Frequency spectrum prediction method based on time-frequency fusion attention network

Abstract

The invention relates to the technical field of electromagnetic frequency spectrums, and particularly discloses a frequency spectrum prediction method based on a time-frequency fusion attention network. Comprising the steps of constructing a spectrum data prediction network model, setting initial learning parameters, collecting spectrum data for periodic analysis, and determining time period attributes of the spectrum data; extracting time domain features of the spectrum data by using a time convolution network, and transmitting the spectrum data to a graph convolution layer network by using a linear gating unit; extracting frequency domain features of the spectrum data by using a graph convolutional network; dynamically weighting the extracted time domain features and frequency domain features by using an attention mechanism to obtain predicted spectrum data; calculating an average absolute error according to the predicted spectrum data and the collected spectrum data; reversely training the spectrum data prediction network model according to the mean absolute error, optimizing learning parameters, and completing training of the spectrum data prediction network model; and inputting the to-be-processed spectrum data into the trained spectrum data prediction network model to obtain a prediction result.

Description

Technical Field

[0001] The present invention relates to the field of electromagnetic spectrum technology, and in particular to a spectrum prediction method based on a time-frequency fusion attention network. Background Art

[0002] With the development of wireless networks and communication technologies, the number of frequency-using devices and their users continues to increase, and the demand for spectrum resources in wireless communications continues to grow rapidly. Traditional static spectrum allocation strategies are difficult to cope with the dynamic spectrum demand in this complex environment, resulting in low utilization, which further exacerbates the shortage of spectrum resources. Spectrum prediction technology uses the characteristic information hidden in historical spectrum data and combines it with relevant information of other channels in the same service to obtain parameters such as the occupancy status, duty cycle, and power spectrum of future spectrum channels. It plays an important role in optimizing spectrum utilization, spectrum resource allocation, and supporting spectrum sharing. Analyzing changes in historical spectrum data, summarizing spectrum usage patterns, and predicting users' future spectrum usage can better achieve efficient spectrum resource management and intelligent spectrum switching strategies.

[0003] Traditional model-driven methods include autoregressive analysis models, hidden Markov models, Bayesian models, etc. However, in actual spectrum environments, the state transition probability may change over time, and this assumption limits the applicability of the model. The spatial, frequency, and temporal dimensions of spectrum utilization patterns are random, depending on factors such as the wireless propagation environment, transmission power, and user mobility. Traditional spectrum prediction methods face challenges such as uneven distribution and dynamic changes of spectrum resources. It is necessary to study spectrum prediction methods that adapt to complex and dynamic spectrum environments.

[0004] Since deep learning can capture the potential complex dependencies in nonlinear spectrum data, more and more neural networks are applied to spectrum prediction tasks, such as convolutional neural networks, recurrent neural networks, and long short-term memory networks. In actual spectrum utilization, due to the periodicity of user behavior, spectrum data presents different characteristics in different time periods, thus forming an obvious periodic law, and the spectrum occupancy state presents significant periodicity. There are two specific problems with existing deep learning methods when performing spectrum prediction tasks: some methods only focus on time dimension feature extraction and fail to fully utilize the important feature information contained in the frequency dimension of spectrum data. Although it can capture changes in time, it lacks consideration of frequency domain features and cannot fully reflect the dynamic complexity of spectrum data. On the other hand, although some methods try to combine the time dimension and frequency dimension for prediction, they ignore the influence of time periodicity, which limits the model's accurate capture of data changes. Summary of the invention

[0005] In view of the above problems, the purpose of the present invention is to provide a spectrum prediction method based on time-frequency fusion attention network, to construct a spectrum prediction model that comprehensively characterizes the correlation between time domain and frequency domain, and to optimize the problem of insufficient prediction accuracy of the spectrum prediction model based on deep learning.

[0006] The technical solution adopted by the present invention is: a spectrum prediction method based on time-frequency fusion attention network, comprising:

[0007] Step S01: constructing a spectrum data prediction network model, setting initial learning parameters, inputting the collected spectrum data into the spectrum data prediction network model, performing periodic analysis on the spectrum data, and determining the time period attribute of the spectrum data;

[0008] Step S02: extracting time domain features of the spectrum data using a temporal convolutional network, and transferring the spectrum data to a graph convolutional layer network using a linear gating unit according to the time domain features;

[0009] Step S03: extracting frequency domain features of the spectrum data using a graph convolutional network;

[0010] Step S04: dynamically weighting the extracted time domain features and the frequency domain features using an attention mechanism to obtain predicted spectrum data;

[0011] Step S05: Calculate the mean absolute error based on the predicted spectrum data and the collected spectrum data;

[0012] Step S06: reversely training the spectrum data prediction network model according to the mean absolute error, optimizing the learning parameters, and completing the training of the spectrum data prediction network model;

[0013] Step S07: input the spectrum data to be processed into the trained spectrum data prediction network model to obtain a prediction result.

[0014] Preferably, the method for determining the time period attribute in step S01 includes: determining the minimum historical data length corresponding to the same evolution trend according to the evolution trend of the spectrum data in different time slots, and determining the time period attribute of the spectrum data according to the time period corresponding to the historical data length.

[0015] Preferably, the temporal convolutional network in step S02 includes three dilated causal convolutions, and the three dilated causal convolutions are respectively used to receive spectrum data of different time period attributes.

[0016] Preferably, the time period attributes include proximity trend, daily trend and weekly trend.

[0017] Preferably, the linear gating unit in step S02 includes a gating mechanism, and the gating mechanism is used to filter and selectively transmit information of the spectrum data.

[0018] Preferably, the method for extracting the frequency domain features in step S03 includes: constructing a correlation coefficient matrix according to the frequency band correlation of the spectrum data, and extracting the frequency domain features by utilizing the ability of a graph convolutional network to aggregate node features.

[0019] Preferably, the dynamic weighting process in step S04 includes: summing up the time domain features and frequency domain features corresponding to the spectrum data of different time period attributes, and using an attention layer to assign different weights to the time domain features and frequency domain features corresponding to different spectrum data.

[0020] Preferably, the predicted spectrum data in step S04 is output through an output layer, and the output layer includes two convolutional layers for deepening the depth of the spectrum data prediction network model, and adjusting and controlling the output dimension.

[0021] Preferably, the learning parameters in step S01 include the number of iterations, the learning rate and the batch size.

[0022] Preferably, when the number of iterations set in the learning parameters is completed, the training of the spectrum data prediction network model is completed.

[0023] Beneficial effects of the above technical solution:

[0024] The present invention discloses a spectrum prediction method based on a time-frequency fusion attention network, which solves the problems of insufficient consideration of data features, weak model robustness, and insufficient prediction accuracy in existing frequency domain prediction algorithms. First, based on a convolutional network, an expansion factor is introduced to increase the modeling range, so that the network level considers longer-term time dependencies and fully extracts time domain features; then, the ability of graph convolutional networks to aggregate node features is used to fully extract frequency domain features, and finally, the attention mechanism is used to dynamically weight time domain features and frequency domain features, thereby improving the prediction accuracy of the model while fully considering the multi-dimensional features of spectrum data. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of a spectrum prediction method based on a time-frequency fusion attention network provided by an embodiment of the present invention;

[0026] Figure 2 is a proximity trend graph of spectrum data provided by an embodiment of the present invention;

[0027] Figure 3 is a trend diagram of spectrum data transformation with a daily cycle provided by an embodiment of the present invention;

[0028] Figure 4It is a trend diagram of spectrum data transformation with a weekly period provided by an embodiment of the present invention;

[0029] Figure 5 is a spectrum prediction network framework diagram provided by an embodiment of the present invention;

[0030] Figure 6 is a comparison diagram of the spectrum prediction method provided by an embodiment of the present invention and other existing spectrum prediction methods;

[0031] Figure 7 is a comparison chart of the training convergence speed of the spectrum prediction method provided by the embodiment of the present invention and other existing spectrum prediction methods;

[0032] Figure 8 This is a comparison chart of the training speeds of the spectrum prediction method provided by an embodiment of the present invention and other existing spectrum prediction methods. DETAILED DESCRIPTION

[0033] The following is a further detailed description of the implementation methods of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0034] The terms "first", "second", etc. (if any) in the specification and claims are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way are interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0035] It should be understood that the term "and / or" used in the present invention is only a description of the association relationship of associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in the present invention generally indicates that the associated objects before and after are in an "or" relationship.

[0036] Figure 1 This is a flow chart of the spectrum prediction method based on the time-frequency fusion attention network implemented by the present invention. Figure 1 , the method of the present invention comprises:

[0037] Step S01: constructing a spectrum data prediction network model, setting initial learning parameters, inputting the collected spectrum data into the spectrum prediction network model, performing periodic analysis on the spectrum data, and determining the time period attribute of the spectrum data;

[0038] Step S02: extracting time domain features of the spectrum data using a temporal convolutional network, and transferring the spectrum data to a graph convolutional layer network using a linear gating unit according to the time domain features;

[0039] Step S03: extracting frequency domain features of the spectrum data using a graph convolutional network;

[0040] Step S04: dynamically weighting the extracted time domain features and the frequency domain features using an attention mechanism to obtain predicted spectrum data;

[0041] Step S05: Calculate the mean absolute error based on the predicted spectrum data and the collected spectrum data;

[0042] Step S06: reversely training the spectrum data prediction network model according to the mean absolute error, optimizing the learning parameters, and completing the training of the spectrum data prediction network model;

[0043] Step S07: input the spectrum data to be processed into the trained spectrum data prediction network model to obtain a prediction result.

[0044] The above spectrum prediction method is further described below with reference to the accompanying drawings.

[0045] Step 1: construct a spectrum data prediction network model, collect spectrum data, input the spectrum data into the spectrum data prediction network model, perform periodic analysis on the spectrum data, and determine the time period attribute of the spectrum data.

[0046] The radio spectrum state of a frequency band has regularity and randomness in its time evolution. The reason why the spectrum state can be predicted by historical data is that the spectrum state is correlated in the time domain and the frequency domain. From the collected measured spectrum data, various time period attributes can be observed, namely, proximity trend, daily trend, and weekly trend. The following formula is used to determine the input vector of different time period attributes:

[0047]

[0048] Among them, I W ,I D ,I C are the characteristic vectors of weekly trend, daily trend and proximity trend, respectively, W t , D t , C tare the lengths of historical data corresponding to the trend, daily trend, and proximity trend, t is the current moment, I t is the spectrum data at time t, p is the spectrum sensing frequency, X W ∈{1,...,XW} is an integer representing the week. X d ∈{1,...,XD} represents the number of days.

[0049] Perform time domain analysis on the measured data, such as Figure 2 It is the trend diagram of the proximity of spectrum data, that is, the ratio curve of SNR. The x-axis is the spectrum sensing time slot interval, and the y-axis is the average ratio between the spectrum data sensed at different time intervals. It can be seen that the data ratio increases with the increase of the time interval, and the similarity between the data decreases accordingly; Figure 3 The trend chart of spectrum data changing with daily cycle shows the sensing data of all time slots in a week. The spectrum data shows obvious periodic characteristics with daily cycle. Figure 4 The figure is a trend chart of spectrum data transformation with a weekly period. It gives the transformation trend of spectrum data with a weekly period. The data has the same trend in the same time period.

[0050] Step 2: Use the temporal convolutional network to obtain the time domain features of the spectrum data.

[0051] The time domain input vector obtained in step 1 is input into the time convolution network, and the dilated causal convolution is used to combine causality and dilatation, so that the spectrum data prediction network model can effectively extract long-term dependent features without violating the time order. Different from the traditional spectrum prediction method, which only considers the closest historical data as prediction dependency, the spectrum prediction method proposed in the present invention uses three dilated causal convolutions (DCC) to receive the time domain feature vectors of proximity trend, daily trend, and periodic trend respectively, such as Figure 5 The calculation process is as follows:

[0052]

[0053] Among them, W i is the weight of the convolution kernel, K is the parameter of the convolution kernel (i.e. the number of weights), d is the expansion factor, which is used to control the sampling interval, x t-(i-1)×d Represents the value of the input sequence at time t-(i-1)×d.

[0054] Use linear gating units to control information transmission. The Gated Linear Unit (GLU) selectively transmits and filters information by introducing a gating mechanism, thereby improving the model's expressiveness and training efficiency. The extracted time domain features are passed to the next level of the network through the gated linear unit. The calculation process is as follows:

[0055] h T = (W1*x+b1)⊙σ(W2*x+b2)

[0056] Among them, W1, W2, b1, b2 are learnable parameters, x is the input feature, ⊙ is the element-by-element product between matrices, σ is the sigmoid activation function, and h T is the output of the linear gating unit.

[0057] Step 3: Use graph convolutional network to obtain the frequency domain features of spectrum data.

[0058] The correlation coefficient between two random variables can be expressed by the Pearson coefficient:

[0059]

[0060] Among them, x n 、x m are spectrum data of different frequency bands, P∈R N×N is the correlation coefficient matrix, cov(,) is the covariance function, and σ(·) is the standard deviation. The larger the correlation coefficient, the stronger the correlation.

[0061] Therefore, we use the ability of graph convolutional network (GCN) to aggregate node features and extract frequency domain features using frequency band correlation. The calculation process is as follows:

[0062]

[0063] Among them, h G is the output of the graph convolutional network, X is the feature matrix, representing the output data of GLU. θ is the learnable convolution kernel. E node ∈R N×d represents a learnable graph node embedding matrix, I N is the identity matrix.

[0064] Using E node and Multiply them together to get the adaptive adjacency matrix A G , A G Instead of the widely used similarity adjacency matrix, nodes can automatically obtain the main dependencies. Figure 5As shown in Figure 2, the output of the gated linear unit is used as the input of the graph convolutional network.

[0065] Step 4: Get the predicted output through dynamic weighting through the attention mechanism.

[0066] like Figure 5 As shown, the attention mechanism is introduced to dynamically weight the extracted time domain features and frequency domain features to accelerate the convergence process. Preferably, all time domain features and frequency domain features in each time-frequency block (TF Bolck) are summed up and then different weights are assigned to multiple features through an attention layer. The calculation process is as follows:

[0067]

[0068] Among them, q is the query vector, α(q,k i ) is the query vector q and the key vector k i The similarity score between i is the output of the linear gating unit or graph convolution, and h is the weighted sum result calculated by the attention mechanism.

[0069] Use the normalized weights to compare the corresponding values ​​h i Perform weighted summation to obtain the final representation h.

[0070] Sum all the outputs h obtained from different time-frequency blocks to get h out , through an output layer, which contains two convolutional layers, to increase the depth of the model, thereby improving the generalization ability of the spectrum data prediction network model, and gradually adjusting and controlling the output dimension. The calculation process is as follows:

[0071]

[0072] Where W c1 , W c2 、b c1 、b c2 are the learnable weights and biases in the convolutional layer, and ReLU is the activation function.

[0073] Step 5: Calculate the mean absolute error based on the predicted spectrum data and the collected spectrum data.

[0074] First, all learnable parameters in the spectrum data prediction network model are randomly initialized, and the number of iterations, learning rate, and batch size are set. The model selects Adam as the optimizer. The number of initial iterations is 1, the maximum number of iterations is 500, the learning rate is 0.0001, and the batch size is 64. The training data set is input into the spectrum data prediction network model in batches for training. The training error of each batch is back-propagated to optimize all learnable parameters. The error loss uses the mean absolute error (MAE) between the predicted value and the true value. The calculation process is as follows:

[0075]

[0076] Where N is the number of data points, Y i is the i-th true value, is the i-th and predicted value, is the absolute error of the ith data.

[0077] When all the data in the training data have completed forward output and back propagation, one iteration is completed.

[0078] Step 6: Use the training set data to train the spectrum data prediction network model.

[0079] Determine whether the current number of iterations has reached the maximum number of iterations. If so, complete the training and output the prediction results. If not, add 1 to the number of iterations and return to continue training the model.

[0080] Step 7: Input the test data set into the trained spectrum data prediction network model and output the prediction result.

[0081] The effect of the present invention is further described below in conjunction with simulation experiments.

[0082] 1. Simulation conditions and parameter settings

[0083] The simulation experiment of the present invention is carried out on the simulation platform of python3.7 and Pytorch1.9. The computer CPU model is Intel(R) Xeon(R) CPU E5-2680 v4, equipped with an independent graphics card model Nvidia GeForce RTX 3060.

[0084] The maximum number of iterations of the network is 500, the learning rate is 0.0001, the Adam optimization algorithm is selected as the network training optimizer, and the data size of each batch is 64.

[0085] 2. Simulation content

[0086] Figure 6Comparison diagram of the spectrum prediction method of the embodiment of the present invention and other existing spectrum prediction methods. TFFAN is the present invention, and the method without adding the attention network is called TFFN, and the other parameter settings are the same. The first 10 time steps are used to predict the input, and the output is the last 5 time steps. In order to cover the length of the input sequence, the dilation factors of the dilated causal convolution are set to 1, 2, 1, 2, 1, 2.

[0087] Figure 6 The horizontal axis represents different models, and the vertical axis represents the predicted root mean square error and mean absolute error. It can be seen from the figure that compared with the TFFN method that adds an attention network, the present invention improves MAE and RMSE by 5.6% and 5.4% respectively, which verifies the effectiveness of the proposed scheme in considering time diversity and frequency domain features. Compared with traditional recurrent neural networks (RNN), long short-term memory networks (LSTM) and gated recurrent units (Gated Recurrent Unit), the MAE is improved by 25.9%, 22% and 5.7% respectively.

[0088] Table 1 lists the prediction error values ​​(RMSE, MAE, MAPE) of the five models.

[0089] Table 1 Numerical results of TFFAN model prediction error

[0090]

[0091] It can be observed from Table 1 that the prediction errors of all models are relatively small, but compared with other methods, the TFFAN method proposed in the present invention shows the best prediction performance in all evaluation indicators. Multi-step time series prediction requires the model to have good long-term memory ability and be able to capture long-term dependencies in time series. Due to the lack of a gating mechanism, RNN cannot selectively update the state in multiple time steps, which leads to the easy neglect of important information or retention of unnecessary information when processing complex time series patterns; although LSTM has a strong ability to capture long-term dependencies, its complex gating mechanism will lead to excessive filtering of information or the propagation of redundant information; GRU has a simpler structure and is easier to find the optimal parameter configuration during training, thereby reducing the accumulation of prediction errors. The TFFN method simply sums the extracted time features and frequency features and obtains the predicted value through the output layer. After removing the attention mechanism, the model loses the ability of dynamic weighting, insufficient information extraction, and reduced ability to capture complex patterns. In general, the proposed method has a MAE 5.6% to 25.9% lower, a RMSE 5.5% to 28.2% lower, and a MAPE 4.7% to 25.4% lower than that of the benchmark algorithm, with average improvements of 15.8%, 16.9%, and 15.1%, respectively.

[0092] In order to further verify the training stability and training efficiency of the model, this paper records the changes in loss values ​​and total training time during the training process of each model, such as Figure 7 As shown in the figure, compared with the TFFN network, the loss of the TFFAN network with the attention network is significantly reduced faster, but it is worth noting that compared with other methods, the convergence speed of the TFFAN network is slower. This is because it considers both time features and frequency features at the same time, and needs to adapt and adjust the weights and relationships of the two types of features. But at the same time, MAE has been on a downward trend, indicating the continuous optimization ability of the model.

[0093] like Figure 8 The figure shows a comparison of training efficiency. It can be seen that the training efficiency of TFFN without adding the attention mechanism is only 6.4% higher than that of the proposed TFFAN method, but the prediction accuracy is reduced by 5.7%; compared with GRU and LSTM, the training efficiency of the TFFAN model is improved by 36.8% and 48.1% respectively. In addition, although the training efficiency of RNN is only 7.3% lower than that of TFFAN, the prediction accuracy of the RNN model is reduced by 25.9%.

[0094] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not limitations on the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made on the basis of the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the protection scope of the present invention.

Claims

1. A spectrum prediction method based on time-frequency fusion attention network, characterized in that: include: Step S01: constructing a spectrum data prediction network model, setting initial learning parameters, inputting the collected spectrum data into the spectrum data prediction network model, performing periodic analysis on the spectrum data, and determining the time period attribute of the spectrum data; Step S02: extracting time domain features of the spectrum data using a temporal convolutional network, and transferring the spectrum data to a graph convolutional layer network using a linear gating unit according to the time domain features; Step S03: extracting frequency domain features of the spectrum data using a graph convolutional network; Step S04: dynamically weighting the extracted time domain features and the frequency domain features using an attention mechanism to obtain predicted spectrum data; Step S05: Calculate the mean absolute error based on the predicted spectrum data and the collected spectrum data; Step S06: reversely training the spectrum data prediction network model according to the mean absolute error, optimizing the learning parameters, and completing the training of the spectrum data prediction network model; Step S07: input the spectrum data to be processed into the trained spectrum data prediction network model to obtain a prediction result.

2. The spectrum prediction method based on time-frequency fusion attention network according to claim 1 is characterized in that: The method for determining the time period attribute in step S01 includes: determining the minimum historical data length corresponding to the same evolution trend according to the evolution trend of the spectrum data in different time slots, and determining the time period attribute of the spectrum data according to the time period corresponding to the historical data length.

3. The spectrum prediction method based on time-frequency fusion attention network according to claim 1 is characterized in that: The temporal convolutional network in step S02 includes three dilated causal convolutions, and the three dilated causal convolutions are respectively used to receive spectrum data of different time period attributes.

4. The spectrum prediction method based on time-frequency fusion attention network according to claim 3 is characterized in that: The time period attributes include proximity trend, daily trend and weekly trend.

5. The spectrum prediction method based on time-frequency fusion attention network according to claim 1 is characterized in that: The linear gating unit in step S02 includes a gating mechanism, and the gating mechanism is used to filter and selectively transmit information of the spectrum data.

6. The spectrum prediction method based on time-frequency fusion attention network according to claim 1 is characterized in that: The method for extracting the frequency domain features in step S03 includes: constructing a correlation coefficient matrix according to the frequency band correlation of the spectrum data, and extracting the frequency domain features by utilizing the ability of the graph convolutional network to aggregate node features.

7. The spectrum prediction method based on time-frequency fusion attention network according to claim 1 is characterized in that: The dynamic weighting process in step S04 includes: summing up the time domain features and frequency domain features corresponding to the spectrum data of different time period attributes, and using an attention layer to assign different weights to the time domain features and frequency domain features corresponding to different spectrum data.

8. The spectrum prediction method based on time-frequency fusion attention network according to claim 7 is characterized in that: In step S04, the predicted spectrum data is output through an output layer, and the output layer includes two convolutional layers, which are used to deepen the depth of the spectrum data prediction network model and adjust and control the output dimension.

9. The spectrum prediction method based on time-frequency fusion attention network according to claim 1 is characterized in that: The learning parameters in step S01 include the number of iterations, the learning rate and the batch size.

10. The spectrum prediction method based on time-frequency fusion attention network according to claim 1 is characterized in that: When the number of iterations set in the learning parameters is completed, the training of the spectrum data prediction network model is completed.

Images