Multi-dimensional spatial spectrum prediction method and system based on graph convolutional neural network

Abstract

The application relates to the field of spectrum sensing prediction, and particularly discloses a multi-dimensional space spectrum prediction method and system based on a graph convolutional neural network. First, dual weights are calculated by combining electromagnetic attenuation mechanism and time sequence evolution similarity, and a dynamic spectrum topology network is adaptively constructed; then, four types of heterogeneous spectrum features, namely, space, time sequence, frequency band and transient, are hierarchically decoupled and embedded and coded, and mechanism-level feature fusion is completed. Meanwhile, a double-layer nested residual graph convolutional network is built to hierarchically mine local and global spectrum correlation features, the convolutional weights are adaptively adjusted in combination with an electromagnetic environment, and a steady-state-transient double-branch time sequence attention mechanism is used to calibrate time sequence features and correct prediction lag bias. Finally, a steady-state and transient joint loss constraint mechanism is used to optimize the model, and multi-dimensional spectrum prediction results are output. The application takes into account spectrum steady-state evolution and sudden disturbance characteristics, can adapt to complex electromagnetic environments, and realizes high-precision global multi-dimensional spectrum prediction.

Description

Technical Field

[0001] This invention relates to the field of spectrum sensing and prediction technology, specifically to a multidimensional spatial spectrum prediction method and system based on graph convolutional neural networks. Background Technology

[0002] With the rapid iteration and widespread adoption of wireless communication technologies, and the dense deployment of multi-standard, multi-band wireless services such as 5G, IoT, and satellite communication, the electromagnetic spectrum space exhibits complex characteristics including multi-dimensional heterogeneity, dynamic time-varying, regional coupling, and significant transient disturbances. As a core strategic resource for wireless communication, the efficient utilization and precise management of spectrum resources rely on accurate prediction of the spectrum status across the entire space, all time periods, and multiple frequency bands. This allows for early perception of the evolutionary patterns of spectrum occupancy, energy distribution, and interference changes, providing crucial data support for dynamic spectrum allocation, interference avoidance, and electromagnetic environment management.

[0003] Current mainstream spectrum prediction technologies are mostly based on traditional time-series prediction models, relying solely on single-point spectrum time-series data for time-dimensional prediction, which has significant technical shortcomings: First, they ignore the spatial electromagnetic propagation correlation and spectrum state coupling characteristics between global spectrum monitoring nodes, failing to uncover cross-regional and long-distance spectrum linkage patterns, and lacking spatial dimension modeling capabilities; Second, they fail to distinguish the differentiated physical attributes of spectrum spatial steady-state characteristics, temporal continuous characteristics, frequency band heterogeneous characteristics, and transient interference characteristics, resulting in feature redundancy and submersion of effective features due to mixed encoding of multiple heterogeneous features, limiting modeling accuracy; Third, traditional models lack the ability to adapt to the differentiated characteristics of steady-state spectrum baselines and transient spectrum abrupt changes, exhibiting poor adaptive performance in complex electromagnetic scenarios such as dense urban areas, sparse suburban areas, and frequency band heterogeneity, resulting in problems such as time-series modeling lag, transient disturbance prediction distortion, and uneven global prediction accuracy; Fourth, single loss constraint models cannot simultaneously consider the overall accuracy of long-term steady-state spectrum evolution and the single-point accuracy of short-term sudden changes, making it difficult to adapt to complex and dynamic real electromagnetic environments.

[0004] Graph convolutional neural networks (GNNs), with their advantage in unstructured data topology modeling, can effectively mine inter-node correlation features, providing a new approach for global spectrum spatial modeling. However, existing spectrum prediction schemes based on graph convolution generally suffer from problems such as poor topology fit, weak decoupling ability of heterogeneous features, insufficient feature mining hierarchy, low temporal calibration accuracy, and lack of scene adaptability, failing to achieve high-precision, full-scene adaptive prediction of multidimensional spatial spectra. Therefore, this invention proposes a multidimensional spatial spectrum prediction method and system based on graph convolutional neural networks. Summary of the Invention

[0005] The purpose of this invention is to provide a multidimensional spatial spectrum prediction method and system based on graph convolutional neural networks, which solves the technical problems of traditional spectrum prediction technology, such as lack of spatial modeling, heterogeneous feature mixing, poor scene adaptability, imbalance between steady-state and transient prediction accuracy, and lag in time-series modeling, and achieves high-precision multidimensional intelligent spectrum prediction that considers the entire domain, multiple frequency bands, and both long-term and short-term.

[0006] To achieve the above objectives, the technical solution provided by this invention is: a multidimensional spatial spectrum prediction method based on graph convolutional neural networks, comprising: S1. Construct an adaptive spectrum topology driven by electromagnetic mechanisms: Using each spectrum monitoring node deployed in a globally distributed manner as a topology node, calculate the spatial electromagnetic correlation weight between nodes by combining the radio electromagnetic spatial attenuation propagation mechanism, calculate the spectrum state coupling weight by combining the historical spectrum temporal evolution law of each node, adaptively fuse the spatial electromagnetic correlation weight and the spectrum state coupling weight to generate a dynamic weighted adjacency matrix, and remove invalid topology connections without electromagnetic coupling correlation to obtain a spectrum topology network that fits the real electromagnetic propagation characteristics. S2. Multidimensional Spectral Heterogeneous Feature Layered Decoupling Embedding: Collect four types of core multidimensional spectral features corresponding to each topological node: spatial steady-state features, temporal continuous features, frequency band heterogeneous features, and transient interference features. Configure independent dedicated coding branches to complete differentiated embedding coding, retain the independent physical attributes of each type of feature, and then complete feature splicing and dimension alignment through a mechanism-level feature fusion strategy to generate a high-dimensional unified spectral fusion feature vector. S3. Feature mining of dual residual adaptive graph convolutional layer: Construct an adaptive graph convolutional neural network with a dual-layer nested residual structure. The first layer of local residual graph convolutional layer mines local spectral correlation features in the neighborhood of nodes, and the second layer of global residual graph convolutional layer mines global spectral coupling and linkage features across nodes and regions. At the same time, the convolutional feature weights are dynamically and adaptively adjusted according to the real-time electromagnetic environment to select and enhance effective spectral features and suppress redundant noise features. S4. Steady-state-transient dual-branch time series dynamic calibration: A dual-branch time series attention mechanism is introduced, which assigns constant benchmark weights to the long-term steady-state baseline features of the spectrum and dynamically adaptive weights to the transient features of short-term sudden disturbances and spectrum jumps, correcting the lag bias of time series modeling and completing the fine calibration of spectrum time series features. S5. Joint Loss Constraint Model Iterative Optimization and Prediction Output: A joint loss constraint mechanism combining steady-state loss function and transient loss function is adopted to perform reverse iterative training on the model until the model converges. Based on the converged optimal model, real-time multidimensional spectrum data is input, and multidimensional spectrum prediction results for each spatial node and each target frequency band within a preset time period are output.

[0007] Furthermore, in step S1, the spatial electromagnetic correlation weight is calculated as follows: the Euclidean spatial distance is calculated based on the latitude and longitude coordinates of the spectrum monitoring node, and the spatial distance is converted into an electromagnetic attenuation coefficient by combining the wireless signal free space attenuation model. The normalized spatial electromagnetic correlation weight is obtained by mapping the magnitude of the electromagnetic attenuation coefficient. The closer the spatial distance and the smaller the electromagnetic attenuation, the greater the corresponding spatial electromagnetic correlation weight.

[0008] Further, in step S1, the calculation method of the spectrum state coupling weight is as follows: extract the spectrum energy time series sequence and frequency band occupancy time series sequence within a preset historical time period of any two spectrum monitoring nodes, calculate the evolution similarity of the two time series sequences through the cosine similarity algorithm, and obtain the spectrum state coupling weight after normalization. The higher the consistency of the spectrum evolution trend of the two nodes, the greater the spectrum state coupling weight.

[0009] Furthermore, in step S1, the adaptive fusion specifically involves: for densely deployed nodes in urban areas, reducing the proportion of spatial weight and increasing the proportion of spectrum coupling weight to avoid the problem of weight redundancy of nearby nodes; for sparsely deployed nodes in suburban areas, increasing the proportion of spatial weight and retaining weak spectrum coupling weight to maintain the correlation characteristics of the entire spectrum baseline; strengthening the coupling weight for nodes in the same frequency band and weakening the coupling weight for nodes in different frequency bands to achieve adaptive matching of topology weight scenarios.

[0010] Furthermore, in step S2, the encoding methods for the four types of core multidimensional spectral features specifically include: topological embedding encoding of spatial steady-state features to solidify the inherent spatial correlation attributes of nodes; smooth sequential embedding encoding of temporally continuous features to extract the evolution law of the spectral day-night baseline; differentiated hot embedding encoding of frequency band heterogeneous features to distinguish the differences in penetration and bandwidth characteristics of different frequency bands; and high-frequency feature enhancement embedding encoding of transient interference features to retain the detailed features of sudden noise, transient interference, and frequency band jumps.

[0011] Further, in step S3, the double-layer nested residual structure is specifically as follows: each graph convolutional layer is provided with a residual bypass connection, which fuses the original features input to the convolutional layer with the depth features output by the convolutional layer element by element; the local residual graph convolutional layer focuses on mining the features of neighborhood spectrum interference superposition, resource reuse, and synchronization fluctuation, while the global residual graph convolutional layer focuses on mining the features of cross-regional interference diffusion, global spectrum baseline shift, and long-distance spectrum linkage.

[0012] Furthermore, in step S3, the logic for dynamically and adaptively adjusting the convolutional feature weights in real-time electromagnetic environment is as follows: when the variance of the spectral temporal fluctuation is less than the preset stability threshold, it is determined to be a steady-state electromagnetic scenario, and the weights of random noise features are weakened and the weights of steady-state features of the spectral baseline are strengthened; when the variance of the spectral temporal fluctuation is greater than the preset abrupt change threshold, it is determined to be a non-stationary electromagnetic scenario, and the weights of high-frequency disturbance features are amplified and the ability to represent transient abrupt change features is strengthened.

[0013] Further, in step S4, the dual-branch temporal attention mechanism is specifically implemented as follows: a steady-state attention branch and a transient attention branch are constructed. The steady-state attention branch generates fixed weights based on long-period temporal statistical features. The transient attention branch generates dynamic weights based on short-term temporal gradient change features to adaptively adapt to sudden fluctuations in the spectrum and instantaneous occupancy jump scenarios. The fixed weights and dynamic weights are fused to complete the temporal feature calibration.

[0014] Furthermore, in step S5, the joint loss constraint mechanism is constructed as follows: the mean square error is used as the steady-state loss function to constrain the overall prediction accuracy of the long-term evolution of the spectrum baseline; the mean absolute error is used as the transient loss function to constrain the single-point prediction accuracy of the spectrum sudden change point; and the two types of losses are balanced by weighting coefficients to achieve bidirectional optimization of the prediction accuracy of steady-state and non-stationary scenarios.

[0015] This invention also provides a multidimensional spatial spectrum prediction system based on graph convolutional neural networks, used to implement the above-mentioned multidimensional spatial spectrum prediction method, comprising: The adaptive topology construction module is used to collect spatial location data and historical spectrum data of the whole-domain spectrum monitoring nodes. Based on the electromagnetic attenuation mechanism and the dual weights calculated by spectrum time series similarity, it adaptively fuses and generates a dynamic weighted topology adjacency matrix to construct a spectrum topology network that fits the real electromagnetic environment. The multi-dimensional feature decoupling and embedding module is used to decompose four types of spectral features: spatial steady state, temporal continuous, frequency band heterogeneous, and transient interference. Differentiated embedding is completed through independent coding branches to achieve mechanism-level fusion of multi-dimensional heterogeneous features and generate a unified high-dimensional feature vector. The dual residual graph convolution feature mining module is used to hierarchically mine local and global spectral coupling features through a double-layer nested residual graph convolutional network, and adaptively adjust the convolution weights according to the real-time electromagnetic scene to filter and optimize core spectral features. The time-series dynamic calibration module is used to perform differentiated weighted calibration of deep spectral features through a steady-state-transient dual-branch time-series attention mechanism, correcting the time-series prediction lag bias and adapting to steady-state spectral evolution and sudden disturbance scenarios. The model training and prediction output module is used to iteratively optimize the model through a steady-state-transient joint loss function, save the optimal model after convergence, and realize the prediction of the multi-dimensional spectrum state of the whole domain based on real-time input data, and output the prediction results of spectrum occupancy, energy intensity and interference intensity of each spatial node and each target frequency band.

[0016] The advantages of this invention compared to the prior art are: This invention constructs an adaptive spectrum topology network that integrates electromagnetic mechanisms and data-driven approaches. It calculates spatial correlation weights by combining wireless free space attenuation mechanisms and calculates spectrum coupling weights based on temporal similarity. At the same time, it achieves adaptive weight fusion for urban areas, suburbs, and scenarios with the same or different frequency bands, eliminates invalid topology connections, and ensures that the topology structure fully conforms to the real electromagnetic propagation characteristics. This solves the problems of fixed weights, poor scenario adaptability, and distorted correlation modeling in traditional topology construction.

[0017] This invention employs a hierarchical decoupling embedding strategy for four types of multidimensional spectral heterogeneous features. It configures independent coding branches for the differentiated physical attributes of four types of features: spatial, temporal, frequency band, and transient. It retains the inherent physical characteristics of each type of feature and achieves efficient fusion of heterogeneous features through mechanism-level fusion. This avoids the feature redundancy and effective information overload problems caused by traditional feature hybrid coding and significantly improves the accuracy of feature representation.

[0018] This invention designs a double-layer nested residual adaptive graph convolutional network to hierarchically mine local neighborhood spectrum correlations and global cross-regional spectrum coupling features. It avoids the feature degradation problem of deep networks by relying on the residual structure. At the same time, it realizes adaptive weight adjustment of electromagnetic scenes based on the variance of spectrum temporal fluctuations. In steady-state scenes, it strengthens baseline features, and in non-steady-state scenes, it strengthens transient change features, effectively screening effective spectrum features and suppressing redundant noise.

[0019] This invention introduces a steady-state-transient dual-branch temporal attention mechanism. By using fixed baseline weights, it ensures the stability of long-term steady-state spectrum prediction. By using dynamic adaptive weights, it adapts to short-term spectrum bursts and frequency band jumps, effectively correcting the lag bias of single temporal modeling, achieving fine calibration of spectrum temporal features, and balancing the accuracy of long-period baseline prediction with the accuracy of short-period transient change prediction.

[0020] This invention constructs a joint steady-state and transient loss constraint mechanism. It constrains the global steady-state evolution accuracy through mean square error loss and constrains the single-point accuracy of transient mutation through mean absolute error loss. It balances the prediction errors of different electromagnetic scenarios in both directions, resulting in better model training convergence and stronger generalization ability. It can achieve high-precision multi-dimensional spectrum intelligent prediction that covers the entire space, multiple frequency bands, and both short and long-term scenarios, and is suitable for complex and dynamic real electromagnetic environments. Attached Figure Description

[0021] Figure 1 This is a flowchart of the multidimensional spatial spectrum prediction method based on graph convolutional neural networks according to the present invention.

[0022] Figure 2 This is a system block diagram of the multidimensional spatial spectrum prediction system based on graph convolutional neural networks according to the present invention.

[0023] Figure 3 This is a flowchart of the adaptive fusion process for spectrum topology weights.

[0024] Figure 4 This is a schematic diagram of a double-nested residual graph convolutional network structure.

[0025] Figure 5 This is a steady-state and transient two-branch sequential attention flowchart. Detailed Implementation

[0026] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0027] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0028] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0029] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0030] The following detailed description of the multidimensional spatial spectrum prediction method and system based on graph convolutional neural networks of the present invention, with reference to the accompanying drawings, provides further insight.

[0031] Combined with appendix Figure 1-5 The specific implementation process of the multidimensional spatial spectrum prediction method and system based on graph convolutional neural networks of this invention is as follows: The first aspect of this invention provides a method for multidimensional spatial spectrum prediction based on graph convolutional neural networks, comprising the following steps: S1. Construct an adaptive spectral topology driven by electromagnetic mechanisms. Using the spectrum monitoring nodes deployed in a globally distributed manner as topology nodes, the spatial electromagnetic correlation weights between nodes are calculated by combining the radio electromagnetic spatial attenuation propagation mechanism, and the spectrum state coupling weights are calculated by combining the historical spectrum temporal evolution law of each node. The spatial electromagnetic correlation weights and spectrum state coupling weights are adaptively fused to generate a dynamic weighted adjacency matrix, and invalid topology connections without electromagnetic coupling are eliminated to obtain a spectrum topology network that conforms to the real electromagnetic propagation characteristics.

[0032] The spatial electromagnetic correlation weight is calculated based on the node spatial distance and the free-space attenuation model of wireless signals. The specific calculation method is as follows: First, based on the two spectrum monitoring nodes , Calculating Euclidean spatial distance using latitude and longitude coordinates: ; in, , They are nodes , The latitude and longitude are mapped to plane coordinates; the electromagnetic attenuation coefficient is obtained by combining the free-space electromagnetic attenuation model: ; in, Predict the target frequency band. The spatial electromagnetic correlation weights are obtained by normalizing the electromagnetic attenuation coefficients. ; in, , These represent the maximum and minimum electromagnetic attenuation coefficients for all node pairs in the global domain. The closer the spatial distance, the smaller the electromagnetic attenuation coefficient, corresponding to the spatial electromagnetic correlation weight. The larger.

[0033] The spectral state coupling weight is calculated based on the similarity of the historical spectral temporal evolution of nodes. The specific calculation method is as follows: extract any two spectral monitoring nodes. , Spectral energy time series within a preset historical period , The temporal evolution similarity between the frequency band occupancy time series and the frequency band occupancy time series is calculated using the cosine similarity algorithm: ; After normalizing the similarity, the spectral state coupling weights are obtained: ; in, , These represent the maximum and minimum values ​​of the temporal similarity between nodes across the entire domain. The higher the consistency of the temporal evolution trend of the spectrum between two nodes, the greater the spectral state coupling weight. The larger.

[0034] The specific method for adaptive fusion of dual-weighted scenarios and generation of dynamically weighted adjacency matrices is as follows: Constructing Adaptive Fusion Weights , And satisfy Two spectrum monitoring nodes , Inter-fusion weights: ; For densely deployed nodes in urban areas, set <0.5、 >0.5, reduce the spatial weight ratio and increase the spectrum coupling weight ratio to avoid the weight redundancy problem of nearby nodes; for sparsely deployed nodes in suburban areas, set >0.5、 <0.5, increase the spatial weight ratio, retain weak spectrum coupling weight, and maintain the correlation characteristics of the global spectrum baseline; strengthen the topology nodes in the same frequency band. Value selection, weakening of different frequency band topology nodes Values ​​are assigned to achieve adaptive matching of weights for different scenarios. An initial dynamic weighted adjacency matrix is ​​constructed based on the fusion weights from all spectrum monitoring nodes. ,in This represents the total number of spectrum monitoring nodes across the entire region.

[0035] The specific method for eliminating invalid topology connections is as follows: preset an electromagnetic coupling association threshold. Iterate through all weight elements in the adjacency matrix. Determine the node , No effective electromagnetic coupling correlation, set By eliminating the corresponding topological connections, a dynamic spectrum topology network that conforms to the actual electromagnetic propagation characteristics is finally obtained.

[0036] S2, Multidimensional Spectral Heterogeneous Feature Hierarchical Decoupling Embedding Four core multidimensional spectral features corresponding to each topological node are collected: spatial steady-state features, temporal continuous features, frequency band heterogeneous features, and transient interference features. Each feature is configured with an independent dedicated coding branch to complete differentiated embedding coding, preserving the independent physical properties of each feature. Then, a mechanism-level feature fusion strategy is used to complete feature splicing and dimension alignment, generating a high-dimensional unified spectral fusion feature vector.

[0037] The spatial steady-state characteristics include the planar coordinates of each monitoring node, deployment density, and the inherent properties of spatial electromagnetic baseline attenuation. Topological embedding encoding is employed, and the encoding formula is as follows: ; in, The dynamically weighted adjacency matrix constructed in step S1, This is the original matrix of spatial steady-state characteristics. , These are the spatially encoded learnable weight matrix and the bias vector, respectively. It is a non-linear activation function. It is a spatial steady-state encoding feature that solidifies the inherent spatial association physical attributes of nodes.

[0038] The temporally continuous features include the temporal evolution data of node long-period spectral energy and frequency band occupancy. Smooth serialization embedding encoding is used, and feature mapping is completed after suppressing temporal noise through moving average filtering. The encoding formula is: ; in, It is a time-series continuous original feature sequence. This is a time-series smoothing filter function. , These are the temporal coding weight matrix and the bias vector, respectively. For layer normalization operation, To extract the evolution pattern of the spectral day-night periodic baseline, the feature is encoded as a time-series continuous encoding feature.

[0039] The heterogeneous features of frequency bands include heterogeneous attributes such as bandwidth, penetration loss, and propagation loss across different frequency bands. Differential one-hot embedding coding is employed, first performing one-hot coding on the frequency band category and parameters, and then completing the feature upscaling mapping through a fully connected layer. The coding formula is as follows: ; in, This represents the original characteristics of frequency band heterogeneity. It is a one-hot encoding function. It is a fully connected mapping layer. It provides heterogeneous coding features for frequency bands, accurately distinguishing the inherent differences in propagation characteristics of different frequency bands.

[0040] Transient interference features include short-time spectral burst noise, transient interference, and frequency band jump data. High-frequency feature enhancement embedding coding is employed. High-frequency abrupt change features are extracted through high-pass filtering and then encoded. The encoding formula is as follows: ; in, The original characteristics of transient interference, This is a high-frequency filtering function. , These are the transient encoding weight matrix and the bias vector, respectively. It is a transient interference coding feature that fully preserves the physical details of instantaneous changes in the spectrum.

[0041] The mechanism-level feature fusion strategy is as follows: based on the physical dimension attributes of four types of encoded features, a dimension mapping matrix is ​​used to... , , , Unified mapping to the same feature dimension D Then, a mechanism-weighted splicing method is used to complete feature fusion. The fusion formula is as follows: ; in, , , , The four types of feature fusion weights, which are adapted to the electromagnetic propagation mechanism, satisfy the following: ; The feature is concatenated and processed dimension by dimension. This is the final high-dimensional unified spectral fusion feature vector.

[0042] S3, Dual Residual Adaptive Convolutional Integral Layer Feature Mining An adaptive graph convolutional neural network with a two-layer nested residual structure is constructed. The first layer of local residual graph convolutional layer mines local spectral correlation features in the neighborhood of nodes, and the second layer of global residual graph convolutional layer mines global spectral coupling and linkage features across nodes and regions. At the same time, the convolutional feature weights are dynamically and adaptively adjusted according to the real-time electromagnetic environment to select and enhance effective spectral features and suppress redundant noise features.

[0043] Local residual map convolutional layers are used to mine local spectral correlation features within the neighborhood of nodes. The formula for local residual map convolution is: ; in, This is the dynamically weighted adjacency matrix after normalization in step S1. The high-dimensional unified spectrum fusion feature vector output from step S2, The weight matrix is ​​a locally convolutional learnable matrix. It is a non-linear activation function. As the neighborhood features output by the local residual convolution, this layer focuses on mining the local correlation features of spectral interference superposition, spectral resource reuse, and spectral synchronization fluctuations between neighborhood nodes.

[0044] Global residual graph convolutional layers are used to mine cross-node, cross-region global spectral coupling and linkage features. The formula for global residual graph convolution operation is: ; in, The global convolutional learnable weight matrix. The global coupling features are the output of the global residual convolution. Based on the local features, this layer further mines the global coupling features of cross-regional interference diffusion, global spectrum baseline shift, and long-distance spectrum linkage.

[0045] Simultaneously, based on the real-time electromagnetic environment's spectral temporal fluctuation variance, the convolutional feature weights are dynamically and adaptively adjusted. First, the fluctuation variance of the spectral feature temporal sequence within a preset short-time window is calculated: ; in, For the first in the window The spectral characteristic values ​​at each time point, The mean of the spectral features within the window. Number of sampling points in the window. Preset stability threshold. With mutation threshold And satisfy ;when The scenario is determined to be a steady-state electromagnetic scene, and a steady-state weighting coefficient is set. Noise suppression coefficient The optimized features are: (1) Enhanced steady-state feature weights of the spectral baseline and weakened random noise feature weights. ; when The scene is determined to be a non-stationary electromagnetic scene, and a transient enhancement weighting coefficient is set. By amplifying the weights of high-frequency disturbances and transient changes, the ability to represent spectral features in non-stationary scenarios is enhanced, resulting in optimized features: ; in, To extract spectral redundancy noise features, To address the high-frequency abrupt changes in the spectrum, the system ultimately achieves adaptive feature mining that effectively enhances spectral features and suppresses redundant noise features.

[0046] S4, Steady-state-transient dual-branch timing dynamic calibration A dual-branch temporal attention mechanism is introduced, which assigns constant baseline weights to the long-term steady-state baseline features of the spectrum and dynamically adaptive weights to the transient features of short-term sudden disturbances and spectrum jumps, thereby correcting the lag bias in temporal modeling and completing the fine calibration of the spectral temporal features.

[0047] In the two-branch temporal attention mechanism, the steady-state temporal attention branch is used to capture long-term steady-state baseline features of the spectrum. It generates constant baseline attention weights based on the long-period temporal statistical mean, and the weight calculation formula is as follows: ; in, For constant steady-state benchmark weights, To preset the long-period time-series sampling window length, for The global coupled spectrum characteristics output at time step S3 have a fixed weight over a long period, which solidifies the stable evolution characteristics of the spectrum baseline.

[0048] The transient temporal attention branch in the two-branch temporal attention mechanism is used to capture short-term sudden disturbances and spectral jump transient features. It generates dynamic adaptive attention weights based on the gradient change rate of short-term temporal features. First, the gradient of the temporal features within the short-term window is calculated: ; The dynamic transient weights are then obtained by normalizing the gradient fluctuation values: ; in, The dynamic transient adaptive weights at time t. To preset the short-time sampling window length, It is dynamically updated in real time to detect sudden changes and disturbances in the time spectrum.

[0049] After completing the calculation of the two-branch weights, the time series features are calibrated using a two-weight fusion strategy. The final calibrated time series feature formula is as follows: ; in, To refine the spectral timing characteristics after calibration, This represents the long-term steady-state baseline characteristics of the spectrum. To address short-term spectral bursts and frequency band jumps, a constant baseline weight is used to ensure the stability of the time series baseline prediction. Dynamic adaptive weights are used to adapt to instantaneous spectral change scenarios, effectively correcting the lag bias in single time series modeling and completing the fine calibration of spectral time series features.

[0050] S5. Iterative Optimization and Prediction Output of Joint Loss Constraint Model A joint loss constraint mechanism combining steady-state and transient loss functions is adopted to perform reverse iterative training on the model until the model converges. Based on the converged optimal model, real-time multidimensional spectrum data is input, and multidimensional spectrum prediction results for each spatial node and each target frequency band within a preset time period are output.

[0051] A joint loss constraint mechanism is constructed by fusing the steady-state loss function and the transient loss function. The steady-state loss function adopts the mean square error loss function to constrain the overall prediction accuracy of the long-term steady-state baseline evolution of the spectrum. The calculation formula is as follows: ; in, This represents the total number of spectrum monitoring nodes across the entire region. The total number of time-series sampling steps during training. For the first The node The true value of the spectral characteristics at time t. This represents the predicted spectral characteristics of the model at the corresponding time point.

[0052] The transient loss function uses the mean absolute error loss function to constrain the single-point prediction accuracy under scenarios of short-term sudden changes in the spectrum, frequency band jumps, and transient interference. The calculation formula is as follows: ; in, It is a set of transient moments characterized by abrupt changes in gradient and abrupt changes in spectral state.

[0053] Construct a joint loss function based on two types of basic loss functions: ; in, , These are the steady-state loss weights and transient loss weights, respectively, satisfying... The global prediction error in steady-state scenarios is balanced with the single-point prediction error in non-stationary transient scenarios by using weighting coefficients.

[0054] The joint loss function is used as the objective of model iteration optimization. The gradient descent algorithm is used to backpropagate and update all learnable weight parameters of the adaptive graph convolutional neural network. The maximum number of iterations and the loss convergence threshold are set. When the fluctuation of the joint loss value in multiple consecutive iterations is less than the preset convergence threshold or the maximum number of iterations is reached, the model is determined to have converged and the optimal model parameters are saved.

[0055] The multidimensional spectrum data of each spatial node and each target frequency band collected in real time are input into the converged optimal model. After the entire process of topology graph construction, heterogeneous feature decoupling and embedding, hierarchical graph convolution feature mining, and dual-branch time-series calibration, the multidimensional spectrum prediction results of each spatial monitoring node and each target frequency band within a preset time period are output, including core spectrum state parameters such as spectrum occupancy rate, energy intensity, and interference intensity.

[0056] The second aspect of the present invention provides a multidimensional spatial spectrum prediction system based on graph convolutional neural networks, for implementing the above-mentioned multidimensional spatial spectrum prediction method, including an adaptive topology construction module, a multidimensional feature decoupling and embedding module, a dual residual graph convolutional feature mining module, a temporal dynamic calibration module, and a model training and prediction output module.

[0057] The adaptive topology construction module is used to collect spatial location data and historical spectrum data of the whole-domain spectrum monitoring nodes. Based on the electromagnetic attenuation mechanism and the dual weights calculated by spectrum time series similarity, it adaptively fuses and generates a dynamic weighted topology adjacency matrix to construct a spectrum topology network that fits the real electromagnetic environment.

[0058] The multidimensional feature decoupling and embedding module is used to split four types of spectral features: spatial steady state, temporal continuous, frequency band heterogeneous, and transient interference. Differentiated embedding is completed through independent coding branches to achieve mechanism-level fusion of multidimensional heterogeneous features and generate a unified high-dimensional feature vector.

[0059] The dual residual graph convolution feature mining module is used to hierarchically mine local and global spectral coupling features through a double-layer nested residual graph convolutional network, and adaptively adjust the convolution weights according to the real-time electromagnetic scene to select and optimize core spectral features.

[0060] The time-series dynamic calibration module is used to perform differentiated weighted calibration of deep spectral features through a steady-state-transient dual-branch time-series attention mechanism, correcting the lag bias of time-series prediction and adapting to steady-state evolution and sudden disturbance scenarios of the spectrum.

[0061] The model training and prediction output module is used to iteratively optimize the model through a steady-state-transient joint loss function, save the optimal model after convergence, and realize the prediction of the multi-dimensional spectrum state of the whole domain based on real-time input data, and output the prediction results of spectrum occupancy, energy intensity and interference intensity of each spatial node and each target frequency band.

[0062] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A multidimensional spatial spectrum prediction method based on graph convolutional neural networks, characterized in that, include: S1. Construct an adaptive spectrum topology driven by electromagnetic mechanisms: Using each spectrum monitoring node deployed in a globally distributed manner as a topology node, calculate the spatial electromagnetic correlation weight between nodes by combining the radio electromagnetic spatial attenuation propagation mechanism, calculate the spectrum state coupling weight by combining the historical spectrum temporal evolution law of each node, adaptively fuse the spatial electromagnetic correlation weight and the spectrum state coupling weight to generate a dynamic weighted adjacency matrix, and remove invalid topology connections without electromagnetic coupling correlation to obtain a spectrum topology network that fits the real electromagnetic propagation characteristics. S2. Multidimensional Spectral Heterogeneous Feature Layered Decoupling Embedding: Collect four types of core multidimensional spectral features corresponding to each topological node: spatial steady-state features, temporal continuous features, frequency band heterogeneous features, and transient interference features. Configure independent dedicated coding branches to complete differentiated embedding coding, retain the independent physical attributes of each type of feature, and then complete feature splicing and dimension alignment through a mechanism-level feature fusion strategy to generate a high-dimensional unified spectral fusion feature vector. S3. Feature mining of dual residual adaptive graph convolutional layer: Construct an adaptive graph convolutional neural network with a dual-layer nested residual structure. The first layer of local residual graph convolutional layer mines local spectral correlation features in the neighborhood of nodes, and the second layer of global residual graph convolutional layer mines global spectral coupling and linkage features across nodes and regions. At the same time, the convolutional feature weights are dynamically and adaptively adjusted according to the real-time electromagnetic environment to select and enhance effective spectral features and suppress redundant noise features. S4. Steady-state-transient dual-branch time series dynamic calibration: A dual-branch time series attention mechanism is introduced, which assigns constant benchmark weights to the long-term steady-state baseline features of the spectrum and dynamically adaptive weights to the transient features of short-term sudden disturbances and spectrum jumps, correcting the lag bias of time series modeling and completing the fine calibration of spectrum time series features. S5. Joint Loss Constraint Model Iterative Optimization and Prediction Output: A joint loss constraint mechanism combining steady-state loss function and transient loss function is adopted to perform reverse iterative training on the model until the model converges. Based on the converged optimal model, real-time multidimensional spectrum data is input, and multidimensional spectrum prediction results for each spatial node and each target frequency band within a preset time period are output.

2. The multidimensional spatial spectrum prediction method based on graph convolutional neural networks according to claim 1, characterized in that: In step S1, the spatial electromagnetic correlation weight is calculated as follows: the Euclidean spatial distance is calculated based on the latitude and longitude coordinates of the spectrum monitoring node, and the spatial distance is converted into an electromagnetic attenuation coefficient by combining the wireless signal free space attenuation model. The normalized spatial electromagnetic correlation weight is obtained by mapping the magnitude of the electromagnetic attenuation coefficient. The closer the spatial distance and the smaller the electromagnetic attenuation, the greater the corresponding spatial electromagnetic correlation weight.

3. The multidimensional spatial spectrum prediction method based on graph convolutional neural networks according to claim 2, characterized in that: In step S1, the spectrum state coupling weight is calculated as follows: extract the spectrum energy time series and frequency band occupancy time series within a preset historical time period of any two spectrum monitoring nodes, calculate the evolution similarity of the two time series using the cosine similarity algorithm, and obtain the spectrum state coupling weight after normalization. The higher the consistency of the spectrum evolution trend of the two nodes, the greater the spectrum state coupling weight.

4. The multidimensional spatial spectrum prediction method based on graph convolutional neural networks according to claim 3, characterized in that: In step S1, the adaptive fusion specifically involves: for densely deployed nodes in urban areas, reducing the proportion of spatial weight and increasing the proportion of spectrum coupling weight to avoid the problem of weight redundancy of nearby nodes; for sparsely deployed nodes in suburban areas, increasing the proportion of spatial weight and retaining weak spectrum coupling weight to maintain the correlation characteristics of the whole domain spectrum baseline; strengthening the coupling weight for nodes in the same frequency band and weakening the coupling weight for nodes in different frequency bands to achieve adaptive matching of topology weight scenarios.

5. The multidimensional spatial spectrum prediction method based on graph convolutional neural networks according to claim 1, characterized in that: In step S2, the encoding methods for the four types of core multidimensional spectral features specifically include: topological embedding encoding of spatial steady-state features to solidify the inherent spatial correlation attributes of nodes; smooth sequential embedding encoding of temporally continuous features to extract the evolution law of the spectral day-night baseline; differentiated hot embedding encoding of frequency band heterogeneous features to distinguish the differences in penetration and bandwidth characteristics of different frequency bands; and high-frequency feature enhancement embedding encoding of transient interference features to retain the detailed features of sudden noise, transient interference, and frequency band jumps.

6. The multidimensional spatial spectrum prediction method based on graph convolutional neural networks according to claim 1, characterized in that: In step S3, the double-layer nested residual structure is specifically as follows: each graph convolutional layer is provided with a residual bypass connection, which fuses the original features input to the convolutional layer with the depth features output by the convolutional layer element by element; the local residual graph convolutional layer focuses on mining the features of neighborhood spectrum interference superposition, resource reuse, and synchronization fluctuation, while the global residual graph convolutional layer focuses on mining the features of cross-regional interference diffusion, global spectrum baseline shift, and long-distance spectrum linkage.

7. The multidimensional spatial spectrum prediction method based on graph convolutional neural networks according to claim 6, characterized in that: In step S3, the logic for dynamically and adaptively adjusting the convolutional feature weights in real-time electromagnetic environment is as follows: when the variance of the spectral temporal fluctuation is less than the preset stability threshold, it is determined to be a steady-state electromagnetic scenario, and the weights of random noise features are weakened and the weights of steady-state features of the spectral baseline are strengthened; when the variance of the spectral temporal fluctuation is greater than the preset abrupt change threshold, it is determined to be a non-stationary electromagnetic scenario, and the weights of high-frequency disturbance features are amplified and the ability to represent transient abrupt changes is strengthened.

8. The multidimensional spatial spectrum prediction method based on graph convolutional neural networks according to claim 1, characterized in that: In step S4, the dual-branch temporal attention mechanism is specifically implemented as follows: constructing a steady-state attention branch and a transient attention branch, wherein the steady-state attention branch generates fixed weights based on long-period temporal statistical features; The transient attention branch generates dynamic weights based on short-term temporal gradient change features, which adaptively adapt to sudden fluctuations in the spectrum and instantaneous occupancy jumps. The fixed weights and dynamic weights are fused to complete the temporal feature calibration.

9. The multidimensional spatial spectrum prediction method based on graph convolutional neural networks according to claim 1, characterized in that: In step S5, the joint loss constraint mechanism is constructed as follows: mean square error is used as the steady-state loss function to constrain the overall prediction accuracy of the long-term evolution of the spectrum baseline; mean absolute error is used as the transient loss function to constrain the single-point prediction accuracy of the spectrum sudden change point; and the two types of losses are balanced by weighting coefficients to achieve bidirectional optimization of the prediction accuracy of steady-state and non-stationary scenarios.

10. A multidimensional spatial spectrum prediction system based on graph convolutional neural networks, characterized in that, The method for implementing the multidimensional spatial spectrum prediction method according to any one of claims 1-9 includes: The adaptive topology construction module is used to collect spatial location data and historical spectrum data of the whole-domain spectrum monitoring nodes. Based on the electromagnetic attenuation mechanism and the dual weights calculated by spectrum time series similarity, it adaptively fuses and generates a dynamic weighted topology adjacency matrix to construct a spectrum topology network that fits the real electromagnetic environment. The multi-dimensional feature decoupling and embedding module is used to decompose four types of spectral features: spatial steady state, temporal continuous, frequency band heterogeneous, and transient interference. Differentiated embedding is completed through independent coding branches to achieve mechanism-level fusion of multi-dimensional heterogeneous features and generate a unified high-dimensional feature vector. The dual residual graph convolutional feature mining module is used to hierarchically mine local and global spectral coupling features through a double-layer nested residual graph convolutional network, and adaptively adjust the convolution weights according to the real-time electromagnetic scene to filter and optimize core spectral features. The time-series dynamic calibration module is used to perform differentiated weighted calibration of deep spectral features through a steady-state-transient dual-branch time-series attention mechanism, correcting the time-series prediction lag bias and adapting to steady-state spectral evolution and sudden disturbance scenarios. The model training and prediction output module is used to iteratively optimize the model through a steady-state-transient joint loss function, save the optimal model after convergence, and realize the prediction of the multi-dimensional spectrum state of the whole domain based on real-time input data, and output the prediction results of spectrum occupancy, energy intensity and interference intensity of each spatial node and each target frequency band.

Images