Multi-band frequency spectrum prediction method based on dynamic graph

By constructing a multi-band dynamic graph and time-frequency dual-channel spectrum prediction model of sliding time window, combining dynamic graph convolution and LSTM network, the problem of insufficient feature capture of multi-band spectrum data in the existing technology is solved, and higher accuracy is achieved. Spectrum prediction.

CN119095065AActive Publication Date: 2024-12-06ARMY ENG UNIV OF PLA
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411275995.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-12
Publication Date
2024-12-06
Estimated Expiration
2044-09-12

AI Technical Summary

Technical Problem

The prior art is difficult to effectively capture global features in multi-band spectrum data, resulting in insufficient spectral prediction accuracy. Especially when processing multi-band data, the static graph structure cannot fully capture the time-varying features of the spectrum.

Method used

The sliding time window is used to construct a multi-band dynamic graph, combining a time-frequency dual-channel spectral prediction model based on historical data and a dynamic graph, and extract global and local features through dynamic graph convolution and LSTM network, and perform adaptive fusion. A multi-band dynamic graph structure of sliding time window is constructed, time-varying features on multiple frequency bands are captured, and the similarity between bands is measured using the k-nearest neighbor algorithm and Euclidean distance are used to measure the frequency bands, and a dynamic graph structure is constructed for prediction.

Benefits of technology

It improves the accuracy and generalization ability of spectrum prediction, can more effectively capture the global and local features of multi-band spectrum data, overcomes the limitations of single feature extraction, and significantly improves the accuracy of spectrum prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119095065A_ABST
    Figure CN119095065A_ABST
Patent Text Reader

Abstract

The invention provides a multi-band frequency spectrum prediction method based on a dynamic graph. The method comprises the following steps: step 1, constructing a multi-band dynamic graph structure of a sliding time window; step 2, constructing a multi-band frequency spectrum tensor based on historical data; step 3, a time-frequency dual-channel spectrum prediction model based on the dynamic graph; and step 4, implementing the multi-band frequency spectrum prediction method based on the dynamic graph. According to the method, the dynamic graph is constructed for the multi-band frequency spectrum correlation of different time windows, and multi-band prediction is performed by adopting a time feature and frequency band feature dual-channel extraction method, so that the problem of relatively poor model generalization ability caused by neglecting the dynamic characteristics of the frequency spectrum environment in a traditional prediction method is effectively avoided.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of wireless communication technology, and in particular to a multi-band spectrum prediction method based on dynamic graphs. Background Art

[0002] With the advancement of wireless communication technology, the demand for spectrum is also growing rapidly. In recent decades, communication technology has experienced rapid development from 3G to 4G and then to 5G, and more and more advanced communication devices have been designed and widely used. This has led to the scarcity of spectrum resources. On the other hand, the low frequency spectrum utilization of the existing static allocation strategy has further aggravated the shortage of spectrum resources. How to dynamically utilize scarce spectrum resources has become the focus of researchers. In order to dynamically and effectively utilize idle spectrum resources, researchers have proposed the concept of dynamic spectrum access, that is, underutilized licensed bands are dynamically shared by a large number of unlicensed users without introducing interference to achieve the reuse of spectrum resources. Accurately predicting spectrum usage is a key prerequisite for achieving spectrum sharing and dynamic spectrum access. Spectrum prediction is a method that infers future channel occupancy status, duty cycle, power spectrum and other information based on historical spectrum detection data. It is a widely accepted method to improve the overall utilization of spectrum by predicting spectrum holes and temporarily allocating them to unlicensed users. Existing research results show that the use of spectrum has high temporal correlation and frequency domain correlation. Making full use of the temporal and frequency correlations obtained from historical observation data is the key to achieving high-precision spectrum prediction.

[0003] In spectrum analysis, a single-band spectrum can be considered as a data sequence with temporal characteristics, while a multi-band spectrum is a collection of multiple such temporal data sequences. Initially, researchers predicted spectrum trends by combining time-frequency domain methods with convolutional neural network (CNN) models. Later, researchers found that this approach did not produce satisfactory results when dealing with multi-band data. Therefore, researchers proposed using a graph structure to associate frequency bands that are far apart.

[0004] Although this approach has shown some improvements, due to the openness of the spectrum environment and the inherent temporal dynamics of spectrum data, e.g. Figure 1As shown. Three frequency bands were randomly selected from the Electrosense data of 790-820MHz for analysis. In time window 1, there was a positive correlation between the three frequency bands 1, 2 and 3. In time window 2, the trends between band 1 and band 3 were relatively similar, but they showed uncorrelated or even negatively correlated behavior with band 2. It can be seen from the figure that the relationship between multiple frequency bands presents time-varying characteristics, but the simple static graphs in previous studies ignored this feature and were not sufficient to fully capture the time-varying global features in multi-band spectrum data. Therefore, the current research focus has shifted to effectively extracting global features with time-varying characteristics between multi-band frequencies, which is an important direction to improve the accuracy of spectrum prediction. Summary of the invention

[0005] The present application provides a multi-band spectrum prediction method based on dynamic graphs, which can be used to solve the problem of ignoring the dynamic information of multi-band correlation in existing research.

[0006] The present application provides a multi-band spectrum prediction method based on a dynamic graph, the method comprising:

[0007] Step 1, construct a multi-band dynamic graph structure of a sliding time window;

[0008] Step 2, construct a multi-band spectrum tensor based on historical data;

[0009] Step 3: a time-frequency dual-channel spectrum prediction model based on dynamic graph;

[0010] Step 4: Implement a multi-band spectrum prediction method based on dynamic graph.

[0011] Furthermore, a multi-band dynamic graph structure of a sliding time window is constructed, including:

[0012] The purpose of this step is to capture the time-varying features on multiple frequency bands and construct a dynamic graph using a sliding window method. The sliding window is applied to the frequency bands, and a dynamic graph that changes over time is constructed by calculating the similarity relationship between frequency bands in each window. This dynamic graph can not only intuitively show the changing trend of the similarity between frequency bands, but also provide in-depth insights into the time-varying features of multiple frequency bands.

[0013] Since the correlation between multiple frequency bands in different time windows changes dynamically, a dynamic graph is constructed for multiple frequency bands by sliding the event window to capture the dynamic relationship of multiple frequency bands.

[0014] Constructing dynamic graphs through time sliding windows in represents a set of nodes, ε represents a set of edges, and each frequency band f i are all considered a node v in the dynamic graph i, the similarity between the frequency bands forms the edge of the graph; the correlation between the calculated frequency band and all frequency bands except this frequency band is calculated, and then the k frequency bands with the highest correlation with this frequency band are selected to establish the connection edge; the frequency bands f1 and f n In the time window w i The similarity on is:

[0015]

[0016] w i =[t p ,t p +1,...,t q ]

[0017] in and Respectively represent the frequency bands f1 and f n In the time window w i The spectrum state on the spectrum; use the Euclidean distance as a metric to measure the similarity between two frequency bands, where α defaults to 2, eps represents a small number to avoid division by zero, and is set to 1e-6 by default; t p to q Represents the size of the time window.

[0018] Furthermore, step 2 constructs a multi-band spectrum tensor based on historical data, including:

[0019] Since the spectrum usage has a certain temporal regularity in adjacent time periods, the power spectrum values ​​of each frequency band are smoothly arranged in time as a time tensor for processing;

[0020] The frequency range is A-BMHz, with CMHz as a frequency band width, forming a total of (BA) / C frequency band data; then, the power spectrum value of each frequency band is analyzed in time order, with T minutes as a time unit, and a size of X is formed in each time unit containing all frequency bands. i ∈R n×c A vector of , where c = 1, n = (BA) / C, is then formed into a tensor smoothly arranged in time:

[0021]

[0022] Where m represents time.

[0023] Furthermore, in step 3, a time-frequency dual-channel spectrum prediction model based on a dynamic graph is established, including:

[0024] The time-frequency dual-channel spectrum prediction model based on dynamic graph first uses the k-nearest neighbor algorithm to model the similarity relationship between multiple frequency bands, and uses the sliding window method to construct a dynamic graph to capture the time-varying features on multiple frequency bands. Subsequently, the dual-channel feature extraction module is used to extract global and local features respectively. Finally, the two types of features are weighted and fused using an adaptive fusion mechanism. The model separates the two modules to ensure that they do not interfere with each other, avoids feature redundancy, and effectively solves the problem of making full use of global and local features between frequency bands.

[0025] The establishment of a time-frequency dual-channel spectrum prediction model based on dynamic graphs includes three modules, namely, a dynamic graph construction module, a global feature extraction module, and a local feature extraction module;

[0026] The dynamic graph construction model generates a graph structure for the multi-bands in each time window to represent the similarity between the multi-bands. The multi-band spectrum is divided into several bands according to the frequency band interval of CMHz, and the power spectrum value of each band is arranged in time order. Then, the sliding time window t is divided. p to q , the multi-band power spectrum in each time window is subjected to correlation analysis, and the most relevant k nodes are selected using the k nearest neighbor algorithm to establish a connection relationship. In this way, each time window dynamically forms a graph structure G=<V,E> ;

[0027] Then the constructed graph structure G =<V,E> and node features Input to two parallel modules, namely global feature extraction module and local feature extraction module;

[0028] The extracted global features and local features are fused, and finally the time-frequency dual-channel spectrum prediction model is output;

[0029] In the global feature extraction module, the dynamic graph convolution module is used to adaptively learn and update the graph structure, effectively process and analyze the dynamic changes between frequency bands, and the long short-term memory (LSTM) network is used to extract temporal features. In the local feature extraction module, local features of adjacent areas between multiple frequency bands are extracted. The gate fusion mechanism is used to weight and fuse these two types of features. Finally, the power spectral density value of the next d time units is output.

[0030] Furthermore, the global feature extraction module includes three dynamic graph convolutional layers and two LSTM layers;

[0031] The global feature extraction module is used to extract the global level features of multiple frequency bands within the graph structure; global correlation refers to the extensive and non-local interdependence between all time-frequency objects in the frequency domain and time dimension. Interdependence not only considers the interaction between adjacent frequency bands in the frequency domain, but also considers the mutual influence of all frequency bands on a long time scale. When modeling the global features of multiple frequency bands, a dynamic graph structure is adopted, which changes dynamically according to the multiple frequency bands in different time windows. In view of the powerful ability of graph convolutional networks in the field of graph learning, a dynamic graph convolutional network (DGCN) is used to capture the information in the graph structure, thereby extracting the graph structure features of multiple frequency bands.

[0032] The dynamic graph convolution layer aggregates the features of adjacent nodes and updates the features of nodes in different time windows, as shown below:

[0033]

[0034] in It is a graph structure updated and aggregated by a layer of dynamic convolution operations; ψ agg and ψ update are the learnable weights for aggregation and update operations, respectively;

[0035] For each node, the aggregation and update operations are as follows:

[0036]

[0037] Yes i The set of neighbor nodes, and v j Yes i A neighbor node of ; In order to adaptively update the node features, a fine-grained edge attention module is used to weight and aggregate the node edges, and the operation is as follows:

[0038]

[0039] in represents v through the attention module i , Att() represents the function of node weighted aggregation; where Att(v i ) represents the weighted aggregation of its own features, as follows:

[0040]

[0041] in, Represents the original features of the node, μ i,i represents the attention weight; ζ() represents the agg Convolutional layer with parameterized 1×1 convolution kernel; v i The weighted aggregation of attention of the neighboring nodes is expressed as

[0042]

[0043] in Represents the connection between two features; the attention weights are as follows:

[0044]

[0045] where u() is represented by ψ update Convolutional layer with parameterized 1×1 convolution kernel;

[0046] In addition, considering that the feature carried by each node is the power spectrum value of the corresponding frequency band arranged in time series, the time information is extracted from the corresponding node features through two LSTM layers to obtain the feature vector χ g :

[0047]

[0048] where ⊙ represents the Hadamard product operator, c i-1 represents the memory information of the previous time step, c i Represents the memory information of the current time step, h i is the output of the LSTM unit, h i-1 is the output of the previous time step; δ represents the weight matrix, and b represents the bias vector;

[0049] Through the global feature extraction module, dynamic graph convolution is responsible for processing the structural information of the graph, while the LSTM layer is used for the temporal evolution of node features, thereby comprehensively capturing the global features of multi-band data with time changes.

[0050] Furthermore, the local feature extraction module is used to extract local feature information of the multi-band spectrum data, and the local feature extraction module includes two convolutional layers and two LSTM layers;

[0051] The multi-band feature vector Input to the convolution layer, and use the convolution kernel to traverse the feature vector to generate a new feature vector η, effectively extracting the information of the current frequency band and its surrounding frequency bands. The feature vector generation is as follows:

[0052]

[0053] Among them, δ represents the convolution kernel weight vector, b represents the offset vector, and λ represents the activation function. Represents the convolution operation;

[0054] Then, the feature vector η is input into LSTM to extract the time information and obtain the feature X c ;

[0055] Finally, the adaptive gating mechanism is used to fuse the multi-band global features extracted by the global feature extraction module with the local features extracted by the local feature extraction module to obtain feature X f :

[0056]

[0057] Among them, γ represents the learned allocation weight, represents the Sigmoid function, χ g and χ c They represent the weights of multi-band global features and multi-band local features respectively, and b represents the deviation.

[0058] Furthermore, in step 4, a multi-band spectrum prediction method based on a dynamic graph is implemented, comprising:

[0059] The task of this application is to predict the power spectral density values ​​of n frequency bands in the future d time units The power spectral density value is determined by training the prediction model F:

[0060]

[0061] First, the frequency band bandwidth is selected as A-BMHz, and then the power spectrum values ​​of the past m time units are used to predict the power spectrum values ​​of the next d time units; in the process of constructing the graph structure, the k nearest neighbor algorithm is used to select k nodes with the strongest correlation to construct the connection relationship; after continuous iterative optimization, the algorithm converges and then outputs the prediction result; the power spectrum values ​​of the future d time units obtained by the multi-band spectrum prediction module based on dynamic graphs are compared with the actual power spectrum values ​​of the current time unit, and the prediction accuracy is evaluated by calculating the error between the actual value and the predicted value. The prediction performance evaluation indicators include Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE);

[0062] The root mean square error is a commonly used indicator to measure the difference between the model prediction value and the actual observation value. It is obtained by calculating the mean of the square of the difference between the prediction value and the actual observation value and taking the square root:

[0063]

[0064] The mean absolute error is used to assess how well the model fits the given data; it is calculated by averaging the absolute values ​​of the differences between the predicted values ​​and the actual observed values:

[0065]

[0066] The mean absolute error is sensitive to the relative error and refers to the mean absolute error between the predicted value and the true value. The determination method is as follows:

[0067]

[0068] X m+i represents the true value of the spectrum of the m+ith time unit, represents the spectrum prediction value of the m+ith time unit, and s represents the prediction time step; the smaller the values ​​of the three evaluation indicators are, the smaller the error between the prediction result obtained by the model and the true value is, which further proves that the prediction effect of the proposed model is better; the prediction effect of the model is obtained by calculating the three error indicators.

[0069] Compared with the prior art, the present invention has the following significant advantages:

[0070] (1) We construct a dynamic graph to capture the temporal dynamics of global features in multiple frequency bands and introduce a dynamic graph convolution module that can adaptively learn and update the graph. For each time window, a dedicated graph is dynamically learned, thereby improving the generalization ability of the model.

[0071] (2) The present invention adopts a dual-channel network that can simultaneously extract global and local features of multiple frequency bands, thereby overcoming the limitation of relying only on a single type of feature. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 A dynamic correlation diagram between multiple frequency bands provided in an embodiment of the present application.

[0073] Figure 2 A multi-band prediction model diagram based on a dynamic graph provided in an embodiment of the present application.

[0074] Figure 3 A long short-term memory network diagram provided for an embodiment of the present application.

[0075] Figure 4 A node selection analysis diagram for constructing a dynamic graph based on k-nearest neighbors provided in an embodiment of the present application.

[0076] Figure 5 The results of the multi-band spectrum prediction method proposed by the present invention at different prediction step sizes are provided in the embodiments of the present application. DETAILED DESCRIPTION

[0077] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0078] The following first introduces the embodiments of the present application in conjunction with the accompanying drawings.

[0079] The following first introduces the embodiments of the present application in conjunction with the accompanying drawings.

[0080] The present invention uses a directed graph model to describe the spatial correlation between user service behaviors and considers the importance differences of different regions. In addition, a spatiotemporal fusion structure is used to simultaneously extract spatial and temporal features in order to more comprehensively understand the user service behavior characteristics and make more accurate predictions. The present invention uses a large amount of historical user service behavior data from multiple days and multiple regions to predict the user service behavior status for multiple days in the future.

[0081] This application constructs a dynamic graph for the multi-band correlation of different time windows. In the graph, each frequency band is regarded as a node, and the correlation between each frequency band and other frequency bands is used to construct edges. According to the k-nearest neighbor algorithm, the k frequency bands with the highest correlation are selected as having connectivity, otherwise, there is no connection relationship. And the historical spectrum data of each node is arranged in chronological order to form a multi-band time correlation tensor. Then, the graph structure and the node's time correlation tensor are input into the dual-channel prediction model, and finally the prediction result is output.

[0082] The steps of using the method of the present invention to predict multi-band spectrum are as follows: 1. Construct a multi-band dynamic graph structure with a sliding time window; step 2, construct a multi-band spectrum tensor based on historical data; step 3, a time-frequency dual-channel spectrum prediction model based on the dynamic graph; step 4, evaluate the performance of the multi-band spectrum prediction method based on the dynamic graph. Each step is described in detail below:

[0083] Step 1: Construct a multi-band dynamic graph structure of a sliding time window.

[0084] The purpose of this step is to capture the time-varying features on multiple frequency bands and construct a dynamic graph using a sliding window method. The sliding window is applied to the frequency bands, and a dynamic graph that changes over time is constructed by calculating the similarity relationship between frequency bands in each window. This dynamic graph can not only intuitively show the changing trend of the similarity between frequency bands, but also provide in-depth insights into the time-varying features of multiple frequency bands.

[0085] Since the correlation between multiple frequency bands in different time windows is dynamically changing, a dynamic graph is constructed for multiple frequency bands by sliding the event window to capture the dynamic relationship of multiple frequency bands; a dynamic graph is constructed by sliding the time window. in represents a set of nodes, ε represents a set of edges, specifically, each frequency band f i are all considered a node v in the dynamic graph i, the similarity between frequency bands forms the edge of the graph. Specifically, the correlation between the frequency band and all frequency bands except this frequency band is calculated, and then the k frequency bands with the highest correlation with this frequency band are selected to establish the connection edge. n In the time window w i The similarity formula on is:

[0086]

[0087] w i =[t p ,t p +1,...,t q ]

[0088] in and Respectively represent the frequency bands f1 and f n In the time window w i The similarity between two frequency bands is measured using the Euclidean distance as a metric, where α defaults to 2 and eps represents a small number to avoid division by zero, which is set to 1e-6 by default. p to q represents the size of the time window. Then, for each node v i , find its k most similar nodes to add edges.

[0089] Step 2: Construct a multi-band spectrum tensor based on historical data.

[0090] Since the spectrum usage has a certain temporal regularity in adjacent time periods, the power spectrum values ​​of each frequency band are smoothly arranged in time as a time tensor for processing;

[0091] The frequency range is A-BMHz, with CMHz as a frequency band width, forming a total of (BA) / C frequency band data. Then, the power spectrum value of each frequency band is analyzed in time order, with T minutes as a time unit, and a size of X is formed in each time unit, which contains all frequency bands. i ∈R n×c A vector of , where c = 1 and n = 59, is then formed into a tensor smoothly arranged in time:

[0092]

[0093] Where m represents time.

[0094] Step 3: Establish a time-frequency dual-channel spectrum prediction model based on dynamic graphs.

[0095] The time-frequency dual-channel spectrum prediction model based on dynamic graph first uses the k-nearest neighbor algorithm to model the similarity relationship between multiple frequency bands, and uses the sliding window method to construct a dynamic graph to capture the time-varying features on multiple frequency bands. Subsequently, the dual-channel feature extraction module is used to extract global and local features respectively. In the global feature extraction module, the dynamic graph convolution module is used to adaptively learn and update the graph structure to effectively process and analyze the dynamic changes between frequency bands, while LSTM is used to extract temporal features. The local feature extraction module is used to extract the features of adjacent regions between multiple frequency bands. Finally, the two types of features are weighted and fused using an adaptive fusion mechanism. The model separates the two modules to ensure that they do not interfere with each other, avoids feature redundancy, and effectively solves the problem of making full use of global and local features between frequency bands.

[0096] The time-frequency dual-channel spectrum prediction model based on dynamic graph is as follows Figure 2 As shown in the figure, the DGCN-LSTM part represents the global feature extraction module, which is used to extract the global features of multiple frequency bands; the CNN-LSTM part represents the local feature extraction module, which is used to extract the local features of multiple frequency bands; the feature fusion part represents the fusion module, which is used to adaptively fuse the multi-band global information and local information features extracted by the global feature extraction module and the local feature extraction module.

[0097] The following is a detailed description of the various parts of the time-frequency dual-channel spectrum prediction model based on dynamic graphs:

[0098] Global feature extraction module:

[0099] The global feature extraction module is used to extract the global level features of multiple frequency bands within the graph structure; global correlation refers to the extensive and non-local interdependence between all time-frequency objects in the frequency domain and time dimension. This global correlation not only considers the interaction between adjacent frequency bands in the frequency domain, but also considers the mutual influence of all frequency bands on a long time scale. When modeling the global features of multiple frequency bands, a dynamic graph structure is used to dynamically change according to the multiple frequency bands in different time windows. In view of the powerful ability of graph convolutional networks in the field of graph learning, a dynamic graph convolutional network (DGCN) is used to capture the information in the graph structure, thereby extracting the graph structure features of multiple frequency bands.

[0100] Specifically, the graph structure and feature matrix are input into the dynamic graph convolution to adaptively update and aggregate node features in different time windows. The specific process is described as follows: First, the features of adjacent nodes are aggregated through the dynamic graph convolution layer:

[0101]

[0102] in It is a graph structure updated and aggregated by a layer of dynamic convolution operations.agg and ψ update are the learnable weights for aggregation and update operations, respectively.

[0103] Specifically, for each node, the aggregation and update operations are as follows:

[0104]

[0105] Yes i The set of neighbor nodes, and v j Yes i In order to adaptively update the node features, a fine-grained edge attention module is used to weight and aggregate the node edges, as follows:

[0106]

[0107] in represents v through the attention module i , Att() represents the function of node weighted aggregation. i ) represents the weighted aggregation of its own features, which is specifically expressed as follows:

[0108]

[0109] in, Represents the original features of the node, μ i,i represents the attention weight. ζ() represents the agg Convolutional layer with parameterized 1×1 convolution kernel. i The weighted aggregation of attention of the neighboring nodes is expressed as

[0110]

[0111] in Indicates the connection between two features. The attention weight is specifically expressed as follows:

[0112]

[0113] where u() is represented by ψ update In addition, considering that the features carried by each node are the power spectrum values ​​of the frequency bands arranged in time series, the time information is extracted from these node features through two LSTM layers to obtain the feature vector χ g .

[0114]

[0115] where ⊙ represents the Hadamard product operator, c i-1 represents the memory information of the previous time step, c i Represents the memory information of the current time step, h i is the output of the LSTM unit, h i-1 is the output of the previous time step. δ represents the weight matrix and b represents the bias vector. Through the global feature extraction module, the dynamic graph convolution is responsible for processing the structural information of the graph, while the LSTM layer focuses on the time evolution of node features, thereby fully capturing the global features of multi-band data with time changes.

[0116] Local feature extraction module:

[0117] In the local feature extraction module, convolutional layers and long short-term memory units are used to extract features in the local range of multiple frequency bands. It includes two convolutional layers and two long short-term memory units.

[0118] The multi-band feature vector Input to the convolution layer, and use the convolution kernel to traverse the feature vector to generate a new feature vector η, effectively extracting the information of the current frequency band and its surrounding frequency bands. The specific process of feature vector generation is as follows:

[0119]

[0120] Among them, δ represents the convolution kernel weight vector, b represents the offset vector, and λ represents the activation function. Then, the feature vector η is input into LSTM to extract the temporal information and obtain the feature χ c .

[0121] Feature fusion module:

[0122] Finally, the adaptive gating mechanism is used to fuse the multi-band global features extracted by the global feature extraction module with the local features extracted by the local feature extraction module to obtain the feature χ f :

[0123]

[0124] Among them, γ represents the learned allocation weight, represents the Sigmoid function, χ g and χ c They represent the weights of multi-band global features and multi-band local features respectively, and b represents the deviation.

[0125] Step 4: Implement a multi-band spectrum prediction method based on dynamic graph.

[0126] The task of this application is to predict the power spectral density value of these n frequency bands in the next d time units This problem is solved by training the prediction model F. The specific operation process is as follows:

[0127]

[0128] The power spectrum value obtained by the multi-band spectrum prediction module based on dynamic graph in the future d time units is compared with the real power spectrum value of this time unit, and the prediction accuracy is evaluated by calculating the error between the real value and the predicted value. The prediction performance evaluation indicators include Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE);

[0129] The root mean square error is a commonly used indicator to measure the difference between the model prediction value and the actual observation value. It is obtained by calculating the mean of the square of the difference between the prediction value and the actual observation value and taking the square root:

[0130]

[0131] The mean absolute error is used to assess how well the model fits the given data; it is calculated by averaging the absolute values ​​of the differences between the predicted values ​​and the actual observed values:

[0132]

[0133] The mean absolute error is sensitive to the relative error and refers to the mean absolute error between the predicted value and the true value. The specific calculation is as follows:

[0134]

[0135] Among these three indicators, X m+i represents the true value of the spectrum of the m+ith time unit, represents the spectrum prediction value of the m+ith time unit, and s represents the prediction time step; the smaller the values ​​of the three evaluation indicators are, the smaller the error between the prediction result obtained by the model and the true value is, which further proves that the prediction effect of the proposed model is better.

[0136] The present application is further described below in conjunction with specific embodiments.

[0137] In this example, the data we used comes from the Electrosense open source data platform. The frequency band bandwidth we selected is 790-820MHz, the data acquisition time resolution is 15 minutes, and the frequency resolution is 0.5MHz. Therefore, there are a total of 59 frequency bands in the data set. In this experiment, we use the power spectrum values ​​of the past 12 hours to predict the power spectrum values ​​of the next 3 hours. We divide the data set into training set, validation set, and test set in a ratio of 7:1:2. In the experiment, the hyperparameters of the model are set to MiniBatchSize, MaxEpochs, 200, InitialLearnRate, and Optimizers, Adam.

[0138] In the process of building the graph structure, the k nearest neighbor algorithm is used to select k nodes with the strongest correlation to construct the connection relationship, so the selection of k is very important. According to the calculation principle of graph convolution, the value of k should not be too large, so we experimented with k ranging from 3 to 18. The experimental results are as follows: Figure 3 As shown in the figure, when k is 12, the prediction effect is the best, so for each node, the top 12 nodes with the strongest correlation are selected to set the connectivity. After continuous iterative optimization, the algorithm converges and outputs the prediction results.

[0139] Comparison scheme: The method proposed in this invention is compared with the models commonly used in prediction tasks. The comparison models are: GCN-LSTM, CNN-LSTM, LSTM, and CONVLSTM.

[0140] The GCN-LSTM model mainly includes two layers of graph convolution and two layers of long short-term memory units. The difference from the global feature extraction module of the present invention is that the GCN-LSTM model has no dynamic characteristics and can only extract features from static graphs.

[0141] The CNN-LSTM model consists of two convolutional layers and two layers of long short-term memory units.

[0142] The LSTM model contains only two layers of long short-term memory units.

[0143] The CONVLSTM model is a model used to extract the appearance of two-dimensional matrix features and is often used in prediction tasks.

[0144] The prediction effect achieved by the model proposed in this application is compared with other models, and the comparison results are reflected by prediction evaluation indicators. The prediction effects and comparison indicators achieved by the five models are shown in Table 1. It can be seen from the table that the model proposed in this application achieves the best prediction effect. In order to more intuitively demonstrate the advantages of the proposed model over other models, the comparison under different prediction steps is shown, such as Figure 4 shown.

[0145] Table 1: Comparison results of prediction indicators between dynamic multi-band prediction model and comparison model

[0146]

[0147]

[0148] In order to verify the effectiveness of dynamic graphs, an ablation analysis is performed on the spectrum prediction method based on dynamic graphs proposed in this paper. The spectrum prediction based on dynamic graphs and that based on static graphs are compared. The comparison results are shown in Table 2. The results show that dynamic graphs can better capture the characteristics of multi-band spectra.

[0149] Table 2: Comparison of prediction indicators between dynamic multi-band prediction model and static graph model

[0150] MAE RMSE MAPE Static Image 0.7848 1.4393 6.6491% Dynamic graph 0.7477 1.4022 6.4177%

[0151] The significant advantages of the present invention are:

[0152] (1) We construct a dynamic graph to capture the temporal dynamics of global features in multiple frequency bands and introduce a dynamic graph convolution module that can adaptively learn and update the graph. For each time window, a dedicated graph is dynamically learned, thereby improving the generalization ability of the model.

[0153] (2) The present invention adopts a dual-channel network that can simultaneously extract global and local features of multiple frequency bands, thereby overcoming the limitation of relying only on a single type of feature.

[0154] The above-described embodiments of the present application do not constitute a limitation on the protection scope of the present application.

Claims

1. A multi-band spectrum prediction method based on dynamic graph, characterized in that: The method comprises: Step 1, construct a multi-band dynamic graph structure of a sliding time window; Step 2, construct a multi-band spectrum tensor based on historical data; Step 3: a time-frequency dual-channel spectrum prediction model based on dynamic graph; Step 4: Implement a multi-band spectrum prediction method based on dynamic graph.

2. The method according to claim 1, characterized in that Construct a multi-band dynamic graph structure of a sliding time window, including: Since the correlation between multiple frequency bands in different time windows changes dynamically, a dynamic graph is constructed for multiple frequency bands by sliding the event window to capture the dynamic relationship of multiple frequency bands. Constructing dynamic graphs through time sliding windows in represents a set of nodes, ε represents a set of edges, and each frequency band f i are considered a node v in the dynamic graph i , the similarity between the frequency bands forms the edge of the graph; the correlation between the calculated frequency band and all frequency bands except this frequency band is calculated, and then the k frequency bands with the highest correlation with this frequency band are selected to establish the connection edge; the frequency bands f1 and f n In the time window w i The similarity on is: w i =[t p ,t p +1,...,t q ] in and Respectively represent the frequency bands f1 and f n In the time window w i The spectrum state on the spectrum; use the Euclidean distance as a metric to measure the similarity between two frequency bands, where α defaults to 2, eps represents a small number to avoid division by zero, and is set to 1e-6 by default; t p to q Represents the size of the time window.

3. The method according to claim 1, characterized in that Step 2: construct a multi-band spectrum tensor based on historical data, including: The frequency range is A-BMHz, with CMHz as a frequency band width, forming a total of (BA) / C frequency band data; then, the power spectrum value of each frequency band is analyzed in time order, with T minutes as a time unit, and a size of X is formed in each time unit containing all frequency bands. i ∈R n×c A vector of , where c = 1, n = (BA) / C, is then formed into a tensor smoothly arranged in time: Where m represents time.

4. The method according to claim 1, characterized in that Step 3: Establish a time-frequency dual-channel spectrum prediction model based on dynamic graph, including: The establishment of a time-frequency dual-channel spectrum prediction model based on dynamic graphs includes three modules, namely, a dynamic graph construction module, a global feature extraction module, and a local feature extraction module; The dynamic graph construction model generates a graph structure for the multi-bands in each time window to represent the similarity between the multi-bands. The multi-band spectrum is divided into several bands according to the frequency band interval of CMHz, and the power spectrum value of each band is arranged in time order. Then, the sliding time window t is divided. p to q , the multi-band power spectrum in each time window is subjected to correlation analysis, and the most relevant k nodes are selected using the k nearest neighbor algorithm to establish a connection relationship. In this way, each time window dynamically forms a graph structure G=<V,E> ; Then the constructed graph structure G =<V,E> and node features χ={X1,X2,...,X m }∈R n×m×c Input to two parallel modules, namely global feature extraction module and local feature extraction module; The extracted global features and local features are fused, and finally the time-frequency dual-channel spectrum prediction model is output; In the global feature extraction module, the dynamic graph convolution module is used to adaptively learn and update the graph structure, and the long short-term memory (LSTM) network is used to extract temporal features. In the local feature extraction module, local features of adjacent areas between multiple frequency bands are extracted. The gate fusion mechanism is used to weight and fuse these two types of features. Finally, the power spectral density value of the next d time units is output.

5. The method according to claim 4, characterized in that The global feature extraction module includes three dynamic graph convolution layers and two LSTM layers; The dynamic graph convolution layer aggregates the features of adjacent nodes and updates the features of nodes in different time windows, as shown below: in It is a graph structure updated and aggregated by a layer of dynamic convolution operations; ψ agg and ψ update are the learnable weights for aggregation and update operations, respectively; For each node, the aggregation and update operations are as follows: Yes i The set of neighbor nodes, and v j Yes i A neighbor node of ; In order to adaptively update the node features, a fine-grained edge attention module is used to weight and aggregate the node edges, and the operation is as follows: in represents v through the attention module i Node features, Att() represents the function of node weighted aggregation; where Att(v i ) represents the weighted aggregation of its own features, as follows: in, Represents the original features of the node, μ i,i represents the attention weight; ζ() represents the agg Convolutional layer with parameterized 1×1 convolution kernel; v i The weighted aggregation of attention of the neighboring nodes is expressed as in Represents the connection between two features; the attention weights are as follows: where u() is represented by ψ update Convolutional layer with parameterized 1×1 convolution kernel; Two LSTM layers are used to extract time information from the corresponding node features to obtain the feature vector χ g : where ⊙ represents the Hadamard product operator, c i-1 represents the memory information of the previous time step, c i Represents the memory information of the current time step, h i is the output of the LSTM unit, h i-1 is the output of the previous time step; δ represents the weight matrix, and b represents the bias vector; Through the global feature extraction module, dynamic graph convolution is responsible for processing the structural information of the graph, while the LSTM layer is used for the temporal evolution of node features, thereby comprehensively capturing the global features of multi-band data with time changes.

6. The method according to claim 4, characterized in that The local feature extraction module is used to extract local feature information of multi-band spectrum data. The local feature extraction module includes two convolutional layers and two LSTM layers; The multi-band feature vector Input to the convolution layer, and use the convolution kernel to traverse the feature vector to generate a new feature vector η. The feature vector generation is as follows: Among them, δ represents the convolution kernel weight vector, b represents the offset vector, and λ represents the activation function. Represents the convolution operation; Then, the feature vector η is input into LSTM to extract the temporal information and obtain the feature χ c ; Finally, the adaptive gating mechanism is used to fuse the multi-band global features extracted by the global feature extraction module with the local features extracted by the local feature extraction module to obtain the feature Among them, γ represents the learned distribution weight, θ represents the Sigmoid function, and χ g and χ c They represent the weights of multi-band global features and multi-band local features respectively, and b represents the deviation.

7. The method according to claim 1, characterized in that Step 4, implementing a multi-band spectrum prediction method based on a dynamic graph, including: Predict the power spectral density values ​​of n frequency bands in the future d time units The power spectral density value is determined by training the prediction model F: First, the frequency band bandwidth is selected as A-BMHz, and then the power spectrum values ​​of the past m time units are used to predict the power spectrum values ​​of the next d time units; in the process of constructing the graph structure, the k nearest neighbor algorithm is used to select k nodes with the strongest correlation to construct the connection relationship; after continuous iterative optimization, the algorithm converges and then outputs the prediction result; finally, the power spectrum value obtained by the multi-band spectrum prediction module based on dynamic graph in the future d time units is compared with the actual power spectrum value of the current time, and the prediction accuracy is evaluated by calculating the error between the actual value and the predicted value. The prediction performance evaluation indicators include mean absolute error, root mean square error and mean absolute error; The root mean square error is calculated by taking the mean of the squares of the differences between the predicted values ​​and the actual observed values ​​and taking the square root: The mean absolute error is calculated by averaging the absolute values ​​of the differences between the predicted values ​​and the actual observed values: The mean absolute error is sensitive to the relative error and refers to the mean absolute error between the predicted value and the true value. The determination method is as follows: X m+i represents the true value of the spectrum of the m+ith time unit, It represents the spectrum prediction value of the m+ith time unit, and s represents the prediction time step. The prediction effect of the model is obtained by calculating three error indicators.

Citation Information

Patent Citations

  • Multi-band base station equipment signal transmitting and receiving method and device

    CN117119620A

  • Multi-modal feature fusion emotion recognition method based on gating cross-attention mechanism

    CN117370828A

  • Spectrum prediction algorithm based on hypergraph convolutional neural network

    CN118473561A

  • Audiovisual device and method

    US20230156272A1