High-latitude ionospheric phase flicker prediction method based on segmented recurrent neural network

Abstract

The invention provides a high-latitude ionosphere phase flicker prediction method based on a segmented recurrent neural network, and belongs to the technical field of aviation navigation. According to the method, four kinds of heterogeneous data of a super dual aurora radar, an altimeter, a geomagnetic solar activity index and ionospheric scintillation are integrated, and a minute-level fusion data set is constructed through space-time registration; complementing missing values by using a self-attention neural network and keeping data distribution characteristics; a segmented recurrent neural network model is innovatively designed, cross-day scale long-range dependence and minute-level sudden flicker evolution characteristics are synchronously captured through a double-path state transfer architecture, and continuous high-precision prediction of flicker phases is achieved. A time decay weighted evaluation mechanism is adopted to carry out non-linear penalty on the deviation of the starting moment and the ending moment of flicker, and the robustness of the model in complex space weather is verified. The support efficiency of the dynamic suppression strategy of the navigation receiver is remarkably improved, and the positioning stability in the polar region ionosphere environment is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of navigation technology, specifically relating to a high-latitude ionospheric phase scintillation prediction method based on a segmented recurrent neural network, which is used to improve the positioning reliability of Global Navigation Satellite System (GNSS) in complex ionospheric environments. Background Technology

[0002] Ionospheric phase scintillation is a rapid fluctuation phenomenon in radio signals caused by irregularities in the electron density of the Earth's ionosphere, primarily occurring in the equatorial and high-latitude regions. When GNSS signals traverse such irregular structures, it can lead to signal fading, phase jitter, and even receiver lock-off, severely impairing navigation accuracy and reliability. Ionospheric scintillation in high-latitude polar regions is particularly complex, its dynamic evolution driven by multiple factors including solar wind energy injection, particle deposition, and geomagnetic activity.

[0003] Current prediction methods can be divided into three categories: empirical models (such as HAPEE) rely on the statistical relationship between solar wind parameters and scintillation index, and can only provide probabilistic predictions; physical models (such as WBMOD and GISM) are based on ionospheric physical mechanisms, but due to insufficient accuracy and timeliness of input parameters, they are difficult to capture sudden scintillation events; data-driven models (such as LSTM and random forest) can achieve hourly predictions, but are limited by defects such as long sequence gradient vanishing, memory explosion and weak ability to capture minute-level mutations.

[0004] Existing technologies suffer from three major bottlenecks: 1) Traditional interpolation methods (linear / spline) disrupt the data distribution pattern, leading to deviations in physical laws; 2) LSTM / GRU cannot model long-range dependencies across daily scales due to gradient vanishing, and also consumes a lot of GPU memory; 3) There is a lack of a multi-dimensional evaluation mechanism for prediction bias at the start and end times of scintillation. A solution that can simultaneously achieve long-sequence modeling, minute-level resolution prediction, and efficient evaluation is urgently needed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a high-latitude ionospheric phase scintillation prediction method, which solves the problems of poor long sequence fitting ability, weak minute-level mutation capture, and large memory usage, thereby improving the stability of navigation systems in complex ionospheric environments.

[0006] To solve the above-mentioned technical problems, the technical solution proposed in this application is as follows: This invention provides a method for predicting phase scintillation in high-latitude ionospheric regions based on a segmented recurrent neural network, comprising the following steps: (1) Spatiotemporal alignment of data from super dual auroral radar, altimeter, geomagnetic solar activity index and ionospheric phase scintillation was performed to unify the resolution to 1 minute; (2) Calculate the dynamic window size of the radar vector field data to suppress noise; (3) Use the SAITS neural network based on the self-attention mechanism to complete the missing data in the fusion dataset; (4) Divide the dataset into training set, validation set and test set in a ratio of 7:1:2; (5) Construct a segmented recurrent neural network model, input historical L minutes of data, and output the future flicker phase sequence; (6) Use time-decayed weighted root mean square error Evaluate model performance.

[0007] Furthermore, in step (1): The first valid record is extracted every minute from the super dual-polarization radar data; Altimeter data is extended to a 1-minute resolution by copying; The geomagnetic solar activity index and ionospheric phase scintillation data retain the original time series.

[0008] Furthermore, the radar vector field data mentioned in step (2) includes power p_l, matching velocity v, fitted spectral width w_l, and elevation angle elv.

[0009] Furthermore, in step (3), the SAITS neural network learns the data distribution through the Transformer encoder and uses a masking mechanism to generate missing value imputation results.

[0010] Furthermore, the segmented recurrent neural network model described in step (5) includes an encoding layer and a decoding layer: The coding layer includes: Piecewise projection layer, input sequence Segmented According to the formula Projection compressed to ; in, The original input sequence has a length of L. The input matrix after segmentation, Let n be the number of segments and c be the number of features. These are the common weight matrix and the common bias vector, respectively. This represents the time step, and d is the projected feature dimension. For a data domain space of dimension L × c, It is a data domain space with dimensions n × w × c; Recursive coding layer: Contains n GRU units, outputs coding vectors hn ; The decoding layer includes: Location embedding layer: generates location codes ; Parallel decoding layer: This is the initial state. The input is a GRU unit, which outputs the predicted sequence in parallel.

[0011] Furthermore, the GRU unit of the recursive coding layer is operated according to the following formula: Reset Door: ; Update Gate: ; Candidate state: ; Final state: ; Where t is the time step, For the state of the previous moment Input at this time The concatenated vector For each weight matrix, To reset the gate vector, To update the gate vector, It is the sigmoid activation function. Let be the candidate hidden state vector. Let be the current hidden state vector. It is the hyperbolic tangent activation function.

[0012] Furthermore, the position encoding ,in, Encoding for relative positions; Encodes the channel.

[0013] in, It is the first The position code corresponding to the time step For the first The channel encoding corresponding to the time step. For the first The relative position encoding corresponding to the time step.

[0014] Furthermore, ensure Predict the number of subsequences Where H is the total prediction length; for parallel decoding layers, As the initial state and the previous state of each unit, The input for each time step is passed in parallel through the GRU gated recurrent unit corresponding to the encoding stage to obtain the output for the first m time steps. The final predicted sequence is obtained by using a Dropout layer to prevent overfitting and a fully connected layer to restore dimensionality. .

[0015] Furthermore, in step (6) The calculation formula is:

[0016] in, Here, Δt is the prediction time bias, τ is the decay constant (τ < 120 s), i is the time step, and N is the total number of samples. For the first i Predicted values ​​at each time point For the first i The actual value at each point in time.

[0017] On the other hand, this application claims protection for an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the program to implement the steps of any of the aforementioned methods.

[0018] Compared with the prior art, the present invention achieves the following beneficial technical effects: This invention constructs a high-fidelity ionospheric feature sequence through multi-source heterogeneous data fusion and data completion using a self-attention mechanism. It innovatively employs a segmented recurrent neural network architecture, overcoming the gradient vanishing and memory bottlenecks of traditional models in ultra-long time-series modeling, simultaneously capturing cross-diurnal scale evolution patterns and extracting high-frequency features of minute-level sudden scintillation events. Combined with a time decay evaluation mechanism, it performs reinforcement learning on the deviation between scintillation start and end times, significantly improving the physical consistency and spatiotemporal generalization ability of the prediction results. This method effectively solves the shortcomings of existing technologies, such as weak long-range dependency modeling, delayed response to sudden fluctuations, and high hardware resource consumption. It provides a high-precision, low-latency phase scintillation dynamic suppression strategy for navigation receivers in high-latitude regions, greatly enhancing the continuity and reliability of GNSS positioning under complex space weather conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 The flowchart illustrates a high-latitude ionospheric phase scintillation prediction method based on a segmented recurrent neural network, as provided in this embodiment of the invention.

[0021] Figure 2 This is a detailed flowchart of a segmented recurrent neural network according to an embodiment of this application.

[0022] Figure 3This is a structural block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] To make the technical solution, innovations, and implementation effects of this invention clearer, the high-latitude ionospheric phase scintillation prediction method based on a piecewise recurrent neural network will be described in detail below, combining specific application scenarios and model details. For example... Figure 1 As shown, it includes the following steps: (1) Spatiotemporal alignment of data from super dual auroral radar, altimeter, geomagnetic solar activity index and ionospheric phase scintillation was performed to unify the resolution to 1 minute; First, multi-source observation data, including super dual auroral radar, altimeter, geomagnetic solar activity index, and ionospheric phase scintillation, are integrated to establish a feature matrix based on a unified time reference. Smooth interpolation is used to upsample low-frequency data and downsample high-frequency data to ensure an overall 1-minute resolution of the data.

[0025] In this application, optionally, in step (1), the first valid record per minute is extracted from the super dual auroral radar data (high resolution), the altimeter data (15-minute resolution) is extended to 1-minute resolution by copying, and the original time series of the geomagnetic-solar activity index and ionospheric phase scintillation data (both 1-minute resolution) are preserved to ensure that all data are unified to a 1-minute reference.

[0026] (2) Calculate the dynamic window size of the radar vector field data to suppress noise; In this application, in step (2), for the power p_l, the matching velocity v, the fitted spectral width w_l and the elevation angle elv vector field characteristic data in the radar data, the middle value of the dynamic window is selected as the data at the current time.

[0027] (3) Use the SAITS neural network based on the self-attention mechanism to complete the missing data in the fusion dataset; In this application, in step (3), the SAITS (Self-Attention-based Imputation for TimeSeries) neural network learns the data distribution of the electronic altimeter through a self-attention mechanism to generate the missing value imputation result. (4) Divide the dataset into training set, validation set and test set in a ratio of 7:1:2; (5) Construct a segmented recurrent neural network model, input historical L minutes of data, and output the future flicker phase sequence; In this application, the prediction model in step (5) includes two parts: an encoding layer and a decoding layer.

[0028] The encoding layer contains, Segmented projection layer: The input sequence is segmented and activated using a common weight matrix and ReLU function; Recursive coding layer: processes the projected subsequence; The decoding layer contains, Location embedding layer: identifies the timeline position and feature index of the output subsequence; Parallel decoding layer: Outputs all prediction results in parallel.

[0029] The process involves segmenting the original time series, performing dimensionality transformation, and activation. The resulting fused dataset serves as the input sequence for the segmented recurrent neural network model, and is first segmented.

[0030] The original input sequence has a length of L. The input matrix after segmentation, Let n be the number of segments and c be the number of features for each segment. Then, projection is performed, and the calculation formula is as follows:

[0031] These are the common weight matrix and the common bias, respectively. This represents the time step, and d is the projected feature dimension. For a data domain space of dimension L × c, The data domain space has dimensions n × w × c; the recursive coding layer contains n GRU gated recurrent units, and each unit module and its calculation formula are as follows: Reset Door:

[0032] Update Gate:

[0033] Candidate state:

[0034] Final state:

[0035] Where t is the time step, For the state of the previous moment Input at this time The concatenated vector For each weight matrix, To reset the gate vector, To update the gate vector, It is the sigmoid activation function. Let be the candidate hidden state vector. Let be the current hidden state vector. This is the hyperbolic tangent activation function. The output of this layer... Encode the information of the entire input sequence. This also corresponds to the final output of the encoding layer; for the decoding layer, the calculation formula for the position embedding layer is:

[0036] in, It is the first The position code corresponding to the time step includes the relative position code.

[0037] and channel encoding,

[0038] ensure Additionally, predict the number of subsequences. Where H is the total prediction length; for parallel decoding layers, As the initial state and the previous state of each unit, The input for each time step is passed in parallel through the GRU gated recurrent unit corresponding to the encoding stage to obtain the output for the first m time steps. The final predicted sequence is obtained by using a Dropout layer to prevent overfitting and a fully connected layer to restore dimensionality. .

[0039] (6) Use time-decay weighted root mean square error Evaluate model performance.

[0040] In this application, the model evaluation in step (6) uses time-decayed weighted root mean square error to evaluate the model, characterized in that the formula for calculating the weighted root mean square error is:

[0041] in, Here, Δt is the prediction time bias, τ is the decay constant (τ < 120 s), i is the time step, and N is the total number of samples. For the first i Predicted values ​​at each time point For the first i The actual value at each point in time.

[0042] The following example illustrates a method for predicting the phase scintillation of the ionosphere in high latitudes based on a piecewise recurrent neural network. This method is used to predict the future intensity of the ionospheric phase scintillation in high-latitude regions. The specific steps are as follows: 1. Construct a fused dataset For high-resolution super dual-aurolas radar data, the first valid record per minute is extracted to match a 1-minute timeframe. For altimeter data with a 15-minute resolution, a replication method is used to extend it to a 1-minute resolution. Geomagnetic solar activity index and ionospheric phase scintillation data (both with a 1-minute resolution) are directly preserved in their original time series. For the vector field characteristic data of the super dual-aurolas radar (including power p_l, coordination velocity v, spectral width w_l, and elevation angle elv), a dynamic window midpoint is used to calculate the current time-time data to improve feature robustness and suppress outlier interference. Through the above methods, all data are unified to a 1-minute timeframe while preserving key physical features, providing high-quality input for subsequent modeling.

[0043] 2. Use a self-attention-based time-series missing value interpolation neural network to complete the missing data of the electronic altimeter. A self-attention-based time-series missing value imputation neural network is used to impute missing values ​​in the raw data collected by an electronic altimeter. This network learns the temporal distribution characteristics of the data through a self-attention mechanism, generating imputed results consistent with the actual data distribution while preserving the original distribution characteristics. Specifically, the network captures the temporal dependencies of the data through a multi-layer Transformer encoder, combines this with a masking mechanism to learn the missing locations, and finally outputs a complete dataset.

[0044] 3. Divide the data into training, validation, and test sets in a 7:1:2 ratio. The preprocessed dataset is divided into training, validation, and test sets in a 7:1:2 ratio. The training set is used for learning model parameters, the validation set is used for tuning hyperparameters (such as learning rate and network depth), and the test set is used to evaluate the final performance of the model and ensure its generalization ability.

[0045] 4. Construct a prediction model based on a segmented recurrent neural network, input historical L minutes of data, and output the future flicker phase sequence.

[0046] like Figure 2 As shown, a prediction model is constructed that includes a segmented projection layer, a recursive coding layer, a positional embedding layer, and a parallel decoding layer: Segmented Projection Layer: This layer segments and projects the input L-minute historical data sequence. The core calculation formula is as follows:

[0047] The original input sequence has a length of L. The input matrix after segmentation, Let n be the number of segments and c be the number of features for each segment. Then, projection is performed, and the calculation formula is as follows:

[0048] These are the common weight matrix and the common bias, respectively. This represents the time step, and d is the projected feature dimension. For a data domain space of dimension L × c, It is a data domain space with dimensions n × w × c.

[0049] Recursive coding layer: This layer achieves efficient compression and feature extraction of long sequence information through segmented GRU. The core calculation formula for each unit is as follows: Reset Door:

[0050] Update Gate:

[0051] Candidate state:

[0052] Final state:

[0053] Where t is the time step, For the state of the previous moment Input at this time The concatenated vector For each weight matrix, To reset the gate vector, To update the gate vector, It is the sigmoid activation function. Let be the candidate hidden state vector. Let be the current hidden state vector. This is the hyperbolic tangent activation function. The output of this layer... Encode the information of the entire input sequence. This also corresponds to the final output of the encoding layer.

[0054] Location embedding layer: Resolves timing confusion and multivariate interference issues; the core calculation formula is: Location coding

[0055] in, It is the first The position encoding corresponding to the time step, where

[0056]

[0057] ensure Additionally, predict the number of subsequences. , where H is the predicted total length.

[0058] Parallel decoding layer: This layer efficiently transforms the hidden state vector into a temporal prediction output in parallel. Its core features include the use of shared GRU gating units and the same previous state. .

[0059] 5. Evaluation Model RMSE The formula for calculating the weighted root mean square error is:

[0060] in, Here, Δt is the prediction time bias, τ is the decay constant (τ < 120 s), i is the time step, and N is the total number of samples. For the first i Predicted values ​​at each time point For the first i The actual value at each point in time.

[0061] Figure 3 This is a structural block diagram of an electronic device according to an embodiment of this application, such as... Figure 3 As shown, the electronic device may include: one or more (only one is shown in the figure) processors 401, memory 403, and transmission devices 405, such as... Figure 3 As shown, the electronic device may also include an input / output device 407.

[0062] The memory 403 can be used to store software programs and modules, such as the program instructions / modules corresponding to the high-latitude ionospheric phase scintillation prediction method based on a segmented recurrent neural network in this embodiment. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 403, thereby realizing the aforementioned high-latitude ionospheric phase scintillation prediction method based on a segmented recurrent neural network. The memory 403 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 403 may further include memory remotely located relative to the processor 401, and these remote memories can be connected to electronic devices via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0063] The aforementioned transmission device 405 is used to receive or send data via a network, and can also be used for data transfer between the processor and memory. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 405 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 405 is a radio frequency (RF) module used for wireless communication with the Internet.

[0064] Specifically, memory 403 is used to store application programs.

[0065] The processor 401 can invoke the application program stored in the memory 403 through the transmission device 405 to execute the steps of the aforementioned method.

[0066] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments, and will not be repeated here.

[0067] Those skilled in the art will understand that Figure 3 The structure shown is for illustrative purposes only. Electronic devices can be smartphones (such as Android phones, iOS phones, etc.), tablets, PDAs, mobile internet devices (MIDs), PADs, and other electronic devices. Figure 3 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 3 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 3 The different configurations shown.

[0068] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware of an electronic device. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0069] Embodiments of this application also provide a storage medium. Optionally, in this embodiment, the storage medium can be used to execute program code for a high-latitude ionospheric phase scintillation prediction method based on a segmented recurrent neural network.

[0070] Optionally, in this embodiment, the storage medium may be located on at least one of the network devices in the network shown in the above embodiment.

[0071] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting phase scintillation in high-latitude ionospheric regions based on a piecewise recurrent neural network, characterized in that, Includes the following steps: (1) Spatiotemporal alignment of data from super dual auroral radar, altimeter, geomagnetic solar activity index and ionospheric phase scintillation was performed to unify the resolution to 1 minute; (2) Calculate the dynamic window size of the radar vector field data to suppress noise; (3) Use the SAITS neural network based on the self-attention mechanism to complete the missing data in the fusion dataset; (4) Divide the dataset into training set, validation set and test set in a 7:1:2 ratio; (5) Construct a segmented recurrent neural network model, input historical L minutes of data, and output the future flicker phase sequence; (6) Use time-decayed weighted root mean square error Evaluate model performance.

2. The method according to claim 1, characterized in that, In step (1): The first valid record is extracted every minute from the super dual-polarization radar data; Altimeter data is extended to a 1-minute resolution by copying; The geomagnetic solar activity index and ionospheric phase scintillation data retain the original time series.

3. The method according to claim 1, characterized in that, The radar vector field data mentioned in step (2) includes power p_l, matching velocity v, fitted spectral width w_l, and elevation angle elv.

4. The method according to claim 1, characterized in that, In step (3), the SAITS neural network learns the data distribution through the Transformer encoder and uses a masking mechanism to generate missing value imputation results.

5. The method according to claim 1, characterized in that, The segmented recurrent neural network model described in step (5) includes an encoding layer and a decoding layer: The coding layer includes: Piecewise projection layer, input sequence Segmented According to the formula Projection compressed to ; in, The original input sequence has a length of L. The input matrix after segmentation, Let n be the number of segments and c be the number of features. These are the common weight matrix and the common bias vector, respectively. This represents the time step, and d is the projected feature dimension. For a data domain space of dimension L × c, It is a data domain space with dimensions n × w × c; Recursive coding layer: Contains n GRU units, outputs coding vectors hn ; The decoding layer includes: Location embedding layer: generates location codes ; Parallel decoding layer: This is the initial state. The input is a GRU unit, which outputs the predicted sequence in parallel.

6. The method according to claim 5, characterized in that, The GRU unit of the recursive coding layer is calculated according to the following formula: Reset Door: ; Update Gate: ; Candidate state: ; Final state: ; Where t is the time step, For the state of the previous moment Input at this time The concatenated vector For each weight matrix, To reset the gate vector, To update the gate vector, It is the sigmoid activation function. Let be the candidate hidden state vector. This is the current hidden state vector. It is the hyperbolic tangent activation function.

7. The method according to claim 5, characterized in that, The location code ,in, Encoding for relative positions; Encoding for the channel. in, It is the first The position code corresponding to the time step For the first The channel encoding corresponding to the time step. For the first The relative position encoding corresponding to the time step.

8. The method according to claim 5, characterized in that, ensure Predict the number of subsequences Where H is the total prediction length; for parallel decoding layers, As the initial state and the previous state of each unit, The input for each time step is passed in parallel through the GRU gated recurrent unit corresponding to the encoding stage to obtain the output for the first m time steps. The final predicted sequence is obtained by using a Dropout layer to prevent overfitting and a fully connected layer to restore dimensionality. .

9. The method according to claim 1, characterized in that, In step (6) The calculation formula is: in, Here, Δt is the prediction time bias, τ is the decay constant (τ < 120 s), i is the time step, and N is the total number of samples. For the first i Predicted values ​​at each time point For the first i The actual value at each point in time.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1-9.

Images