Ionospheric d-layer parameter gridding inversion method and device based on multi-modal deep learning

Abstract

The scheme provides an ionospheric D layer parameter gridding inversion method and device based on a multi-modal deep learning, comprising: deploying a pre-trained ionospheric D layer parameter inversion model, and inputting multiple VLF sequence data of a target inversion region to the ionospheric D layer parameter inversion model to obtain a final ionospheric D layer parameter distribution of the target inversion region, wherein the ionospheric D layer parameter inversion model comprises a feature processing branch, a local correction branch, a global correction branch and a fine iteration branch. The scheme projects sparse irregular path point features onto a grid through bilinear interpolation, and performs multi-head cross attention calculation from the grid to the path point to improve the inversion accuracy of the area far away from the observation path.

Description

Technical Field

[0001] This application relates to the field of computational electromagnetics, and in particular to a method and apparatus for gridded inversion of ionospheric D-layer parameters based on multimodal deep learning. Background Technology

[0002] The D layer of the ionosphere is the region with the lowest degree of ionization in the Earth's atmosphere and plays a decisive role in the propagation of very low frequency (VLF) radio waves. Its two-dimensional spatial distribution of reference altitude h' and steepness β are key parameters for analyzing the impact of space weather events such as solar flares and geomagnetic storms, and for ensuring the stable operation of space-based navigation and communication systems. Currently, the industry commonly uses ensemble Kalman filtering (such as LETKF) data assimilation methods for ionospheric inversion. This method acquires multipath signal amplitude and phase observations through a VLF transmitter-receiver network, grids the ionospheric parameters, and, based on the assumptions of linear observation operators and Gaussian error distribution, assimilates the observation data to update the grid parameter estimates, thus achieving large-area ionospheric state inversion.

[0003] Traditional LETKF assimilation methods suffer from insurmountable technical drawbacks: the relationship between ionospheric parameters and VLF observations is highly nonlinear; observation noise and environmental disturbances often exhibit non-Gaussian distributions, and the linear Gaussian assumption can lead to significant inversion biases, resulting in a substantial decrease in accuracy under complex ionospheric conditions; the VLF receiver network has limited coverage, and observational information is spatially highly sparse and irregularly distributed, making it difficult to optimally configure the LETKF error covariance localization parameter, which can easily lead to insufficient information diffusion or spurious correlations, causing a sharp drop in inversion accuracy in areas far from the observation path; at the same time, the iterative calculations and model integration of this method are time-consuming, failing to meet the high-efficiency requirements of real-time space weather monitoring. Therefore, there is an urgent need for a new modeling method that can fundamentally address nonlinear and non-Gaussian problems, effectively utilize sparse and irregular observations, and achieve rapid and accurate inversion. Summary of the Invention

[0004] This application provides a method and apparatus for gridded inversion of ionospheric D-layer parameters based on multimodal deep learning. It projects sparse and irregular path point features onto a grid through bilinear interpolation and performs multi-head cross-attention calculation from the grid to the path points to improve the inversion accuracy of regions far from the observation path.

[0005] In a first aspect, embodiments of this application provide a gridded inversion method for ionospheric D-layer parameters based on multimodal deep learning, the method comprising:

[0006] Deploy a pre-trained ionospheric D-layer parameter inversion model and input multiple VLF sequence data of the target inversion region into the ionospheric D-layer parameter inversion model. The ionospheric D-layer parameter inversion model includes a feature processing branch, a local correction branch, a global correction branch, and a fine iteration branch. The feature processing branch encodes each VLF sequence data and the corresponding path static information to obtain the corresponding path-level features and path-point-level features. The local correction branch includes a bilinear interpolation unit and a feature enhancement unit. The bilinear interpolation unit interpolates the features at each path point level into the target two-dimensional grid to obtain an interpolated observation feature map. The target two-dimensional grid is the two-dimensional grid of the target inversion region. The feature enhancement unit performs multi-head cross-attention calculation on the features at each grid point in the interpolated observation feature map and the features at each path point level to obtain a context feature map. The global correction branch generates a global correction field based on each path-level feature, and uses the global correction field to correct the prior ionospheric D-layer parameter distribution to obtain the inverted ionospheric D-layer parameter initial distribution. The fine iterative branch uses a multimodal tensor as input to iteratively correct the initial distribution of the inverted ionospheric D-layer parameters to obtain the final ionospheric D-layer parameter distribution of the target inversion region. The multimodal tensor includes a context feature map, the position information of the target two-dimensional grid, and an interpolated observation feature map.

[0007] Secondly, embodiments of this application provide a gridded inversion device for ionospheric D-layer parameters based on multimodal deep learning, comprising: The deployment module is used to deploy a pre-trained ionospheric D-layer parameter inversion model and input multiple VLF sequence data of the target inversion region into the ionospheric D-layer parameter inversion model. The ionospheric D-layer parameter inversion model includes a feature processing branch, a local correction branch, a global correction branch, and a fine iteration branch. The feature extraction module uses a feature processing branch to encode each VLF sequence data and the corresponding path static information to obtain the corresponding path-level features and path-point-level features. The local correction module includes a local correction branch, which includes a bilinear interpolation unit and a feature enhancement unit. The bilinear interpolation unit interpolates the features at each path point level into the target two-dimensional grid to obtain an interpolated observation feature map. The target two-dimensional grid is the two-dimensional grid of the target inversion region. The feature enhancement unit performs multi-head cross-attention calculation on the features at each grid point in the interpolated observation feature map and the features at each path point level to obtain a context feature map. The global correction module uses a global correction branch to generate a global correction field based on the features of each path level, and uses the global correction field to correct the prior ionospheric D-layer parameter distribution to obtain the initial distribution of the inverted ionospheric D-layer parameters. The fine iteration module uses a fine iteration branch to iteratively correct the initial distribution of the inverted ionospheric D-layer parameters using a multimodal tensor as input to obtain the final ionospheric D-layer parameter distribution of the target inversion region. The multimodal tensor includes a context feature map, the position information of the target two-dimensional grid, and an interpolated observation feature map.

[0008] Thirdly, embodiments of this application provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute a gridded inversion method for ionospheric D-layer parameters based on multimodal deep learning.

[0009] The main contributions and innovations of this invention are as follows: This application's embodiments accurately fit the nonlinear, non-Gaussian mapping relationship between ionospheric D-layer parameters and VLF observations to construct the training dataset, eliminating inversion bias in complex ionospheric environments at its source and improving the accuracy of model predictions. The local correction branch of this scheme uses a combination of bilinear interpolation, deformable convolution, and multi-head cross-attention to project sparse, irregular path point features onto a two-dimensional grid and model global dependencies, solving the problem of sparse spatial distribution and uneven coverage of VLF observations, and effectively diffusing observation information to uncovered areas, significantly improving the inversion accuracy of grid points far from the observation path. The fine iterative branch of this scheme adopts a multi-scale iterative structure of coarse-scale downsampling iteration and fine-scale upsampling correction, combined with grid weight maps to gating constraints on the residuals, refining the parameter distribution layer by layer, preserving the high-frequency detailed structure of the ionosphere, suppressing invalid corrections in unobserved areas, and making the final inversion field more closely resemble the real ionospheric morphology.

[0010] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a gridded inversion method for ionospheric D-layer parameters based on multimodal deep learning, according to an embodiment of this application. Figure 2 This is a schematic diagram of a VLF transmission and reception path within a target inversion area according to an embodiment of this application; Figure 3 This is a structural block diagram of an ionospheric D-layer parameter gridding inversion device based on multimodal deep learning according to an embodiment of this application; Figure 4This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0012] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0013] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0014] Example 1 This application provides a gridded inversion method for ionospheric D-layer parameters based on multimodal deep learning. It projects sparse, irregular path point features onto a grid using bilinear interpolation and performs multi-head cross-attention calculation from the grid to the path points to improve inversion accuracy in regions far from the observation path. Specifically, refer to... Figure 1 The method includes: Deploy a pre-trained ionospheric D-layer parameter inversion model and input multiple VLF sequence data of the target inversion region into the ionospheric D-layer parameter inversion model. The ionospheric D-layer parameter inversion model includes a feature processing branch, a local correction branch, a global correction branch, and a fine iteration branch. The feature processing branch encodes each VLF sequence data and the corresponding path static information to obtain the corresponding path-level features and path-point-level features. The local correction branch includes a bilinear interpolation unit and a feature enhancement unit. The bilinear interpolation unit interpolates the features at each path point level into the target two-dimensional grid to obtain an interpolated observation feature map. The target two-dimensional grid is the two-dimensional grid of the target inversion region. The feature enhancement unit performs multi-head cross-attention calculation on the features at each grid point in the interpolated observation feature map and the features at each path point level to obtain a context feature map. The global correction branch generates a global correction field based on each path-level feature, and uses the global correction field to correct the prior ionospheric D-layer parameter distribution to obtain the inverted ionospheric D-layer parameter initial distribution. The fine iterative branch iteratively corrects the initial distribution of the inverted ionospheric D-layer parameters based on the context feature map and the target two-dimensional grid to obtain the final ionospheric D-layer parameter distribution of the target inversion region.

[0015] In the current embodiment, the target inversion region is divided into two-dimensional grids using a preset fine granularity to obtain the target two-dimensional grid.

[0016] Furthermore, the coordinates of each grid point in the target two-dimensional grid are normalized to obtain the latitude and longitude coordinates of each grid point.

[0017] In the current embodiment, the distribution of prior ionospheric D-layer parameters is obtained based on the temporal information and geographical location of the target inversion region.

[0018] Specifically, the prior ionospheric D-layer parameter distribution consists of the prior height and prior steepness at each grid point in the target two-dimensional grid. The prior ionospheric D-layer parameter distribution of the target inversion region is calculated based on the Ferguson empirical formula, expressed as:

[0019]

[0020] in, The prior height of the grid points, The prior steepness of the grid points. The zenith angle of the sun. Geographical latitude, For months.

[0021] In the current embodiment, each VLF transceiver path within the target inversion region is acquired, and VLF sequence data is obtained by observing multiple equidistant sampling points on the VLF transceiver path. The VLF sequence data includes the coordinates, coordinate difference, path distance, amplitude, phase, amplitude prior residual, phase prior residual, and prior ionospheric D-layer parameters of each equidistant sampling point.

[0022] Specifically, the VLF transceiver path is the path between multiple transceiver stations within the target inversion area. L equidistant sampling points are constructed between each group of transceiver stations, and prediction is performed at each equidistant sampling point to obtain VLF sequence data. A schematic diagram of the VLF transceiver path within the target inversion area is shown below. Figure 2 As shown.

[0023] Specifically, in VLF sequence data, the coordinate difference is the coordinate difference between the current equidistant sampling point and the previous adjacent equidistant sampling point, and the path distance is the ratio of the path length from the current equidistant sampling point to the corresponding launch station to the total path length. It is worth mentioning that the coordinates, coordinate differences, and path distance are normalized results to ensure the scale uniformity of parameters in different dimensions.

[0024] Specifically, the amplitude and phase are represented by the three components measured at the current equidistant sampling points. For example, the three-component amplitude is represented as follows: The three-component phase is represented as .

[0025] Specifically, the amplitude prior residual is the difference between the observed amplitude and the prior amplitude at the current equidistant sampling point, and the phase prior residual is the difference between the observed phase and the prior phase at the current equidistant sampling point. The prior amplitude and prior phase at each equidistant sampling point are simulated based on the prior ionospheric D-layer parameter distribution using models such as LWPC.

[0026] Specifically, the a priori ionospheric D-layer parameters for each equidistant sampling point are obtained by interpolating the distribution of the a priori ionospheric D-layer parameters.

[0027] Specifically, an input tensor is constructed based on multiple VLF sequence data. The input is used to retrieve the parameters of the D layer of the ionosphere, where P is the number of VLF transmit / receive paths within the target retrieval region, and L is the number of equidistant sampling points on each VLF transmit / receive path. It consists of multi-dimensional VLF sequence data.

[0028] In the current embodiment, the distribution of ionospheric D-layer parameters in any region is obtained as the first sample. Gaussian perturbation and local Blob perturbation are added to each training sample to obtain the second sample. The second sample is numerically simulated to generate VLF sequence data of the corresponding region as the third sample. Multiple third samples are obtained to form a training sample set, and an ionospheric D-layer parameter inversion architecture is constructed. The ionospheric D-layer parameter inversion architecture is trained with the training sample set to obtain a trained ionospheric D-layer parameter inversion model.

[0029] Specifically, the Ferguson empirical formula is used to obtain the ionospheric D-layer parameter distribution in any region as the first sample. To make the first sample more realistic, Gaussian perturbation and local Blob perturbation are added to the first sample.

[0030] Furthermore, a weighted summation of Gaussian random fields with different spatial correlation lengths is performed to obtain the highly perturbed field. Correlation coefficients were randomly constructed for height and steepness in the first sample. The steepness perturbation field is calculated based on the correlation coefficient and the height perturbation field. The formula is expressed as:

[0031] in, For steep perturbation fields, The correlation coefficient is... For highly disturbed fields, An independent unit perturbation field, used to avoid and same.

[0032] Next, the perturbation coefficients of height and steepness are randomly sampled, and combined with the height perturbation field and steepness perturbation field to obtain the height perturbation value and steepness perturbation value, expressed by the formula:

[0033] in, The height perturbation coefficient is a randomly sampled value. For highly disturbed fields, This represents the height disturbance value. This represents the steepness disturbance value. The steepness perturbation coefficient is a randomly sampled value. This represents a steeply oriented perturbation field.

[0034] For example, since space weather events such as solar flares, geomagnetic storms, and proton events can cause drastic changes in altitude and steepness, this scheme constructs Gaussian random fields at spatial correlation lengths of 1200 km, 600 km, and 300 km to simulate these changes and approximate the real environment. Correlation coefficients are constructed to ensure a negative correlation between altitude and steepness. The value is between -0.7 and -0.4.

[0035] Furthermore, to simulate the local anomalous structure in the D layer of the ionosphere, a random number of two-dimensional Gaussian local spots are superimposed within the target region, and the center, scale, and amplitude of each two-dimensional Gaussian local spot are random variables, as expressed by the formula:

[0036] in, coordinates The local disturbance value at point N is the total number of two-dimensional Gaussian local spots, and k is the index of the two-dimensional Gaussian local spots. Let the amplitude of the k-th two-dimensional Gaussian local spot be denoted as . , Let the coordinates be the center coordinates of the k-th two-dimensional Gaussian local spot. Let be the scale radius of the k-th two-dimensional Gaussian local spot, and exp be an exponential function.

[0037] Therefore, the formulas for the height and steepness of the Gaussian perturbation and the local Blob perturbation superimposed in the second sample can be expressed as follows:

[0038]

[0039] in, To superimpose the heights of Gaussian perturbations and local Blob perturbations, The height of the corresponding position in the first sample. For the corresponding height disturbance value, This corresponds to the local disturbance value; To incorporate the steepness of the superimposed Gaussian perturbation and the local Blob perturbation, This represents the steepness at the corresponding position in the first sample. This corresponds to the steepness disturbance value. This represents the corresponding local disturbance value.

[0040] Furthermore, using precise numerical models such as Long Wavelength Propagation Capability (LWPC) or Long Wavelength Mode Propagator (LMP), the coordinates of the second sample are analyzed. and The amplitude and phase of each equidistant sampling point on each VLF transmit / receive path are simulated as input to the model for training.

[0041] Specifically, the model is trained using a physics-inspired method to generate complex perturbations, enabling it to learn the real evolution patterns of the ionosphere. Even when faced with complex perturbation scenarios not seen in the training set (such as multi-scale stacking and local blobs), it maintains good inversion performance, demonstrating strong generalization ability and robustness.

[0042] In the current embodiment, the total loss function of the ionospheric D-layer parameter inversion model during training includes the weighted sum of the principal loss, correction loss, prior regularization loss, high-frequency detail loss, multi-stage supervision loss, and total variational loss.

[0043] Specifically, the main loss integrates the global loss and the neighboring path weighted loss, modulating the weights through observation coverage to prioritize accuracy near the observation path; the correction loss supervises the consistency of parameter residual predictions; the prior regularization loss makes the predicted values ​​in unobserved areas converge to the prior; the high-frequency detail loss includes Sobel gradient loss and Laplace loss, constraining the prediction field to retain fine structure; the multi-stage supervised loss supervises the output of fine iterative branches and intermediate iterative steps; and the total variational loss smooths the prediction field, suppressing noise and ensuring physical rationality.

[0044] Specifically, this scheme, through deep nonlinear mapping and a multi-physics constraint loss function, can accurately invert complex ionospheric structures from sparse observations. Experiments show that on datasets containing multi-scale perturbations, the model can reduce the MAE for height from the prior 2.08 km to 0.72 km, and the MAE for steepness from... Reduce to The inversion accuracy far exceeds that of traditional methods, and once the model is trained, it only takes milliseconds to complete the inversion of a sample (covering an area of ​​thousands of kilometers) under GPU acceleration, which is more than three orders of magnitude faster than the traditional LETKF method, truly realizing real-time monitoring.

[0045] In the current embodiment, the feature processing branch includes a path observation encoder and a linear modulation layer. The path observation encoder encodes each VLF sequence data to obtain global path features and local path point features. The linear modulation layer modulates the global path features and local path point features based on the corresponding path static information to obtain path-level features and path point-level features.

[0046] Specifically, the path observation encoder consists of a multi-layer Transformer encoder and a multi-scale 1D convolution, thereby performing deep encoding on the input VLF sequence data, and then using a linear modulation layer to inject the corresponding path static information into the global path features and local path point features for modulation.

[0047] Furthermore, the path static information includes the latitude and longitude of the transceiver station, its operating frequency, and its transmission power for the corresponding VLF transceiver path.

[0048] In other words, the path-level features obtained through the linear modulation layer characterize the global signal trend, geometric properties, and inherent link characteristics of the entire VLF link; the path-point-level features characterize the fused local features of VLF observations, spatial geometry, prior parameters, and path static properties of each equidistant sampling point on each path.

[0049] In the current embodiment, the bilinear interpolation unit includes an interpolation layer and a deformable observation mixer. The interpolation layer interpolates path point-level features into the target two-dimensional grid using bilinear interpolation to obtain an initial grid feature map, and generates a grid weight map based on the degree to which each grid point is covered by path point-level features. The initial grid feature map and the grid weight map are concatenated and then input into the deformable observation mixer. The deformable observation mixer uses multiple cascaded deformable convolutional layers to perform feature processing to obtain the interpolated observation feature map.

[0050] Specifically, the interpolation layer accurately projects the irregular, discrete path point-level features in the target inversion region onto the target two-dimensional grid, and generates a grid weight map based on the degree to which each grid point is covered by the path point-level features. That is, the higher the coverage of a grid point, the more VLF transmission and reception paths pass through this grid point, thus increasing the weight of this grid point. However, not all grid points in the target two-dimensional grid are covered by the path point-level features, and these uncovered points are also affected by other VLF signals. Therefore, the deformable observation mixer learns the sampling offset of each grid point based on the initial grid feature map and the grid weight map, so that the receptive field is adaptively stretched along the path direction, thereby more effectively spreading information from the path point to the surrounding area and enhancing the ability to fill the gaps in the observation coverage.

[0051] In the current embodiment, in the multi-head cross-attention calculation step of the feature enhancement unit, a query is constructed using the grid position information in the target two-dimensional grid, the corresponding prior ionospheric D-layer parameters, the corresponding path point-level features in the interpolated observation feature map, and the corresponding grid point weights in the grid weight map. A key value is constructed using the path point-level features output by the feature processing branch and the corresponding path point coordinates.

[0052] For example, the grid position information in the target two-dimensional grid, the corresponding path point-level features in the interpolated observation feature map, and the corresponding grid point weights in the grid weight map are concatenated and projected as the query Q. Similarly, the path point-level features output by the feature processing branch and the concatenated projection of the corresponding path point coordinates are used to construct the key value.

[0053] Specifically, the detailed multi-head attention calculation method is the same as that in existing technologies, and will not be elaborated upon in this solution.

[0054] Specifically, this scheme performs multi-head attention calculation on features from grid points to path points, enabling each grid point to directly focus on all path point-level features, thereby capturing global dependencies.

[0055] Furthermore, to improve computational efficiency, multi-head cross-attention calculation is performed on the downsampled target 2D grid, and then the result of the multi-head cross-attention calculation is upsampled and aligned with the target 2D grid of the original resolution size to obtain the context feature map.

[0056] Specifically, each grid point in the context feature map incorporates the global correlation between that location and all VLF path observations across the entire region, the overall trend of ionospheric perturbations across the entire region, the global physical constraints of unobserved areas, and the global distribution information of observation coverage.

[0057] In the current embodiment, each path-level feature is weighted and aggregated in the global correction branch to obtain a global context vector. A low-order polynomial is constructed to represent the distribution law of the ionospheric D-layer parameter deviation in the target two-dimensional grid. The global context vector is processed by MLP to obtain the low-order polynomial coefficients. The low-order polynomial coefficients are used to adjust the distribution law of the ionospheric D-layer parameter deviation in the target two-dimensional grid to obtain the global correction field.

[0058] Specifically, the global correction field includes a global height correction field and a global steepness correction field. The global height correction field is used to correct the a priori height distribution in the a priori ionospheric D-layer parameters, and the global steepness correction field is used to correct the a priori steepness distribution in the a priori ionospheric D-layer parameters. The corrected a priori height distribution and a priori steepness distribution are used as the initial distribution for inverting the ionospheric D-layer parameters.

[0059] Specifically, since the ionospheric parameters are low-frequency components, these low-frequency components exhibit large-scale and smooth physical continuity changes within the target inversion region. Therefore, a low-order polynomial is constructed to characterize the distribution law of the ionospheric D-layer parameter deviation. Then, the low-order polynomial coefficients are generated through MLP prediction to adjust the low-order polynomial, thereby completing the global low-frequency correction of the ionospheric D-layer parameters and correcting large-scale, systematic deviations.

[0060] In the current embodiment, the fine-scale iteration branch includes sequentially connected coarse-scale iteration units and fine-scale iteration units. In the coarse-scale iteration unit, the multimodal tensor at the original scale and the initial distribution of the inverted ionospheric D-layer parameters are downsampled to a first iteration scale. The coarse-scale iteration unit iterates based on the multimodal tensor at the first iteration scale and the initial distribution of the inverted ionospheric D-layer parameters to obtain a first residual. The first residual obtained in each iteration is superimposed with the initial distribution of the inverted ionospheric D-layer parameters at the first iteration scale to obtain a coarse-scale iteration result. In the fine-scale iteration unit, the coarse-scale iteration result is upsampled to the original scale. The fine-scale iteration unit iterates based on the multimodal tensor at the original scale and the initial distribution of the inverted ionospheric D-layer parameters to obtain a second residual. The second residual obtained in each iteration is superimposed with the coarse-scale iteration result at the original scale to obtain a fine-scale iteration result. After the iteration is completed, the fine-scale iteration result of the last iteration is output as the final ionospheric D-layer parameter distribution.

[0061] Specifically, the coarse-scale iterative unit and the fine-scale iterative unit adopt an encoder-decoder structure based on U-Net for iterative updates. Combined with the Transformer attention of the bottleneck layer in the U-Net encoder-decoder structure, multi-scale features are extracted and multi-scale residual maps are output to form the first residual or the second residual.

[0062] Specifically, the number of iterations for the coarse-scale iteration unit and the fine-scale iteration unit is preset. In this scheme, the number of iterations for the coarse-scale iteration unit is 1, and the number of iterations for the fine-scale iteration unit is 6.

[0063] Furthermore, the multimodal tensor also includes a grid weight map. When the multimodal tensor includes a grid weight map, the fine-scale iterative unit gates the second residual based on the grid weight map, and uses the gated result as the new second residual.

[0064] Specifically, gating the second residual with a grid weighted graph can prevent invalid corrections in unobserved areas, making the final ionospheric D-layer parameter distribution more accurate.

[0065] Specifically, the final distribution of ionospheric D-layer parameters is the height distribution and steepness distribution within the target inversion region.

[0066] Specifically, the deformable convolution, cross attention, and neighboring path weighted loss modules in the ionospheric D-layer parameter inversion model work together to enable the model to extract and diffuse information from irregular and sparse VLF path observations to the maximum extent, significantly improving the accuracy of the inversion field in regions far from the path and reducing artifacts.

[0067] Example 2 Based on the same concept, referencing Figure 3 This application also proposes a gridded inversion device for ionospheric D-layer parameters based on multimodal deep learning, comprising: The deployment module is used to deploy a pre-trained ionospheric D-layer parameter inversion model and input multiple VLF sequence data of the target inversion region into the ionospheric D-layer parameter inversion model. The ionospheric D-layer parameter inversion model includes a feature processing branch, a local correction branch, a global correction branch, and a fine iteration branch. The feature extraction module uses a feature processing branch to encode each VLF sequence data and the corresponding path static information to obtain the corresponding path-level features and path-point-level features. The local correction module includes a local correction branch, which includes a bilinear interpolation unit and a feature enhancement unit. The bilinear interpolation unit interpolates the features at each path point level into the target two-dimensional grid to obtain an interpolated observation feature map. The target two-dimensional grid is the two-dimensional grid of the target inversion region. The feature enhancement unit performs multi-head cross-attention calculation on the features at each grid point in the interpolated observation feature map and the features at each path point level to obtain a context feature map. The global correction module uses a global correction branch to generate a global correction field based on the features of each path level, and uses the global correction field to correct the prior ionospheric D-layer parameter distribution to obtain the initial distribution of the inverted ionospheric D-layer parameters. The fine iteration module uses a fine iteration branch to iteratively correct the initial distribution of the inverted ionospheric D-layer parameters using a multimodal tensor as input to obtain the final ionospheric D-layer parameter distribution of the target inversion region. The multimodal tensor includes a context feature map, the position information of the target two-dimensional grid, and an interpolated observation feature map.

[0068] Example 3 This embodiment also provides an electronic device, see reference. Figure 4 It includes a memory 404 and a processor 402, the memory 404 storing a computer program and the processor 402 being configured to run the computer program to perform the steps in any of the above method embodiments.

[0069] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0070] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0071] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0072] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the ionospheric D-layer parameter gridding inversion methods based on multimodal deep learning in the above embodiments.

[0073] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0074] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0075] The input / output device 408 is used to input or output information. In this embodiment, the input information may be multiple VLF sequence data of the target inversion region, and the output information may be the final ionospheric D-layer parameter distribution of the target inversion region.

[0076] Optionally, in this embodiment, the processor 402 can be configured to perform the following steps via a computer program: Deploy a pre-trained ionospheric D-layer parameter inversion model and input multiple VLF sequence data of the target inversion region into the ionospheric D-layer parameter inversion model. The ionospheric D-layer parameter inversion model includes a feature processing branch, a local correction branch, a global correction branch, and a fine iteration branch. The feature processing branch encodes each VLF sequence data and the corresponding path static information to obtain the corresponding path-level features and path-point-level features. The local correction branch includes a bilinear interpolation unit and a feature enhancement unit. The bilinear interpolation unit interpolates the features at each path point level into the target two-dimensional grid to obtain an interpolated observation feature map. The target two-dimensional grid is the two-dimensional grid of the target inversion region. The feature enhancement unit performs multi-head cross-attention calculation on the features at each grid point in the interpolated observation feature map and the features at each path point level to obtain a context feature map. The global correction branch generates a global correction field based on each path-level feature, and uses the global correction field to correct the prior ionospheric D-layer parameter distribution to obtain the inverted ionospheric D-layer parameter initial distribution. The fine iterative branch uses a multimodal tensor as input to iteratively correct the initial distribution of the inverted ionospheric D-layer parameters to obtain the final ionospheric D-layer parameter distribution of the target inversion region. The multimodal tensor includes a context feature map, the position information of the target two-dimensional grid, and an interpolated observation feature map.

[0077] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0078] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0079] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 4 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0080] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0081] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A gridded inversion method for ionospheric D-layer parameters based on multimodal deep learning, characterized in that, Includes the following steps: Deploy a pre-trained ionospheric D-layer parameter inversion model and input multiple VLF sequence data of the target inversion region into the ionospheric D-layer parameter inversion model. The ionospheric D-layer parameter inversion model includes a feature processing branch, a local correction branch, a global correction branch, and a fine iteration branch. The feature processing branch encodes each VLF sequence data and the corresponding path static information to obtain the corresponding path-level features and path-point-level features. The local correction branch includes a bilinear interpolation unit and a feature enhancement unit. The bilinear interpolation unit interpolates the path point-level features into the target two-dimensional grid to obtain an interpolated observation feature map. The bilinear interpolation unit includes an interpolation layer and a deformable observation mixer. The interpolation layer interpolates the path point-level features into the target two-dimensional grid using bilinear interpolation to obtain an initial grid feature map, and generates a grid weight map based on the degree to which each grid point in the initial grid feature map is covered by the path point-level features. The initial grid feature map and the grid weight map are concatenated and input into the deformable observation mixer. The deformable observation mixer uses multiple cascaded deformable convolutional layers for feature processing to obtain the interpolated observation feature map. The target two-dimensional grid is the two-dimensional grid of the target inversion region. The feature enhancement unit performs multi-head cross-attention calculation on the features of each grid point in the interpolated observation feature map and each path point-level feature to obtain a context feature map. The global correction branch generates a global correction field based on each path-level feature, and uses the global correction field to correct the prior ionospheric D-layer parameter distribution to obtain the initial distribution of inverted ionospheric D-layer parameters. The fine-scale iterative branch uses a multimodal tensor as input to iteratively correct the initial distribution of the inverted ionospheric D-layer parameters to obtain the final ionospheric D-layer parameter distribution of the target inversion region. The fine-scale iterative branch includes sequentially cascaded coarse-scale iterative units and fine-scale iterative units. In the coarse-scale iterative unit, the original-scale multimodal tensor and the initial distribution of the inverted ionospheric D-layer parameters are downsampled to a first iteration scale. The coarse-scale iterative unit iterates based on the multimodal tensor and the initial distribution of the inverted ionospheric D-layer parameters at the first iteration scale to obtain a first residual. The first residual obtained in each iteration is then compared with the inversion scale of the first iteration. The initial distribution of ionospheric D-layer parameters is superimposed to obtain a coarse-scale iterative result. In the fine-scale iterative unit, the coarse-scale iterative result is upsampled to the original scale. The fine-scale iterative unit iterates based on the multimodal tensor of the original scale and the initial distribution of inverted ionospheric D-layer parameters to obtain a second residual. The second residual obtained in each iteration is superimposed with the coarse-scale iterative result of the original scale to obtain a fine-scale iterative result. After the iteration is completed, the fine-scale iterative result of the last iteration is output as the final ionospheric D-layer parameter distribution. The multimodal tensor includes a context feature map, the position information of the target two-dimensional grid, and the interpolated observation feature map.

2. The method for gridded inversion of ionospheric D-layer parameters based on multimodal deep learning according to claim 1, characterized in that, The distribution of ionospheric D-layer parameters in any region is obtained as the first sample. Gaussian perturbation and local Blob perturbation are added to each training sample to obtain the second sample. The second sample is used to generate VLF sequence data of the corresponding region through numerical simulation as the third sample. Multiple third samples are obtained to form a training sample set. The ionospheric D-layer parameter inversion architecture is constructed. The ionospheric D-layer parameter inversion architecture is trained with the training sample set to obtain the trained ionospheric D-layer parameter inversion model.

3. The method for gridded inversion of ionospheric D-layer parameters based on multimodal deep learning according to claim 1, characterized in that, The feature processing branch includes a path observation encoder and a linear modulation layer. The path observation encoder encodes each VLF sequence data to obtain global path features and local path point features. The linear modulation layer modulates the global path features and local path point features based on the corresponding path static information to obtain path-level features and path point-level features.

4. The method for gridded inversion of ionospheric D-layer parameters based on multimodal deep learning according to claim 1, characterized in that, The path static information includes the latitude and longitude of the transceiver station, the operating frequency, and the transmission power of the corresponding VLF transceiver path.

5. The method for gridded inversion of ionospheric D-layer parameters based on multimodal deep learning according to claim 1, characterized in that, In the multi-head cross-attention calculation step in the feature enhancement unit, a query is constructed using the grid position information in the target two-dimensional grid, the corresponding prior ionospheric D-layer parameters, the corresponding path point-level features in the interpolated observation feature map, and the corresponding grid point weights in the grid weight map. The key value is constructed using the path point-level features output by the feature processing branch and the corresponding path point coordinates.

6. The method for gridded inversion of ionospheric D-layer parameters based on multimodal deep learning according to claim 1, characterized in that, In the global correction branch, each path-level feature is weighted and aggregated to obtain a global context vector. A low-order polynomial is constructed to represent the distribution law of the ionospheric D-layer parameter deviation in the target two-dimensional grid. The global context vector is processed by MLP to obtain the low-order polynomial coefficients. The low-order polynomial coefficients are used to adjust the distribution law of the ionospheric D-layer parameter deviation in the target two-dimensional grid to obtain the global correction field.

7. A gridded inversion device for ionospheric D-layer parameters based on multimodal deep learning, characterized in that, include: The deployment module is used to deploy a pre-trained ionospheric D-layer parameter inversion model and input multiple VLF sequence data of the target inversion region into the ionospheric D-layer parameter inversion model. The ionospheric D-layer parameter inversion model includes a feature processing branch, a local correction branch, a global correction branch, and a fine iteration branch. The feature extraction module uses a feature processing branch to encode each VLF sequence data and the corresponding path static information to obtain the corresponding path-level features and path-point-level features. The local correction module includes a local correction branch, which comprises a bilinear interpolation unit and a feature enhancement unit. The bilinear interpolation unit interpolates the path point-level features into the target two-dimensional grid to obtain an interpolated observation feature map. The bilinear interpolation unit includes an interpolation layer and a deformable observation mixer. The interpolation layer interpolates the path point-level features into the target two-dimensional grid using bilinear interpolation to obtain an initial grid feature map, and generates a grid weight map based on the degree to which each grid point in the initial grid feature map is covered by the path point-level features. The initial grid feature map and the grid weight map are concatenated and input into the deformable observation mixer, which uses multiple cascaded deformable convolutional layers for feature processing to obtain the interpolated observation feature map. The target two-dimensional grid is the two-dimensional grid of the target inversion region. The feature enhancement unit performs multi-head cross-attention calculation on the features of each grid point in the interpolated observation feature map and each path point-level feature to obtain a context feature map. The global correction module uses a global correction branch to generate a global correction field based on the features of each path level, and uses the global correction field to correct the prior ionospheric D-layer parameter distribution to obtain the initial distribution of the inverted ionospheric D-layer parameters. The fine iteration module employs a fine iteration branch that uses a multimodal tensor as input to iteratively correct the initial distribution of inverted ionospheric D-layer parameters to obtain the final ionospheric D-layer parameter distribution of the target inversion region. This fine iteration branch includes sequentially cascaded coarse-scale iteration units and fine-scale iteration units. In the coarse-scale iteration unit, the original-scale multimodal tensor and the initial distribution of inverted ionospheric D-layer parameters are downsampled to a first iteration scale. The coarse-scale iteration unit iterates based on the first iteration scale's multimodal tensor and the initial distribution of inverted ionospheric D-layer parameters to obtain a first residual. The first residual obtained in each iteration is then compared with the first iteration... The initial distribution of the inverted ionospheric D-layer parameters is superimposed to obtain the coarse-scale iterative result. In the fine-scale iterative unit, the coarse-scale iterative result is upsampled to the original scale. The fine-scale iterative unit iterates based on the multimodal tensor of the original scale and the initial distribution of the inverted ionospheric D-layer parameters to obtain the second residual. The second residual obtained in each iteration is superimposed with the coarse-scale iterative result of the original scale to obtain the fine-scale iterative result. After the iteration is completed, the fine-scale iterative result of the last iteration is output as the final ionospheric D-layer parameter distribution. The multimodal tensor includes the context feature map, the position information of the target two-dimensional grid, and the interpolated observation feature map.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to execute the ionospheric D-layer parameter gridding inversion method based on multimodal deep learning as described in any one of claims 1-6.

Images