiotemporally non-uniformly discrete spectrum data; Step 2, constructing a spectrum knowledge graph based on the collected data, and representing the multidimensional associations of the spectrum situation through "entity-relationship-attribute" triples; Step 3, extracting graph structure features from the knowledge graph; Step 4, fusing the graph structure features with runaway attributes and inputting the result into an improved machine learning model for situation reconstruction; Step 5, outputting the potential spectrum situation at unobserved times. This application significantly improves the accuracy and robustness of situation reconstruction by mining deep associations in spectrum data through knowledge graphs, and reduces the dependence on continuous observation.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication technology, specifically relating to a method for reconstructing frequency usage status based on spectrum knowledge graphs. Background Technology

[0002] Low-Earth orbit (LEO) satellites, with their advantages of low orbital altitude and low deployment cost, have become an important carrier for space-based spectrum sensing. The frequency usage patterns formed by their sensing data can reflect the spectrum usage patterns of different spatiotemporal regions and radiation sources, providing crucial support for spectrum sharing and resource management. However, existing situation characterization methods struggle to effectively depict the spatiotemporal behavioral relationships between spectrum situation and time, radiation sources, and observation platforms. Due to the limited sensing time, rapid switching of observation perspectives, and irregular revisit cycles of LEO satellites over specific areas, their sensing results are essentially spatiotemporally non-uniform discrete samples of actual spectrum usage. Traditional methods are prone to information redundancy or cognitive gaps, making it difficult to support in-depth situation analysis.

[0003] Knowledge graphs, as a structured information relation descriptive semantic network, can achieve a unified representation of spatiotemporal and frequency-based multidimensional spectral situations through two types of triples: "entity-relationship-attribute" and "entity-relationship-entity." This allows previously scattered and isolated spectral data to form an organically connected whole. Knowledge graphs can represent spectral situations from multiple dimensions such as time, space, and frequency, revealing the deep-seated relationships hidden behind spectral data, thereby effectively characterizing the spatiotemporal behavioral relationships of spectral situations.

[0004] Current low-Earth orbit (LEO) satellite constellations suffer from spatiotemporally non-uniform discrete sampling of spectral sensing data, making it difficult for traditional interpolation or machine learning methods to model global correlations, resulting in insufficient accuracy in situation reconstruction. While knowledge graphs can characterize the multidimensional relationships in spectral data, they have not yet been effectively applied to situation reconstruction. Therefore, there is an urgent need for a method combining knowledge graphs and machine learning to improve the predictive ability of the situation at unobserved times. Summary of the Invention

[0005] To address the lack of spectrum situational awareness for low-Earth orbit satellites in existing technologies, this application provides a frequency situational awareness reconstruction method based on a spectrum knowledge graph. This method significantly improves the accuracy and robustness of situational awareness reconstruction by mining deep correlations in spectrum data through knowledge graph mining, and reduces the dependence on continuous observation.

[0006] To achieve the above objectives, this application employs the following technical solution:

[0007] This application presents a method for reconstructing frequency usage patterns based on a spectral knowledge graph, which includes the following steps:

[0008] Step 1: Deploy a low-Earth orbit satellite constellation to collect spatiotemporally non-uniform discrete spectrum data and form a spectrum sensing dataset;

[0009] Step 2: Construct a spectrum knowledge graph based on the collected spectrum data, and represent the spatiotemporal frequency multidimensional correlation of the spectrum situation through triples;

[0010] Step 3: Extract graph structure features from the spectrum knowledge graph, including node degree features, centrality features, and shortest path features;

[0011] Step 4: Fuse the graph structure features extracted in Step 3 with the spatiotemporal attributes of the spectrum data, and input them into the improved spectrum situation reconstruction model for situation reconstruction.

[0012] Step 5: Output the potential spectral situation at unobserved times.

[0013] A further improvement of this application is that step 1 includes the following steps:

[0014] Step 1.1: Deploy a low-Earth orbit satellite constellation to receive signals transmitted from ground stations;

[0015] Step 1.2: Record the spatiotemporal attributes of the received signal's spectral data, including timestamp, hourly cycle characteristics, weekday characteristics, latitude, longitude, and altitude;

[0016] Step 1.3: Collect the physical characteristics of the signal, including link ID, transmit power, receive frequency, and Doppler frequency offset;

[0017] Step 1.4: Label the spectral interference status of the signal to form a labeled spectral sensing dataset.

[0018] A further improvement of this application is that step 2 includes the following steps:

[0019] Step 2.1: The collected spectrum data is semi-structured spectrum data. Knowledge extraction is performed on the semi-structured spectrum data to generate entities and relationships.

[0020] Step 2.2: Construct the ontology;

[0021] Step 2.3: Map the entities and relations generated in Step 2.1 to the ontology in Step 2.2 through entity linking and knowledge mapping, perform knowledge fusion and knowledge storage, and form a spectrum knowledge graph by representing the spatiotemporal frequency multidimensional association of the spectrum situation through the "entity-relationship-attribute" triple.

[0022] A further improvement of this application is that step 3 includes the following steps:

[0023] Step 3.1: Extract node degree features: Calculate the out-degree and in-degree of each ground station node to reflect the connectivity strength of the ground station node in the spectral knowledge graph. Calculate the degree of the ground station node:

[0024]

[0025] in, This indicates an edge with the "OCCURRED_AT" relation. For time nodes; Given a set of time points, if So Otherwise, it is 0;

[0026] Step 3.2: Extract shortest path features: Calculate time nodes With ground station nodes Shortest path length between :

[0027]

[0028] Shortest path features are obtained by normalization using a Gaussian kernel function. :

[0029] ;

[0030] in, Indicates time node With ground station nodes A complete measure of the shortest path length between two points, including both cases where the path exists and cases where the path does not exist. The raw numerical metric representing the shortest path;

[0031] Step 3.3: Extract centrality features: Calculate the degree centrality of nodes to measure the centrality of ground station nodes. Importance in spectral knowledge graphs

[0032]

[0033] in, This represents the degree centrality of nodes.

[0034] A further improvement of this application is that step 4 includes the following steps:

[0035] Step 4.1: Process the timestamp, hourly cycle features, weekday features, latitude, longitude, and altitude of the spatiotemporal attributes;

[0036] Step 4.2: Concatenate the node degree features, centrality features, and shortest path features with the spatiotemporal attributes such as timestamp, hourly cycle features, weekday features, latitude, longitude, and altitude to form an enhanced feature vector;

[0037] Step 4.3: Input the enhanced feature vector formed in Step 4.2 into the improved spectral situation reconstruction model for training and prediction to complete the situation reconstruction. The improved spectral situation reconstruction model includes KG-LSTM model, KG-Transformer model and KG-XGBoost model.

[0038] A further improvement of this application is that step 4.1 processes the timestamp, hourly cycle feature, weekday feature, latitude, longitude, and altitude of the spatiotemporal attributes, including the following steps:

[0039] Step 4.1.1: Time alignment processing. The original time string in the spatiotemporal attributes is converted into a Unix timestamp value, and the time period features are extracted. The time period features include hourly period features. and week characteristics Establish time series ,in:

[0040]

[0041] in, For timestamps, It has hourly cycle characteristics. As a characteristic of the week, As a dimension, Longitude For height;

[0042] Step 4.1.2: Interference tag generation and definition of binarized interference detection function:

[0043]

[0044] in, For the detection threshold, For interference power, At that time, interference power It is 1 if it is true, otherwise it is 0. For carrier power, Noise power;

[0045] Step 4.1.3, Outlier Removal: Filtering rules based on time continuity detection: Retain data. ;

[0046] in, As an empirical threshold, This represents the standard deviation of the historical difference.

[0047] A further improvement in this application is that, in step 4.2, the formation of the enhanced feature vector is achieved in three ways:

[0048] The first method involves generating an enhanced feature vector using a KG-LSTM model: The KG-LSTM model fuses the extracted node degree features, shortest path features, and centrality features with spatiotemporal attributes through concatenation to construct a 9-dimensional enhanced feature vector.

[0049]

[0050] in, For node degree features, For the shortest path feature, The centrality feature of a node For transpose;

[0051] The second approach involves generating an enhanced feature vector using the KG-Transformer model. The KG-Transformer model employs a feature-level fusion strategy, inputting node degree features, shortest path features, and centrality features along with spatiotemporal attributes to form a 9-dimensional enhanced feature vector.

[0052]

[0053] in, For the spectral feature matrix:

[0054]

[0055] in, Indicates the degree of a node. Indicates the path length. Indicates node centrality;

[0056] The third method involves forming an enhanced feature vector using the KG-XGBoost model: The KG-XGBoost model integrates node degree features, centrality features, and shortest path features with spatiotemporal attributes through feature concatenation, forming a 9-dimensional enhanced feature vector.

[0057] A further improvement of this application is that the KG-LSTM model includes:

[0058] Dual-branch structure: a spatiotemporal feature branch with 6-dimensional input and a spectral feature branch with 3-dimensional input. The LSTM layer uses 128 units, and the gating calculation is as follows:

[0059]

[0060]

[0061]

[0062] in, For input gate, For the Gate of Oblivion For output gate, It is the Sigmoid activation function. Here is the weight matrix of the input gate. Here is the weight matrix for the forget gate. This is the weight matrix of the output gate. This is the hidden state from the previous moment. For the current input, For the bias term of the input gate, For the bias term of the forget gate, This is the bias term for the output gate;

[0063] Feature mapping layer: projects node degree features, shortest path features, and centrality features into an 8-dimensional space.

[0064]

[0065]

[0066] in, For node degree features, This represents the shortest path between nodes. The centrality feature of a node As a mapping feature, the 3D feature vector is projected into an 8D vector. This is the weight matrix. The input is the feature vector of the graph. It is the bias vector;

[0067] Gating fusion mechanism: fusing gating values and spectral eigenvectors The calculation method is as follows:

[0068]

[0069]

[0070] in, The weight matrix of the gating fusion mechanism, Spatiotemporal characteristics, For the bias term of the gating fusion mechanism;

[0071] Loss function: Weighted cross-entropy loss is used, and topological consistency constraints are introduced. The formula is as follows:

[0072]

[0073] in, The number of samples in the training set used for loss calculation, For positive sample weights, For the true label of the sample, To predict probabilities, This represents the category weight.

[0074] A further improvement of this application is that the KG-Transformer model includes:

[0075] Topology-enhanced attention mechanism: By introducing graph feature vectors into the self-attention computation, the topology-enhanced attention computation is as follows:

[0076]

[0077] in, For adjustment coefficients, For the first Each input feature vector For the first Each input feature vector For KG-enhanced query projection matrix, For KG-enhanced bond projection matrix, The dimension of the key / query vector. For the first in the spectrum knowledge graph 1 node For the first in the spectrum knowledge graph 1 node Topological relation functions;

[0078] Feature-aware projection: A linear transformation is applied to the spectral features, represented as:

[0079]

[0080] in, This indicates that the blocks are joined diagonally. The projection matrix of the spectral features. The query projection matrix for spatiotemporal features. This is the query projection matrix for the spectral features. The key projection matrix is ​​a spatiotemporal feature. The key projection matrix represents the spectral features;

[0081] Loss function: Combining cross-entropy loss and topology consistency loss, the topology consistency constraint term is:

[0082]

[0083] in, For topology consistency loss, For the first one extracted directly from the spectral knowledge graph The original spectral features of each sample are the true values ​​of the spectral features. The feature representation predicted by the KG-Transformer model is learned internally by the Transformer model for the th... The latent feature representation of each sample, Indicating the batch size, the final loss function is expressed as:

[0084]

[0085] in, This is the binary cross-entropy loss function.

[0086] A further improvement of this application is that the KG-XGBoost model includes:

[0087] Add topology consistency constraints:

[0088]

[0089] in: For feature fusion weights, For learnable projection matrices, For real labels, the first The true value of each sample For the first The original predicted values ​​for each sample. The total number of samples, The total number of decision trees, For decision tree indexing, For sample index, indicating the first... training samples, For the first The regularization term of a decision tree measures the complexity of the tree. For the first Prediction function for each tree; The original spectral feature matrix extracted directly from the spectral knowledge graph without KG-XGBoost model transformation;

[0090] Feature-first strategy: During the decision tree generation stage, graph features are prioritized for splitting, and the candidate features are ranked as follows:

[0091]

[0092] Improved splitting gain calculation: Introduce weighting coefficients for spectral features, as shown in the following formula:

[0093]

[0094] For input samples and corresponding spectral features Feature concatenation is performed to obtain the enhanced feature vector. :

[0095]

[0096] Perform tree ensemble prediction, and the output is represented as:

[0097] ,

[0098] in, For learning rate, For input samples The predicted value;

[0099] By using the sigmoid function Convert to probability:

[0100]

[0101] Regularization optimization: Adjust the weight of leaf nodes and the depth of the tree.

[0102] in, For data utility gain, This represents the reward coefficient for the map features.

[0103] The beneficial effects of this application are:

[0104] 1. To address the lack of situational awareness in low-Earth orbit satellite spectrum sensing scenarios and the performance limitations of traditional data-driven situational awareness reconstruction methods in modeling complex spatiotemporal relationships, this paper proposes a method that mines deep correlations in spectrum data using knowledge graphs, achieving an accuracy improvement of 5%-7% compared to traditional methods.

[0105] 2. Make full use of the association information stored in the knowledge graph, and enhance the model's reasoning ability for the situation at unobserved times by extracting graph structure features such as the number of neighbors, centrality, and shortest path, so as to effectively explore the potential situation at unobserved times;

[0106] 3. Compared with traditional situation reconstruction methods, this application can better model the correlation and temporal dependency between situations, significantly improve the accuracy of situation reconstruction, and is suitable for low-orbit satellite discontinuous sampling scenarios, reducing the dependence on dense observations. Attached Figure Description

[0107] Figure 1 This is a flowchart of this application.

[0108] Figure 2 This is a schematic diagram of the KG-LSTM situation reconstruction results.

[0109] Figure 3 This is a schematic diagram of the situation reconstruction results using the KG-Transformer.

[0110] Figure 4 This is a schematic diagram of the KG-XGBoost situation reconstruction results. Detailed Implementation

[0111] To make the objectives, technical solutions, and advantages of this application clearer, the application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0112] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the structures and / or processing steps closely related to the solution of this application are shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0113] like Figure 1 As shown, this application presents a frequency usage situation reconstruction method based on a spectrum knowledge graph. Addressing the lack of frequency usage situation data in low-Earth orbit satellite scenarios and the performance limitations of traditional data-driven situation reconstruction methods, this application studies a frequency usage situation reconstruction method based on a spectrum knowledge graph, including the following steps:

[0114] Step 1: Deploy a low-Earth orbit satellite constellation to collect spatiotemporally non-uniformly discrete spectrum data, forming a spectrum sensing dataset. This includes the following steps:

[0115] Step 1.1: Deploy a low-Earth orbit satellite constellation to receive signals transmitted from ground stations;

[0116] Step 1.2: Record the spatiotemporal attributes of the received signal's spectral data, including timestamp, hourly cycle characteristics, weekday characteristics, latitude, longitude, and altitude;

[0117] Step 1.3: Collect the physical characteristics of the signal, including link ID, transmit power, receive frequency, and Doppler frequency offset;

[0118] Step 1.4: Label the spectral interference status of the signal to form a labeled spectral sensing dataset.

[0119] Step 2: Construct a spectrum knowledge graph based on the collected spectrum data, using triples to represent the spatiotemporal-frequency multidimensional associations of the spectrum situation; a triple (subject, verb, object) is a structured representation. The spatiotemporal-frequency multidimensional associations refer to the fact that the knowledge graph, as a structured information relationship description model, has the ability to integrate massive amounts of data, associating data across time, space, and frequency ranges. The specific steps for constructing a spectrum knowledge graph include the following:

[0120] Step 2.1: The collected spectrum data is semi-structured spectrum data. Knowledge extraction is performed on the semi-structured spectrum data to generate entities and relationships.

[0121] Step 2.2: Construct the ontology;

[0122] Step 2.3: Map the entities and relations generated in Step 2.1 to the ontology in Step 2.2 through entity linking and knowledge mapping, perform knowledge fusion and knowledge storage, and form a spectrum knowledge graph by representing the spatiotemporal frequency multidimensional association of the spectrum situation through the "entity-relationship-attribute" triple.

[0123] Step 3: Extract graph structure features from the spectral knowledge graph. These features include node degree, centrality, and shortest path characteristics. Specifically, this includes the following steps:

[0124] Step 3.1: Extract node degree features: Calculate the out-degree and in-degree of each ground station node to reflect the connectivity strength of the ground station node in the spectral knowledge graph. Calculate the degree of the ground station node:

[0125]

[0126] in, This represents the "OCCURRED_AT" relation edge, where "OCCURRED_AT" is the relation type in the spectral knowledge graph, describing the node. At the time point The association that occurs or appears For time nodes; Given a set of time points, if So Otherwise, it is 0. The degree of a node is calculated to represent the number of connections a node has with other nodes in the graph. Nodes with higher degrees usually have higher weight in the graph, and nodes with more neighbors are usually more important in the graph.

[0127] Step 3.2: Extract shortest path features: Calculate time nodes With ground station nodes Shortest path length between :

[0128]

[0129] Shortest path features are obtained by normalization using a Gaussian kernel function. :

[0130] ;

[0131] in, Indicates time node With ground station nodes A complete measure of the shortest path length between two points, including both cases where the path exists and cases where the path does not exist. Shortest path is an abbreviation for the shortest path length. The original numerical measure of the shortest path refers to converting the "original path length" into a quantitative value that reflects the closeness of the relationships between nodes. is the scaling factor for the Gaussian kernel function, controlling the decay rate of normalization. The larger, right The more significant the impact, the better in this embodiment. .

[0132] Step 3.3: Extract centrality features: Calculate the degree centrality of nodes to measure the centrality of ground station nodes. Importance in spectral knowledge graphs

[0133]

[0134] in, This represents the degree centrality of nodes.

[0135] Step 4: Fuse the graph structure features extracted in Step 3 with the spatiotemporal attributes of the spectrum data, and input them into the improved spectrum situation reconstruction model for situation reconstruction.

[0136] Step 5: Output the potential spectral situation at unobserved times. Based on the spectral knowledge graph, infer the uncollected spectral data. Unobserved times refer to the periods when satellite sensors failed to effectively collect spectral data.

[0137] Step 4 includes the following steps:

[0138] Step 4.1: Process the timestamp, hourly cycle features, weekday features, latitude, longitude, and altitude of the spatiotemporal attributes.

[0139] Step 4.1 Process the spatiotemporal attributes such as timestamp, hourly cycle feature, weekday feature, latitude, longitude, and altitude, including the following steps:

[0140] Step 4.1.1: Time alignment processing. The original time string in the spatiotemporal attributes is converted into a Unix timestamp value, and the time period features are extracted. The time period features include hourly period features. and week characteristics Establish time series ,in:

[0141]

[0142] in, For timestamps, It has hourly cycle characteristics. As a characteristic of the week, As a dimension, Longitude For height;

[0143] Step 4.1.2: Interference tag generation and definition of binarized interference detection function:

[0144]

[0145] in, For the detection threshold, For interference power, At that time, interference power It is 1 if it is true, otherwise it is 0. For carrier power, Noise power;

[0146] Step 4.1.3, Outlier Removal: Filtering rules based on time continuity detection: Retain data. ;

[0147] in, As an empirical threshold, =0.5, This represents the standard deviation of the historical difference.

[0148] Step 4.2: Concatenate the node degree features, centrality features, and shortest path features with the spatiotemporal attributes such as timestamp, hourly cycle features, weekday features, latitude, longitude, and altitude to form an enhanced feature vector.

[0149] In step 4.2, the enhanced feature vector is formed in three ways:

[0150] The first method involves generating an enhanced feature vector using a KG-LSTM model: The KG-LSTM model fuses the extracted node degree features, shortest path features, and centrality features with spatiotemporal attributes through concatenation to construct a 9-dimensional enhanced feature vector.

[0151]

[0152] in, For node degree features, For the shortest path feature, The centrality feature of a node This is a transpose. For example... Figure 2 As shown.

[0153] The second method involves forming an enhanced feature vector using the KG-Transformer model: The KG-Transformer model employs a feature-level fusion strategy, combining node degree features, shortest path features, and centrality features with spatiotemporal attributes as input to the KG-Transformer model, forming a 9-dimensional enhanced feature vector. For example... Figure 3 As shown.

[0154]

[0155] in, For the spectral feature matrix:

[0156]

[0157] in, Indicates the degree of a node. Indicates the path length. Indicates node centrality;

[0158] The third method involves using the KG-XGBoost model to create enhanced feature vectors: The KG-XGBoost model fuses node degree features, centrality features, and shortest path features with spatiotemporal attributes through feature concatenation, forming a 9-dimensional enhanced feature vector. For example... Figure 4 As shown.

[0159] Step 4.3: Input the enhanced feature vector formed in Step 4.2 into the improved spectral situation reconstruction model for training and prediction to complete the situation reconstruction. The improved spectral situation reconstruction model includes KG-LSTM model, KG-Transformer model and KG-XGBoost model.

[0160] The KG-LSTM model includes:

[0161] Dual-branch structure: a spatiotemporal feature branch with 6-dimensional input and a spectral feature branch with 3-dimensional input. The LSTM layer uses 128 units, and the gating calculation is as follows:

[0162]

[0163]

[0164]

[0165] in, For input gate, For the Gate of Oblivion For output gate, The sigmoid activation function has an output range of [0,1] and is used for probabilistic gating control. Here is the weight matrix of the input gate. Here is the weight matrix for the forget gate. This is the weight matrix of the output gate. This is the hidden state from the previous time step; it represents the state of the LSTM at the previous time step. The output contains the sequence history information up to the previous time step. For the current input, For the bias term of the input gate, For the bias term of the forget gate, This is the bias term for the output gate. , , The weight matrices of the three gates are used to weight the concatenated input vector. Perform a linear transformation. , , : Bias terms corresponding to the three gates. Used to adjust the activation threshold of the gates.

[0166] Feature mapping layer: projects node degree features, shortest path features, and centrality features into an 8-dimensional space.

[0167]

[0168]

[0169] in, For node degree features, This represents the shortest path between nodes. The centrality feature of a node As a mapping feature, the 3D feature vector is projected into an 8D vector. This is the weight matrix. The input is the feature vector of the graph. It is the bias vector;

[0170] Gating fusion mechanism: fusing gating values and spectral eigenvectors The calculation method is as follows:

[0171]

[0172]

[0173] in, The weight matrix of the gating fusion mechanism, Spatiotemporal characteristics, For the bias term of the gating fusion mechanism, To integrate the gating values ​​(dynamic weights);

[0174] Loss function: Weighted cross-entropy loss is used, and topological consistency constraints are introduced. The formula is as follows:

[0175]

[0176] in, =0.6, =0.4, The number of samples in the training set used for loss calculation, For positive sample weights, The weighting coefficient (representing interference), The true label of the sample is 0, which indicates no interference, while a value of 1 indicates interference. To predict the probability, the sample is predicted to be positive. This represents the category weight.

[0177] The KG-Transformer model includes:

[0178] Topology-enhanced attention mechanism: By introducing graph feature vectors into the self-attention computation, the topology-enhanced attention computation is as follows:

[0179]

[0180] in, =0.5, which is the adjustment coefficient. For the first Each input feature vector For the first Each input feature vector For KG-enhanced query projection matrix, For KG-enhanced bond projection matrix, The dimension of the key / query vector. For the first in the spectrum knowledge graph 1 node For the first in the spectrum knowledge graph 1 node Topological relation functions;

[0181] Feature-aware projection: A linear transformation is applied to the spectral features, represented as:

[0182]

[0183] in, This indicates that the blocks are joined diagonally. The projection matrix of the spectral features. The query projection matrix for spatiotemporal features. This is the query projection matrix for the spectral features. The key projection matrix is ​​a spatiotemporal feature. The key projection matrix represents the spectral features;

[0184] Loss function: Combining cross-entropy loss and topology consistency loss, the topology consistency constraint term is:

[0185]

[0186] in, For topology consistency loss, For the first one extracted directly from the spectral knowledge graph The original spectral features of each sample are the true values ​​of the spectral features. The feature representation predicted by the KG-Transformer model is learned internally by the Transformer model for the th... The latent feature representation of each sample, Indicating the batch size, the final loss function is expressed as:

[0187]

[0188] in, This is the binary cross-entropy loss function.

[0189] The KG-XGBoost model includes:

[0190] Add topology consistency constraints:

[0191]

[0192] in: For feature fusion weights, For learnable projection matrices, For real labels, the first The true value of each sample For the first The original model predictions for each sample. The total number of samples, The total number of decision trees, For the decision tree index, the k-th tree in the gradient boosting iteration, For sample index, indicating the first... training samples, For the first The regularization term of a decision tree measures the complexity of the tree. For the first Prediction function for each tree; The original spectral feature matrix extracted directly from the spectral knowledge graph without KG-XGBoost model transformation;

[0193] Feature-first strategy: During the decision tree generation stage, graph features are prioritized for splitting, and the candidate features are ranked as follows:

[0194]

[0195] Improved splitting gain calculation: Introduce weighting coefficients for spectral features, as shown in the following formula:

[0196]

[0197] For input samples and corresponding spectral features Feature concatenation is performed to obtain the enhanced feature vector. :

[0198]

[0199] Perform tree ensemble prediction, and the output is represented as:

[0200] ,

[0201] in, The learning rate, which controls the contribution of each tree to the final result, is a key adjustment knob for preventing overfitting and improving the model's generalization ability. , It is the model's response to the input samples The predicted value;

[0202] By using the sigmoid function Convert to probability:

[0203]

[0204] Regularization optimization: Adjust the weight of leaf nodes and the depth of the tree.

[0205] in, For standard data utility gain, This represents the reward coefficient for the map features.

[0206] In summary, this application addresses the lack of situational awareness in low-Earth orbit satellite spectrum sensing scenarios and the performance limitations of traditional data-driven situational awareness reconstruction methods in modeling complex spatiotemporal relationships. It proposes a frequency-based situational awareness reconstruction method based on a spectrum knowledge graph. This method fully utilizes the association information stored in the knowledge graph, effectively mining potential situations at unobserved times by extracting graph structure features such as neighbor count, centrality, and shortest path. Compared to traditional situational awareness reconstruction methods, the proposed scheme can better model the associations and temporal dependencies between situations, significantly improving the accuracy of situational awareness reconstruction while reducing dependence on continuous observation time.

[0207] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for reconstructing frequency usage patterns based on a spectral knowledge graph, characterized in that: The frequency usage situation reconstruction method includes the following steps: Step 1: Deploy a low-Earth orbit satellite constellation to collect spatiotemporally non-uniform discrete spectrum data and form a spectrum sensing dataset; Step 2: Construct a spectrum knowledge graph based on the collected spectrum data, and represent the spatiotemporal frequency multidimensional correlation of the spectrum situation through triples; Step 3: Extract graph structure features from the spectrum knowledge graph, including node degree features, centrality features, and shortest path features; Step 4: Fuse the graph structure features extracted in Step 3 with the spatiotemporal attributes of the spectrum data, and input them into the improved spectrum situation reconstruction model for situation reconstruction. Step 5: Output the potential spectral situation at unobserved times, where: Step 2 includes the following steps: Step 2.1: The collected spectrum data is semi-structured spectrum data. Knowledge extraction is performed on the semi-structured spectrum data to generate entities and relationships. Step 2.2: Construct the ontology; Step 2.3: Map the entities and relations generated in Step 2.1 to the ontology in Step 2.2 through entity linking and knowledge mapping, perform knowledge fusion and knowledge storage, and form a spectrum knowledge graph by representing the spatiotemporal frequency multidimensional association of the spectrum situation through entity-relationship-attribute triples. Step 3 includes the following steps: Step 3.1: Extract node degree features: Calculate the out-degree and in-degree of each ground station node to reflect the connection strength of the ground station node in the spectral knowledge graph. For each ground station node x∈V x Calculate the degree of the ground station node: Among them, E occ The "OCCURRED_AT" relation edge represents the time node; V t Let E be a set of time points, if (t,x)∈E occ ,So Otherwise, it is 0; Step 3.2: Extract the shortest path feature: Calculate the shortest path length ρ(t,x) between time node t and ground station node x: Shortest path features are obtained by normalization using a Gaussian kernel function. Where ρ(t,x) represents the complete measure of the shortest path length between time node t and ground station node x, including both cases where the path exists and where the path does not exist, SPL(.) represents the original numerical measure of the shortest path, and λ is the scaling factor of the Gaussian kernel function; Step 3.3: Extract centrality features: Calculate the degree centrality of nodes to measure the importance of ground station node x in the spectral knowledge graph. Where c(x) represents the degree centrality of nodes.

2. The method for reconstructing frequency usage patterns based on a spectral knowledge graph according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1: Deploy a low-Earth orbit satellite constellation to receive signals transmitted from ground stations; Step 1.2: Record the spatiotemporal attributes of the received signal's spectral data, including timestamp, hourly cycle characteristics, weekday characteristics, latitude, longitude, and altitude; Step 1.3: Collect the physical characteristics of the signal, including link ID, transmit power, receive frequency, and Doppler frequency offset; Step 1.4: Label the spectral interference status of the signal to form a labeled spectral sensing dataset.

3. The method for reconstructing frequency usage patterns based on a spectral knowledge graph according to claim 2, characterized in that: Step 4 includes the following steps: Step 4.1: Process the timestamp, hourly cycle features, weekday features, latitude, longitude, and altitude of the spatiotemporal attributes; Step 4.2: Concatenate the node degree features, centrality features, and shortest path features with the spatiotemporal attributes such as timestamp, hourly cycle features, weekday features, latitude, longitude, and altitude to form an enhanced feature vector; Step 4.3: Input the enhanced feature vector formed in Step 4.2 into the improved spectral situation reconstruction model for training and prediction to complete the situation reconstruction. The improved spectral situation reconstruction model includes KG-LSTM model, KG-Transformer model and KG-XGBoost model.

4. The method for reconstructing frequency usage patterns based on a spectral knowledge graph according to claim 3, characterized in that: Step 4.1 involves processing the timestamp, hourly cycle features, weekday features, latitude, longitude, and altitude of the spatiotemporal attributes, including the following steps: Step 4.1.1: Time alignment processing. The original time string in the spatiotemporal attributes is converted into a Unix timestamp value, and time period features are extracted. These time period features include hourly period features h. t and weekday characteristics d t Establish time series X t ∈R T×6 ,in: Where, τ t For timestamps, h t For hourly periodic characteristics, d t As a characteristic of the week, For dimension, λ t For longitude, α t For height; Step 4.1.2: Interference tag generation and definition of binarized interference detection function: y = I{C / NC / (N+I)≥Δ} Where Δ is the detection threshold, I is the interference power, when C / NC / (N+I)≥Δ, the interference power I is 1, otherwise it is 0, C is the carrier power, N is the noise power; Step 4.1.3, Outlier Removal: Filtering rules based on time continuity detection: Retain data. Where δ is the empirical threshold, σ Δ This represents the standard deviation of the historical difference.

5. The method for reconstructing frequency usage patterns based on a spectral knowledge graph according to claim 3, characterized in that: In step 4.2, the enhanced feature vector is formed in three ways: The first method involves generating an enhanced feature vector using a KG-LSTM model: The KG-LSTM model fuses the extracted node degree features, shortest path features, and centrality features with spatiotemporal attributes through concatenation to construct a 9-dimensional enhanced feature vector. Among them, deg(x) t ) represents the node degree feature. For the shortest path feature, c(x) t ) represents the centrality feature of a node, and T represents the transpose; The second approach involves generating an enhanced feature vector using the KG-Transformer model. The KG-Transformer model employs a feature-level fusion strategy, inputting node degree features, shortest path features, and centrality features along with spatiotemporal attributes to form a 9-dimensional enhanced feature vector. Among them, F KG For the spectral feature matrix: F KG =[k t ,r t ,c t ] Where, k t ρ represents the degree of a node. t c represents the path length. t Indicates node centrality; The third method involves forming an enhanced feature vector using the KG-XGBoost model: The KG-XGBoost model integrates node degree features, centrality features, and shortest path features with spatiotemporal attributes through feature concatenation, forming a 9-dimensional enhanced feature vector.

6. The method for reconstructing frequency usage patterns based on a spectral knowledge graph according to claim 5, characterized in that: The KG-LSTM model includes: Dual-branch structure: a spatiotemporal feature branch with 6-dimensional input and a spectral feature branch with 3-dimensional input. The LSTM layer uses 128 units, and the gating calculation is as follows: i t =σ(W i ·[h t-1 ,x t ]+b i ) f t =σ(W f ·[h t-1 ,x t ]+b f ) the t =σ(W o ·[h t-1 ,x t ]+b o ) Among them, i t For the input gate, f t For the Gate of Oblivion, o t For the output gate, σ is the Sigmoid activation function, and W... i W is the weight matrix of the input gate. f W is the weight matrix of the forget gate. o h is the weight matrix of the output gate. t-1 x is the hidden state from the previous time step. t For the current input, b i For the bias term of the input gate, b f b is the bias term for the forget gate. o This is the bias term for the output gate; Feature mapping layer: projects node degree features, shortest path features, and centrality features into an 8-dimensional space. With t =ReLU(W g g t +b g ) Among them, deg(x) t ) represents the node degree feature. For the shortest path between nodes, c(x) t ) represents the centrality feature of a node, z t For mapping features, the 3D feature vector is projected into an 8D vector, W. g Let g be the weight matrix. t Let b be the input spectral feature vector. g It is the bias vector; Gating fusion mechanism: fusing the gating value u t and the spectral eigenvector g t The calculation method is as follows: u t =σ(W u [h t ;z t ]+b u ) Mr t =in t ⊙h t +(1-u t )⊙z t Among them, W u h is the weight matrix of the gating fusion mechanism. t As a spatiotemporal feature, b u For the bias term of the gating fusion mechanism; Loss function: Weighted cross-entropy loss is used, and topological consistency constraints are introduced. The formula is as follows: Where N is the number of samples in the training set used for loss calculation, w p For positive sample weights, y t For the true label of the sample, To predict the probability, w n For category weights.

7. The method for reconstructing frequency usage patterns based on a spectral knowledge graph according to claim 5, characterized in that: The KG-Transformer model includes: Topology-enhanced attention mechanism: Graph feature weights are introduced into the self-attention calculation. The topology-enhanced attention calculation is as follows: Where λ is the adjustment coefficient, x i Let x be the feature vector of the i-th input. j Let j be the feature vector of the input. For KG-enhanced query projection matrix, For the KG-enhanced bond projection matrix, d k v represents the dimension of the key / query vector. i v is the i-th node in the spectrum knowledge graph. j Let φ be the j-th node in the spectrum knowledge graph, and let φ be the topological relation function. Feature-aware projection: A linear transformation is applied to the spectral features, represented as: in, Indicates diagonal splicing of blocks, U * ∈R 3×8 W is the projection matrix of the spectral features. Q U is the query projection matrix for spatiotemporal features. Q W is the query projection matrix of the spectral features. K U is the key projection matrix of spatiotemporal features. K The key projection matrix represents the spectral features; Loss function: Combining cross-entropy loss and topology consistency loss, the topology consistency constraint term is: Among them, L KG For topology consistency loss, This refers to the original spectral features of the i-th sample extracted directly from the spectral knowledge graph, i.e., the true value of the spectral features. Here, B represents the features predicted by the KG-Transformer model, and B represents the batch size. The final loss function is expressed as: L total =L BCE +L KG Among them, L BCE This is the binary cross-entropy loss function.

8. The method for reconstructing frequency usage patterns based on a spectral knowledge graph according to claim 5, characterized in that: The KG-XGBoost model includes: Add topology consistency constraints: Where: α is the feature fusion weight, W kg ∈R 3×3 W orig ∈R 3×3 For the learnable projection matrix, y i For the true label, the true value of the i-th sample. Let be the original predicted value of the i-th sample, n be the total number of samples, K be the total number of decision trees, k be the decision tree index, and i be the sample index, representing the i-th training sample. Ω(f k f is the regularization term for the k-th decision tree. k Let F be the prediction function for the k-th tree. kg The original spectral feature matrix extracted directly from the spectral knowledge graph without KG-XGBoost model transformation; Feature-first strategy: During the decision tree generation stage, graph features are prioritized for splitting, and the candidate features are ranked as follows: Improved splitting gain calculation: Introduce weighting coefficients for spectral features, as shown in the following formula: For input sample x∈R 6 and the corresponding spectral features f kg ∈R 3 Feature concatenation is performed to obtain the enhanced feature vector. Perform tree ensemble prediction, and the output is represented as: Where η is the learning rate. It is the predicted value of the input sample x; By using the sigmoid function Convert to probability: Regularization optimization: Adjust the weight of leaf nodes and the depth of the tree. Among them, G split (f j ) represents the data utility gain, and β represents the reward coefficient for the graph features.

Images