A spectrum situation prediction method based on conditional diffusion model

Abstract

The present invention belongs to the field of intelligent wireless communications and discloses a spectrum situation prediction method based on a conditional diffusion model. First, the global features of time, frequency and space dimensions are extracted from historical spectrum data through a multi-head attention mechanism, and a unified conditional vector is formed through cross-attention fusion; then, the conditional vector is input into the conditional diffusion model, and noise is gradually injected into the future spectrum data in the forward process to provide samples for network training and learn the noise evolution law; in the backward process, Gaussian noise is used as the starting point, and the noise estimation network is used to gradually denoise and restore the future spectrum state. This method can achieve high-precision prediction in scenarios where spectrum observation is missing, and provide forward-looking support for wireless resource management tasks such as spectrum pre-allocation, interference avoidance and power control.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent wireless communications and relates to a spectrum situation prediction method based on a conditional diffusion model. Background Art

[0002] With the advancement of 6G technology, immersive applications and services such as the massive Internet of Things (IoT) are further consuming limited spectrum resources. The frequency usage characteristics of these massive services are highly dynamic across time, frequency bands, and space. Currently, mobile communication networks primarily rely on a "sense-and-schedule" strategy, where base stations first detect the current spectrum situation and then instantly issue power control or frequency band switching commands. This passive approach often fails to fully adjust links when traffic surges or interference emerges, resulting in a sharp drop in throughput and even session interruptions. Therefore, if future spectrum conditions can be predicted in advance based on historical spectrum measurements, base stations can proactively switch frequency bands or reallocate power, upgrading from passive response to proactive resource management, significantly improving spectrum utilization and service reliability.

[0003] Traditional spectrum situation prediction methods primarily rely on statistical modeling, first fitting parameters to historical spectrum data and then extrapolating future data using a recursive formula. For example, patent CN107070569A discloses a spectrum situation prediction method based on a hidden Markov model prediction model. This method uses the hidden Markov model to first establish a probability matrix for spectrum "occupied / idle" state transitions, then uses this matrix to infer the spectrum state for the next time slot, achieving real-time spectrum situation prediction. Patent CN106231603B discloses a spectrum prediction method based on a statistical model. This method first calculates the correlation coefficient between channels based on historical occupancy status and clusters them. Within each channel cluster, a subset of detection channels is selected. Using the hidden Markov model and the Viterbi algorithm, the current state of the detected channels is used to infer the occupancy state of the remaining channels in the same cluster, thereby improving prediction efficiency and reducing detection overhead. However, these algorithms are suitable for capturing short-term variations in historical data but struggle to adapt to the highly dynamic characteristics of the spectrum in time, frequency, and space.

[0004] In recent years, with the rise of deep learning technology, the feature extraction capabilities of deep networks have significantly enhanced the representation of historical data, thereby achieving more accurate prediction results. Researchers have widely introduced models such as LSTM, RNN, and Transformer to model the time, frequency, and spatial features in historical spectrum data to predict future spectrum trends, and have achieved high accuracy on a variety of public or private data sets. For example, patent CN119155802A discloses a spectrum prediction method based on a time-frequency feature fusion network. First, the time domain features of the spectrum data are extracted through a time convolutional network, and then the frequency domain features are extracted using a graph convolutional network. The two types of features are then dynamically fused using an attention mechanism to achieve high-precision prediction of future spectrum states. Patent CN116744455A discloses an interpretable spectrum trend prediction method based on wavelet decomposition and LSTM network. First, the spectrum data is subjected to wavelet decomposition to obtain multi-scale subsequences. Then, the subsequence features are fused and predicted using LSTM networks, achieving interpretable and more accurate spectrum trend prediction. However, these methods generally assume that monitoring nodes can scan the full bandwidth during each sampling period and therefore possess complete historical data. In actual deployments, to reduce RF costs, power consumption, and sampling time, nodes often employ a partial frequency sweep approach, scanning only a portion of the broadband subbands at a time. This cost-effective approach naturally results in a large amount of missing observations in historical spectrum data. Furthermore, most intelligent prediction models still focus on modeling the time dimension; while a few works consider the dynamics of time, frequency, and space, they do not deeply integrate multidimensional features. Therefore, how to jointly mine and fuse the three-dimensional time-frequency-space features based on partially observed historical spectrum data and accurately infer future spectrum status has become a key challenge in current intelligent proactive spectrum management and a core technical issue that needs to be overcome.

[0005] To this end, the present invention proposes a spectrum situation prediction method based on the conditional diffusion model for partial observed historical spectrum data. Specifically, a "time-frequency-space" feature extraction network is first constructed, and the time, frequency and space features are respectively extracted from the zero-filled historical spectrum data through a multi-dimensional attention mechanism, and the "time-frequency-space" features are fused into a unified conditional vector using a cross-attention mechanism; the conditional vector is then injected into the conditional diffusion model to guide it to learn the distribution of added noise in the forward diffusion stage, and to continuously provide a priori guidance in the backward generation process, so as to achieve gradual denoising and fine reconstruction of future spectrum. This method ensures detail generation through multi-layer attention modeling of global correlation and conditional diffusion model, and can output high-precision spectrum situation prediction results even in partial frequency sweep scenarios where historical spectrum data is missing, providing forward-looking decision support for spectrum resource management tasks such as dynamic frequency band switching, power control and interference avoidance. Summary of the Invention

[0006] The present invention considers that in a mobile communication network, in order to achieve forward-looking prediction of spectrum situation, the system is deployed according to a preset density. spectrum monitoring nodes, using partial frequency sweeping to The frequency bands are monitored and the measurement results are uploaded to the base station or edge node in real time. Subsequently, the base station or edge node performs zero padding and data processing on some of the observed spectrum data from different monitoring nodes and constructs a data set for model training. When constructing the data set, the continuous The data of each sampling period is the historical window , followed by The data of the sampling period is used as the prediction window This allows us to construct a large number of sample pairs , performing offline training at the base station or edge node. After model training is complete, the base station can receive spectrum data continuously uploaded by monitoring nodes online, generate the latest historical window in real time, and input it into the model to predict future spectrum status, enabling online prediction and dynamic forward-looking spectrum resource management.

[0007] The technical solution of the present invention:

[0008] A spectrum situation prediction method based on conditional diffusion model, such as Figure 1 As shown, it consists of a "time-frequency-space" feature extraction network and a spectrum situation prediction generation network. First, through the "time-frequency-space" feature extraction network, deep correlation modeling of historical spectrum data is performed in multiple dimensions of time, frequency and space to extract multi-dimensional fusion features; then, the extracted features are used as conditional inputs to guide the spectrum situation prediction generation network to gradually restore future spectrum data, thereby achieving accurate prediction of spectrum evolution trends and detailed recovery. This method effectively combines the global modeling capabilities of the multi-dimensional attention mechanism and the high-fidelity generation capabilities of the conditional diffusion model, significantly improving the accuracy of spectrum situation prediction. The specific steps are as follows:

[0009] (1) Constructing a “time-frequency-space” feature extraction network;

[0010] The “time-frequency-space” feature extraction network is used to extract the historical spectrum data In the process of fully exploring the correlation between time dimension, frequency dimension and space dimension, a high-dimensional fusion feature is constructed; is the number of historical sampling cycles, is the number of monitoring nodes, is the frequency point number;

[0011] (1.1) Temporal feature extraction;

[0012] The multi-head attention mechanism is used to model the time dimension of historical spectrum data to fully capture the global time domain correlation of historical spectrum data. First, the historical spectrum data is Flatten along the time dimension to obtain a two-dimensional matrix ; Subsequently, position encoding is introduced to preserve historical spectrum data The time sequence information is obtained to get the matrix after position embedding:

[0013] (1)

[0014] in, represents the position encoding matrix; then, Each attention head is generated by three sets of linear transformations The query matrix , key matrix , value matrix :

[0015] (2)

[0016] in, 、 、 is a learnable linear mapping matrix; then, multi-head attention calculation is performed to obtain the output of each attention head: :

[0017] (3)

[0018] in, Represents the bond matrix The transpose of Indicates the normalization operation of the attention score. is the dimension of each attention head; the output of all attention heads is feature fused and input into the feedforward neural network with activation function to further improve the expression ability of the "time-frequency-space" feature extraction network; finally, in order to ensure training stability and accelerate convergence, a normalization operation is added after the residual connection to obtain the global feature of the time dimension ;

[0019] (1.2) Frequency feature extraction;

[0020] To capture historical spectrum data in different geographical locations The correlation between the frequency points in the historical spectrum data The self-attention mechanism is introduced in the frequency dimension to model the frequency dimension; the historical spectrum data Flatten along the frequency dimension to obtain a two-dimensional matrix , and perform position encoding, multi-head attention calculation and feature fusion, the same as the operation process of step (1.1), to obtain the global features of the frequency dimension ;

[0021] (1.3) Spatial feature extraction;

[0022] In order to capture the spatial correlation between different geographical locations, the historical spectrum data The self-attention mechanism is introduced into the spatial dimension to model the frequency dimension; then, the historical spectrum data Flatten along the spatial dimension to obtain a two-dimensional matrix , and perform position encoding, multi-head attention calculation and feature fusion, the same as the operation process of step (1.1), to obtain the global features of spatial dimension correlation ;

[0023] (1.4) “Time-frequency-space” feature fusion;

[0024] In order to unify the features extracted from the time dimension, frequency dimension and space dimension, a cross-attention mechanism is introduced to fuse the global features of the time dimension with the global features containing the correlation of the frequency dimension and the space dimension. First, the query matrix required by the cross-attention mechanism is generated by linear mapping. , key matrix Sum Matrix :

[0025] (4)

[0026] in, 、 、 is the learnable weight matrix; finally, 、 、 Perform attention calculation to obtain "time-frequency-space" fusion features ;

[0027] (2) Constructing a spectrum situation prediction generation network;

[0028] In the forward process, by feeding the real future spectrum data Gaussian noise is injected stepwise to construct noisy data corresponding to different diffusion time steps. This data is then used to supervise the noise estimation network in learning the distribution of the injected noise at different diffusion time steps. Subsequently, the noise estimation network is trained to accurately fit the noise distribution. After the noise estimation network is trained, the trained noise estimation network is used in a backward process, starting with standard Gaussian noise, to gradually estimate and remove the noise corresponding to each step, ultimately restoring high-fidelity future spectral data.

[0029] (2.1) Forward process;

[0030] The forward process provides training samples for the noise estimation network, from real future spectrum data Starting from , Gaussian noise is gradually injected into it according to the following diffusion formula to generate noisy data:

[0031] (5)

[0032] in, represents the diffusion time step The corresponding noisy data, represents the diffusion time step; represents the cumulative retention coefficient, represents the noise intensity added at each step, is the preset noise scheduling parameter; represents the real noise that follows Gaussian distribution; is the unit covariance matrix;

[0033] The noise estimation network adopts the U-Net structure, and its input is the diffusion time step , the corresponding noisy data And the "time-frequency-space" fusion features obtained by the "time-frequency-space" feature extraction network , the output is the estimated noise at the current time step :

[0034] (6)

[0035] in, Estimate network parameters for the noise; during training, the optimization goal is to minimize the mean square error between the estimated noise and the true noise:

[0036] (7)

[0037] By performing gradient descent on Equation (7), all learnable parameters of the noise estimation network and the “time-frequency-space” feature extraction network are updated simultaneously;

[0038] (2.2) Backward process;

[0039] After the noise estimation network and the “time-frequency-space” feature extraction network are trained, the “time-frequency-space” feature extraction network is first used to extract the historical spectrum data. The “time-frequency-space” fusion feature is extracted as the conditional input of the noise estimation network. Starting from pure noise, the trained noise estimation network is used to gradually denoise and reconstruct future spectrum data. The backward sampling process of each step is:

[0040] (8)

[0041] in, is the noise scaling coefficient in the backward process, Represents standard Gaussian noise; repeat the above process until , that is, the final predicted future spectrum data is obtained .

[0042] Beneficial effects of the present invention:

[0043] (1) The present invention can achieve accurate spectrum situation prediction even when the actual spectrum monitoring nodes adopt a partial frequency sweep strategy and there are a lot of missing historical observation data. It breaks through the limitation of traditional models' dependence on full-frequency observation data and enhances the deployment adaptability and practicality in real communication networks.

[0044] (2) The present invention introduces a multi-dimensional attention mechanism to extract time-frequency-space features, and combines it with a conditional diffusion model to generate high-fidelity predictions, significantly improving the accuracy of spectrum prediction and the ability to restore details, providing strong support for forward-looking spectrum management tasks such as frequency band pre-allocation, power control, and interference avoidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is an algorithm flow chart of the spectrum situation prediction method based on the conditional diffusion model in an example of the present invention. DETAILED DESCRIPTION

[0046] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0047] 1. Online Measurement and Dataset Construction

[0048] In the target mobile communication network, three spectrum monitoring nodes (lightweight sensors) were deployed at a preset density. Each node sampled 250 frequency points within the 600MHz to 850MHz band at a 1MHz sampling resolution and a 1-minute sampling period, ensuring the integrity of the original spectrum data in the time, frequency, and space dimensions. To simulate the partial frequency scanning strategy used by spectrum monitoring nodes in actual deployments, a masking process was applied to the complete spectrum data during the dataset construction phase. Specifically, at each time step and each monitoring node location, several sub-frequency bands were selected as observable frequency points, while the remaining frequency points were treated as missing data and zero-padded. The masked spectrum data generated in this way effectively approximated the observations obtained by partial frequency scanning by monitoring nodes in actual operation, ensuring consistency between the training data and the actual deployment scenario. Subsequently, the spectrum data from all monitoring nodes was concatenated into a unified three-dimensional tensor in the "time-position-frequency" order. When constructing training samples, 50 minutes of zero-padded spectrum data is used as the historical window, and the following 10 minutes of complete spectrum data is used as the prediction target. The window is slid once a minute to continuously generate a large number of <historical window, prediction window> sample pairs for supervised learning.

[0049] 2. Offline Network Training

[0050] In the offline phase, the model is optimized end-to-end on a cloud server. For each training sample, the historical window is first input into the "time-frequency-space" feature extraction network to extract and fuse time, frequency, and spatial dependencies to obtain a conditional vector; the true prediction target is gradually injected with Gaussian noise according to a fixed noise schedule to form noisy data corresponding to all diffusion time steps. The noisy data, diffusion time steps, and conditional vectors are input into the noise estimation network together to estimate the noise of each diffusion step. The feature extraction network and the noise estimation network as a whole use the mean square error as the loss function, and the loss function is subjected to backward gradient descent to update all learnable parameters in the noise estimation network and the feature extraction network until the training set error converges.

[0051] 3. Model Deployment and Real-time Prediction

[0052] After model training is complete, it can be deployed on base stations or edge servers to support real-time spectrum prediction. In actual operation, due to cost and energy constraints, front-end spectrum monitoring nodes only perform partial frequency sweep sampling, observing a subset of frequency sub-bands every minute. After uploading this partial observation data to the edge server, the system performs zero-filling for missing frequency points and reconstructs a three-dimensional input tensor structure consistent with the training phase. This tensor is first fed into a feature extraction network to generate a conditional vector. Subsequently, the system initializes the spectrum estimates for future time slices with Gaussian noise. Backdiffusion is performed through the noise estimation network to gradually remove the noise, ultimately generating a complete spectrum prediction for the next 10 minutes. Experimental results show that the mean squared error (MSE) between the predicted spectrum state and the actual spectrum state in the test set is significantly lower than that of comparable methods, demonstrating a more accurate reconstruction of the future spectrum state. This further demonstrates the stable and efficient spectrum prediction capabilities of the proposed method even in the absence of observations, demonstrating its superior practicality and engineering reliability. The prediction results can be directly used in proactive spectrum resource management, including decision modules such as frequency band pre-allocation, power control, and interference avoidance, significantly improving spectrum utilization efficiency and network communication quality.

Claims

1. A spectrum situation prediction method based on a conditional diffusion model, characterized in that: Here are the steps: (1) Constructing a "time-frequency-space" feature extraction network; "Time-frequency-space" feature extraction network is used to extract historical spectrum data In the process of fully exploring the correlation between time dimension, frequency dimension and space dimension, a high-dimensional fusion feature is constructed; is the number of historical sampling cycles, is the number of monitoring nodes, is the frequency point number; The implementation process is as follows: (1.1) Temporal feature extraction; The multi-head attention mechanism is used to model the time dimension of historical spectrum data to fully capture the global time domain correlation of historical spectrum data. First, the historical spectrum data is Flatten along the time dimension to obtain a two-dimensional matrix ; Subsequently, position encoding is introduced to preserve historical spectrum data The time sequence information is obtained to get the matrix after position embedding: （1）； in, represents the position encoding matrix; then, Each attention head is generated by three sets of linear transformations The query matrix , key matrix , value matrix : （2）； in, 、 、 is a learnable linear mapping matrix; then, multi-head attention calculation is performed to obtain the output of each attention head: : （3）； in, Represents the bond matrix The transpose of Indicates the normalization operation of the attention score. is the dimension of each attention head; the output of all attention heads is feature fused and input into the feedforward neural network with activation function; finally, in order to ensure training stability and accelerate convergence, a normalization operation is added after the residual connection to obtain the global feature of the time dimension ; (1.2) Frequency feature extraction; To capture historical spectrum data in different geographical locations The correlation between the frequency points in the historical spectrum data The self-attention mechanism is introduced in the frequency dimension to model the frequency dimension; the historical spectrum data Flatten along the frequency dimension to obtain a two-dimensional matrix , and perform position encoding, multi-head attention calculation and feature fusion, the same as the operation process of step (1.1), to obtain the global features of the frequency dimension ; (1.3) Spatial feature extraction; In order to capture the spatial correlation between different geographical locations, the historical spectrum data The self-attention mechanism is introduced into the spatial dimension to model the frequency dimension; then, the historical spectrum data Flatten along the spatial dimension to obtain a two-dimensional matrix , and perform position encoding, multi-head attention calculation and feature fusion, the same as the operation process of step (1.1), to obtain the global features of spatial dimension correlation ; (1.4) "Time-frequency-space" feature fusion; In order to unify the features extracted from the time dimension, frequency dimension and space dimension, a cross-attention mechanism is introduced to fuse the global features of the time dimension with the global features containing the correlation of the frequency dimension and the space dimension. First, the query matrix required by the cross-attention mechanism is generated by linear mapping. , key matrix Sum Matrix : （4）； in, 、 、 is the learnable weight matrix; finally, 、 、 Perform attention calculation to obtain "time-frequency-space" fusion features ; (2) Construct spectrum situation prediction generation network; In the forward process, by feeding the real future spectrum data Gaussian noise is injected stepwise to construct noisy data corresponding to different diffusion time steps. This is used to supervise the noise estimation network to learn the distribution of the injected noise at different diffusion time steps. Subsequently, the noise estimation network is trained to accurately fit the noise distribution. After the noise estimation network is trained, in the backward process, standard Gaussian noise is used as the starting point. The trained noise estimation network is used to gradually estimate and remove the noise corresponding to each step, ultimately restoring high-fidelity future spectrum data. The implementation process is as follows: (2.1) Forward process; The forward process provides training samples for the noise estimation network, from real future spectrum data Starting from , Gaussian noise is gradually injected into it according to the following diffusion formula to generate noisy data: （5）； in, represents the diffusion time step The corresponding noisy data, represents the diffusion time step; represents the cumulative retention coefficient, represents the noise intensity added at each step, is the preset noise scheduling parameter; represents the real noise that follows Gaussian distribution; is the unit covariance matrix; The noise estimation network adopts the U-Net structure, and its input is the diffusion time step , the corresponding noisy data And the "time-frequency-space" fusion features obtained by the "time-frequency-space" feature extraction network , the output is the estimated noise at the current time step : （6）； in, Estimate network parameters for the noise; during training, the optimization goal is to minimize the mean square error between the estimated noise and the true noise: （7）； By performing gradient descent on Equation (7), all learnable parameters of the noise estimation network and the "time-frequency-space" feature extraction network are updated simultaneously; (2.2) Backward process; After the noise estimation network and the "time-frequency-space" feature extraction network are trained, the "time-frequency-space" feature extraction network is first used to extract the historical spectrum data. Extract the "time-frequency-space" fusion features as the conditional input of the noise estimation network. Starting from pure noise, use the trained noise estimation network to gradually denoise and reconstruct the future spectrum data. The backward sampling process of each step is: （8）； in, is the noise scaling coefficient in the backward process, Represents standard Gaussian noise; repeat the above process until , that is, the final predicted future spectrum data is obtained .

Images