A method, system, device, and medium for three-dimensional spectrum prediction based on multi-scale graph convolutional networks.

Abstract

This invention provides a three-dimensional spectrum prediction method, system, device, and medium based on multi-scale graph convolutional networks, belonging to the field of spectrum prediction technology. The method involves continuously acquiring received power data from multiple frequency points at a preset time resolution to form three-dimensional tensor data and constructing a feature matrix. The feature matrix is ​​transformed using a learnable parameter matrix of preset dimensions. An adaptive threshold is set based on the median score of all nodes to generate a weighted adjacency matrix. First-order and second-order graph convolution operations are performed on the weighted adjacency matrix, and feature transformations are performed using multi-dimensional learnable weight matrices to form multi-dimensional fused features. These multi-dimensional fused features are expanded according to a time series, and a gating mechanism is constructed through element-wise multiplication and residual connections. The prediction result is output through a linear transformation layer. This improves the accuracy of spectrum prediction, enabling the model to output more accurate prediction results when processing long-term spectrum data.

Description

Technical Field

[0001] This invention belongs to the field of spectrum prediction technology, specifically relating to a three-dimensional spectrum prediction method, system, device, and medium based on multi-scale graph convolutional networks. Background Technology

[0002] With the rapid development of wireless communication technology, efficient management and accurate prediction of spectrum resources have become increasingly important. Three-dimensional spectrum prediction, as a spectrum analysis technique that simultaneously considers time, space, and frequency dimensions, has broad application prospects in wireless communication network planning, spectrum sensing, and resource allocation. Currently, three-dimensional spectrum prediction methods have become a research hotspot. These methods construct spectrum monitoring data into a graph structure and process the graph structure data using graph convolution operations to achieve spectrum prediction.

[0003] However, in related technologies, the 3D spectrum prediction method based on graph convolutional networks has the following problems:

[0004] The relationship between nodes in the graph is described by constructing an adjacency matrix using a fixed threshold. This fixed adjacency matrix construction method cannot dynamically respond to changes in the electromagnetic environment. For example, in a real electromagnetic environment, sudden interference may occur, causing drastic changes in the spectral characteristics of certain areas; or equipment may move, altering the spatial relationships between devices. Because the static graph structure cannot capture these changes in a timely manner, the prediction results are lagging and cannot accurately reflect the real-time spectral state.

[0005] Standard graph convolution operations model only one type of neighborhood relationship, considering only the direct neighbors of nodes. This single-scale graph convolution approach ignores multi-scale spatial correlations. In real-world spectral data, spatial information at different scales plays a crucial role in spectrum prediction. For example, at a local scale, the spectral characteristics of adjacent nodes may exhibit strong correlations; while at a global scale, distant nodes may have indirect connections. Because existing methods cannot effectively integrate multi-scale spatial information, the models perform poorly when handling complex spatial structures, limiting the accuracy and reliability of spectrum prediction. Summary of the Invention

[0006] This invention provides a three-dimensional spectrum prediction method based on multi-scale graph convolutional networks. The method achieves dynamic response to changes in the electromagnetic environment, automated mining of implicit correlations, and effective modeling of multi-scale spatial correlations.

[0007] The methods include:

[0008] S101: Continuously collect received power data from multiple frequency points at a preset time resolution to form three-dimensional tensor data; calculate eight features for each location and frequency point, including spectrum occupancy, mean received power, standard deviation, interquartile range, autocorrelation coefficient, entropy, skewness, and abundance, and construct a feature matrix;

[0009] S102: Calculate the association score between nodes through bilinear transformation, transform the feature matrix using a learnable parameter matrix of preset dimensions, and generate an association score matrix in the 0-1 interval after activation by the Sigmoid function; set an adaptive threshold based on the median of the scores of all nodes to generate a weighted adjacency matrix.

[0010] S103: Perform first-order and second-order graph convolution operations on the weighted adjacency matrix, and perform feature transformation using multidimensional learnable weight matrices respectively. After ReLU activation, the features are concatenated to form multidimensional fused features. The normalized adjacency matrix is ​​obtained by multiplying the inverse square root diagonal matrix of the degree matrix with the original adjacency matrix.

[0011] S104: The multidimensional fused features are expanded according to the time series, and a one-dimensional convolution with an expansion rate of 4 is used to generate input gates and memory gates. A gating mechanism is constructed through element-wise multiplication and residual connection. The three-dimensional prediction results of time, location and frequency are output through a linear transformation layer.

[0012] It should be further noted that after step S104, the following steps are also included: weighted summation of the prediction error based on preset frequency point weight coefficients, and the learnable parameters are constrained by L2 regularization term.

[0013] The regularization coefficient is set to 0.0001;

[0014] The method also calculates the mean squared error of the true and predicted values, the frequency weight product, and the parameter 2 norm penalty term based on the loss function.

[0015] It should be further noted that the method divides the power data received in step S101 into a training set and a test set in a 7:3 ratio;

[0016] Set the learning rate to 0.0002 and the batch size to 128 for 500 iterations of training;

[0017] Each iteration executes three processes: forward propagation to calculate predicted values, adaptive weighted loss calculation, and backpropagation parameter update. Finally, the model performance is evaluated on the test set.

[0018] It should be further explained that step S101 specifically includes:

[0019] Based on monitoring devices at N fixed spatial locations, with a time resolution of 1 minute, the received power data of F frequency points are continuously collected for a total duration T as the raw data;

[0020] The original data is stored as a three-dimensional tensor. N represents the number of monitoring devices, F represents the number of frequency points, and T represents the total duration of data collection;

[0021] For each location i and frequency point f, calculate 8 expert features, including: spectrum occupancy, mean received power, standard deviation, interquartile range, autocorrelation coefficient, entropy, skewness, and abundance.

[0022] Constructing the feature matrix The feature vector corresponding to each frequency point f in each row M i It is a two-dimensional matrix used to store the feature vectors of all frequency points at position i.

[0023] It should be further explained that step S102 specifically includes:

[0024] The association score between nodes i and j is calculated using a bilinear transformation:

[0025]

[0026] in, It is a learnable parameter matrix, Flatten(M) i ) and Flatten (M j ) are respectively the characteristic matrix M i and M j The flattened vector has σ as the Sigmoid function, which restricts the score to the interval [0,1].

[0027] Generate a weighted adjacency matrix based on the scores:

[0028]

[0029] Where γ is the adaptive threshold, calculated as the threshold value for all s. ij of the median.

[0030] It should be further explained that step S103 specifically includes:

[0031] The calculation method for constructing a first-order graph convolution is as follows:

[0032] Z (1) =ReLU(ÂMW (1) )

[0033] Where Â=D -1 / 2 AD -1 / 2 It is the normalized adjacency matrix, D is the degree matrix, D ii =∑ j A ij , It is a learnable weight matrix, W (1) These are the parameters that the model needs to learn, used to transform the features;

[0034] The calculation method for constructing a second-order graph convolution is as follows:

[0035] Z (2) =ReLU(Â 2 MW (2) )

[0036] Among them, 2 =Â× is a second-order adjacency matrix. It is a learnable weight matrix;

[0037] Execution features are fused into:

[0038]

[0039] By concatenating the results of first-order and second-order graph convolutions, multi-scale spatial features Z are obtained.

[0040] It should be further noted that step S104 specifically includes: reconstructing the fused feature Z according to the time series...

[0041] ;

[0042] Construct gated dilated convolutions, and compute the input gate G in the following manner. t With memory gate G m :

[0043] G t =σ(Conv1D(H))

[0044] G m =tanh(Conv1D(H))

[0045] Where tanh is the hyperbolic tangent function, Conv1D is a one-dimensional convolution with a dilation rate of 4, and the input gate G is... t Memory gate G is used to control the inflow of input information. m Used to control the updating of memory information;

[0046] The output features are:

[0047] H out =G t ⊙G m +H

[0048] Here, ⊙ represents element-wise multiplication;

[0049] Construct a frequency domain prediction layer to obtain the final prediction result:

[0050]

[0051] Among them, Linear is a linear transformation layer that maps the features after gated convolution to the predicted spectral dimension, thus obtaining the final three-dimensional spectral prediction result.

[0052] This application also provides a three-dimensional spectrum prediction system based on a multi-scale graph convolutional network, the system comprising:

[0053] The feature extraction module is used to continuously collect received power data from multiple frequency points at a preset time resolution to form three-dimensional tensor data; for each location and frequency point, it calculates eight features: spectrum occupancy, mean received power, standard deviation, interquartile range, autocorrelation coefficient, entropy, skewness, and abundance, and constructs a feature matrix.

[0054] The adaptive construction module is used to calculate the association score between nodes through bilinear transformation. It transforms the feature matrix using a learning parameter matrix of preset dimensions, and generates an association score matrix in the 0-1 interval after activation by the Sigmoid function. An adaptive threshold is set according to the median of the scores of all nodes to generate a weighted adjacency matrix.

[0055] The feature fusion module is used to perform first-order and second-order graph convolution operations on the weighted adjacency matrix, respectively using multi-dimensional learnable weight matrices for feature transformation, and concatenating them after ReLU activation to form multi-dimensional fused features; the normalized adjacency matrix is ​​obtained by multiplying the inverse square root diagonal matrix of the degree matrix with the original adjacency matrix;

[0056] The temporal prediction module is used to expand the multidimensional fused features according to the time series, and use one-dimensional convolution with an expansion rate of 4 to generate input gates and memory gates. It constructs a gating mechanism through element-wise multiplication and residual connection, and outputs three-dimensional prediction results of time, location and frequency through a linear transformation layer.

[0057] According to another embodiment of this application, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the three-dimensional spectrum prediction method based on a multi-scale graph convolutional network.

[0058] According to another embodiment of this application, a storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the three-dimensional spectrum prediction method based on a multi-scale graph convolutional network.

[0059] As can be seen from the above technical solutions, the present invention has the following advantages:

[0060] The 3D spectrum prediction method based on multi-scale graph convolutional networks provided in this application calculates learnable correlation weights between nodes through bilinear transformation. The learnable parameter matrix W is continuously optimized during training, enabling the correlation score calculation to adapt to different data characteristics. This overcomes the limitations of constructing adjacency matrices with fixed thresholds and allows for dynamic adjustment of node correlations based on changes in the actual electromagnetic environment. The adaptive threshold γ is calculated based on the median of all correlation scores and can be automatically adjusted according to the characteristics of the data itself, ensuring good adaptability of the model across different electromagnetic environments and datasets, avoiding the subjectivity and tediousness of manual parameter tuning. Dynamic graph structure learning effectively solves the problem that static graph structures cannot respond to environmental changes in a timely manner, enhancing the model's adaptability to complex electromagnetic environments and ensuring that the model can capture the correlations between different locations in real time and accurately, thereby improving the prediction results' ability to reflect real-time spectrum conditions.

[0061] This application constructs a multi-scale graph convolution, where the first-order graph convolution effectively models the direct neighbor information of nodes through a normalized adjacency matrix and a learnable weight matrix, capturing the spectral correlation between adjacent nodes at a local scale. For example, in a certain region, the spectral characteristics of neighboring nodes often exhibit similarities, and the first-order graph convolution can accurately learn this local association, enabling the model to be keenly aware of local spectral changes. The second-order graph convolution utilizes a second-order adjacency matrix to capture the indirect neighbor information of nodes, thereby uncovering the indirect relationships between distant nodes at a global scale. It can utilize spatial domain information at different scales, enhancing the modeling ability for unstructured spectral data and improving the accuracy and reliability of spectral prediction. A spatiotemporal gated convolutional network is constructed, unfolding the features output by the multi-scale graph convolution in the time dimension. This processing method arranges the spatial features according to the time series, creating favorable conditions for modeling in the time domain, enabling the model to fully consider the changing trends of spectral data over time and uncover potential patterns in the time series. By introducing gated dilated convolution to compute the input gate and memory gate, the model state can be dynamically adjusted according to the input information, effectively alleviating the gradient vanishing problem in long-term time-series prediction. After processing through the gating mechanism, the features are mapped to the predicted spectral dimension through the frequency domain prediction layer, ultimately obtaining a three-dimensional spectral prediction result. This method improves the accuracy of spectral prediction, enabling the model to output more accurate prediction results when processing long-term spectral data. Attached Figure Description

[0062] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1The flowchart shows a 3D spectrum prediction method based on multi-scale graph convolutional networks.

[0064] Figure 2 This is a schematic diagram of a 3D spectrum prediction system based on a multi-scale graph convolutional network.

[0065] Figure 3 This is a schematic diagram of an electronic device. Detailed Implementation

[0066] Existing 3D spectrum prediction methods based on graph convolutional networks have significant limitations when dealing with complex electromagnetic environments and diverse spectral data. To improve the accuracy and adaptability of 3D spectrum prediction, this application provides a 3D spectrum prediction method based on multi-scale graph convolutional networks, which achieves dynamic response to changes in the electromagnetic environment, automated mining of implicit correlations, and effective modeling of multi-scale spatial correlations.

[0067] The following describes in detail the three-dimensional spectrum prediction method based on multi-scale graph convolutional networks involved in this application. Specific details, such as particular system structures and techniques, are presented for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details.

[0068] It should be understood that, when used in this specification, the term "comprising" indicates the presence of the described feature, integral, step, operation, element, and / or component, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0069] The terms "one embodiment" or "some embodiments" used in this application mean that one or more embodiments of this application include the specific features, structures, or characteristics described in that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this application do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0070] In embodiments of the present invention, computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or power server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (exemplarily using an Internet service provider for Internet connection).

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] Please see Figure 1 The diagram shows a flowchart of a 3D spectrum prediction method based on a multi-scale graph convolutional network in a specific embodiment. The method includes:

[0073] S101: Continuously collect received power data from multiple frequency points at a preset time resolution to form three-dimensional tensor data; calculate eight features for each location and frequency point, including spectrum occupancy, mean received power, standard deviation, interquartile range, autocorrelation coefficient, entropy, skewness, and abundance, and construct a feature matrix.

[0074] In some embodiments, received power data at multiple frequency points are continuously collected at a preset time resolution to form three-dimensional tensor data. The three dimensions of this three-dimensional tensor data are time, location, and frequency. Subsequently, eight features are calculated for each location and frequency.

[0075] Specifically, spectrum occupancy reflects the frequency of a signal occurring in a specific frequency band; mean received power is the average received power over a period of time; standard deviation reflects the dispersion of received power data relative to the mean; interquartile range shows the dispersion of the middle 50% of the data; autocorrelation coefficient measures the correlation of received power data at different points in time; entropy represents the uncertainty of the signal; skewness describes the asymmetry of the received power distribution; and abundance is a measure of signal complexity. By calculating these features, a feature matrix is ​​finally constructed.

[0076] S102: Calculate the association score between nodes through bilinear transformation, transform the feature matrix using a learnable parameter matrix of preset dimensions, and generate an association score matrix in the 0-1 interval after activation by the Sigmoid function; generate a weighted adjacency matrix by setting an adaptive threshold based on the median of the scores of all nodes.

[0077] In some embodiments, a bilinear transformation method is used to calculate the association scores between nodes. Specifically, a transformation operation is performed on the feature matrix using a learnable parameter matrix of a preset dimension, followed by activation by the Sigmoid function, thereby generating an association score matrix in the 0-1 interval. Then, an adaptive threshold is set based on the median of the scores of all node pairs. When the score of a node pair is greater than the threshold, a corresponding weight is assigned to it in the adjacency matrix; otherwise, the weight is 0, thus generating a weighted adjacency matrix.

[0078] S103: Perform first-order graph convolution and second-order graph convolution operations on the weighted adjacency matrix, respectively use multidimensional learnable weight matrices for feature transformation, and then concatenate them after ReLU activation to form multidimensional fused features; the normalized adjacency matrix is ​​obtained by multiplying the inverse square root diagonal matrix of the degree matrix with the original adjacency matrix.

[0079] In some embodiments, first-order and second-order graph convolution operations are performed on the weighted adjacency matrix. During the operation, a multidimensional learnable weight matrix is ​​used to transform the features, which are then processed by the ReLU activation function. The processed results are then concatenated to form a multidimensional fused feature. The normalized adjacency matrix here is obtained by multiplying the inverse square root diagonal matrix of the degree matrix with the original adjacency matrix, and its function is to normalize the information transmission between nodes.

[0080] S104: The multidimensional fused features are expanded according to the time series, and a one-dimensional convolution with an expansion rate of 4 is used to generate input gates and memory gates. A gating mechanism is constructed through element-wise multiplication and residual connection. The three-dimensional prediction results of time, location and frequency are output through a linear transformation layer.

[0081] In some embodiments, the multidimensional fused features are unfolded according to the time series, and then a one-dimensional convolution with a dilation rate of 4 is used to generate input gates and memory gates. Element-wise multiplication is used to allow the input gates to control the amount of input information entering the system, and the memory gates to control the amount of historical information retained. These are then combined with residual connections to construct a gating mechanism. This gating mechanism can effectively capture long-distance temporal dependencies, and finally, a linear transformation layer outputs a three-dimensional prediction result for time, location, and frequency.

[0082] S105: Adaptive weighted loss function, which uses manually set frequency weight coefficients to weight and sum the prediction errors, combined with L2 regularization to constrain the learnable parameters, wherein the regularization coefficient is set to 0.0001; the loss function calculation includes the mean square error between the true value and the predicted value, the frequency weight product term, and the parameter L2 norm penalty term.

[0083] In some embodiments, the adaptive weighted loss function uses manually set frequency weight coefficients to perform a weighted summation of prediction errors. Simultaneously, an L2 regularization term is used to constrain the learnable parameters; here, the regularization coefficient is set to 0.0001. The loss function calculation encompasses the mean squared error between the true and predicted values, the frequency weight product term, and the parameter L2 norm penalty term, thereby balancing the model's fitting ability and generalization ability.

[0084] S106: Perform model training and testing. Divide the dataset into training and testing sets in a 7:3 ratio. Set the learning rate to 0.0002 and the batch size to 128 for 500 iterations of training. Each iteration performs three processes: forward propagation to calculate the predicted value, adaptive weighted loss calculation, and backpropagation to update the parameters. Finally, evaluate the model performance on the test set.

[0085] As can be seen, the three-dimensional tensor data in this embodiment can comprehensively represent the relationship between the three dimensions of time, location, and frequency. Calculating eight features can characterize the spectrum from multiple perspectives. The bilinear transformation in step S102 can capture complex nonlinear relationships between nodes. The learnable parameter matrix allows the model to automatically learn appropriate association patterns based on the data. The Sigmoid function maps scores to the 0-1 interval, making the association relationships clearer and more explicit. The adaptive threshold setting can automatically adjust the sparsity of the adjacency matrix according to the characteristics of the data. First-order graph convolution mainly focuses on the direct neighbor information of nodes, capturing local structural features; second-order graph convolution can capture indirect neighbor information, uncovering deeper structural relationships. By concatenating graph convolution results of different orders, multi-scale structural information can be fused. One-dimensional convolution combined with a dilation rate of 4 can effectively capture long-distance temporal dependencies without increasing too many parameters. The gating mechanism can dynamically control the flow of information according to the input, preventing gradient vanishing or exploding, thus better processing time series data. The frequency weight coefficient can highlight the error of important frequencies, allowing the model to focus more on the prediction accuracy of these key frequencies. L2 regularization can prevent overfitting, improve the model's generalization ability, and enable the model to perform well when faced with unknown data.

[0086] Finally, this embodiment divides the dataset into training and testing sets, enabling an objective evaluation of the model's generalization ability. Setting an appropriate learning rate ensures stable convergence of the model during training. The choice of batch size affects the efficiency and stability of model training. Multiple iterations of training allow the model to fully learn the patterns and regularities in the data.

[0087] Based on the above embodiments, in order to further improve the reliability of the prediction effect of the three-dimensional spectrum prediction method based on multi-scale graph convolutional networks provided in the above embodiments, the following is a more specific implementation method, including the following specific steps:

[0088] Step S301: Collect spectrum detection data to obtain raw spectrum monitoring data and extract valuable features from it.

[0089] This embodiment can deploy monitoring devices at N=28 fixed spatial locations to continuously collect received power data from F=31 frequency points (790-820MHz, 1MHz intervals), with a time resolution of 1 minute and a total duration of T=30 days. N represents the number of monitoring devices, determined by the actual deployment; F represents the number of frequency points, determined based on the spectrum range and intervals of interest; and T represents the total duration of data collection, in minutes.

[0090] The original data is stored as a three-dimensional tensor. Where T = 43200 (30 days × 24 hours × 60 minutes), N = 28, and F = 31. D is a three-dimensional tensor used to store the received power data for all times, all locations, and all frequencies. For example, the spectrum data for location i = 5 at time t = 1200 (the 20th hour) is a vector. Here, i represents the location index, t represents the time index, and P... f This represents the received power at frequency point f.

[0091] For each location i and frequency f, eight expert features are calculated: spectrum occupancy, mean received power, standard deviation, interquartile range, autocorrelation coefficient, entropy, skewness, and abundance. These features can describe the characteristics of the spectrum from different perspectives; for example, spectrum occupancy reflects the use of the spectrum, and mean received power reflects the average power level.

[0092] Finally, construct the feature matrix. The feature vector corresponding to each frequency point f in each row For example, the characteristic matrix at position i=5. , where the f=10th row represents the 8-dimensional feature at the 800MHz frequency point. M i It is a two-dimensional matrix used to store the feature vectors of all frequency points at position i.

[0093] Step S302: Define dynamic graph structure learning to achieve adaptive modeling of spatial relationships, enabling the model to better capture the correlations between different locations. Its input is the set of feature matrices {M} output from step one. i Output the dynamically weighted adjacency matrix. .

[0094] The dynamic graph construction in this embodiment involves the following methods:

[0095] First, calculate the learnable relevance weights: use a bilinear transformation to calculate the association score between nodes i and j: .

[0096] in, It is a learnable parameter matrix (248 = 31 × 8), Flatten(M) i ) and Flatten (M j ) are respectively the characteristic matrix M i and M j The flattened vector has σ as its sigmoid function, which restricts the score to the [0,1] interval. W represents the parameters the model needs to learn, optimized through training to find a suitable way to calculate the inter-node correlation scores.

[0097] Generate a weighted adjacency matrix based on the scores:

[0098]

[0099] Where γ is the adaptive threshold, calculated as the threshold value for all s. ij The median. Adaptive calculation of γ avoids the hassle of manual parameter tuning, automatically adjusting the threshold according to the actual data, thus improving the model's generalization ability.

[0100] Step S303: To fuse local and global spatial features and enhance the model's ability to model unstructured data, graph convolutions at different scales can capture spatial information of varying ranges, thereby improving model performance. Its input is the feature matrix {M} output from step one. i The output is the multi-scale spatial features, along with the adjacency matrix A generated in step two. .

[0101] Specifically, the first-order graph convolution (1-hop GCN) is constructed and calculated as follows:

[0102] Z (1) =ReLU(ÂMW (1) )

[0103] Where Â=D -1 / 2 AD -1 / 2 It is the normalized adjacency matrix (D is the degree matrix, D ii =∑ j A ij ),

[0104] It is a learnable weight matrix. A normalized adjacency matrix  can make the computation of graph convolution more stable. W (1) These are the parameters that the model needs to learn, used to transform the features.

[0105] The second-order graph convolution (2-hop GCN) is constructed and its calculation method is as follows:

[0106] Z (2) =ReLU(Â 2 MW (2) )

[0107] Among them, 2 =Â× is a second-order adjacency matrix that captures indirect spatial relationships. It is a learnable weight matrix. Second-order graph convolution can capture information about the indirect neighbors of a node.

[0108] Execution feature fusion: .

[0109] By concatenating the results of first-order and second-order graph convolutions, multi-scale spatial features Z are obtained.

[0110] Step S304: Given the vanishing gradient problem that easily occurs in long-term time-series prediction, the purpose of this step is to jointly model the time-domain and frequency-domain dependencies. By introducing a gating mechanism, the vanishing gradient problem is solved, improving the accuracy of spectral prediction. Its input is the feature output from step three. Output three-dimensional spectrum prediction results. .

[0111] Its network structure is as follows:

[0112] Expand Z along the time dimension: reconstruct Z as a time series. This allows spatial features to be expanded along the time dimension, facilitating subsequent temporal modeling.

[0113] Construct gated dilated convolution and compute the input gate G. t With memory gate G m :

[0114] G t =σ(Conv1D(H))

[0115] G m =tanh(Conv1D(H))

[0116] Where σ is the Sigmoid function, tanh is the hyperbolic tangent function, and Conv1D is a one-dimensional convolution with a dilation rate of 4. Input gate G t Memory gate G is used to control the inflow of input information. m Used to control the updating of memory information.

[0117] Output features are obtained:

[0118] H out =G t ⊙G m +H

[0119] Here, ⊙ represents element-wise multiplication. Through gating mechanisms, the model's state can be dynamically adjusted based on input information, thereby mitigating the vanishing gradient problem.

[0120] Construct a frequency domain prediction layer to obtain the final prediction result:

[0121] Among them, Linear is a linear transformation layer that maps the features after gated convolution to the predicted spectral dimension, thus obtaining the final three-dimensional spectral prediction result.

[0122] Step S305: Execute adaptive weighted loss.

[0123] This embodiment takes into account that different frequency points may have varying degrees of importance; for example, the prediction accuracy for high-frequency interference frequencies may be more critical. Therefore, an adaptive weighted loss function is constructed to improve prediction accuracy for high-frequency interference frequencies while preventing model overfitting. Its input is the actual spectrum data Y. t,i,f and predicted spectrum data t,i,f And the learnable parameters W of the model. Output loss value. .

[0124] Loss function:

[0125]

[0126] Where, ω f It is a frequency weight, manually set according to the importance of frequency bands. This is the L2 regularization term, used to prevent overfitting. T, N, and F represent the number of time points, locations, and frequencies, respectively. Y t,i,f It is real spectrum data. t,i,f Here, W represents the predicted spectral data, W is the learnable parameter of the model, and λ is the regularization coefficient, used to control the weight of the regularization term. By setting different weights for different frequency points, the model can pay more attention to the prediction error of important frequencies, thereby improving the overall prediction accuracy. At the same time, the L2 regularization term can constrain the model parameters and avoid overfitting.

[0127] Step S306: Divide the data collected and feature-extracted in step S301 into a training set and a test set in a 7:3 ratio. Set the learning rate to 0.0002, the number of training iterations to 500, and the batch size to e=128. Train the model in the following order:

[0128] Step S3061: Obtain a batch of data from the training set and input it into the model for forward propagation. The data will then pass through dynamic graph structure learning, multi-scale graph convolution, and spatiotemporal gated convolutional networks in sequence, outputting the 3D spectral prediction result. .

[0129] Step S3062: Calculate the loss: based on the prediction results Using the real spectrum data Y, an adaptive weighted loss function is applied. Calculate the loss value for the current batch.

[0130] Step S3063: Backpropagation: Perform backpropagation, calculate the gradient of the loss function with respect to the model parameters, and update the model parameters.

[0131] After 500 training iterations, the model is finally evaluated using the test set data. The loss value and evaluation metrics for the test set are calculated to assess the model's performance on unseen data.

[0132] As can be seen, the representation of 3D tensor data makes the data organization more structured, facilitating subsequent processing and analysis. Multi-feature extraction can more comprehensively characterize spectral features, providing rich input information for the model. The adaptive adjacency matrix generation method can automatically adjust according to the actual situation of the data, capturing the relationships between nodes more flexibly than a fixed-structure graph, thereby improving the model's adaptability to different data. Multi-scale graph convolution integrates structural information at different levels, enabling a more comprehensive understanding of the graph's structure and allowing the model to have a deeper grasp of the complex relationships in spectral data, thus improving prediction accuracy. Gating mechanisms combined with dilated convolution can effectively capture long-distance temporal dependencies, making them very suitable for processing the temporal characteristics of spectral data and better predicting future spectral changes. The adaptive weighted loss function can differentiate errors based on the importance of frequency points, while L2 regularization can prevent overfitting, improving the model's generalization ability and robustness, enabling the model to work stably in different application scenarios. Reasonable dataset partitioning, learning rate and batch size settings, and sufficient iterations ensure that the model fully learns the patterns in the data, improving model performance and stability.

[0133] The following are embodiments of a three-dimensional spectrum prediction system based on multi-scale graph convolutional networks provided in this disclosure. This system and the three-dimensional spectrum prediction method based on multi-scale graph convolutional networks in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the three-dimensional spectrum prediction system based on multi-scale graph convolutional networks, please refer to the embodiments of the three-dimensional spectrum prediction method based on multi-scale graph convolutional networks described above.

[0134] like Figure 2 As shown, the system includes: a construction and feature extraction module, which is used to continuously collect received power data of multiple frequency points at a preset time resolution to form three-dimensional tensor data; and to calculate eight features for each location and frequency point, including spectrum occupancy, mean received power, standard deviation, interquartile range, autocorrelation coefficient, entropy, skewness, and abundance, to construct a feature matrix.

[0135] The adaptive construction module is used to calculate the association score between nodes through bilinear transformation. It transforms the feature matrix using a learning parameter matrix of preset dimensions, and generates an association score matrix in the 0-1 interval after activation by the Sigmoid function. An adaptive threshold is set according to the median of the scores of all nodes to generate a weighted adjacency matrix.

[0136] The feature fusion module is used to perform first-order and second-order graph convolution operations on the weighted adjacency matrix, respectively using multi-dimensional learnable weight matrices for feature transformation, and concatenating them after ReLU activation to form multi-dimensional fused features; the normalized adjacency matrix is ​​obtained by multiplying the inverse square root diagonal matrix of the degree matrix with the original adjacency matrix;

[0137] The temporal prediction module is used to expand the multidimensional fused features according to the time series, and use one-dimensional convolution with an expansion rate of 4 to generate input gates and memory gates. It constructs a gating mechanism through element-wise multiplication and residual connection, and outputs three-dimensional prediction results of time, location and frequency through a linear transformation layer.

[0138] like Figure 3 As shown, this application also provides an electronic device, including a display module 103, a memory 102, a processor 101, and a computer program stored in the memory and executable on the processor 101. When the processor 101 executes the program, it implements the steps of a three-dimensional spectrum prediction method based on a multi-scale graph convolutional network.

[0139] In embodiments of the present invention, electronic devices include, but are not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments described and / or claimed herein.

[0140] In this embodiment, processor 101 may be implemented using at least one of an application-specific integrated circuit, a programmable logic device, a field-programmable gate array, a processor, a controller, a microcontroller, a microprocessor, or an electronic unit designed to perform the functions described herein. In some cases, such an implementation may be implemented within a controller. For software implementation, implementations such as processes or functions may be implemented with separate software modules that allow the performance of at least one function or operation. Software code may be implemented by a software application (or program) written in any suitable programming language, and the software code may be stored in memory and executed by the controller.

[0141] The display module 103 is used to display information input by the user or information provided to the user. The display module 103 may include a display panel, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like.

[0142] The memory 102 can be used to store software programs and various data. The memory 102 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0143] This application also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a three-dimensional spectrum prediction method based on a multi-scale graph convolutional network.

[0144] The storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0145] In a storage medium, a readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying readable program code. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0146] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A three-dimensional spectrum prediction method based on multi-scale graph convolutional networks, characterized in that, The methods include: S101: Continuously acquire received power data at multiple frequency points with a preset time resolution to form three-dimensional tensor data; For each location and frequency point, eight features are calculated: spectrum occupancy, mean received power, standard deviation, interquartile range, autocorrelation coefficient, entropy, skewness, and abundance, and a feature matrix is ​​constructed. Step S101 specifically includes: Based on monitoring devices at N fixed spatial locations, with a time resolution of 1 minute, the received power data of F frequency points are continuously collected for a total duration T as the raw data; The original data is stored as a three-dimensional tensor. N represents the number of monitoring devices, F represents the number of frequency points, and T represents the total duration of data collection; For each position i and frequency point f, calculate the eight features; Constructing the feature matrix The feature vector corresponding to each frequency point f in each row M i It is a two-dimensional matrix used to store the feature vectors of all frequency points at position i; S102: Calculate the association score between nodes through bilinear transformation, transform the feature matrix using a learnable parameter matrix of preset dimensions, and generate an association score matrix in the 0-1 interval after activation by the Sigmoid function; set an adaptive threshold based on the median of the scores of all nodes to generate a weighted adjacency matrix. Step S102 specifically includes: The association score between nodes i and j is calculated using a bilinear transformation: in, It is a learnable parameter matrix, Flatten(M) i ) and Flatten (M j ) are respectively the feature matrix M i and M j The flattened vector has σ as the Sigmoid function, which restricts the score to the interval [0,1]. Generate a weighted adjacency matrix based on the scores: Where γ is the adaptive threshold, calculated as the threshold value for all s. ij the median; S103: Obtain the normalized adjacency matrix by multiplying the diagonal matrix of the inverse square root of the degree matrix with the weighted adjacency matrix. Perform first-order graph convolution and second-order graph convolution operations on the normalized adjacency matrix, respectively use multidimensional learnable weight matrices for feature transformation, and then concatenate them after ReLU activation to form multidimensional fusion features. Step S103 specifically includes: The calculation method for constructing a first-order graph convolution is as follows: Z (1) =ReLU(ÂMW (1) ) Where Â=D -1 / 2 AD -1 / 2 It is the normalized adjacency matrix, D is the degree matrix, D ii =∑ j A ij , It is a learnable weight matrix, W (1) These are the parameters that the model needs to learn, used to transform the features; The calculation method for constructing a second-order graph convolution is as follows: Z (2) =ReLU( 2 MW (2) ) Among them, 2 =Â× is a second-order adjacency matrix. It is a learnable weight matrix; Execution features are fused into: By concatenating the results of first-order and second-order graph convolutions, multi-scale spatial features Z are obtained. S104: The multidimensional fusion features are expanded according to the time series, and a one-dimensional convolution with an expansion rate of 4 is used to generate input gates and memory gates. A gating mechanism is constructed through element-wise multiplication and residual connections. The three-dimensional prediction results of time, location and frequency are output through a linear transformation layer. Step S104 specifically includes: reconstructing the fused feature Z according to the time series as follows: ； Construct gated dilated convolutions, and compute the input gate G in the following manner. t With memory gate G m : G t =σ(Conv1D(H)) G m =tanh(Conv1D(H)) Where tanh is the hyperbolic tangent function, Conv1D is a one-dimensional convolution with a dilation rate of 4, and the input gate G is... t Memory gate G is used to control the inflow of input information. m Used to control the updating of memory information; The output features are: H out =G t ⊙G m +H Here, ⊙ represents element-wise multiplication; Construct a frequency domain prediction layer to obtain the final prediction result: Among them, Linear is a linear transformation layer that maps the features after gated convolution to the predicted spectral dimension, thus obtaining the final three-dimensional spectral prediction result.

2. The three-dimensional spectrum prediction method based on multi-scale graph convolutional networks according to claim 1, characterized in that, Step S104 is followed by the following steps: The prediction error is weighted and summed based on the preset frequency point weight coefficients, and the learnable parameters are constrained by the L2 regularization term. The regularization coefficient is set to 0.0001; The method also calculates the mean squared error of the true and predicted values, the frequency weight product, and the parameter 2 norm penalty term based on the loss function.

3. The three-dimensional spectrum prediction method based on multi-scale graph convolutional networks according to claim 2, characterized in that, The method divides the power data received in step S101 into a training set and a test set in a 7:3 ratio; The learning rate was set to 0.0002 and the batch size to 128 for 500 iterations of training. Each iteration executes three processes: forward propagation to calculate predicted values, adaptive weighted loss calculation, and backpropagation parameter update. Finally, the model performance is evaluated on the test set.

4. A three-dimensional spectrum prediction system based on multi-scale graph convolutional networks, characterized in that, The system is used to implement the three-dimensional spectrum prediction method based on multi-scale graph convolutional networks as described in any one of claims 1 to 3; the system includes: The feature extraction module is used to continuously collect received power data from multiple frequency points at a preset time resolution to form three-dimensional tensor data; for each location and frequency point, it calculates eight features: spectrum occupancy, mean received power, standard deviation, interquartile range, autocorrelation coefficient, entropy, skewness, and abundance, and constructs a feature matrix. The adaptive construction module is used to calculate the association score between nodes through bilinear transformation. It transforms the feature matrix using a learning parameter matrix of preset dimensions, and generates an association score matrix in the 0-1 interval after activation by the Sigmoid function. An adaptive threshold is set according to the median of the scores of all nodes to generate a weighted adjacency matrix. The feature fusion module is used to perform first-order and second-order graph convolution operations on the weighted adjacency matrix, respectively using multi-dimensional learnable weight matrices for feature transformation, and concatenating them after ReLU activation to form multi-dimensional fused features; the normalized adjacency matrix is ​​obtained by multiplying the inverse square root diagonal matrix of the degree matrix with the original adjacency matrix; The temporal prediction module is used to expand the multidimensional fused features according to the time series, and use one-dimensional convolution with an expansion rate of 4 to generate input gates and memory gates. It constructs a gating mechanism through element-wise multiplication and residual connection, and outputs three-dimensional prediction results of time, location and frequency through a linear transformation layer.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the three-dimensional spectrum prediction method based on multi-scale graph convolutional networks as described in any one of claims 1 to 3.

6. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the three-dimensional spectrum prediction method based on multi-scale graph convolutional networks as described in any one of claims 1 to 3.

Images