Federal learning short wave highest available frequency prediction method for space-time coupling

Abstract

The invention discloses a space-time coupling-oriented federated learning short wave highest available frequency prediction method, which comprises the following steps of: A, data preprocessing and data acquisition; step B, actually measured data sequence decomposition: performing time sequence decomposition on the MUF to obtain a trend component and applying the trend component to subsequent steps; c, observing mutual information change between models in federated learning by using a MINE method; step D, analyzing the spatial relationship between the federated learning local models; and E, selecting a personalized fusion strategy to complete an MUF prediction task in federated learning. According to the method, the similarity of MUFs of all stations is evaluated, trend terms in the MUFs are used as federal learning training data, mutual information between local models of all the stations is evaluated, and potential spatial correlation consistent with geomagnetic latitude information in the data is determined. The invention provides a personalized federal learning framework based on mutual information, the model is complete, the physical significance is clear, and the accuracy of short-wave highest available frequency spatio-temporal reasoning is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of shortwave highest available frequency prediction, and in particular relates to a method for predicting shortwave highest available frequency through federated learning oriented to spatiotemporal coupling. Background Art

[0002] Shortwave (HF) communications have been widely used in commercial and military fields due to their flexibility, mobility, and long-distance transmission capabilities without relying on fixed facilities. However, with the increasingly urgent demand for timely shortwave communication services, the effective prediction of available propagation frequencies has become a pressing issue. According to the recommendations of the International Telecommunication Union (ITU), the maximum usable frequency (MUF) is a key factor in determining the appropriate operating frequency for shortwave communications. Accurate prediction of MUF can speed up the frequency selection process, enhance transmission capacity, and improve communication quality. Unfortunately, due to the limited distribution of actual sounding stations, MUF data is scarce and cannot meet the needs of shortwave communications in unfamiliar areas. Therefore, using existing sounding data to infer and predict the MUF of shortwave propagation in unfamiliar locations through spatiotemporal reasoning can help alleviate the problems of resource waste, poor channel quality, and long link establishment time caused by blind frequency selection under cold start conditions.

[0003] Currently, the modeling and broadcasting of the ionosphere mainly rely on ionospheric data assimilation methods. Data assimilation involves incorporating sounding data with irregular spatial and temporal distributions in order to match them with physical empirical models and reduce errors. These methods have achieved positive results in near-real-time broadcasting of the ionosphere. However, current ionospheric assimilation methods have two main limitations. First, they have difficulty accurately predicting future states, which affects their ability to provide MUF information of sufficient length for shortwave equipment. Second, existing methods mainly use elliptical Gaussians of geomagnetic coordinates to estimate spatial correlations in the horizontal direction, which does not accurately reflect the non-isotropic correlations actually observed. This discrepancy is particularly problematic for shortwave communication needs, because it is very important to understand the state of the ionosphere at the refraction point. Summary of the invention

[0004] The present invention aims to solve the problems raised in the background technology and proposes a method for predicting the highest available shortwave frequency through federated learning oriented to spatiotemporal coupling.

[0005] In order to achieve the purpose of the present invention, the present invention discloses a method for predicting the shortwave maximum available frequency by federated learning for spatiotemporal coupling, comprising the following steps:

[0006] Step A, data preprocessing: measuring the parameter foF2 of the F2 layer in different regions and the propagation factor M(3000)F2 of the F2 layer in the ionosphere at 3000 km, obtaining the highest available frequency MUF of the data, and preliminarily observing the spatial relationship of different regions;

[0007] Step B, decomposition of measured data series: select the additive decomposition model to decompose the MUF of a region into time series, and observe the original data, long-term trend component, diurnal component, and residual component;

[0008] Step C: Observe the changes in mutual information between models in federated learning; use the MINE method to observe the changes in mutual information between models in federated learning;

[0009] Step D: Spatial relationship analysis between local models of federated learning: The spatial relationship between local models is studied by analyzing the mutual information in the federated learning training process;

[0010] Step E: Complete the MUF prediction task; select a personalized fusion strategy to complete the MUF prediction task in federated learning.

[0011] Furthermore, step A specifically includes the following steps:

[0012] Step A1, selection of vertical measurement stations and vertical measurement data and data reading; using vertical measurement data of urban ionosphere vertical measurement stations at different longitudes and latitudes to obtain the vertical detection critical frequency foF2 of the M(3000)F2 and F2 layers;

[0013] Step A2, calculate F2(3000)MUF, the specific formula is F2(3000)MUF=M(3000)F2×foF2.

[0014] Furthermore, in step B, the measured data sequence is decomposed: the decomposition method based on STL in the classical decomposition method is used to perform additive decomposition on the MUF obtained in step A to observe the proportions of different components; wherein, the additive decomposition model for MUF is expressed as follows:

[0015] y t =D t +T t +R t

[0016] Where yt represents the time series, Dt represents the diurnal term, Tt represents the trend term, and Rt represents the residual term. The diurnal term represents the short-period diurnal characteristics of the ionosphere, indicating the major changes that occur within a day due to the rotation of the earth. The trend term represents the long-term changes affected by factors such as solar activity. The existence of the residual term will hinder the accurate identification of the spatiotemporal attributes of the original data.

[0017] Furthermore, step C specifically includes the following steps:

[0018] Step C1, mutual information experiments are conducted between different sites and the global model. The input data is the original MUF data. The data set is divided into three parts: training set, validation set and test set. The original data is divided into the following proportions: 70% for training set, 20% for validation set and 10% for test set. FedAvg is used as the benchmark federated learning method. The fusion strategy includes merging and averaging all local models. The LSTM model is selected as the local model of each site. Global training rounds and local training rounds are performed 30 times each.

[0019] Step C2: Use the MINE method to observe the change of mutual information between the global model and each local model during the federated learning training process; where the mutual information is defined as follows:

[0020]

[0021] in, is the joint probability distribution, and is the marginal probability distribution.

[0022] Furthermore, the MINE method effectively estimates the lower limit of mutual information by evaluating the expected value of samples; since mutual information can be expressed as the KL (Kullback-Leibler) divergence between the joint probability distribution and the marginal probability distribution; KL divergence and the powerful modeling ability of neural networks can estimate the minimum value of mutual information between continuous variables; therefore, mutual information is expressed as the KL divergence between the joint probability distribution and the marginal probability distribution:

[0023]

[0024] Among them, X and Z are independent random variables, D KL It is expressed as:

[0025]

[0026] Among them, the distribution Relative to the distribution It is absolutely continuous;

[0027] In order to solve the problem that joint probability distribution and marginal probability distribution are difficult to obtain in actual data, the dual representation of KL divergence is used:

[0028]

[0029] If all functions T have finite upper limits, then both expected values ​​are also finite; when a family of functions is specified for T When , we can get the lower bound of mutual information; therefore, the MINE method also mentions the f-divergence representation of KL divergence:

[0030]

[0031] The MINE method uses the fitting ability of deep neural networks to define functions In addition, we obtain the bounds of the mutual information:

[0032] I(X; Z)≥I Θ (X,Z)

[0033] Among them, I Θ (X, Z) refers to estimating the mutual information between random variables by building a deep neural network; therefore, I Θ (X,Z) is specifically expressed as:

[0034]

[0035] When training deep neural networks, the actual data distribution cannot be obtained and can only be estimated through data sampling; Θ (X,Z) is further expressed as:

[0036]

[0037] Among them, for the distribution is the distribution of n independent and identically distributed samples;

[0038] A two-layer fully connected neural network model is constructed to estimate mutual information, and the activation function used is ReLu.

[0039] Furthermore, in step D, the spatial relationship between the local models of federated learning is analyzed: the trend term of the MUF data in step B is used as the input of each local model, and the magnetic inclination data of the 2018 International Geomagnetic Reference Field (IGRF) and the CSES Global Geomagnetic Magnetic Model (CGGM) are considered; the mutual information between the sites of the MUF long-term trend component during the federated learning training process is calculated, and the geographical location and magnetic latitude of the station where each model is located are also taken into account, and the mutual information and site distribution of each model are observed.

[0040] Furthermore, in step E, a personalized fusion strategy is selected to complete the MUF task in federated learning, and the k-NearestNeighbor fusion strategy is adopted, that is, the nearest neighbor model with higher mutual information with the target model is selected from all models; in the above strategy, the selection of neighbors is based on the mutual information of MINE as the correlation measure, and the selection of hyperparameters is the empirical value selected through the experimental process. This personalized fusion strategy is selected to complete the MUF task in federated learning.

[0041] Compared with the prior art, the significant progress of the present invention lies in: the present invention provides a shortwave broadcast frequency scheduling method based on multi-source data processing, which is scientifically designed and mines the implicit spatial correlation in real-world MUF measurements. Based on time series decomposition theory and dynamic time warping, the similarities between different components of the MUF of each site are evaluated to determine the necessary data input requirements to better obtain spatial correlation features; the present invention evaluates the spatial correlation between local models from different sites in a federated learning environment. By using the trend terms in the MUF as training data for federated learning, the potential spatial correlation in the data is determined by evaluating the mutual information between the local models and geomagnetic latitude information of all sites. The present invention proposes a personalized federated learning framework based on mutual information, which has a complete model and clear physical meaning, and effectively improves the accuracy of spatiotemporal reasoning of the highest available frequency of shortwave.

[0042] In order to more clearly illustrate the functional characteristics and structural parameters of the present invention, further description is given below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0044] Figure 1 A schematic diagram of the distributed site joint MUF spatiotemporal reasoning prediction solution based on federated learning provided by the present invention;

[0045] Figure 2 Schematic diagram of the personalized federated learning framework based on mutual information provided by the present invention.

[0046] Figure 3 A flowchart of a method for predicting the shortwave maximum available frequency by federated learning for spatiotemporal coupling provided by the present invention; DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] like Figure 1As shown, in order to solve the above two defects, the present invention proposes a method based on known ionospheric detection stations, which helps unfamiliar regional stations predict MUF by using the spatial correlation of data between stations for spatiotemporal inference. Although the existing ionospheric assimilation models have made significant progress, these models cannot be defined and fitted by linear equations due to the complex physical mechanisms of the ionosphere. In contrast, deep learning has attracted much attention in the field of earth sciences due to its ability to fit data nonlinearly, thereby enabling the discovery of more complex features. However, when using deep learning, it is necessary to consider physical laws to ensure interpretability. Therefore, how to analyze and utilize the non-uniform spatiotemporal correlation in ionospheric multi-source single-variable spatiotemporal reasoning modeling, and then realize the interpretable spatiotemporal reasoning of ionospheric parameters in unfamiliar regions based on deep learning, has become a key issue in improving the prediction performance of the shortwave highest available frequency under the condition of no detection equipment coverage.

[0049] like Figure 2 As shown, the present invention provides a method for predicting the highest available shortwave frequency by federated learning for spatiotemporal coupling. Figure 2 Further improved Figure 1 Details of the proposed framework in . It consists of two federated learning processes that are trained synchronously. Figure 2 The upper part of the figure shows the personalized federated learning framework, and the training data is the original MUF data. However, since the local model uses a naive LSTM model. Figure 2 As shown in the lower part, another federated learning is used to provide a personalized strategy based on mutual information. The training data used in this part is the trend item in the MUF data. The corresponding steps are as follows Figure 3 As shown, the following steps are included:

[0050] Step A, data preprocessing: measure the parameters foF2 and M(3000)F2 of the F2 layer in different regions, obtain MUF data, and preliminarily observe the spatial relationship of different regions;

[0051] Step B, measured data series decomposition: For the MUF of a region, select the additive decomposition model to decompose the time series and observe the original data, long-term trend component, diurnal component, and residual component;

[0052] Step C: Observe the changes in mutual information between models in federated learning: Use the MINE method to observe the changes in mutual information between models in federated learning

[0053] Step D: Spatial relationship analysis between local models in federated learning: The spatial relationship between local models is studied by analyzing the mutual information during federated learning training.

[0054] Step E, complete the MUF prediction task: select a personalized fusion strategy to complete the MUF prediction task in federated learning

[0055] In the present invention, in specific implementation, step A specifically includes the following sub-steps:

[0056] Step A1, selection of vertical measurement station and vertical measurement data and data reading: using the vertical measurement data of urban ionospheric vertical measurement stations at different longitudes and latitudes, obtain the propagation factor M(3000)F2 of the ionosphere F2 layer at 3000 kilometers and the vertical detection critical frequency foF2 of the F2 layer.

[0057] Step A2, calculate F2(3000)MUF, the specific formula is as follows:

[0058] F2(3000)MUF=M(3000)F2×foF2, formula (1);

[0059] Step B, measured data sequence decomposition: Adopt the STL decomposition method based on the classical decomposition method to perform additive decomposition on the MUF obtained in step A and observe the proportion of different components. Among them, the additive decomposition model of MUF can be expressed as follows:

[0060] y t =D t +T t +R t , formula (2);

[0061] Among them, y t represents the time series, D t represents the day and night term, T t represents the trend term, R t Represents the residual term. The diurnal term represents the short-period diurnal characteristics of the ionosphere, which represents the significant changes that occur in a day due to the rotation of the Earth. The trend term represents the long-term changes affected by factors such as solar activity. The existence of the residual term hinders the accurate identification of the spatiotemporal properties of the original data.

[0062] In the present invention, in specific implementation, step C specifically includes the following sub-steps:

[0063] Step C1: Conduct mutual information experiments between different sites and the global model. The input data used in the present invention is the original MUF data, and the data set is divided into three parts: training set, validation set and test set. The original data is divided into the following proportions: 70% for training set, 20% for validation set and 10% for test set. FedAvg is used as the benchmark federated learning method. The fusion strategy includes merging and averaging all local models. The commonly used LSTM model is selected as the local model of each site. 30 rounds of global training and 30 rounds of local training are performed.

[0064] Step C2: Use the MINE method to observe the changes in mutual information between the global model and each local model during the federated learning training process. The mutual information is defined as follows:

[0065]

[0066] in, is the joint probability distribution, and is the marginal probability distribution.

[0067] Among them, the MINE method effectively estimates the lower limit of mutual information by evaluating the expected value of samples. Since mutual information can be expressed as the Kullback-Leibler (KL) divergence between the joint probability distribution and the marginal probability distribution. KL divergence and the powerful modeling ability of neural networks can estimate the minimum value of mutual information between continuous variables. Therefore, mutual information can be expressed as the KL divergence between the joint probability distribution and the marginal probability distribution:

[0068]

[0069] Among them, X and Z are independent random variables, D KL It can be expressed as:

[0070]

[0071] Among them, the distribution Relative to the distribution It is absolutely continuous.

[0072] In order to solve the problem that joint probability distribution and marginal probability distribution are difficult to obtain in actual data, the dual representation of KL divergence is used:

[0073]

[0074] If all functions T have finite upper bounds, then both expected values ​​are also finite. When , we can get the lower bound of mutual information. Therefore, the MINE method also mentions the f-divergence representation of KL divergence:

[0075]

[0076] The MINE method uses the fitting ability of deep neural networks to define functions In addition, the bounds of the mutual information can be obtained:

[0077] I(X; Z)≥I Θ (X, Z), formula (8);

[0078] Among them, IΘ (X, Z) refers to estimating the mutual information between random variables by building a deep neural network. Therefore, I Θ (X,Z) can be specifically expressed as:

[0079]

[0080] When training a deep neural network, the actual data distribution cannot be obtained and can only be estimated through data sampling. Θ (X,Z) can be further expressed as:

[0081]

[0082] Among them, for the distribution is the distribution of n independent and identically distributed samples.

[0083] The present invention constructs a two-layer fully connected neural network model to estimate mutual information, and the activation function adopted is ReLu.

[0084] Step D, spatial relationship analysis between local models of federated learning: Use the trend term of MUF data in step B as the input of each local model, and consider the magnetic inclination data of the 2018 International Geomagnetic Reference Field (IGRF) and the CSES Global Geomagnetic Magnetic Model (CGGM). Calculate the mutual information between each station of the MUF long-term trend component during the federated learning training process, and the geographical location and magnetic latitude of each model station are also taken into account, and observe the mutual information and station distribution of each model.

[0085] Step E, select a personalized fusion strategy to complete the MUF task in federated learning, and use the k-NearestNeighbor fusion strategy, that is, select the first k nearest neighbor models with higher mutual information with the target model from all models. In the above strategy, the selection of neighbors is based on the mutual information of MINE as the correlation measure, and the selection of the hyperparameter k is an empirical value selected through the experimental process. This personalized fusion strategy is selected to complete the MUF task in federated learning.

[0086] In summary, compared with the prior art, the shortwave broadcast frequency scheduling method based on multi-source data processing provided by the present invention is scientifically designed and mines the implicit spatial correlation in real-world MUF measurements. Based on time series decomposition theory and dynamic time warping, the similarities between different components of MUF at each site are evaluated to determine the necessary data input requirements, so as to better obtain spatial correlation features; the present invention evaluates the spatial correlation between local models from different sites in a federated learning environment. By using the trend items in MUF as training data for federated learning, the potential spatial correlation in the data is determined by evaluating the mutual information between the local models and geomagnetic latitude information of all sites. The present invention proposes a personalized federated learning framework based on mutual information. The model is complete, the physical meaning is clear, and the accuracy of spatiotemporal reasoning of the highest available frequency of shortwave is effectively improved.

[0087] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0088] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the shortwave highest available frequency by federated learning for spatiotemporal coupling, characterized in that: The following steps are involved: Step A, data preprocessing: measuring the parameter foF2 of the F2 layer in different regions and the propagation factor M(3000)F2 of the F2 layer in the ionosphere at 3000 km, obtaining the highest available frequency MUF of the data, and preliminarily observing the spatial relationship of different regions; Step B, decomposition of measured data series: select the additive decomposition model to decompose the MUF of a region into time series, and observe the original data, long-term trend component, diurnal component, and residual component; Step C: Observe the changes in mutual information between models in federated learning; use the MINE method to observe the changes in mutual information between models in federated learning; Step D: Spatial relationship analysis between local models of federated learning: The spatial relationship between local models is studied by analyzing the mutual information in the federated learning training process; Step E: Complete the MUF prediction task; select a personalized fusion strategy to complete the MUF prediction task in federated learning.

2. According to the method of claim 1, the method for predicting the shortwave maximum available frequency by federated learning for spatiotemporal coupling is characterized in that: Step A specifically includes the following steps: Step A1, selection of vertical measurement stations and vertical measurement data and data reading; using vertical measurement data of urban ionosphere vertical measurement stations at different longitudes and latitudes to obtain the vertical detection critical frequency foF2 of the M(3000)F2 and F2 layers; Step A2, calculate F2(3000)MUF, the specific formula is F2(3000)MUF=M(3000)F2×foF2.

3. According to the method of claim 1, the method for predicting the shortwave maximum available frequency by federated learning for spatiotemporal coupling is characterized in that: In step B, the measured data sequence is decomposed: the MUF obtained in step A is additively decomposed using the STL-based decomposition method to observe the proportions of different components; the additive decomposition model for MUF is expressed as follows: y t =D t +T t +R t Where yt represents the time series, Dt represents the diurnal term, Tt represents the trend term, and Rt represents the residual term. The diurnal term represents the short-period diurnal characteristics of the ionosphere, indicating the major changes that occur within a day due to the rotation of the earth. The trend term represents the long-term changes affected by factors such as solar activity. The existence of the residual term will hinder the accurate identification of the spatiotemporal attributes of the original data.

4. According to the method of claim 1, the method for predicting the shortwave maximum available frequency by federated learning for spatiotemporal coupling is characterized in that: Step C specifically includes the following steps: Step C1, mutual information experiments are conducted between different sites and the global model. The input data is the original MUF data. The data set is divided into three parts: training set, validation set and test set. The original data is divided into the following proportions: 70% for training set, 20% for validation set and 10% for test set. FedAvg is used as the benchmark federated learning method. The fusion strategy includes merging and averaging all local models. The LSTM model is selected as the local model of each site. Global training rounds and local training rounds are performed 30 times each. Step C2: Use the MINE method to observe the change of mutual information between the global model and each local model during the federated learning training process; where the mutual information is defined as follows: in, is the joint probability distribution, and is the marginal probability distribution.

5. According to claim 4, a method for predicting the shortwave maximum available frequency by federated learning for spatiotemporal coupling is characterized in that: The MINE method effectively estimates the lower bound of mutual information by evaluating the expected value of the samples; the mutual information is expressed as the KL divergence between the joint probability distribution and the marginal probability distribution: Among them, X and Z are independent random variables, D KL It is expressed as: Among them, the distribution Relative to the distribution It is absolutely continuous; In order to solve the problem that joint probability distribution and marginal probability distribution are difficult to obtain in actual data, the dual representation of KL divergence is used: If all functions T have finite upper limits, then both expected values ​​are also finite; when a family of functions is specified for T When , we can get the lower bound of mutual information; therefore, the MINE method also mentions the f-divergence representation of KL divergence: The MINE method uses the fitting ability of deep neural networks to define functions In addition, we obtain the bounds of the mutual information: I(X;Z)≥I Θ (X,Z) Among them, i Θ (X, Z) refers to estimating the mutual information between random variables by constructing a deep neural network; therefore, i Θ (X,Z) is specifically expressed as: When training deep neural networks, the actual data distribution cannot be obtained and can only be estimated through data sampling; Θ (X,Z) is further expressed as: Among them, for the distribution is the distribution of n independent and identically distributed samples; A two-layer fully connected neural network model is constructed to estimate mutual information, and the activation function used is ReLu.

6. The method for predicting the shortwave highest available frequency by federated learning for spatiotemporal coupling according to claim 1, characterized in that: In step D, the spatial relationship between the local models of federated learning is analyzed: the trend term of the MUF data in step B is used as the input of each local model, and the magnetic inclination data is considered; the mutual information between the stations of the MUF long-term trend component during the federated learning training process is calculated, and the geographical location and magnetic latitude of the station where each model is located are also taken into account, and the mutual information and station distribution of each model are observed.

7. The method for predicting the shortwave highest available frequency by federated learning for spatiotemporal coupling according to claim 1, characterized in that: In step E, a personalized fusion strategy is selected to complete the MUF task in federated learning, and the k-NearestNeighbor fusion strategy is adopted, that is, the nearest neighbor model with higher mutual information with the target model is selected from all models; in the above strategy, the selection of neighbors is based on the mutual information of MINE as the correlation measure, and the selection of hyperparameters is the empirical value selected through the experimental process. This personalized fusion strategy is selected to complete the MUF task in federated learning.

Images