Emergency communication spectrum prediction method based on gating adaptive time sequence diagram integrated network

Abstract

The invention provides an emergency communication spectrum prediction method based on a gating adaptive time sequence diagram integrated network, which relates to the technical field of communication, and comprises the following steps: acquiring multi-band historical spectrum data, and dividing a training sample and a prediction label according to a time step; constructing a self-adaptive dynamic graph structure fusing priori knowledge and a learnable matrix, and combining gating time sequence convolution, graph convolution, channel alignment convolution and residual connection to form a gating time sequence graph network; carrying out iterative training by utilizing the training sample and the dynamic graph structure, and obtaining an integrated prediction model through period storage and average weight snapshot; and inputting a to-be-tested sample into the model and outputting a frequency spectrum prediction value. The method has the advantages that spatio-temporal feature extraction is fused, and prediction precision and scene adaptability are improved; the generalization and robustness of the model in the environment change are enhanced by the self-adaptive dynamic graph structure; the parallel gating time sequence convolution is adopted to reduce the calculation complexity and the reasoning time delay, and the millisecond-level response requirement of emergency communication is met.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to an emergency communication spectrum prediction method based on a gated adaptive timing graph integrated network. Background Technology

[0002] Reliable and efficient emergency communications are crucial in key areas such as emergency response, disaster relief, large-scale event management, and public safety. Spectrum prediction, as a core technology in this field, aims to accurately estimate future spectrum occupancy based on historical observation data, thereby achieving dynamic and optimized allocation of spectrum resources. Existing spectrum prediction methods can be broadly categorized into two types: time-series modeling-based methods and graph-based methods.

[0003] (I) Time-series modeling methods characterize the temporal evolution of spectral data along the time dimension, which can be further subdivided into: Traditional statistical models include autoregressive (AR), moving average (MA), autoregressive moving average (ARMA), and autoregressive integral moving average (ARIMA). These model-driven methods heavily rely on the statistical properties of historical data (such as mean and variance), resulting in high model complexity and difficulty in effectively capturing and characterizing the complex nonlinear features in multidimensional spectral data.

[0004] Traditional machine learning methods, such as Hidden Markov Models (HMMs) and Support Vector Regression (SVR), offer a degree of flexibility, but their representational power and generalization performance are often limited when dealing with high-dimensional, long-sequence data.

[0005] Deep learning-based methods, represented by Recurrent Neural Networks (RNNs) and their variants such as Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRUs), can effectively model sequence dependencies. However, they have significant drawbacks when applied to spectrum prediction: First, they are primarily designed for single-dimensional time series and struggle to handle multi-frequency band scenarios effectively. This is because the model structure focuses mainly on temporal correlations while neglecting spatial correlations between different frequency bands. Second, their recurrent structure necessitates sequential processing step-by-step, preventing true parallel computation. This not only leads to high computational complexity but also results in significant delays in training and inference, making it difficult to meet the real-time requirements of scenarios such as emergency communications. Subsequent research has attempted to combine Convolutional Neural Networks (CNNs) with RNNs to simultaneously learn cross-frequency band features in both temporal and "spatial" contexts. However, simple combinations often result in complex model structures, difficult training, and unsatisfactory fusion effects.

[0006] (II) Graph-based modeling methods: To characterize the interactions between frequency bands (i.e., complex spatial dependencies), some studies have introduced graph neural networks (GNNs), particularly graph convolutional networks (GCNs). These methods treat each frequency band in the spectral data as a node in a graph, modeling the interactions between nodes by constructing an adjacency matrix (weighted edges), thereby capturing the complex dependencies between frequency bands. However, most existing graph-based modeling methods employ static graph structures, i.e., constructing the adjacency matrix once based on prior knowledge (such as fixed physical distances between frequency bands or preset correlation thresholds), and keeping it unchanged throughout the prediction process. In complex and variable electromagnetic environments such as emergency communications, the spectral state exhibits highly dynamic changes with time and spatial location, and the interactions between frequency bands are also time-varying. Static graph structures cannot dynamically reflect these dependencies that change with time and environment, making it difficult to capture the potential, dynamically evolving interaction characteristics between frequency bands. This results in insufficient generalization ability of the model when the environment changes abruptly, and the prediction results are prone to failure.

[0007] In summary, the existing technology has the following three main shortcomings: Ignoring spatiotemporal dynamics leads to insufficient prediction accuracy and adaptability: Most time series models only consider the evolution in the time dimension and ignore the spatial correlation between frequency bands. They cannot fully reflect the spatiotemporal dynamic changes of the electromagnetic environment, resulting in decreased prediction accuracy and poor adaptability in complex scenarios.

[0008] Static graph modeling is difficult to handle dynamic environments: Existing graph modeling methods mostly use static adjacency matrices, which cannot dynamically adjust the graph structure as the spectrum state changes, thus limiting the model's ability to accurately characterize time-varying electromagnetic environments.

[0009] Model computational efficiency and real-time bottlenecks: Traditional RNN models suffer from high computational complexity and large inference latency due to their sequence dependencies, making it difficult to achieve efficient deployment in emergency communication scenarios that require rapid response. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention provides an emergency communication spectrum prediction method based on a gated adaptive timing graph ensemble network, comprising: S1. Acquire the collected multi-band historical spectrum data, and divide the multi-band historical spectrum data into training samples and corresponding prediction labels by a sliding window according to the preset historical time step and prediction step. Step S2: Calculate the Pearson correlation coefficient between different frequency bands in the training samples and construct an initial prior adjacency matrix according to a set threshold. Generate an adaptive adjacency matrix by introducing a learnable matrix. Merge the adaptive adjacency matrix with the initial prior adjacency matrix to obtain an adaptive dynamic graph structure. Construct a gated temporal graph network consisting of a gated temporal convolutional layer, a graph convolutional layer, a channel-aligned convolutional layer, and a residual connection structure. Step S3: Input the training samples and the adaptive dynamic graph structure into the gated temporal graph network for iterative training. Perform backpropagation based on the loss function of the obtained model output value and the predicted label and update the model weights. During the training process, save multiple model weight snapshots according to the set snapshot period. After the training is completed, average the multiple model weight snapshots to obtain the integrated spectrum prediction model. Step S4: Input the spectrum sample to be tested into the integrated spectrum prediction model, and output the spectrum status prediction value to provide the emergency communication command and dispatch system for frequency band scheduling of communication services.

[0011] Preferably, step S2 includes: Step S21: The similarity between different frequency bands in the training samples is measured by using the Pearson correlation coefficient, and the initial prior adjacency matrix is ​​constructed according to the set threshold. Step S22: The adaptive adjacency matrix is ​​generated by introducing a learnable matrix; Step S23: The adaptive adjacency matrix is ​​fused with the initial prior adjacency matrix to form the final adaptive dynamic graph structure.

[0012] Preferably, the input-output connection relationship in the gated timing graph network in step S2 is as follows: The training samples are input into the gated temporal convolutional layer, which extracts temporal features through a parallel dual-branch structure with dilated convolution. The temporal features extracted by the gated temporal convolutional layer are input into the graph convolutional layer to perform spatial feature aggregation between frequency bands. The output of the graph convolutional layer is element-wise summed with the original input features of the training samples using a residual connection structure. Then, the input channel-aligned convolutional layer undergoes dimensionality transformation to obtain the predicted value of the model output and the model weights.

[0013] Preferably, the parallel dual-branch structure in the gated temporal convolutional layer includes: The first branch extracts nonlinear features through an activation function; The second branch generates weight coefficients using the Sigmoid function; The time series characteristics are obtained by multiplying the outputs of the two branches element by element.

[0014] Preferably, step S3 includes: Step S31: Input each batch of training samples into the gated temporal graph network for forward computation to obtain the model output value; Step S32: Calculate the deviation between the initial model prediction value and the predicted label using the mean squared error function, and construct the loss function; Step S33: Update the weight parameters in the gated temporal graph network using the optimizer based on the gradient of the loss function; Step S34: Within each preset snapshot period, the learning rate is reduced according to a preset decay factor after each round of training. Step S35: When the current training round reaches an integer multiple of the snapshot period, save the current model weight snapshot and reset the learning rate to the initial learning rate; Step S36: After reaching the total number of training rounds, extract all saved model weight snapshots and perform arithmetic averaging to generate integrated spectrum prediction model parameters.

[0015] Preferably, the adaptive adjacency matrix in step S22 is generated in the following way: An adaptive adjacency matrix is ​​generated by multiplying two learnable matrices, wherein the learnable matrix has a dimension of N×c, where N is the number of frequency bands and c is the feature length.

[0016] This invention also provides an emergency communication spectrum prediction system based on a gated adaptive timing graph ensemble network, which applies the emergency communication spectrum prediction method described above, including: The data processing module is used to acquire the collected multi-band historical spectrum data, and divide the multi-band historical spectrum data into training samples and corresponding prediction labels through a sliding window according to the preset historical time step and prediction step. The model building module is used to build an adaptive dynamic graph structure based on the training samples, and to build a gated temporal graph network consisting of gated temporal convolutional layers, graph convolutional layers, channel-aligned convolutional layers, and residual connection structures. An integrated training module, connecting the data processing module and the model building module, is used to input the training samples and the adaptive dynamic graph structure into the gated temporal graph network for iterative training. It performs backpropagation based on the loss function of the obtained model output value and the predicted label and updates the model weights. During the training process, multiple model weight snapshots are saved according to a set snapshot period. After the training is completed, the multiple model weight snapshots are averaged to obtain the integrated spectrum prediction model. The prediction application module, connected to the integrated training module, is used to input the spectrum sample to be tested into the integrated spectrum prediction model and output the spectrum state prediction value to provide the emergency communication command and dispatch system for frequency band scheduling of communication services.

[0017] Preferably, the model building module includes: The prior adjacency matrix construction unit is used to measure the similarity between different frequency bands in the training samples using the Pearson correlation coefficient, and to construct an initial prior adjacency matrix according to a set threshold. An adaptive adjacency matrix generation unit is used to generate an adaptive adjacency matrix by introducing a learnable matrix. The dynamic graph fusion unit is used to fuse the adaptive adjacency matrix with the initial prior adjacency matrix to form the final adaptive dynamic graph structure.

[0018] Preferably, the adaptive adjacency matrix generation unit generates an adaptive adjacency matrix by performing matrix multiplication on two learnable matrices of dimension N×c, where N is the number of frequency bands and c is the feature length.

[0019] Preferably, the hierarchical connection relationship of the gated sequence graph network in the model construction module is as follows: The gated temporal convolutional layer is connected to the input of the training samples and extracts temporal features through a parallel dual-branch structure with dilated convolution. The input of the graph convolutional layer is connected to the output of the gated temporal convolutional layer and the output of the dynamic graph fusion unit, respectively, for spatial feature aggregation. The residual connection structure connects the output of the graph convolutional layer to the input of the training samples, respectively, for element-wise summation; The channel-aligned convolutional layer is connected to the output of the residual connection structure and is used to perform dimensional transformation and output the model output value.

[0020] The above technical solution has the following advantages or beneficial effects: 1. By fusing gated temporal convolution and graph convolution, deep joint extraction of spatiotemporal features of spectral data is achieved, effectively overcoming the shortcomings of existing methods that only focus on the time dimension or use static graphs while ignoring spatiotemporal dynamic characteristics. This comprehensively reflects the spatiotemporal changes of the electromagnetic environment and significantly improves prediction accuracy and scene adaptability.

[0021] 2. By constructing an adaptive dynamic graph structure and combining prior knowledge with learnable matrices, the graph model can dynamically update the correlation between frequency bands according to the real-time spectrum status, which solves the problem that static graph modeling cannot cope with dynamic environmental changes and enhances the model's generalization ability and robustness in the event of sudden environmental changes.

[0022] 3. By replacing the traditional recurrent neural network structure with a gated temporal convolutional network with parallel computing characteristics, the computational complexity and inference latency of the model are significantly reduced. This effectively breaks through the bottlenecks of existing methods in terms of computational efficiency and real-time performance, and meets the urgent need for millisecond-level rapid response in scenarios such as emergency communication. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an emergency communication spectrum prediction method based on a gated adaptive timing graph integrated network, as shown in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the structure of the gated adaptive timing graph ensemble network in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the sub-process of step S2 in Embodiment 2 of the present invention; Figure 4 This is a schematic diagram of the sub-process of step S3 in Embodiment 4 of the present invention; Figure 5 This is a learning rate reduction curve corresponding to 100 snapshot rounds and 1000 total rounds in Embodiment 4 of the present invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment; other embodiments that conform to the spirit of the present invention may also fall within the scope of the present invention.

[0025] Example 1: This embodiment provides an emergency communication spectrum prediction method based on a gated adaptive timing graph ensemble network, such as... Figure 1 As shown, it includes the following steps: Step S1: Acquire the collected multi-band historical spectrum data, and divide the multi-band historical spectrum data into training samples and corresponding prediction labels by using a sliding window according to the preset historical time step and prediction step.

[0026] Step S2: Calculate the Pearson correlation coefficient between different frequency bands in the training samples and construct an initial prior adjacency matrix according to the set threshold. Generate an adaptive adjacency matrix by introducing a learnable matrix and fuse the two to obtain an adaptive dynamic graph structure. Construct a gated temporal graph network consisting of gated temporal convolutional layers, graph convolutional layers, channel-aligned convolutional layers and residual connection structures.

[0027] Step S3: Iteratively train the training samples and the adaptive dynamic graph structure into the gated temporal graph network. Backpropagate and update the model weights based on the loss function of the model output value and the predicted label. During the training process, save multiple model weight snapshots at the set snapshot period. After training is completed, average the multiple snapshots to obtain the integrated spectrum prediction model.

[0028] Step S4: Input the spectrum sample to be tested into the integrated spectrum prediction model, and output the spectrum status prediction value to provide the emergency communication command and dispatch system for frequency band scheduling of communication services.

[0029] Specifically, this embodiment provides an emergency communication spectrum prediction method based on a Gated Adaptive Temporal-graph Ensemble Network (GATEN). This GATEN network is not a single algorithm, but rather an integrated architecture constructed through specific logical steps, such as... Figure 2 As shown, it mainly consists of the following three functional modules: The adaptive dynamic graph modeling module is used to characterize the complex correlations between frequency bands in spectral data. By combining prior knowledge with learnable parameters, this module can dynamically update the graph structure according to changes in the electromagnetic environment, thus solving the technical problem that traditional static graph models cannot adapt to changing environments.

[0030] Gated Temporal Graph Network (GTN) Module: As the core of feature extraction, this module integrates a gated temporal convolutional network (TCN) and a graph convolutional network (GCN). TCN utilizes parallel computing to replace the traditional recurrent neural network structure, aiming to significantly reduce inference latency; GCN uses the aforementioned dynamic graph structure to extract spatial dimension features; at the same time, a residual structure is introduced to avoid information loss during deep network training.

[0031] Snapshot Ensemble Module: Introducing a snapshot ensemble strategy during the model training phase, this module acquires multiple model snapshots with differences by periodically adjusting the dynamic learning rate and then averages them, thereby improving the robustness and generalization performance of predictions in sudden and complex scenarios.

[0032] The specific implementation process of this method is as follows: through the execution of steps S1-S4, the construction and application of the above-mentioned GATEN network are finally completed: First, step S1 is executed to acquire multi-band historical spectrum data. According to the preset historical time step and prediction step, the multi-band historical spectrum data is divided into training samples and corresponding prediction labels using a sliding window. This step provides the initial data stream for the GATEN network, and the windowed slicing enables the network to capture the patterns of frequency band evolution over time.

[0033] Subsequently, step S2 is executed. This step implements the construction of the adaptive dynamic graph modeling module and the gated sequence graph network module in the GATEN network: 1. Calculate the Pearson correlation coefficients between different frequency bands in the training samples and construct an initial prior adjacency matrix based on a set threshold. Simultaneously, introduce a learnable matrix to generate an adaptive adjacency matrix. The two are then fused to construct an adaptive dynamic graph modeling module. In step S2 of this embodiment, by fusing prior knowledge and a learnable matrix, an adaptive dynamic graph modeling module is constructed. This module can dynamically update the adjacency matrix based on the real-time input spectrum state, capturing the potential, time-evolving interaction features between frequency bands. This dynamic adjustment mechanism enables the model to maintain strong adaptability even when facing sudden environmental changes and frequency band interference switching in emergency communications, solving the problem of lack of generalization ability caused by a fixed model structure.

[0034] 2. A GTN module is constructed, consisting of gated temporal convolutional layers, graph convolutional layers, channel-aligned convolutional layers, and residual connection structures. In this process, the gated temporal convolutional layers filter and extract temporal information in parallel through a dual-branch structure. The parallel computing characteristics of the gated temporal convolutional layers allow the model to process the entire sequence data simultaneously, significantly reducing computational complexity and inference latency. Combined with the fast forward propagation in step S4, this ensures that the system can output prediction results within milliseconds, meeting the urgent needs of real-time spectrum management in critical fields such as emergency communications.

[0035] The graph convolutional layer aggregates spatial features based on the adaptive dynamic graph, achieving deep joint feature modeling in the spatiotemporal dimensions. It organically combines gated temporal convolution with graph convolution. This technique can simultaneously extract the evolutionary features of spectral data along the time axis and the spatial dependence features between frequency bands, thus comprehensively reflecting the spatiotemporal dynamic changes of the electromagnetic environment and effectively solving the problem of insufficient prediction accuracy caused by neglecting spatial correlations in complex emergency scenarios.

[0036] Next, step S3 is executed. This step implements the snapshot ensemble module in the GATEN network through a specific training mechanism. Training samples are input into the constructed network layers for iterative training, and backpropagation is performed based on the loss function of the output values ​​and labels. During training, the learning rate is dynamically adjusted according to a set snapshot period (e.g., every 100 rounds), and the current weight snapshot is saved. After training is completed, the multiple snapshots are arithmetically averaged to obtain the ensembled spectrum prediction model. This approach obtains and fuses multiple locally optimal model snapshots through dynamic learning rate scheduling, leveraging the advantages of ensemble learning to smooth prediction noise, resulting in a final model that exhibits stronger robustness than traditional methods when dealing with highly random and non-stationary spectrum data.

[0037] Finally, step S4 is executed. The spectrum sample to be tested is input into the gated adaptive timing graph ensemble network that has been finally constructed and integrated through steps S1-S3. In this embodiment, the spectrum sample to be tested represents the electromagnetic environment state within the target frequency band range in the current and recent past period.

[0038] This method leverages the spatiotemporal feature extraction capabilities and ensemble robustness learned by a gated adaptive temporal graph ensemble network to output high-precision spectral state predictions. These predictions are data matrices corresponding to future input time steps, and their core parameters include: Predicted target time: The specified future time point (such as the first sampling interval after the current time).

[0039] Frequency band identifier: the predicted frequency points or frequency band ranges.

[0040] Predicted power spectral density value: The expected signal power intensity (unit: dBm) of each frequency band at the predicted target time, which is a direct quantitative basis for determining the frequency band occupancy status.

[0041] It can also include derived occupancy status indicators: by comparing the predicted power value with a preset threshold, a binary status label of "occupied" (1) or "idle" (0) for each frequency band can be further generated to provide an intuitive judgment for decision-making.

[0042] Before the introduction of this method, emergency communication command and dispatch systems could only perceive current or historical spectrum occupancy. When a new emergency communication service (such as establishing a rescue team video connection) was initiated, it could only attempt to access the system based on the current frequency band's availability. If the frequency band was occupied by other services immediately after access, co-channel interference would occur, leading to a decrease in communication quality or even interruption. Allocation strategies were relatively fixed or slow to react, making it impossible to achieve globally optimal dynamic allocation in complex and rapidly changing emergency electromagnetic environments. After introducing this method, the emergency communication command system can proactively act before interference occurs based on the prediction results. For example, if a prediction shows that frequency band A, currently carrying the core voice link, will be interfered with by a sudden high-power signal in 30 seconds, the system can immediately instruct the communication terminal to smoothly and seamlessly switch the link to the predicted available frequency band B before the interference occurs. When newly arriving rescue forces request communication resources, the system can directly allocate suitable future available frequency bands based on the prediction results, ensuring successful access on the first attempt and avoiding conflicts with known future occupancy plans.

[0043] Therefore, the complete technical process from input of the sample to output of the predicted value described in this embodiment tightly embeds the prediction algorithm into the decision-making closed loop of emergency communication, which directly improves the reliability, robustness and resource utilization efficiency of emergency communication in complex and sudden scenarios.

[0044] Example 2: This embodiment specifically defines step S2 in Embodiment 1 and elaborates on the generation logic of the adaptive dynamic graph structure, such as... Figure 3 As shown, step S2 includes: Step S21: The similarity between different frequency bands in the training samples is measured by using the Pearson correlation coefficient, and an initial prior adjacency matrix is ​​constructed according to the set threshold. Step S22 involves generating an adaptive adjacency matrix by introducing a learnable matrix; the method for generating the adaptive adjacency matrix in step S22 is as follows: An adaptive adjacency matrix is ​​generated by multiplying two learnable matrices. The learnable matrix has a dimension of N×c, where N is the number of frequency bands and c is the feature length.

[0045] Step S23: The adaptive adjacency matrix is ​​fused with the initial prior adjacency matrix to form the final adaptive dynamic graph structure.

[0046] Specifically, in order to address the problem that static graph modeling is difficult to handle complex and variable electromagnetic environments, this embodiment designs the graph construction process.

[0047] Traditional graph modeling typically employs a static approach, establishing the graph structure by calculating the correlations between frequency bands. While this method performs well in static scenarios, in complex scenarios such as emergency communications and public safety, the electromagnetic environment exhibits significant dynamism and uncertainty. The spectral state changes continuously over time and space, making it difficult for static modeling to accurately reflect real-time dependencies, thus impacting prediction accuracy.

[0048] To address this, this invention proposes an adaptive dynamic graph modeling method that combines prior knowledge with a learnable matrix to dynamically capture potential feature changes between frequency bands. Specifically, the Pearson correlation coefficient is first used to measure the similarity between different frequency bands, and an initial adjacency matrix is ​​constructed based on a set threshold to ensure the reliability of the prior structure. It is defined as follows:

[0049] in, This represents the Pearson correlation coefficient. The threshold is then set. Subsequently, a learnable matrix is ​​introduced. and To generate adaptive matrices This supplements and dynamically corrects prior correlations. Its expression is:

[0050] in, , This represents the feature length. Finally, the learned adaptive matrix... With the prior adjacency matrix The components are merged to form the final dynamic graph structure:

[0051] This method retains the original frequency band-related knowledge while introducing learnable parameters to achieve dynamic modeling of the time-varying relationship of the spectrum under complex electromagnetic environments, effectively improving the system's environmental adaptability and modeling accuracy.

[0052] Example 3: This embodiment focuses on illustrating the internal feature extraction mechanism of the gated temporal graph network in Embodiment 1. For example... Figure 2 As shown, the input-output connections in the gated timing graph network are as follows: Training samples are input into a gated temporal convolutional layer, which extracts temporal features through a parallel dual-branch structure with dilated convolutions. The parallel dual-branch structure in the gated temporal convolutional layer includes: The first branch extracts nonlinear features through an activation function; The second branch generates weight coefficients using the Sigmoid function; The time series characteristics are obtained by multiplying the outputs of the two branches element by element.

[0053] The temporal features extracted by the gated temporal convolutional layer are input into the graph convolutional layer for spatial feature aggregation between frequency bands. By using a residual connection structure, the output of the graph convolutional layer is element-wise summed with the original input features of the training samples, and then the input channels are aligned to perform dimensionality transformation on the convolutional layer to obtain the model output value and model weights.

[0054] Specifically, gated temporal graph networks are used to jointly capture the temporal and spatial dependencies of spectral signals. In the time dimension, this invention uses gated temporal convolutional layers (TCNs) instead of traditional recurrent neural networks to achieve parallel computation and long sequence modeling. Compared to RNNs or LSTMs, TCNs can process the entire time series simultaneously, avoiding the gradient vanishing problem, and expand the receptive field through dilated convolutions. Its convolution calculation form is as follows:

[0055] in, Indicates the coefficient of thermal expansion. This represents the kernel size.

[0056] To enhance the expressive power of temporal modeling, a gating mechanism is introduced after the TCN layer. Nonlinear features and information weights are extracted through two parallel branches. The former uses an activation function to enhance the feature response, while the latter uses a sigmoid function to adjust the intensity of the information flow. The product of the two is used as the output to achieve adaptive control of the information channel.

[0057] In the spatial dimension, a graph convolutional network (GCN) is used to model the non-Euclidean structure between frequency band nodes and extract multi-scale spatial features. The update formula is as follows:

[0058] in, For degree matrix, It is an adjacency matrix. For the input feature matrix, This is the weight matrix.

[0059] Furthermore, to avoid feature decay caused by gating mechanisms, this invention introduces residual connections into the network, ensuring feature consistency during cross-layer propagation. The stacking of multiple GTN modules captures multi-level spatiotemporal dependencies at different time scales, and finally, feature fusion and prediction result generation are achieved through convolutional output layers.

[0060] Example 4: This embodiment details the specific implementation of snapshot ensemble training in step S3 of Example 1. To address the complexity and suddenness of the electromagnetic environment, and to enhance the robustness and generalization performance of the model while avoiding the high computational cost of multiple independent training iterations in traditional ensemble learning methods, this embodiment employs a snapshot ensemble strategy. Its core lies in: guiding the model to converge to different local optima through periodic learning rate scheduling during a single training session, thereby obtaining a series of differentiated sub-model snapshots, which are ultimately fused to improve overall prediction performance.

[0061] like Figure 4 As shown, the specific implementation details are as follows: Step S31: Input each batch of training samples into the gated temporal graph network for forward computation to obtain the model output value.

[0062] Step S32: The mean squared error function is used to calculate the deviation between the initial model prediction value and the predicted label, and the loss function is constructed.

[0063] Step S33: Update the weight parameters in the gated temporal graph network using the optimizer based on the gradient of the loss function.

[0064] Step S34: Within each preset snapshot period, after each training round, the learning rate is gradually reduced according to a preset decay factor (e.g., 0.95). Figure 5 The figure shows the learning rate decrease curve when the snapshot epoch is 100 and the total epoch is 1000. This gradual decrease causes the model to converge toward a local optimum within the current epoch.

[0065] Step S35: When the current training epoch reaches an integer multiple of the snapshot period, save the model weights at this time as a snapshot. Then, reset the learning rate to its initial value. This reset operation simulates a restart of the training process, allowing the model to escape its current optimization trajectory and begin exploring the parameter space at a new pace, thereby converging to another differentiated local optimum. This achieves the effect of training multiple models simultaneously.

[0066] Step S36: After reaching the total number of training rounds, extract all saved model weight snapshots and generate the integrated final spectrum prediction model parameters through simple arithmetic averaging.

[0067] In summary, this embodiment, through the aforementioned dynamic learning rate adjustment and periodic snapshot saving mechanism, enables the model to learn at different paces at different stages, resulting in diverse parameter distributions. Finally, parameter fusion of multiple snapshot models effectively improves the stability and generalization ability of spectrum prediction in complex and dynamic scenarios such as emergency communications. Experimental data shows that this fusion strategy can reduce the mean absolute error by approximately 2.6%.

[0068] Furthermore, this invention also provides an embodiment in a specific application scenario to illustrate the overall technical solution of this invention.

[0069] Taking the ElectroSense open-source real dataset as an example, the specific steps are as follows: Step 1: Data collection and dataset construction.

[0070] The dataset used in this embodiment comes from the open-source platform ElectroSense, a distributed spectrum monitoring system that collects and analyzes radio frequency spectrum data globally through crowdsourcing. The experiment selected a monitoring sensor located in Madrid, Spain, to collect power spectral density (PSD) data in the frequency range of 500MHz to 800MHz. The sampling period was from May 28, 2021 to June 28, 2021, lasting one month. Details of the dataset are shown in Table 1. Table 1. Dataset Details

[0071] Step 2: Dataset partitioning.

[0072] Step 2.1: Divide the dataset obtained in Step 1 into a training set and a test set in an 8:2 ratio; Step 2.2: Divide the training set and the test set into samples and labels respectively by using a sliding window according to the set historical time step and prediction step. In this embodiment, the historical time step is 16 and the prediction step is 1.

[0073] Step 3: Adaptive graph construction.

[0074] Step 3.1: Based on the training set obtained in Step 2, the similarity between different frequency bands is measured using the Pearson correlation coefficient, and an initial adjacency matrix is ​​constructed according to a set threshold to ensure the reliability of the prior structure. It is defined as:

[0075] in, This represents the Pearson correlation coefficient. The threshold value is used.

[0076] Step 3.2: An adaptive matrix is ​​generated by introducing learnable matrices A and B, thereby supplementing and dynamically correcting prior correlations. Its expression is:

[0077] in, , Indicates the feature length. In this example, Set it to 10.

[0078] Step 3.3: Merge the learned adaptive adjacency matrix with the prior adjacency matrix to form the final dynamic graph structure.

[0079] Step 4: Model training.

[0080] Step 4.1: Input the training set obtained by splitting the dataset and the adaptive adjacency matrix obtained by constructing the adaptive graph into the gated TCN layer of the GATEN network to capture the temporal correlation of the spectral data; Step 4.2: Pass the output of the gated TCN through the GCN layer to obtain the correlation between frequency bands; Step 4.3: Then, channel alignment is performed through convolutional layers, and the final prediction dimension is output through the output module to obtain the initial model prediction value and model weights; Step 4.4: Calculate the loss by combining the obtained predicted values ​​with the labels, and use it for backpropagation of the model; Step 4.5: Save the snapshot weights of the model at the set snapshot rounds. In this example, the snapshot rounds are 100 and the total number of rounds is 1000, which will yield 10 snapshot model weights. Step 4.6: Average the weights of the 10 snapshot models to obtain the final optimal spectrum prediction model weights.

[0081] The experimental setup is shown in Table 2: Table 2

[0082] Step 5: Test.

[0083] The test samples are directly input into the optimal spectrum prediction model obtained in step 4 to obtain the final prediction value.

[0084] In summary, this invention constructs a high-precision spectrum prediction method for complex electromagnetic environments using GATEN. This method comprehensively utilizes adaptive dynamic graph modeling, the multi-dimensional feature representation capabilities of TCN and GCN, and introduces a snapshot ensemble strategy during the training phase, achieving efficient modeling and robust prediction of the spatiotemporal dynamic characteristics of the spectrum. The spectrum prediction framework based on gated adaptive temporal graph ensemble proposed in this invention can achieve high-precision, low-latency prediction in complex and abrupt electromagnetic environments, effectively improving the reliability and resource utilization efficiency of emergency communication and public safety communication systems.

[0085] This invention creatively integrates multiple technologies such as deep learning, dynamic graph modeling, and ensemble learning, overcoming the problems of traditional time-series models relying on time-step calculations and static graph modeling lacking time-varying adaptability, thus achieving dynamic perception and prediction of spectrum states. This method not only significantly improves prediction accuracy and generalization ability while reducing model training and maintenance costs, but also possesses strong scalability and adaptability, making it flexibly applicable to various complex scenarios such as emergency communication, disaster relief, public safety assurance, and communication scheduling for major events, demonstrating promising prospects for engineering promotion and application.

[0086] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. An emergency communication spectrum prediction method based on gated adaptive timing graph ensemble networks, characterized in that, include: S1. Acquire the collected multi-band historical spectrum data, and divide the multi-band historical spectrum data into training samples and corresponding prediction labels by a sliding window according to the preset historical time step and prediction step. Step S2: Calculate the Pearson correlation coefficient between different frequency bands in the training samples and construct an initial prior adjacency matrix according to a set threshold. Generate an adaptive adjacency matrix by introducing a learnable matrix. Merge the adaptive adjacency matrix with the initial prior adjacency matrix to obtain an adaptive dynamic graph structure. Construct a gated temporal graph network consisting of a gated temporal convolutional layer, a graph convolutional layer, a channel-aligned convolutional layer, and a residual connection structure. Step S3: Input the training samples and the adaptive dynamic graph structure into the gated temporal graph network for iterative training. Perform backpropagation based on the loss function of the obtained model output value and the predicted label and update the model weights. During the training process, save multiple model weight snapshots according to the set snapshot period. After the training is completed, average the multiple model weight snapshots to obtain the integrated spectrum prediction model. Step S4: Input the spectrum sample to be tested into the integrated spectrum prediction model, and output the spectrum status prediction value to provide the emergency communication command and dispatch system for frequency band scheduling of communication services.

2. The emergency communication spectrum prediction method according to claim 1, characterized in that, Step S2 includes: Step S21: The similarity between different frequency bands in the training samples is measured by using the Pearson correlation coefficient, and the initial prior adjacency matrix is ​​constructed according to the set threshold. Step S22: The adaptive adjacency matrix is ​​generated by introducing a learnable matrix; Step S23: The adaptive adjacency matrix is ​​fused with the initial prior adjacency matrix to form the final adaptive dynamic graph structure.

3. The emergency communication spectrum prediction method according to claim 1, characterized in that, The input-output connection relationships in the gated timing graph network in step S2 are as follows: The training samples are input into the gated temporal convolutional layer, which extracts temporal features through a parallel dual-branch structure with dilated convolution. The temporal features extracted by the gated temporal convolutional layer are input into the graph convolutional layer to perform spatial feature aggregation between frequency bands. The output of the graph convolutional layer is element-wise summed with the original input features of the training samples using a residual connection structure. Then, the input channel-aligned convolutional layer undergoes dimensionality transformation to obtain the model output value and model weights.

4. The emergency communication spectrum prediction method according to claim 3, characterized in that, The parallel dual-branch structure in the gated temporal convolutional layer includes: The first branch extracts nonlinear features through an activation function; The second branch generates weight coefficients using the Sigmoid function; The time series characteristics are obtained by multiplying the outputs of the two branches element by element.

5. The emergency communication spectrum prediction method according to claim 1, characterized in that, Step S3 includes: Step S31: Input each batch of training samples into the gated temporal graph network for forward computation to obtain the model output value; Step S32: Calculate the deviation between the initial model prediction value and the predicted label using the mean squared error function, and construct the loss function; Step S33: Update the weight parameters in the gated temporal graph network using the optimizer based on the gradient of the loss function; Step S34: Within each preset snapshot period, the learning rate is reduced according to a preset decay factor after each round of training. Step S35: When the current training round reaches an integer multiple of the snapshot period, save the current model weight snapshot and reset the learning rate to the initial learning rate; Step S36: After reaching the total number of training rounds, extract all saved model weight snapshots and perform arithmetic averaging to generate integrated spectrum prediction model parameters.

6. The emergency communication spectrum prediction method according to claim 2, characterized in that, The adaptive adjacency matrix in step S22 is generated as follows: An adaptive adjacency matrix is ​​generated by multiplying two learnable matrices, wherein the learnable matrix has a dimension of N×c, where N is the number of frequency bands and c is the feature length.

7. An emergency communication spectrum prediction system based on a gated adaptive timing graph ensemble network, characterized in that, The emergency communication spectrum prediction method according to any one of claims 1-6 includes: The data processing module is used to acquire the collected multi-band historical spectrum data, and divide the multi-band historical spectrum data into training samples and corresponding prediction labels through a sliding window according to the preset historical time step and prediction step. The model building module is used to build an adaptive dynamic graph structure based on the training samples, and to build a gated temporal graph network consisting of gated temporal convolutional layers, graph convolutional layers, channel-aligned convolutional layers, and residual connection structures. An integrated training module, connecting the data processing module and the model building module, is used to input the training samples and the adaptive dynamic graph structure into the gated temporal graph network for iterative training. It performs backpropagation based on the loss function of the obtained model output value and the predicted label and updates the model weights. During the training process, multiple model weight snapshots are saved according to a set snapshot period. After the training is completed, the multiple model weight snapshots are averaged to obtain the integrated spectrum prediction model. The prediction application module, connected to the integrated training module, is used to input the spectrum sample to be tested into the integrated spectrum prediction model and output the spectrum state prediction value to provide the emergency communication command and dispatch system for frequency band scheduling of communication services.

8. The emergency communication spectrum prediction system according to claim 7, characterized in that, The model building module includes: The prior adjacency matrix construction unit is used to measure the similarity between different frequency bands in the training samples using the Pearson correlation coefficient, and to construct an initial prior adjacency matrix according to a set threshold. An adaptive adjacency matrix generation unit is used to generate an adaptive adjacency matrix by introducing a learnable matrix. The dynamic graph fusion unit is used to fuse the adaptive adjacency matrix with the initial prior adjacency matrix to form the final adaptive dynamic graph structure.

9. The emergency communication spectrum prediction system according to claim 8, characterized in that, The adaptive adjacency matrix generation unit generates an adaptive adjacency matrix by performing matrix multiplication on two learnable matrices of dimension N×c, where N is the number of frequency bands and c is the feature length.

10. The emergency communication spectrum prediction system according to claim 7, characterized in that, The hierarchical connection relationship of the gated sequence graph network in the model construction module is as follows: The gated temporal convolutional layer is connected to the input of the training samples and extracts temporal features through a parallel dual-branch structure with dilated convolution. The input of the graph convolutional layer is connected to the output of the gated temporal convolutional layer and the output of the dynamic graph fusion unit, respectively, for spatial feature aggregation. The residual connection structure connects the output of the graph convolutional layer to the input of the training samples, respectively, for element-wise summation; The channel-aligned convolutional layer is connected to the output of the residual connection structure and is used to perform dimensional transformation and output the model output value.

Images