Frequency spectrum time sequence prediction method based on time-space-frequency domain feature fusion

Abstract

The invention discloses a spectrum time sequence prediction method based on time-space-frequency domain feature fusion, and belongs to the field of spectrum analysis and multivariate time sequence prediction. The method comprises the following steps: predicting multi-frequency-point time slot sequence data; the method comprises the following steps: firstly, mapping multi-frequency-point time slot sequence data of a historical time step length into a channel hybrid embedding representation vector and a channel independent embedding representation vector; constructing a multi-scale time-frequency feature extraction module and an adaptive graph time-space feature extraction module which are respectively used for capturing multi-scale time-frequency features and time-space features of the channel mixed representation vector and the channel independent representation vector; in addition, a bidirectional interaction feature fusion module is constructed and used for aligning captured time-frequency (frequency domain features obtained by time and Fourier transform) and time-space (spatial correlation features between time and multiple frequency points); and finally, designing an output layer to map the features output by the fusion module into a multi-frequency-point time slot sequence data predicted value of a future time step length.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spectrum analysis and multivariate time series prediction, and particularly relates to a spectrum time series prediction method based on time-space-frequency domain feature fusion. BACKGROUND

[0002] Multivariate time series prediction belongs to the cross-research and application of statistics, machine learning, deep learning and other fields. Multivariate time series contains a group of time series, which can be regarded as an N×L matrix, where N represents the number of variables and L represents the length of time steps. In the time dimension, each time series may have periodicity, trend and other characteristics, and the FFT transformation in the time dimension can reflect the frequency domain features. In addition, in the spatial dimension, there may be spatial features of mutual correlation between time series variables. Therefore, making full use of the time-space-frequency domain multidimensional features of multivariate time series and mining the rich features of historical data can help improve the prediction performance of the model. Specifically, multivariate time series prediction is to predict future data N×L2 based on historical data N×L1. L1 and L2 are the historical time window length and the future time window length, respectively. Traditional time series prediction models require heavy computing tasks and a large amount of manpower. Therefore, deep neural networks have entered people's field of vision due to their ability to extract high-order characteristics. Through deep neural networks, complex patterns within and between time series can be identified, and the original time series data set can be identified, reducing the required manpower. The latest Mamba model combines state space theory with advanced deep learning technology, uses selective state representation, which dynamically adjusts according to input data, and its modeling performance has reached the level of the most advanced Transformer model, while maintaining linear scalability with sequence length.

[0003] In actual application scenarios, due to the wide popularity of radio communication equipment, the available spectrum resources have become increasingly scarce. Therefore, the state will control the spectrum resources more and more strictly. However, many unlicensed transmitters will occasionally interfere with the electromagnetic spectrum of the base station for private interests or destructive purposes. If these interferences cannot be detected in time and accurately, it may cause the signal distortion or even loss of the base station, thereby affecting the individual's life and even seriously threatening the national security. How to accurately predict the time slot of the electromagnetic signal helps to provide decision support for spectrum management. Therefore, the present application proposes a spectrum time series prediction method based on time-space-frequency domain feature fusion, which can learn the time dependence-space correlation-multiscale frequency domain periodicity features of the frequency points from the historical time step multichannel time slot sequence data, and calculate the prediction value of the future time step multichannel time slot sequence. SUMMARY

[0004] The present application aims to solve the problem of insufficient prediction accuracy in spectrum time series prediction due to the failure to fully integrate multi-dimensional features in time domain, frequency domain and spatial domain.

[0005] To achieve the prediction of multi-frequency point time slot sequence data, the present application provides a spectrum time series prediction method based on time-space-frequency domain feature fusion: first, map the multi-frequency point time slot sequence data of the historical time step into a channel mixed embedded representation vector and a channel independent embedded representation vector, then construct a multi-scale time-frequency feature extraction module and an adaptive graph time-space feature extraction module for capturing the multi-scale time-frequency features and time-space features of the channel mixed representation vector and the channel independent representation vector respectively; in addition, a bidirectional interaction feature fusion module is constructed for aligning the captured time-frequency (time and frequency domain features obtained by Fourier transform) and time-space (spatial correlation features between time and multi-frequency points); finally, an output layer is designed to map the features output by the fusion module to the prediction value of the multi-frequency point time slot sequence data at the future time step.

[0006] To achieve the above purpose, the present application provides the following technical scheme: a spectrum time series prediction method based on time-space-frequency domain feature fusion, the specific steps of which are as follows:

[0007] Step S10: collect multi-frequency point time slot sequence data of the historical time step for batch processing and normalization;

[0008] Step S20: construct a linear network layer to map the multi-frequency point time slot sequence data to a channel mixed (Channel-Dependent, CD) representation vector and a channel independent (Channel-Independent, CI) representation vector;

[0009] Step S30: construct a multi-scale time-frequency feature extraction module and an adaptive graph time-space feature extraction module for capturing the time-frequency features and time-space features of the multi-frequency point time slot sequence data from the CD representation vector and the CI representation vector respectively; time-frequency (time and frequency domain features obtained by Fourier transform) and time-space (spatial correlation features between time and multi-frequency points);

[0010] Step S40: construct a bidirectional interaction feature fusion module for aligning the captured time-frequency and time-space features;

[0011] Step S50: construct an output layer to align and decode the feature matrix captured by the bidirectional interaction feature fusion module into the prediction value of the multi-frequency point time slot sequence data at the future time step.

[0012] Further, the specific steps of step S10 are as follows:

[0013] Step S101, select the multi-frequency point time slot sequence data, construct the sequence of field strength value changing with time, originally represented as N is the number of selected frequency points (i.e. as a variable number), and the time slot sequence data of N frequency points in L historical time steps is used as the input of the model.

[0014] Step S102, normalize it, and the calculation formula is as follows:

[0015] x i,j =(x i,j -max(x i,j )) / σ j ,

[0016] Where, i={1,...,L}, j={1,...,N}, σ j is the standard deviation of the frequency point j time slot sequence data.

[0017] Further, the specific steps of the step S20 are as follows:

[0018] Step S201, after normalization, the original representation X is mapped through the linear network layer Linear1 to obtain the CD vector representation The calculation formula is as follows:

[0019] X m emb =Linear1(X)+PE

[0020] Where, represents linear mapping, L is the time step, and d1 is the CD vector representation dimension.

[0021] Step S202, construct the linear layer Linear2 to map to obtain the CI vector representation The calculation formula is as follows:

[0022]

[0023] Where, represents linear mapping, N is the number of frequency points, and d2 is the CI vector representation dimension.

[0024] Further, the specific steps of the step S30 are as follows:

[0025] Step S301, in the multi-scale time-frequency feature extraction module, the frequency domain periodicity mode of the frequency point sequence data is calculated using Fast Fourier Transform (FFT), and the calculation formula is as follows:

[0026]

[0027] where Amp(.) denotes the calculation of amplitude value, denotes the average amplitude, which is averaged over d1 dimensions by function Amp(.). i denotes the period corresponding to the first k amplitude values. i denotes the total time steps divided by the period f i the obtained sequence division scale, Topk(.) denotes the top k, denotes the period frequency f corresponding to the top k amplitude values.

[0028] In order to align the input sequence i under different period scales and obtain the vector representation of different periods in the frequency domain The calculation formula is as follows:

[0029]

[0030] where i={1,...,k} denotes different period patterns in the frequency domain, padding(.) denotes the zero padding operation, and Reshape(.) operation realizes aligning the input sequence

[0031] Step S302, further, using a multi-scale convolutional network model to further process capture multi-scale features under different period patterns, the calculation process is as follows:

[0032]

[0033] where, denotes the multi-scale features obtained by convolution operation, and Inception denotes multi-scale product operation.

[0034] Step S303, further, constructing Mamba layer (a state-aware network structure) to enhance the ability of the model to capture long-term time dependence features, the calculation formula is as follows:

[0035]

[0036] where, denotes the feature representation obtained by Mamba Block operation.

[0037] Step S304, further, using SoftMax function to normalize the amplitude, as shown in the following formula:

[0038]

[0039] where f k denotes the period corresponding to the first k amplitude values, denotes the first k amplitude values.

[0040] Step S305, further, reconstructs the feature representation of each time scale s i to a vector with dimension size.

[0041]

[0042] Step S306, further, multiplies the by the corresponding a1, …, a k , and sums all the results. Then, based on the addition residual structure, the multi-scale time-frequency feature is obtained, and the calculation formula is as follows:

[0043]

[0044] wherein, denotes the multi-scale time-frequency feature.

[0045] Step S307, in the adaptive graph spatial-temporal feature extraction module, to learn the spatial dependency relationship between different frequency point sequences, a learnable adjacency matrix A (which can adaptively describe the spatial dependency relationship between different frequency point sequences) is initialized, and is represented by the following formula:

[0046] A = softMax(ReLu(E1(E2) T ))

[0047] wherein, ReLu(.) denotes an activation function, and E1 and E2 denote learnable parameters.

[0048] Step S308, multiplying the CI vector representation feature by the adjacency matrix A, which can be input for dynamic and adaptive learning, so as to capture more rich spatial dependency features between different frequency points, and is represented by the following formula:

[0049]

[0050] wherein, GCN denotes a graph convolutional network operation (Graph Convolutional Network, GCN), A is an N*N adjacency matrix, is an embedding representation of N frequency point sequences, denotes the output spatial dependency feature vector.

[0051] Step S309, a Mamba layer is constructed to enhance the ability of the model to capture long-term time dependency, and a residual structure is constructed to obtain the time-space feature​​ The calculation formula is as follows:

[0052]

[0053] wherein, represents the time-space characteristics between multi-frequency points.

[0054] Further, the specific steps of the step S40 are as follows:

[0055] Step S401, the bidirectional interaction feature fusion module is a multi-head cross attention mechanism (Multi-head Cross Attention, MCA), which enables the model to simultaneously learn time-frequency and time-space dependent features interactively and perform fusion. This mechanism not only enables bidirectional interactive learning of time-frequency and time-space features, but also effectively filters out irrelevant features in time-frequency and time-space features, improving the performance of the model. The calculation process is as follows:

[0056]

[0057] wherein, are used as queries, keys and values in MCA, respectively, represents the time-space feature and the time-frequency feature interactively learned feature representation.

[0058]

[0059] wherein, are used as queries, keys and values in MCA, respectively, represents the time-frequency feature and the time-space feature interactively learned feature representation.

[0060] Step S402, on the basis of obtaining and , the fused time-frequency and time-space dependent features are obtained by using vector concatenation operation, as shown in the following formula:

[0061]

[0062] wherein, represents the vector concatenation operation, X out represents the vector representation of the time-space-frequency feature fusion.

[0063] Further, the specific steps of the step S50 are as follows:

[0064] Step S501, the design output layer is composed of a multi-layer perceptron (MLP), and X out The vector representation representing the time-space-frequency feature fusion is decoded into a multi-frequency point time slot sequence data prediction value of a future time step.

[0065]

[0066] Wherein, is a multi-frequency point time slot sequence data prediction value of a future time step, and MLP is a multi-layer perceptron.

[0067] Step S502, a mean square error (MSE) loss function is used to measure the difference between the multi-frequency point time slot sequence data prediction value of the future time step and the true value. The calculation formula is shown as follows:

[0068]

[0069] Wherein, is a multi-frequency point time slot sequence data prediction value of a future time step, and y is a multi-frequency point time slot sequence data true value of a future time step.

[0070] The beneficial effects of the present application are:

[0071] 1. By constructing a linear network layer in step S20, the multi-frequency point time slot sequence data is mapped into a channel mixed (CD) embedding representation vector and a channel independent (CI) embedding representation vector, effectively solving the technical problem of insufficient extraction of time domain, frequency domain and channel interaction features in multi-frequency point data, and providing an initial representation with high expression ability for subsequent multi-dimensional feature modeling.

[0072] Specifically, the CD vector retains the sequence characteristics of the multi-frequency point data in the time dimension through linear mapping combined with position embedding, and captures the dynamic correlation between different frequency points through the channel mixing mechanism; and the CI vector models the feature distribution of each frequency point in the time dimension independently through the transpose processing of the original data. This separation embedding strategy can avoid the feature redundancy caused by the coupling of inter-channel information in traditional methods, and at the same time provides differentiated input representation for subsequent modules. For example, in multi-scale time-frequency feature extraction, the CD vector can directly reflect the periodic pattern of the frequency point sequence, and the CI vector can independently represent the time evolution law of a single frequency point. This hierarchical representation significantly enhances the decoupling ability of the model for multi-dimensional features, and lays a foundation for cross-dimensional fusion.

[0073] 2. The multi-scale time-frequency feature extraction module and the adaptive graph time-space feature extraction module built in step S30 solve the modeling problems of time-domain periodicity patterns and spatial correlation in multi-frequency point data, respectively, and the bidirectional interactive feature fusion module in step S40 realizes effective alignment and collaborative optimization of cross-dimensional features, significantly improving the modeling ability of the model for complex time-space-frequency domain relationships.

[0074] In the multi-scale time-frequency feature extraction module, the frequency domain periodicity patterns are calculated by FFT, and the features at different time scales are captured by combining a multi-scale convolutional network (Inception structure), which can dynamically adapt to short-period fluctuations and long-period trends that may exist in multi-frequency point data. For example, for the rapid change characteristics of high-frequency signals (such as transient interference), the module can extract short-period patterns; while for the slow change characteristics of low-frequency signals (such as environmental noise), long-period patterns are modeled. Meanwhile, the introduction of the Mamba layer further enhances the model's ability to capture long-term temporal dependencies, avoiding the gradient vanishing problem in traditional RNN / LSTM long sequence modeling.

[0075] In the adaptive graph time-space feature extraction module, the spatial dependency between frequency points is dynamically modeled by a learnable adjacency matrix A, solving the problem that traditional graph structures cannot adapt to dynamic scenarios due to fixed or static adjacency matrices. For example, in spectrum resource allocation, the signal strength of different frequency points may change over time and produce new association patterns. This module, through the combination of GCN operations and Mamba layers, can adaptively learn and update the spatial relationships between frequency points.

[0076] In addition, the bidirectional interactive feature fusion module in step S40 realizes the bidirectional interaction of time-frequency and time-space features through a multi-head cross-attention mechanism (MCA). This module not only filters redundant information through dynamic matching of queries, keys, and values, but also retains key information of the original features through residual connections. For example, in the interaction process between time-frequency features and time-space features, the MCA mechanism can identify cross-dimensional features that have a synergistic effect on the prediction task (such as the periodic changes of a certain frequency point within a specific time window and the spatial correlation of adjacent frequency points), thereby improving the semantic consistency of feature fusion. This bidirectional alignment mechanism effectively avoids information loss that may be caused by one-way fusion, ensuring the alignment and complementarity of multi-dimensional features at the semantic level.

[0077] 3. The output layer designed in step S50 decodes the fused features into multi-frequency point time slot sequence prediction values for future time steps, solving the information compression and distortion problems in high-dimensional output decoding of traditional methods, and realizing accurate prediction of complex spectrum dynamics.

[0078] The output layer adopts a multi-layer perception (MLP) to perform nonlinear mapping on the fused features, which can fully excavate the high-order interaction relationship of cross-dimensional features. For example, in the frequency spectrum prediction task, the MLP can map the combined mode of the time-frequency features and the time-space features into a specific field strength value sequence through multiple nonlinear transformations, avoiding the simplification of complex relationships by a linear decoder. At the same time, the selection of the mean square error (MSE) loss function ensures the minimum deviation between the predicted value and the true value in the numerical space, which is particularly suitable for the continuous change characteristics in the frequency spectrum data. In addition, the design optimizes the parameters of the entire model through end-to-end training, so that the feature extraction, fusion and decoding processes between the modules form a closed-loop feedback, further improving the robustness of the prediction.

[0079] In summary, the present application effectively solves the limitations of traditional methods in multi-dimensional feature modeling in the time-space-frequency domain through the collaborative design of hierarchical embedding, multi-scale modeling and bidirectional fusion. The core advantages are:

[0080] (1) The CD / CI embedding strategy realizes the decoupling and independent modeling of information between channels;

[0081] (2) The multi-scale time-frequency module and the adaptive graph structure capture periodic patterns and spatial correlations, respectively;

[0082] (3) The bidirectional interaction mechanism realizes semantic alignment of cross-dimensional features. The above technical means jointly act to significantly improve the accuracy and generalization ability of the frequency spectrum time series prediction, which is particularly suitable for complex signal analysis tasks in high-dynamic and multi-frequency point scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0083] Figure 1 The method steps flowchart of the present application.

[0084] Figure 2 The multi-scale time-frequency feature extraction module technical principle diagram of the present application.

[0085] Figure 3 The adaptive graph time-space feature extraction module technical principle diagram.

[0086] Figure 4 The frequency spectrum time series prediction method implementation effect schematic diagram based on time-space-frequency domain feature fusion. DETAILED DESCRIPTION

[0087] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0088] Referring to Figure 1 , the present application provides an inventive method step flow chart:

[0089] The present application provides the following technical solutions: a spectrum time sequence prediction method based on time-space-frequency domain feature fusion, and the specific steps of the spectrum time sequence prediction method are as follows:

[0090] Step 1: Collecting historical time step multi-frequency point time slot sequence data for batch processing and normalization processing;

[0091] Step 1-1, selecting multi-frequency point time slot sequence data, constructing a sequence of field strength values changing with time, and the original representation is N is the number of selected frequency points (i.e. variable number), and the time slot sequence data of N frequency points in L historical time steps are used as the input of the model.

[0092] Each x l is a row vector with a length of N, representing the multi-frequency point field strength value of the lth time step, l∈{1,2,...,L}.

[0093] Matrix form:

[0094]

[0095] Step 1-2, normalization processing is performed, and the calculation formula is as follows:

[0096] x i，j =(x i,j -max(x i，j )) / σ j ,

[0097] Wherein, i={1,...,L},j={1,...,N},σ j is the standard deviation of the frequency point j time slot sequence data.

[0098] Step 2: Constructing a linear network layer to map the multi-frequency point time slot sequence data to a channel-dependent (CD) representation vector and a channel-independent (CI) representation vector;

[0099] Step 2-1, after normalization, the original representation X is mapped through the linear network layer Linear1 to obtain the CD vector representation The calculation formula is as follows:

[0100] X m emb =Linear1(X)+PE

[0101] wherein, represents a linear mapping, represents a position embedding, L is a time step, and d1 is a CD vector representation dimension.

[0102] Step 2-2, constructing a linear layer Linear2 mapping to obtain a CI vector representation The calculation formula is as follows:

[0103]

[0104] wherein, represents a linear mapping, N is the number of frequency points, and d2 is a CI vector representation dimension.

[0105] Step 3: constructing a multi-scale time-frequency feature extraction module and an adaptive graph time-space feature extraction module, respectively, for capturing time-frequency features and time-space features of multi-frequency point time slot sequence data from CD representation vectors and CI representation vectors; time-frequency (time and Fourier transform obtained frequency domain features) and time-space (time and spatial correlation features between multiple frequency points);

[0106] Step 3-1, as shown in Figure 2 In the multi-scale time-frequency feature extraction module, the frequency domain periodicity mode of the frequency point sequence data is calculated using the Fast Fourier Transform (FFT), and the calculation formula is as follows:

[0107]

[0108] wherein, Amp(·) represents the calculation of the amplitude value, represents the average amplitude, which is averaged in d1 dimensions by the function Amp(·).f i represents the period corresponding to the first k amplitude values.s i represents the total time step divided by the period f i obtained sequence division scale. In order to align the input sequence i at different period scales s and obtain the vector representation of different periods in the frequency domain The calculation formula is as follows:

[0109]

[0110] wherein, i={1,...,k} represents different period patterns in the frequency domain, padding(.) represents a zero padding operation, and the Reshape(.) operation realizes alignment of the input sequence

[0111] Step 3-2, further, using a multi-scale convolutional network model to further process Capture multi-scale features in different cycle patterns, the calculation process is as follows:

[0112]

[0113] Wherein, The multi-scale feature obtained by convolution operation is represented by Inception, and the multi-scale product operation is represented by Inception.

[0114] Step 3-3, further, construct Mamba layer (a state-aware network structure) to enhance the ability of the model to capture long-term time dependence features, the calculation formula is as follows:

[0115]

[0116] Wherein, The feature representation obtained by Mamba Block operation is represented by Inception.

[0117] Step 3-4, further, using SoftMax function to normalize the amplitude, as shown in the following formula:

[0118]

[0119] Wherein, f i The first k amplitude values correspond to the period, The first k amplitude values are represented by Inception.

[0120] Step 3-5, further, the feature representation of each time scale s i Reconstruct to The vector dimension size.

[0121]

[0122] Step 3-6, further, multiply By the corresponding a1,…,a k And sum all the results. Based on the addition residual structure to obtain multi-scale time-frequency features, the calculation formula is as follows:

[0123]

[0124] Wherein, The multi-scale time-frequency feature is represented by Inception.

[0125] Step 3-7, as Figure 3 ​As shown in the adaptive graph spatio-temporal feature extraction module, to learn the spatial dependency between different frequency sequences, a learnable adjacency matrix A (which can adaptively describe the spatial dependency between different frequency sequences) is initialized, which is represented by the following formula:

[0126] A = softMax(ReLu(E1(E2) T ))

[0127] where ReLu(.) represents an activation function, and E1 and E2 represent learnable parameters.

[0128] Step 3-8, the CI vector represents the feature multiplied by the adjacency matrix A, which can be input for dynamic and adaptive learning, thereby capturing more rich spatial dependency features between different frequencies, represented by the following formula:

[0129]

[0130] where GCN represents a graph convolutional network operation (GCN), and the GCN includes A, which is an N*N adjacency matrix, is considered as an embedding representation of N frequency sequences, represents the output spatial dependency feature vector.

[0131] Step 3-9, build a Mamba layer to enhance the model's ability to capture long-term temporal dependencies, and build a residual structure to obtain the time-space features The calculation formula is as follows:

[0132]

[0133] where, represents the time-space features between multi-frequency sequences.

[0134] Step 4: Build a bidirectional interactive feature fusion module to align the captured time-frequency and time-space features;

[0135] Step 4-1, the bidirectional interactive feature fusion module is a multi-head cross-attention mechanism (MCA), which enables the model to simultaneously learn and fuse time-frequency and time-space dependency features in an interactive manner. This mechanism not only enables bidirectional interactive learning of time-frequency and time-space features, but also effectively filters out irrelevant features in time-frequency and time-space features, improving the performance of the model. The calculation process is as follows:

[0136]

[0137] where, Used as query, key and value in MCA respectively, Representing spatiotemporal features and time-frequency characteristics Feature representation after interactive learning.

[0138]

[0139] in, Used as query, key and value in MCA respectively, Representing time-frequency features and spatiotemporal characteristics Feature representation after interactive learning.

[0140] Step 4-2, after getting and Based on this, the vector concatenation operation is used to obtain the fused time-frequency and time-space dependency features, as shown in the following formula:

[0141]

[0142] in, Represents vector concatenation operation, X out Vector representation representing the fusion of time-space-frequency features.

[0143] Step 5: Construct the output layer to align the feature matrix captured by the bidirectional interactive feature fusion module and decode it into the predicted value of the multi-frequency time slot sequence data for the future time step.

[0144] Step 5-1, design the output layer to consist of a multi-layer perceptron (MLP) and convert X out The vector representing the fusion of time-space-frequency features is decoded into the predicted value of the multi-frequency time slot sequence data of the future time step. The calculation formula is as follows:

[0145]

[0146] in, is the predicted value of the multi-frequency time slot sequence data for the future time step, and MLP is a multi-layer perceptron.

[0147] In step 5-2, the mean square error (MSE) loss function is used to measure the difference between the predicted value and the true value of the multi-frequency time slot sequence data in the future time step. The calculation formula is as follows:

[0148]

[0149] in, is the multi-frequency point time slot sequence data prediction value of the future time step, and y is the multi-frequency point time slot sequence data true value of the future time step.

[0150] The results of the implementation of the present application are shown in Figure 4 Specifically, taking the time slot sequence data of four frequency points as an example, the input of the method and system of the present application is set as: 96-step historical window time slot sequence data, and the output is: 24-step future prediction window time slot sequence data, and the blue line is the true value of the signal time slot sequence data, and the orange line is the prediction value of the signal time slot sequence data. In addition, Figure 4 The true value and the prediction value curves of the middle frequency points 1-4 sequence are basically overlapped, indicating the effectiveness of the implementation results of the present application.

[0151] The method of the present application takes the multi-frequency point signal time slot sequence data to be predicted as the research object. First, the multi-frequency point time slot sequence data of the historical time step is mapped into a channel mixed embedding representation vector and a channel independent embedding representation vector, then a multi-scale time-frequency feature extraction module and an adaptive graph time-space feature extraction module are constructed to capture the multi-scale time-frequency features and time-space features of the channel mixed representation vector and the channel independent representation vector respectively; in addition, a bidirectional interaction feature fusion module is constructed to align the captured time-frequency (frequency domain features obtained by time and Fourier transform) and time-space (spatial correlation features between time and multi-frequency points); finally, an output layer is designed to map the features output by the fusion module to the prediction value of the multi-frequency point time slot sequence data of the future time step.

[0152] In summary, the present application has the following characteristics:

[0153] Firstly, by mapping the multi-frequency point time slot sequence data of the historical time step into a channel mixed embedding representation vector and a channel independent embedding representation vector, the time domain, frequency domain and inter-channel interaction features in the multi-frequency point data can be effectively extracted, providing rich information representation for subsequent prediction.

[0154] Secondly, the designed multi-scale time-frequency feature extraction module and adaptive graph time-space feature extraction module can capture the time-frequency and time-space features of the data respectively, accurately depicting the complex correlation between time, frequency and spatial dimensions. Through the bidirectional interaction feature fusion module, the time-frequency features and the time-space features are effectively aligned, thereby improving the cross-dimensional fusion capability of the model and avoiding the limitations of single-dimensional analysis.

[0155] Finally, the output layer maps the fused features to the prediction value of the multi-frequency time slot sequence data of the future time step, which can accurately predict the future spectrum change and provide support for the reasonable scheduling and management of network spectrum resources. In general, this method overcomes the dependence of traditional prediction methods on a single feature dimension, can consider the interaction of three-dimensional information of time, space and frequency, and provides more accurate prediction results, especially in the processing of high dynamic and multi-frequency spectrum data. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions described in the foregoing embodiments or make equivalent replacements to part of the technical features, as long as they are within the spirit and principles of the present application. Any modification, equivalent replacement, improvement, etc. made shall be included in the protection scope of the present application.

Claims

1. A spectrum time series prediction method based on time-space-frequency domain feature fusion, characterized in that: The following steps are involved: step S10: Collect multi-frequency time slot sequence data of historical time steps for batch processing and normalization; Step S20: Construct a linear network layer to map the multi-frequency time slot sequence data to a CD representation vector and a CI representation vector, where CD represents channel mixing and CI represents channel independence: Step S30: constructing a multi-scale time-frequency feature extraction module and an adaptive graph time-space feature extraction module, which are used to capture the time-frequency features and time-space features of the multi-frequency point time slot sequence data from the CD representation vector and the CI representation vector, respectively; the time-frequency feature refers to the frequency domain feature obtained by time and Fourier transform, and the time-space feature refers to the spatial correlation feature between time and multiple frequency points; Step S40: constructing a bidirectional interactive feature fusion module for aligning the captured time-frequency and time-space features; Step S50: construct an output layer, align the feature matrix captured by the bidirectional interactive feature fusion module, and decode it into a predicted value of the multi-frequency point time slot sequence data of the future time step.

2. The method according to claim 1, wherein: The specific steps of step S10 are as follows: Step S101: Select multi-frequency time slot sequence data and construct a sequence of field strength values ​​varying with time. The original representation is: N is the number of selected frequency points, and the time slot series data of N frequency points within L historical time steps are used as the input of the model; Step S102: normalize the data. The calculation formula is as follows: x i,j =(x i,j -max(x i，j )) / σ j , Among them, i={1,...,L}, j={1,...,N},σ j is the standard deviation of the time slot sequence data at frequency point j.

3. The method according to claim 1, wherein: The specific steps of step S20 are as follows: Step S201: After normalization, the original representation X is mapped through the linear network layer Linear1 to obtain the CD vector representation The calculation formula is as follows: in, represents a linear mapping, represents position embedding, L is the time step, and d1 is the CD vector representation dimension; Step S202: construct a linear layer Linear2 mapping to obtain CI vector representation The calculation formula is as follows: in represents the transpose of X, It is represented as a linear mapping, N is the number of frequency points, and d2 is the dimension of CI vector representation.

4. The method according to claim 1, wherein: The specific steps of step S30 are as follows: Step S301: In the multi-scale time-frequency feature extraction module, the frequency domain periodic pattern of the frequency point sequence data is calculated using the fast Fourier transform (FFT). The calculation formula is as follows: Among them, Amp(·) represents the calculation of the amplitude value, represents the average amplitude, which is averaged in the d1 dimension by the function Amp(·), and f i Indicates the period corresponding to the first k amplitude values, s i Indicates the total time step divided by the period f in the time period corresponding to the i-th amplitude value i The obtained sequence partition scale, Topk(·) means taking the first k, Indicates the periodic frequency f corresponding to the first k amplitudes; In different cycles i Align input sequences under scale And get the vector representation of different periods in the frequency domain The calculation formula is as follows: Where i = {1, ..., k} represents different periodic patterns in the frequency domain, padding (.) indicates the use of zero padding operation, and Reshape (·) operation is implemented Align input sequences Step S302: further process the image using a multi-scale convolutional network model. Capturing multi-scale features in different periodicity modes, the calculation process is as follows: in, Represents the multi-scale features obtained by convolution operation, and Inception represents the multi-scale product operation; Step S303: Further, a Mamba layer is constructed to enhance the model's ability to capture long-term time-dependent features. The calculation formula is as follows: in, Represents the feature representation obtained by Mamba Block operation; Step S304: Use the SoftMax function to normalize the amplitude, as shown in the following formula: where a1,…,a k Indicates the weight of the importance of different amplitude values, f k represents the period corresponding to the first k amplitude values, represents the first k amplitude values; Step S305: Each time scale s i The following feature representation Refactored to The size of its vector dimensions; Step S306, Multiply by the corresponding a1,…,a k , and sum up all the results, and then obtain the multi-scale time-frequency features based on the additive residual structure. The calculation formula is as follows: in, Represent multi-scale time-frequency features; In step S307, in the adaptive graph spatiotemporal feature extraction module, a learnable adjacency matrix A is initialized to learn the spatial dependency between different frequency point sequences, which is expressed by the following formula: A=softMax(ReLu(E1(E2) T )) Among them, ReLu(.) represents the activation function, E1 and E2 represent the learnable parameters; Step S308: CI vector represents the feature The adjacency matrix A is placed in the GCN module for calculation operations, and dynamic and adaptive learning is performed according to the input to capture richer spatial dependency features between different frequency points, which is expressed by the following formula: Among them, GCN represents the graph convolutional network operation, A is the N*N adjacency matrix, is regarded as an embedded representation of a sequence of N frequency points, represents the spatially dependent eigenvector of the output; Step S309: construct the Mamba layer to enhance the model's ability to capture long-term temporal dependencies, and construct a residual structure to obtain spatiotemporal features. The calculation formula is as follows: in, Represents the spatiotemporal characteristics between multiple frequency point sequences.

5. The method according to claim 1, wherein: The specific steps of step S40 are as follows: Step S401: The bidirectional interactive feature fusion module uses the multi-head cross attention mechanism MCA to enable the model to interactively learn time-frequency and time-space dependent features simultaneously and fuse them. The multi-head cross attention mechanism enables the bidirectional interactive learning of time-frequency and time-space features to filter out irrelevant features in time-frequency and time-space features. The calculation process is as follows: in, Used as query, key and value in MCA respectively, Representing spatiotemporal features and time-frequency characteristics Feature representation after interactive learning; in, Used as query, key and value in MCA respectively, Representing time-frequency features and spatiotemporal characteristics Feature representation after interactive learning; Step S402, after obtaining and Based on this, the vector concatenation operation is used to obtain the fused time-frequency and time-space dependency features, as shown in the following formula: in, Represents vector concatenation operation, X out A vector representation representing the fusion of time-space-frequency features.

6. The method according to claim 1, wherein: The specific steps of step S50 are as follows: Step S501: Design the output layer to consist of a multi-layer perceptron MLP. out The vector representing the fusion of time-space-frequency features is decoded into the predicted value of the multi-frequency time slot sequence data of the future time step. The calculation formula is as follows: in, is the predicted value of the multi-frequency time slot sequence data for the future time step, and MLP is a multi-layer perceptron. In step S502, the mean square error (MSE) loss function is used to measure the difference between the predicted value and the true value of the multi-frequency time slot sequence data in the future time step. The calculation formula is as follows: in, is the predicted value of the multi-frequency time slot sequence data in the future time step, y is the true value of the multi-frequency time slot sequence data in the future time step, and n is the number of samples.

Images