Dynamic spectrum prediction method based on time sequence knowledge graph, medium and equipment

Abstract

The invention provides a dynamic spectrum prediction method based on a time sequence knowledge graph, a medium and equipment. The method comprises the following steps: constructing a communication spectrum coordination knowledge graph framework structure fusing static knowledge and dynamic knowledge; constructing a knowledge embedding model based on a cyclic evolution network, and realizing semantic representation of entities and relationships thereof in an electromagnetic spectrum space; through a knowledge fusion method, static knowledge embedding and dynamic knowledge embedding are fused with historical communication node frequency data so as to realize feature interaction of a static knowledge graph and a dynamic knowledge graph, and features of communication nodes in a frequency domain are extracted at the same time; constructing a space-time diagram convolutional neural network model, extracting space correlation characteristics between nodes, capturing a time evolution rule of node attributes, and dynamically predicting communication frequency; according to the method, the characteristics of the communication node frequency are effectively extracted from the communication semantic relationship, and the prediction precision of the communication frequency is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a dynamic spectrum prediction method, medium, and device based on time-series knowledge graphs. Background Technology

[0002] With the rapid development of integrated urban air-ground networks, unmanned aerial vehicle (UAV)-assisted communication systems are gradually becoming a core component of smart city ecosystems. The rise of emerging application scenarios such as UAV delivery coordination, emergency response networks, and aerial base station relay requires seamless coexistence of UAVs and ground infrastructure (such as 5G small base stations and vehicle-to-everything (V2X) devices) in spectrum usage. However, while this dual-domain air-ground collaboration model improves network coverage and flexibility, it also brings unprecedented spectrum management challenges. Current spectrum management mainly relies on static spectrum allocation or rule-based dynamic adjustment mechanisms. Although dynamic spectrum prediction technology provides theoretical support for dynamic allocation by monitoring the electromagnetic environment in real time and predicting future spectrum occupancy, its input dimension is singular, relying only on data such as electromagnetic signal strength. It fails to integrate spatial distribution information (such as UAV trajectories and base station locations), scene infrastructure characteristics (such as building layout and road network density), and user behavior patterns, resulting in the model's inability to accurately capture the spatial correlation of spectrum usage and leading to significant errors in spectrum prediction. Summary of the Invention

[0003] To address the communication spectrum prediction problem in the scenario of integrated air-ground wireless ad hoc networks in complex urban environments, this invention provides a dynamic spectrum prediction method, medium, and device based on time-series knowledge graphs. It can effectively model the spatial correlation of spectrum usage and predict the communication frequencies of different communication devices, significantly reducing frequency prediction errors and improving prediction accuracy and model stability.

[0004] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0005] In a first aspect, the present invention discloses a dynamic spectrum prediction method based on a time-series knowledge graph, the method comprising the following steps:

[0006] Step 1: For the scenario of air-ground integrated wireless ad hoc network in complex urban environment, a communication spectrum prediction knowledge graph framework structure integrating static and dynamic knowledge is constructed in the form of triples of <head entity, relation, tail entity> and quadruples of <head entity, relation, tail entity, time>. The system represents the multi-dimensional complex relationship between communication nodes and complex buildings, interference sources and communication frequencies.

[0007] Step 2: Construct a knowledge graph embedding representation model for communication spectrum prediction. This model obtains the knowledge embedding of the static knowledge graph by passing it through a single-layer relational graph convolutional network. On the other hand, it updates the relation embedding of the dynamic knowledge graph by first passing it through a gated recurrent unit, then aggregating neighbor messages through a multi-layer relation-aware graph convolutional network, and finally obtaining the knowledge embedding of the dynamic knowledge graph through a time-gating mechanism. Then, it calculates the probability scores of entities and relations through a scoring function, establishes a loss function based on these probability scores, and obtains the corresponding embedding vector as the feature vector by minimizing the training loss.

[0008] Step 3: Construct a knowledge fusion model based on a multilayer perceptron. The knowledge fusion model fuses the static knowledge graph embedding vector and the dynamic knowledge graph embedding vector with the frequency data of historical communication nodes, respectively, to fuse multi-source heterogeneous data.

[0009] Step 4: Construct a spatiotemporal graph convolutional neural network model. The input data of the spatiotemporal graph convolutional neural network model are the dynamic feature vectors obtained by the knowledge fusion model and the existing historical node frequency selection sequence. The output data is the predicted communication frequency bands of different nodes. The spatiotemporal graph convolutional neural network model extracts the spatial correlation characteristics between nodes through the graph convolution module and combines the long short-term memory recurrent neural network to capture the temporal evolution law of node attributes and dynamically predict the communication frequency.

[0010] Step 5: Use the trained spatiotemporal graph convolutional neural network model to predict communication frequencies, realizing dynamic spectrum prediction based on temporal knowledge graph.

[0011] Step 1 further includes:

[0012] Step 1.1: Construct a scenario for an integrated air-ground wireless ad hoc network in a complex urban environment. This scenario consists of K single-antenna ground nodes and M single-antenna UAVs, with N = M + K communication nodes.

[0013] Step 1.2: For the scenario of an integrated air-ground wireless ad hoc network in a complex urban environment, construct a triple <head entity, relation, tail entity> that describes the relationship between communication nodes and scenario parameters. The head entity is the communication node, the relation is the scenario parameters including location, density of surrounding buildings and number of neighboring nodes, and the tail entity is the parameter value of the relation.

[0014] Construct a quadruple <head entity, relation, tail entity, time> that describes the real-time frequency usage information of the communication node. The head entity is the communication node, the relation is the real-time frequency usage information including communication frequency, communication transmission quality and data accumulation, and the tail entity is the parameter value of the relation.

[0015] Step 1.3: Construct a communication spectrum prediction knowledge graph based on triples and quadruples, integrating static and dynamic knowledge. Among them, dynamic knowledge graph A timestamped knowledge graph sequence used to describe the dynamic knowledge of real-time frequency usage by communication nodes. For the knowledge graph at time t, For the entity set of a dynamic graph, N D This represents the number of nodes in the dynamic knowledge graph. For the relation set of a dynamic graph, ε t For a set of timestamped quadruples, in ε t Any quadruple in is represented as<h,r,o,t> And h,o∈v and This indicates that at timestamp t, the relationship between the head entity h and the tail entity o is r; static knowledge graph Used to describe static knowledge related to physical scenarios and device parameters. N is the entity set of a static graph. S This represents the number of nodes in the static knowledge graph. For a static graph relation set, ε s For a set of triples in a static graph, in ε s Any triplet in the array is represented as <h s ,r s ,o s >, and h s ,o s ∈v s and Represents the head entity h s and tail entity o s There exists a relation r s .

[0016] Furthermore, in step 2, the process by which the communication spectrum prediction knowledge graph embedding representation model obtains the knowledge embedding of the static knowledge graph through a single-layer relational graph convolutional network includes the following steps:

[0017] Static knowledge graph The entities in the dataset are initialized with entity embedding vectors through random initialization, and the entity embedding matrix is ​​set. This represents the entity embedding vector corresponding to entity q, where entity q includes a head entity and a tail entity; it is obtained through matrix multiplication. By aggregating the neighbor information of entities and introducing the nonlinear activation function ReLU, the structural information in the knowledge graph is mapped to a low-dimensional vector space to obtain the static knowledge embedding of the head entity i: In the formula, ε is the entity embedding vector of the randomly initialized tail entity j; sFor a static graph, the set of triples (i, r) s (j) indicates that there is a relationship r between the head entity i and the tail entity j. s ; For relational graph convolutional networks r s The relation matrix in the model is used to assign different weights to different relations to distinguish their impact on information propagation; ReLU is the activation function used to enhance the model's expressive power; the standardization constant c... i The total number of entities connected to entity i is used to average and normalize the neighbor information to prevent the embedding of higher-order nodes from being over-amplified.

[0018] Furthermore, in step 2, the process of obtaining the knowledge embedding of the dynamic knowledge graph from the communication spectrum prediction knowledge graph embedding representation model includes the following steps:

[0019] Relations in the dynamic knowledge graph are randomly initialized to obtain relation embedding vectors, and a relation embedding matrix is ​​set. in This represents the total number of relationships in the dynamic knowledge graph. For relationship The corresponding relation embedding vector;

[0020] At timestamp t-1, the entity embedding matrix H of the dynamic knowledge graph related to relation r... t-1 Perform average pooling operations to capture the dynamic information of entities related to relation r in the previous time step. t-1 The dynamic information of entities is combined with the static features r of relationships to generate a fused relationship embedding r. t ′=[pooling(Η t-1 );r];By embedding all relations into r t 'Composition of fusion relationship embedding matrix Where d is the embedding dimension; the fusion relationship embedding matrix R is input into the gated recurrent unit. t ', embedding the relationships in the time series from the previous time step t-1 to the current time step t to obtain R. t :R t =GRU(R) t-1 ,R t ′);

[0021] Embed the entity at timestamp t-1 into matrix H t-1 The embedding matrix R is related to the time step t of the gated loop unit output. t As the input to the relation-aware graph convolutional network at timestamp t, the updated entity embedding vector is obtained after w layers of relation-aware graph convolutional network operations: Where l∈[0,w-1] is the number of layers in the relation-aware graph convolutional network; These are the l-th embeddings of entities h and o at timestamp t, respectively. These are the weight matrices for aggregating information about the surrounding nodes and for transmitting information about the nodes themselves, respectively; c is a normalization constant used to normalize the neighbor information to prevent the embedding of higher-order nodes from being over-amplified.

[0022] Using the weight matrix W3 and the bias b, H t-1 Mapping to the gating space yields the time gate U t :U t =σ(H t-1 W3+b); Settings The output embedding matrix is ​​the result of the w-layer relation-aware graph convolutional network operation; where, is the weight matrix for the time gate; b is the bias term, which adjusts the baseline of the gated signal; σ(·) is the Sigmoid function, used to compress the output to [0,1], representing the weight of the current layer information; through the time gating mechanism, the output of the multi-layer relation-aware graph convolutional network at the current time step t is dynamically adjusted. Embedded H at the previous time step t-1 The fusion ratio is used to ultimately obtain the entity embedding matrix H of the dynamic knowledge graph. t :

[0023] Furthermore, the probability scores of entities and relations are calculated using a scoring function, and a loss function is established based on these probability scores. The process of minimizing the training loss to obtain the corresponding embedding vectors as feature vectors includes the following steps:

[0024] Set up a one-dimensional convolutional network and a fully connected network to calculate the probability score p of entity o or relation r. i ((o / r) / h,r,H t ,R t ):

[0025] p i ((o / r) / h,r,H t ,R t )=σ((H t / R t )ConvTransE(h t ;(r t / o t )));

[0026] The entity prediction loss function L e Represented as:

[0027]

[0028] The relationship prediction loss function L r Represented as:

[0029]

[0030] In the formula, T is the number of timestamps in the training set; The binary label for the triple indicates whether entity o or relation r is a triple at time t+1.<h,r,o> The correct tail entity o or relation r is used to guide the model in learning the probability distribution of entity or relation predictions; |V| and |R| represent the number of elements in the entity set and relation set, respectively.

[0031] The static graph constraint loss function L, used to control the association strength between dynamic and static embeddings, is... x Represented as:

[0032]

[0033] In the formula, ρ represents the rate of angle increase, x∈[0,1,...,n], n is the historical prediction step size, used to dynamically adjust the constraint strength; the maximum angle of the two embeddings of the entity is 90°; Represents the cosine similarity between static and dynamic embeddings;

[0034] During the training phase of the knowledge embedding model, entity prediction loss, relation prediction loss, and static graph constraint loss are used to jointly constrain the model training: L = λ1L e +λ2L r +L s λ1 and λ2 are hyperparameters used to balance L e and L r The weights are determined; by minimizing the training loss, the embedding vectors of the corresponding head entity, relation, and tail entity are obtained, where the embedding vector of the head entity is used as the feature vector.

[0035] Step 3 further includes the following steps:

[0036] Embed static knowledge into the feature matrix H s The historical communication node frequency data sequence X is weighted by the weight matrix W. s Perform a weighted summation and add the corresponding bias vector b. s We obtain a weighted input; we set the activation function ReLU(·) to perform a nonlinear transformation on the weighted input to generate the output of the current layer; we recursively map layer by layer to establish the mapping relationship between input features and output values, and obtain the static fusion feature matrix: X s =ReLu(H s XW s +b s );

[0037] Embedding dynamic knowledge into matrix H t Frequency data sequence x of historical communication nodes t Through the weight matrix W d Perform a weighted summation and add the corresponding bias vector b. d We obtain a weighted input; we set the activation function ReLU(·) to perform a nonlinear transformation on the weighted input to generate the output of the current layer; we recursively map layer by layer to establish the mapping relationship between input features and output values, and obtain the dynamic fusion feature matrix: X d =ReLu(H t X t W d +b d );

[0038] X is connected directly. s and X d Connect the nodes to obtain the final dynamic feature matrix of the fused knowledge nodes: X t ′=[X s ,X d ].

[0039] Step 4 further includes the following steps:

[0040] Step 4.1: Construct the spatiotemporal graph convolutional neural network model G KG =(V KG E KG A KG ), node set V KG This represents the set of network nodes in a communication spectrum prediction model for an integrated air-to-ground wireless ad hoc network in a complex urban environment. The number of node sets is Edge set E KG This represents the set of relationships between user equipment and drones, including user equipment ω. m With user equipment ω m User equipment ω m With drones and drones With drones The relationship between them Constructing a spatiotemporal graph network G KG adjacency matrix Used to describe whether there are relationships between graph nodes;

[0041]

[0042] In the formula, A ij =1 indicates whether the m-th communication node (including drones or user equipment) communicates with other communication nodes;

[0043] Step 4.2: Set the node labels of the communication nodes in the spatiotemporal graph network to the graph model G. KG The communication frequency selected by the m-th communication node at time t is used to predict the communication frequency for the next h steps based on the observed communication frequencies of historical communication nodes over T steps:

[0044] X t+h =f KG (G KG ;X;SDKG;W KG ),

[0045] In the formula, W KG f is the trainable parameter matrix of the mapping function; c is the dimension of the feature vectors of the graph nodes; the mapping function f KG (·) is used to store the historical communication node frequency usage data sequence X = (x (t-T+1) ,x (t-T+2) ,...,x (t) Mapped to future communication frequencies in spatiotemporal networks;

[0046] Step 4.3: Based on the graph model's topology, a neighborhood aggregation strategy is used to update the features of graph nodes and their neighbors; the dynamic feature matrix X of the nodes after knowledge fusion is then utilized. t ' is the feature matrix X of the graph convolutional neural network t , represented as X t =X′ t The diagram shows the architecture of a convolutional neural network, represented as: H t =ReLu(A′) KG ReLU(A′ KG X t W1)W2)W3+b; where, in This represents the adjacency matrix after adding self-loops. for The degree matrix, W1, W2, W3 are the weights of two graph convolutional layers and one fully connected layer, and b represents the bias;

[0047] Step 4.4: Extract node attribute features H from the knowledge graph for communication spectrum prediction using a graph convolutional neural network. t The data is then fed into a Long Short-Term Memory Recurrent Neural Network (LSTN) to learn the temporal evolution of node labels and predict communication frequencies. A Graph Convolutional Neural Network (GCNN) is used to extract node relationship information from the graph structure and model the spatial dependence of node features. The LSTN is then used to capture the dynamic changes of node labels over time and generate predicted node communication frequencies.

[0048] f t =σ(W f·[H t ,h t-1 ]+b f ),

[0049] i t =σ(W i ·[H t ,h t-1 ]+b i )

[0050]

[0051] o t =σ(W o ·[H t ,h t-1 ]+b o ),

[0052] h t-1 =o t ⊙tanh(c t ),

[0053] In the formula: i t For the input gate, f t For the Gate of Oblivion, o t The output gate is used to combine the current time step information with the remembered information, the forget gate controls the retention ratio of historical information to prevent the loss of important information, and the output gate determines which information to output; σ(·) is the Sigmoid function, representing the gating signal; c t To remember the state of the cell unit, via f t ⊙c t-1 To preserve the historical cell state, while through To write the cell unit state update value to the current cell state, where h is the update value for the state of the memory cell unit. t b is the output or hidden state of the long short-term memory recurrent neural network at time t; f b i b c b o All are bias terms, ⊙ represents the Hadamard product; the feature matrix h of the last hidden state t As the output of the Long Short-Term Memory recurrent neural network, it is fed into a fully connected layer for calculation and outputs the score value of each communication node in each time slot for different communication frequencies: p t =FC(h) t );

[0054] Step 4.5: Calculate the cross-entropy between the actual and predicted frequencies of known communication nodes, and construct the loss function for the communication frequency prediction model based on the spatiotemporal graph convolutional neural network: In the formula, Let Y be the training sample set. Ω Each entry represents the frequency used by each communication node i in time slot t within time slot t ∈ {1,...,T}, c i,t C is the index; C is the total number of different communication frequencies.

[0055] Furthermore, in step 4.5, for the optimized spatiotemporal graph convolutional neural network model, the accuracy between the true values ​​and predicted values ​​of known graph nodes is used as the evaluation metric for communication frequency prediction: Where M Ω For the test sample set; y i This refers to the actual communication frequency of the node; Predicted communication frequency for nodes; For indicator functions, it means It is 1 if it is true, otherwise it is 0.

[0056] Secondly, the present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program causes a computer to execute the dynamic spectrum prediction method based on time-series knowledge graphs as described above.

[0057] Thirdly, the present invention discloses an electronic device, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the dynamic spectrum prediction method based on time-series knowledge graph as described above.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] This invention presents a dynamic spectrum prediction method, medium, and device based on a temporal knowledge graph. Addressing the communication spectrum prediction problem in complex urban environments with integrated air-to-ground wireless ad hoc networks, the invention constructs a scenario of such a network within an urban environment. By combining the spatial and positional relationships and performance parameters of communication equipment and UAV platforms, a communication channel quality assessment value is simulated, and the optimal communication frequency is allocated to communication nodes based on this value. Secondly, the invention constructs a communication spectrum coordination knowledge graph that integrates static and dynamic knowledge to describe the complex relationships between entities such as user equipment, UAV equipment, buildings, and communication frequencies. An embedding algorithm is used to generate embedded feature vectors describing these complex relationships. Then, the invention extracts node attribute features from the communication spectrum coordination knowledge graph using a graph convolutional neural network and uses these features as input to a long short-term memory recurrent neural network to learn the temporal evolution of node attributes, i.e., communication frequencies, thereby achieving communication frequency prediction. The proposed algorithm can accurately predict communication frequencies, thus enabling coordinated allocation of communication frequencies. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the main steps of a dynamic spectrum prediction method based on time-series knowledge graphs according to the present invention.

[0061] Figure 2 This is a flowchart illustrating the specific operation of a dynamic spectrum prediction method based on time-series knowledge graphs according to the present invention.

[0062] Figure 3 This is an example diagram illustrating the architecture of a dynamic spectrum prediction method based on time-series knowledge graphs according to the present invention.

[0063] Figure 4 This is a scenario example diagram of the integrated air-ground wireless self-organizing network in a complex urban environment according to the present invention.

[0064] Figure 5 This is an example diagram of the static knowledge graph of the present invention.

[0065] Figure 6 This is an example diagram of the dynamic knowledge graph of the present invention.

[0066] Figure 7 This is a diagram of the knowledge graph embedding model based on cyclic evolution networks of the present invention. Detailed Implementation

[0067] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0068] Example 1

[0069] This embodiment proposes a dynamic spectrum prediction method based on temporal knowledge graphs. By constructing a communication spectrum coordination knowledge graph that integrates static and dynamic knowledge, the system represents the multidimensional and complex relationships between ground nodes, UAVs, and communication frequencies. A knowledge graph embedding algorithm based on cyclic evolution networks is used to map entities and relationships to a high-dimensional vector space, generating embedded features that represent semantic associations. Based on this, a spatiotemporal graph convolutional neural network is used to model the embedded features. The spatial association characteristics between nodes are extracted through graph convolution modules, and a long short-term memory recurrent neural network is combined to capture the temporal evolution of node attributes, thereby achieving dynamic prediction of communication frequencies. The main steps, operation process, and architecture are as follows: Figure 1-3 As shown, the specific steps include the following.

[0070] Step 1: This invention addresses the communication spectrum coordination and allocation issues in integrated air-to-ground wireless ad hoc networks, resolving problems such as low spectrum utilization and coordination conflicts. It constructs a scenario for integrated air-to-ground wireless ad hoc networks in complex urban environments, such as... Figure 4As shown. The system consists of K single-antenna ground nodes and M single-antenna UAVs, where the sets of UAVs and ground nodes can be specified as follows: and Therefore, there are N = M + K communication nodes. Ground nodes can communicate with other ground nodes or drones, but their transmission may be affected by building obstruction and interference signals, thus interrupting the communication link.

[0071] To characterize the relationships between user equipment, drones, communication frequencies, and scenarios, this invention constructs a communication spectrum prediction knowledge graph that integrates static and dynamic knowledge. Figure 5 and Figure 6 The text in the knowledge graph is merely for characterizing the knowledge graph and has no actual impact on the technical solution of this invention. By constructing triples describing the relationship between communication nodes and scene parameters, triples such as <communication node (user and drone), location, location value>, <communication node, surrounding building density, density value>, and <communication node, number of surrounding neighbor nodes, quantity value> are formed, systematically describing the complex relationship between communication nodes and the scene, such as... Figure 5 As shown, construct a quadruple describing the real-time frequency usage information of communication nodes. The dynamic knowledge can be represented by the quadruple as: <communication node, communication frequency, frequency value, time t>, <communication node, communication transmission quality, transmission quality value, time t>, <communication node, cumulative data amount, cumulative data amount value, time t>, etc., systematically describing the complex relationship between communication nodes and communication frequencies, such as... Figure 6 As shown.

[0072] A knowledge graph for communication spectrum prediction is constructed based on the fusion of static and dynamic knowledge using triples and quadruples. This includes static knowledge graphs and dynamic knowledge graphs. Dynamic knowledge graphs describe the dynamic knowledge of real-time frequency usage by communication nodes and can be represented using graph models. To express. A knowledge graph sequence with timestamps. Let be the knowledge graph at time t, where For the entity set of a dynamic graph, N D This represents the number of nodes in the dynamic knowledge graph. For the relation set of a dynamic graph, ε t A set of timestamped quadruples. In ε t Any quadruple in can be represented as<h,r,o,t> And h,o∈v and This indicates that at timestamp t, the relationship between the head entity h and the tail entity o is r. This is a static knowledge graph, describing static knowledge related to physical scenes and device parameters, among which... For the entity set of a static graph, the same NS This represents the number of nodes in the static knowledge graph. For a static graph relation set, ε s For a set of triples in a static graph, ε s Represented as a triple <h s ,r s ,o s >, and h s ,o s ∈v s and Represents the head entity h s and tail entity o s There exists a relation r s .

[0073] Step Two: Building upon the framework of Step One, a communication spectrum prediction knowledge graph embedding representation model based on a recurrent evolutionary network is constructed. A single-layer relational graph convolutional network is used to obtain the knowledge embedding of the static knowledge graph. Simultaneously, the dynamic knowledge graph is first updated with relational embeddings through gated recurrent units, then neighbor messages are aggregated through a multi-layer relation-aware graph convolutional network. Finally, a time-gating mechanism is used to address the oversmoothing problem of the multi-layer relational graph convolutional network, generating embedding features that can represent semantic associations, such as... Figure 7 As shown.

[0074] Before using triple information from a knowledge graph for computation, the entities and relations in the triples must first be randomly abstracted into a real-valued vector. For static knowledge graphs, this results in the entity embedding vector h′. s and relation embedding vector r s Set the entity embedding matrix. This represents the entity embedding vector corresponding to entity q, where entity q includes a head entity and a tail entity; [Setting] For relational graph convolutional networks, r s A relation matrix is ​​used to assign different weights to different relations, distinguishing their impact on information propagation; the activation function is set to ReLU to enhance the model's expressive power; and a standardization constant c is set. i The total number of entities connected to entity i is used to average and normalize the neighbor information to prevent the embedding of higher-order nodes from being over-amplified; this is achieved through matrix multiplication. By aggregating the neighbor information of entities and introducing the nonlinear activation function ReLU, the structural information in the knowledge graph is mapped to a low-dimensional vector space to obtain the static knowledge embedding of the head entity i: In the formula, It is the entity embedding vector of the randomly initialized tail entity j. The output entity embedding matrix H s The vector, The input embedding matrix H is randomly initialized.′s The formula updates the head entity vector using the head entity, relation, and tail entity in the triple. Since there are also cases where the tail entity of one triple is the head entity of another triple, the formula can update the embedding vector of all entities.

[0075] Similarly, the entity at time slot t in the dynamic knowledge graph is randomly initialized as an entity embedding vector h. t The relation is randomly initialized as an embedding vector r, and the relation embedding matrix R = [r1, r2, ..., r i ], where i∈{1,2,…,N}; at timestamp t-1, the entity embedding matrix H of the dynamic knowledge graph related to relation r is... t-1 Perform average pooling to capture the dynamic information of entities related to relation r in the previous time step; set r t ′=[pooling(Η t-1 ); r], used to combine the dynamic information of entities with the static features of relations to generate the input for relation updates; set Embedded by all relations r t The GRU is composed of ', where d is the embedding dimension; the relation embedding matrix R is input into the GRU. t-1 and R′ t Embedding the relationships in the time series from the previous time step t-1 to the current time step t yields R. t :R t =GRU(R) t-1 ,R t ′);

[0076] Next, a relation-aware graph convolutional network is used to aggregate neighbor information of nodes in the dynamic knowledge graph. Let l∈[0,w-1] be the number of layers in the relation-aware graph convolutional network; set... Let h and o be the l-th level embeddings of entities h and o on timestamp t, respectively; set These are the weight matrices for aggregating information about the surrounding nodes and for transmitting information about the nodes themselves, respectively; c is a normalization constant used to average and normalize the neighbor information to prevent the embedding of higher-order nodes from being over-amplified; H is the entity embedding matrix at timestamp t-1. t-1 The embedding matrix R is related to the GRU output. t As the input to the relation-aware graph convolutional network at timestamp t, the updated entity embedding vector is obtained after w layers of relation-aware graph convolutional network operations:

[0077] Then, to address the oversmoothing issue in multi-layer relation-aware graph convolutional networks, a time-gated loop component was introduced. W3 is the weight matrix for the time gate; b is set as a bias term to adjust the baseline of the gated signal; σ(·) is set as the Sigmoid function to compress the output to [0,1], representing the weight of the current layer information; H is then processed using the weight matrix W3 and the bias b. t-1 Mapping to the gating space yields the time gate U t :U t =σ(H t-1 W3+b); Settings The output embedding matrix is ​​generated after processing by a w-layer relation-aware graph convolutional network. The output of the multi-layer relation-aware graph convolutional network at the current time step t is dynamically adjusted using a time gating mechanism. Embedded H at the previous time step t-1 The fusion ratio is used to ultimately obtain the entity embedding matrix H of the dynamic knowledge graph. t :

[0078] Set the entity prediction loss function L e Relationship prediction loss function L r This is used to ensure the model's ability to model the structure of triples; T is set to the number of timestamps in the training set; The binary label (0 or 1) of the triple indicates whether entity i or relation i is a triple at time t+1.<h,r,o> The correct tail entity o or relation r is used to guide the model in learning the probability distribution of entity or relation prediction; a one-dimensional convolutional network and a fully connected network are set up, which is replaced by ConvTransE(·), to calculate the probability score p of entity i or relation i. i ((o / r) / h,r,H t ,R t ):p i ((o / r) / h,r,H t ,R t )=σ((H t / R t )ConvTransE(h t ;(r t / o t The entity and relation prediction loss function is defined as follows: Set static graph constraint loss function L x This is used to control the association strength between dynamic and static embeddings; γ represents the rate of angle ascent, x∈[0,1,...,m], where m is the historical prediction step size, used to dynamically adjust the constraint strength; the maximum angle between the two embeddings of an entity is set to 90°; [The remaining text appears to be incomplete and requires further context.] The cosine similarity between static and dynamic embeddings is represented; the static graph constraint loss function can be defined as: During the training phase of the knowledge embedding model, entity prediction loss, relation prediction loss, and static graph constraint loss are used to jointly constrain the model training: L = λ1L e +λ2L r +L s λ1 and λ2 are hyperparameters used to balance L e and L r The weights are determined; by minimizing the training loss, the embedding vectors of the corresponding head entity, relation, and tail entity are obtained, where the embedding vector of the head entity is used as the feature vector.

[0079] Step 3: Embed static knowledge into the feature matrix H s The historical communication node frequency data sequence X is weighted by the weight matrix W. s Perform a weighted summation and add the corresponding bias vector b. s ; Set the activation function ReLU(·) to perform a nonlinear transformation on the weighted input to generate the output of the current layer; recursively map layer by layer to establish the mapping relationship between input features and output values: X s =ReLu(H s XW s +b s The dynamic knowledge embedding matrix H can be obtained through the same operations as above. t Frequency data sequence x of historical communication nodes t The fusion feature matrix: X d =ReLu(H t X t W d +b d Then connect X directly. s and X d Connect the nodes to obtain the final dynamic feature matrix of the fused knowledge nodes: X′ t =[X s ,X d ].

[0080] Step 4: To describe the multi-level relationship network between ground nodes, UAVs, and their communication frequencies, construct a spatiotemporal graph convolutional neural network model G. KG =(V KG E KG A KG ), node set V KG This represents the set of network nodes in a communication spectrum prediction model for an integrated air-to-ground wireless ad hoc network in a complex urban environment. The number of node sets is Edge set E KG This represents the set of relationships between user equipment and drones, including user equipment ω. m With user equipment ω m User equipment ωm With drones and drones With drones The relationship between them Constructing a spatiotemporal graph network G KG adjacency matrix Used to describe whether there are relationships between graph nodes;

[0081]

[0082] In the formula, A ij =1 indicates whether the m-th communication node (including drones or user equipment) communicates with other communication nodes;

[0083] Set the node labels of communication nodes in the spatiotemporal graph network to the graph model G. KG The communication frequency selected by the m-th communication node at time t; the feature vector dimension of the graph node is set to c; the mapping function f is set. KG (·) The historical communication node frequency data sequence X = (x (t-T+1) ,x (t-T+2) ,...,x (t) Mapping to future communication frequencies in a spatiotemporal network; predicting the communication frequency in the next h steps based on the observed communication frequencies of historical communication nodes over T steps:

[0084] X t+h =f KG (G KG ;X;SDKG;W KG ),

[0085] In the formula, W KG The training parameter matrix is ​​the mapping function;

[0086] Based on the graph model's topology, a neighborhood aggregation strategy is employed to update the features of graph nodes and their neighbors; the dynamic feature matrix X′ of the nodes after knowledge fusion is utilized. t X, as the feature matrix of the graph convolutional neural network t , can be represented as X t =X′ t The diagram shows the architecture of a convolutional neural network, represented as: H t =ReLu(A′) KG ReLU(A′ KG X t W1)W2)W3+b; where, in This represents the adjacency matrix after adding self-loops. for The degree matrix, W1, W2, W3 are the weights of two graph convolutional layers and one fully connected layer, and b represents the bias;

[0087] Extracting node attribute features H from the knowledge graph of communication spectrum prediction using a graph convolutional neural network. t The data is then fed into a Long Short-Term Memory Recurrent Neural Network (LSTN) to learn the temporal evolution of node labels and predict communication frequencies. A Graph Convolutional Neural Network (GCNN) is used to extract node relationship information from the graph structure and model the spatial dependence of node features. The LSTN is then used to capture the dynamic changes of node labels over time and generate predicted node communication frequencies.

[0088] f t =σ(W f ·[H t ,h t-1 ]+b f ),

[0089] i t =σ(W i ·[H t ,h t-1 ]+b i )

[0090]

[0091] o t =σ(W o ·[H t ,h t-1 ]+b o ),

[0092] h t-1 =o t ⊙tanh(c t )

[0093] In the formula: i t For the input gate, f t For the Gate of Oblivion, o t The output gate is used to combine the current time step information with the remembered information, the forget gate controls the retention ratio of historical information to prevent the loss of important information, and the output gate determines which information to output; σ(·) is the Sigmoid function, representing the gating signal; c t To remember the state of the cell unit, via f t ⊙c t-1 To preserve the historical cell state, while through To write the cell unit state update value to the current cell state, where h is the update value for the state of the memory cell unit. tb is the output or hidden state of the long short-term memory recurrent neural network at time t; f b i b c b o All are bias terms, and ⊙ represents the Hadamard product. The feature matrix h of the last hidden state... t As the output of the Long Short-Term Memory recurrent neural network, it is fed into a fully connected layer for calculation and outputs the score value of each communication node in each time slot for different communication frequencies: p t =FC(h) t ).

[0094] Calculate the cross-entropy between the actual and predicted frequencies of known communication nodes, and construct the loss function for a communication frequency prediction model based on a spatiotemporal graph convolutional neural network: In the formula, Let Y be the training sample set. Ω Each entry represents the frequency used by each communication node i in time slot t in t∈{1,...,T} (its index is c). i,t ); C represents the total number of different communication frequencies.

[0095] This invention focuses on the scenario of an integrated air-ground wireless ad hoc network in a complex urban environment, and conducts research on communication frequency prediction. This prediction process involves multiple interconnected and highly complex multi-dimensional problems, encompassing temporal, spatial, and spectral characteristics, as well as the interactions between devices. Specifically, the communication network includes ground base stations, aerial drones, and jamming devices, and each device not only possesses specific spatial location information but is also related to various attributes such as buildings and spectrum resources. Therefore, to accurately characterize the complex relationships between multi-dimensional features, a knowledge graph embedding method based on cyclic evolution networks is needed to characterize the features of different relationships, thereby achieving unified modeling of devices, communication frequencies, and various attributes. This invention uses environmental parameters, node attributes, and spectrum status as model inputs, combined with historical communication node frequency usage time series data, and employs a dynamic spectrum prediction method based on time-series knowledge graphs to plan the communication frequencies of future communication nodes.

[0096] Example 2

[0097] This embodiment proposes a computer-readable storage medium storing a computer program that causes a computer to execute the interference resource planning method based on knowledge graphs and spatiotemporal graph convolutional neural networks as described in Embodiment 1.

[0098] Example 3

[0099] This embodiment proposes an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the interference resource planning method based on knowledge graph and spatiotemporal graph convolutional neural network as described in Embodiment 1.

[0100] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0101] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, causing a series of operational steps to be executed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that run on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0104] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0105] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A dynamic spectrum prediction method based on temporal knowledge graphs, characterized in that, The method includes the following steps: Step 1: For the scenario of air-ground integrated wireless ad hoc network in complex urban environment, a communication spectrum prediction knowledge graph framework structure integrating static and dynamic knowledge is constructed in the form of triples of <head entity, relation, tail entity> and quadruples of <head entity, relation, tail entity, time>. The system represents the multi-dimensional complex relationship between communication nodes and complex buildings, interference sources and communication frequencies. Step 2: Construct a knowledge graph embedding representation model for communication spectrum prediction. This model obtains the knowledge embedding of the static knowledge graph by passing it through a single-layer relational graph convolutional network. On the other hand, it updates the relation embedding of the dynamic knowledge graph by first passing it through a gated recurrent unit, then aggregating neighbor messages through a multi-layer relation-aware graph convolutional network, and finally obtaining the knowledge embedding of the dynamic knowledge graph through a time-gating mechanism. Then, it calculates the probability scores of entities and relations through a scoring function, establishes a loss function based on these probability scores, and obtains the corresponding embedding vector as the feature vector by minimizing the training loss. Step 3: Construct a knowledge fusion model based on a multilayer perceptron. The knowledge fusion model fuses the static knowledge graph embedding vector and the dynamic knowledge graph embedding vector with the frequency data of historical communication nodes, respectively, to fuse multi-source heterogeneous data. Step 4: Construct a spatiotemporal graph convolutional neural network model. The input data of the spatiotemporal graph convolutional neural network model are the dynamic feature vectors obtained by the knowledge fusion model and the existing historical node frequency selection sequence. The output data is the predicted communication frequency bands of different nodes. The spatiotemporal graph convolutional neural network model extracts the spatial correlation characteristics between nodes through the graph convolution module and combines the long short-term memory recurrent neural network to capture the temporal evolution law of node attributes and dynamically predict the communication frequency. Step 5: Use the trained spatiotemporal graph convolutional neural network model to predict communication frequencies, realizing dynamic spectrum prediction based on temporal knowledge graph.

2. The dynamic spectrum prediction method based on time-series knowledge graphs according to claim 1, characterized in that, Step 1 further includes: Step 1.1: Construct a scenario for an integrated air-ground wireless ad hoc network in a complex urban environment. This scenario consists of K single-antenna ground nodes and M single-antenna UAVs, with N = M + K communication nodes. Step 1.2: For the scenario of an integrated air-ground wireless ad hoc network in a complex urban environment, construct a triple <head entity, relation, tail entity> that describes the relationship between communication nodes and scenario parameters. The head entity is the communication node, the relation is the scenario parameters including location, density of surrounding buildings and number of neighboring nodes, and the tail entity is the parameter value of the relation. Construct a quadruple <head entity, relation, tail entity, time> that describes the real-time frequency usage information of the communication node. The head entity is the communication node, the relation is the real-time frequency usage information including communication frequency, communication transmission quality and data accumulation, and the tail entity is the parameter value of the relation. Step 1.3: Construct a communication spectrum prediction knowledge graph based on triples and quadruples, integrating static and dynamic knowledge. Among them, dynamic knowledge graph A timestamped knowledge graph sequence used to describe the dynamic knowledge of real-time frequency usage by communication nodes. For the knowledge graph at time t, For the entity set of a dynamic graph, N D This represents the number of nodes in the dynamic knowledge graph. For the relation set of a dynamic graph, ε t For a set of timestamped quadruples, in ε t Any quadruple in is represented as<h,r,o,t> And h,o∈v and This indicates that at timestamp t, the relationship between the head entity h and the tail entity o is r; static knowledge graph Used to describe static knowledge related to physical scenarios and device parameters. N is the entity set of a static graph. S This represents the number of nodes in the static knowledge graph. For a static graph relation set, ε s For a set of triples in a static graph, in ε s Any triplet in the array is represented as <h s ,r s ,o s >, and h s ,o s ∈v s and Represents the head entity h s and tail entity o s There exists a relation r s .

3. The dynamic spectrum prediction method based on time-series knowledge graphs according to claim 1, characterized in that, In step 2, the communication spectrum prediction knowledge graph embedding representation model obtains the knowledge embedding of the static knowledge graph by passing it through a single-layer relational graph convolutional network, including the following steps: Static knowledge graph The entities in the dataset are initialized with entity embedding vectors through random initialization, and the entity embedding matrix is ​​set. This represents the entity embedding vector corresponding to entity q, where entity q includes a head entity and a tail entity; it is obtained through matrix multiplication. By aggregating the neighbor information of entities and introducing the nonlinear activation function ReLU, the structural information in the knowledge graph is mapped to a low-dimensional vector space to obtain the static knowledge embedding of the head entity i: In the formula, ε is the entity embedding vector of the randomly initialized tail entity j; s For a static graph, the set of triples (i, r) s (j) indicates that there is a relationship r between the head entity i and the tail entity j. s ; For relational graph convolutional networks r s The relation matrix in the model is used to assign different weights to different relations to distinguish their impact on information propagation; ReLU is the activation function used to enhance the model's expressive power; the standardization constant c... i The total number of entities connected to entity i is used to average and normalize the neighbor information to prevent the embedding of higher-order nodes from being over-amplified.

4. The dynamic spectrum prediction method based on time-series knowledge graphs according to claim 1, characterized in that, In step 2, the process of obtaining the knowledge embedding of the dynamic knowledge graph from the communication spectrum prediction knowledge graph embedding representation model includes the following steps: The relations in the dynamic knowledge graph are randomly initialized to obtain relation embedding vectors, and the relation embedding matrix R = [r1, r2, ..., r...]. i* ],in r represents the total number of relations in the dynamic knowledge graph. i* For relation r i* The corresponding relation embedding vector; At timestamp t-1, the entity embedding matrix H of the dynamic knowledge graph related to relation r... t-1 Perform average pooling operations to capture the dynamic information of entities related to relation r in the previous time step. t-1 The dynamic information of entities is combined with the static features r of relationships to generate a fused relationship embedding r. t ′=[pooling(Η t-1 );r];By embedding all relations into r t 'Composition of fusion relationship embedding matrix Where d is the embedding dimension; the fusion relationship embedding matrix R is input into the gated recurrent unit. t ', embedding the relationships in the time series from the previous time step t-1 to the current time step t to obtain R. t :R t =GRU(R) t-1 ,R t ′); Embed the entity at timestamp t-1 into matrix H t-1 The embedding matrix R is related to the time step t of the gated loop unit output. t As the input to the relation-aware graph convolutional network at timestamp t, the updated entity embedding vector is obtained after w layers of relation-aware graph convolutional network operations: Where l∈[0,w-1] is the number of layers in the relation-aware graph convolutional network; These are the l-th embeddings of entities h and o at timestamp t, respectively. These are the weight matrices for aggregating information about the surrounding nodes and for transmitting information about the nodes themselves, respectively; c is a normalization constant used to normalize the neighbor information to prevent the embedding of higher-order nodes from being over-amplified. Using the weight matrix W3 and the bias b, H t-1 Mapping to the gating space yields the time gate U t :U t =σ(H t-1 W3+b); Settings The output embedding matrix is ​​the result of the w-layer relation-aware graph convolutional network operation; where, is the weight matrix for the time gate; b is the bias term, which adjusts the baseline of the gated signal; σ(·) is the Sigmoid function, used to compress the output to [0,1], representing the weight of the current layer information; through the time gating mechanism, the output of the multi-layer relation-aware graph convolutional network at the current time step t is dynamically adjusted. Embedded H at the previous time step t-1 The fusion ratio is used to ultimately obtain the entity embedding matrix H of the dynamic knowledge graph. t :

5. The dynamic spectrum prediction method based on time-series knowledge graphs according to claim 1, characterized in that, The process of calculating probability scores for entities and relationships using a scoring function, establishing a loss function based on these probability scores, and minimizing the training loss to obtain the corresponding embedding vectors as feature vectors includes the following steps: Set up a one-dimensional convolutional network and a fully connected network to calculate the probability score p of entity o or relation r. i ((o / r) / h,r,H t ,R t ): p i ((o / r) / h,r,H t ,R t )=σ((H t / R t )ConvTransE(h t ;(r t / o t ))); The entity prediction loss function L e Represented as: The relationship prediction loss function L r Represented as: In the formula, T is the number of timestamps in the training set; The binary label for the triple indicates whether entity o or relation r is a triple at time t+1.<h,r,o> The correct tail entity o or relation r is used to guide the model in learning the probability distribution of entity or relation predictions; |V| and |R| represent the number of elements in the entity set and relation set, respectively. The static graph constraint loss function L, used to control the association strength between dynamic and static embeddings, is... x Represented as: In the formula, ρ represents the rate of angle increase, x∈[0,1,...,n], n is the historical prediction step size, used to dynamically adjust the constraint strength; the maximum angle of the two embeddings of the entity is 90°; Represents the cosine similarity between static and dynamic embeddings; During the training phase of the knowledge embedding model, entity prediction loss, relation prediction loss, and static graph constraint loss are used to jointly constrain the model training: L = λ1L e +λ2L r +L s λ1 and λ2 are hyperparameters used to balance L e and L r The weights are determined; by minimizing the training loss, the embedding vectors of the corresponding head entity, relation, and tail entity are obtained, where the embedding vector of the head entity is used as the feature vector.

6. The dynamic spectrum prediction method based on time-series knowledge graph according to claim 1, characterized in that, Step 3 further includes the following steps: Embed static knowledge into the feature matrix H s The historical communication node frequency data sequence X is weighted by the weight matrix W. s Perform a weighted summation and add the corresponding bias vector b. s We obtain a weighted input; we set the activation function ReLU(·) to perform a nonlinear transformation on the weighted input to generate the output of the current layer; we recursively map layer by layer to establish the mapping relationship between input features and output values, and obtain the static fusion feature matrix: X s =ReLu(H s XW s +b s ); Embedding dynamic knowledge into matrix H t Frequency data sequence x of historical communication nodes t Through the weight matrix W d Perform a weighted summation and add the corresponding bias vector b. d We obtain a weighted input; we set the activation function ReLU(·) to perform a nonlinear transformation on the weighted input to generate the output of the current layer; we recursively map layer by layer to establish the mapping relationship between input features and output values, and obtain the dynamic fusion feature matrix: X d =ReLu(H t X t W d +b d ); X is connected directly. s and X d Connect the nodes to obtain the final dynamic feature matrix of the fused knowledge nodes: X t =[X s ,X d ].

7. The dynamic spectrum prediction method based on time-series knowledge graphs according to claim 1, characterized in that, Step 4 further includes the following steps: Step 4.1: Construct the spatiotemporal graph convolutional neural network model G KG =(V KG E KG A KG ), node set V KG This represents the set of network nodes in a communication spectrum prediction model for an integrated air-to-ground wireless ad hoc network in a complex urban environment. The number of node sets is Edge set E KG This represents the set of relationships between user equipment and drones, including user equipment ω. m With user equipment ω m User equipment ω m With drones and drones With drones The relationship between them Constructing a spatiotemporal graph network G KG adjacency matrix Used to describe whether there are relationships between graph nodes; In the formula, A ij =1 indicates whether the m-th communication node (including drones or user equipment) communicates with other communication nodes; Step 4.2: Set the node labels of the communication nodes in the spatiotemporal graph network to the graph model G. KG The communication frequency selected by the m-th communication node at time t is used to predict the communication frequency for the next h steps based on the observed communication frequencies of historical communication nodes over T steps: X t+h =f KG (G KG ;X;SDKG;W KG ), In the formula, W KG f is the trainable parameter matrix of the mapping function; c is the dimension of the feature vectors of the graph nodes; the mapping function f KG (·) is used to store the historical communication node frequency usage data sequence X = (x (t-T+1) ,x (t-T+2) ,...,x (t) Mapped to future communication frequencies in spatiotemporal networks; Step 4.3: Based on the graph model's topology, a neighborhood aggregation strategy is used to update the features of graph nodes and their neighbors; the dynamic feature matrix X of the nodes after knowledge fusion is then utilized. t ' is the feature matrix X of the graph convolutional neural network t , represented as X t =X′ t The diagram shows the architecture of a convolutional neural network, represented as: H t =ReLu(A′) KG ReLU(A′ KG X t W1)W2)W3+b; where, in This represents the adjacency matrix after adding self-loops. for The degree matrix, W1, W2, W3 are the weights of two graph convolutional layers and one fully connected layer, and b represents the bias; Step 4.4: Extract node attribute features H from the knowledge graph for communication spectrum prediction using a graph convolutional neural network. t The data is then fed into a Long Short-Term Memory Recurrent Neural Network (LSTN) to learn the temporal evolution of node labels and predict communication frequencies. A Graph Convolutional Neural Network (GCNN) is used to extract node relationship information from the graph structure and model the spatial dependence of node features. The LSTN is then used to capture the dynamic changes of node labels over time and generate predicted node communication frequencies. f t =σ(W f ·[H t ,h t-1 ]+b f ), i t =σ(W i ·[H t ,h t-1 ]+b i ) o t =σ(W o ·[H t ,h t-1 ]+b o ), h t-1 =o t ⊙tanh(c t ), In the formula: i t For the input gate, f t For the Gate of Oblivion, o t The output gate is used to combine the current time step information with the remembered information, the forget gate controls the retention ratio of historical information to prevent the loss of important information, and the output gate determines which information to output; σ(·) is the Sigmoid function, representing the gating signal; c t To remember the state of the cell unit, through f t ⊙c t-1 To preserve historical cell states, while through To write the cell unit state update value to the current cell state, where h is the update value for the state of the memory cell unit. t b is the output or hidden state of the long short-term memory recurrent neural network at time t; f b i b c b o All are bias terms, ⊙ represents the Hadamard product; the feature matrix h of the last hidden state t As the output of the Long Short-Term Memory recurrent neural network, it is fed into a fully connected layer for calculation and outputs the score value of each communication node in each time slot for different communication frequencies: p t =FC(h t ); Step 4.5: Calculate the cross-entropy between the actual and predicted frequencies of known communication nodes, and construct the loss function for the communication frequency prediction model based on the spatiotemporal graph convolutional neural network: In the formula, Let Y be the training sample set. Ω Each entry represents the frequency used by each communication node i in time slot t within time slot t ∈ {1,...,T}, c i,t C is the index; C is the total number of different communication frequencies.

8. The dynamic spectrum prediction method based on time-series knowledge graph according to claim 7, characterized in that, In step 4.5, for the optimized spatiotemporal graph convolutional neural network model, the accuracy between the true values ​​and predicted values ​​of known graph nodes is used as the evaluation metric for communication frequency prediction. Where M Ω For the test sample set; y i This refers to the actual communication frequency of the node; Predicted communication frequency for nodes; For indicator functions, it means It is 1 if it is true, otherwise it is 0.

9. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the dynamic spectrum prediction method based on time-series knowledge graphs as described in any one of claims 1-8.

10. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the dynamic spectrum prediction method based on time-series knowledge graph as described in any one of claims 1-6.

Images