Ionized layer frequency height map enhancement method and system based on time-frequency cooperation

Abstract

The invention provides an ionized layer frequency height map enhancement method and system based on time-frequency cooperation, and relates to the technical field of frequency height map processing. The method comprises the following steps: step 1, collecting first and second frequency height map sequences; 2, preprocessing the image to generate first and second low-frequency-spectrum image features, potential codes and a mask template; 3, processing the low-frequency-spectrum image features by using an improved context modeling method to obtain a time-frequency planning control signal; 4, processing the first frequency height graph sequence based on a time-frequency planning control signal to obtain a low-frequency-spectrum interpolation image; 5, processing the low-frequency-spectrum interpolation image or the second frequency height map sequence based on an improved residual error migration repairing strategy to obtain a high-frequency-spectrum-resolution frequency height map sequence; and 6, performing super-resolution reconstruction on the low-frequency-spectrum interpolation image or the second frequency-height image sequence based on the high-frequency-spectrum-resolution frequency-height image sequence and the mask template to obtain a target frequency-height image sequence. According to the method and the system, time domain and frequency domain information is cooperatively processed, and the quality of the frequency height graph can be optimized without changing hardware of the altimeter.

Description

Technical Field

[0001] The present invention relates to the technical field of ionogram processing, and particularly relates to a method and system for enhancing ionospheric ionogram based on time-frequency collaboration. Background Art

[0002] The ionosphere is a key region in the Earth's atmosphere, located at an altitude range of approximately 60 kilometers to 1000 kilometers. The ionosphere contains a large number of free electrons and ions. Space weather research mainly focuses on the plasma density changes in the ionosphere, which have a significant impact on the propagation of radio waves and further affect modern technical systems such as satellite communication, Global Positioning System (GPS), and navigation.

[0003] The ionosonde (abbreviated as sounding rocket, also known as vertical sounder) is one of the main instruments for studying the ionosphere. By sweeping and transmitting high-frequency radio waves to the ionosphere and recording the echoes returned from different altitudes, an ionogram (also known as ionization diagram) is generated. The ionogram is a two-dimensional image of frequency-altitude, showing the situation where radio waves are reflected at different altitudes in the ionosphere at different frequencies. Since the frequency of radio wave reflection is related to the characteristic frequency corresponding to the plasma density, the ionogram provides important information about the structure of the ionospheric plasma density, especially the electron density distribution in the F layer (the ionosphere is mainly divided into the D layer, E layer, and F layer from bottom to top, and under solar ultraviolet radiation, the F layer can be further divided into the F1 layer and F2 layer). This is crucial for understanding ionospheric disturbances and their impact on space weather.

[0004] By interpreting the ionogram, scientists can study various ionospheric phenomena, including Spread F (a phenomenon caused by inhomogeneities in the F layer, usually manifested as irregular spatial and temporal fluctuations in the electron density in the ionosphere), the critical frequency of the F2 layer (foF2) (the lowest frequency at which the F2 layer can reflect electromagnetic waves during the day or near the day), and plasma layer disturbances. Due to the complexity of the ionosphere, the ionogram contains a large amount of implicit information that helps to understand the behavior of the ionosphere. Through modern data analysis techniques, researchers can automatically identify and analyze the key patterns in the ionogram, thereby better understanding ionospheric disturbance phenomena and further improving the monitoring and forecasting of space weather.

[0005] In ionospheric sounding, an ionosonde collects reflection data by sequentially transmitting electromagnetic waves of different frequencies within a certain frequency range. The frequency step determines the number of sampling points and the scanning time. Commonly used ionosondes include the DPS-4D (Digisonde-Portable-Sounder-4D). Using a smaller frequency step (such as 0.025 MHz) will increase the number of sampling points, thus prolonging the scanning time. Therefore, when the acquisition interval is short (for example, 5 minutes or 10 minutes), in order to complete the scanning within a limited time and promptly capture the dynamic changes of the ionosphere, a larger frequency step (such as 0.05 MHz) must be adopted, but this will result in loss of spectral details and even make it difficult to calculate the plasma drift velocity using the Doppler effect. On the contrary, when the acquisition interval is long (such as 30 minutes), the instrument has enough time for a more detailed scanning, and a smaller frequency step is used to improve the frequency resolution to capture more subtle structural information in the ionosphere.

[0006] Currently, the traditional method for collecting ionograms in the field of ionospheric sounding is widely adopted. An ionosonde sequentially transmits electromagnetic waves within a preset frequency range and obtains ionospheric reflection data by scanning with a fixed frequency step (such as 0.025 MHz or 0.05 MHz). In these methods, to improve the time resolution, some systems adopt a larger frequency step when the acquisition interval is short; while a smaller frequency step is adopted when the acquisition interval is long. At the same time, existing research has attempted to optimize the time-domain information and frequency-domain information respectively through traditional interpolation (such as linear interpolation or spline interpolation) and super-resolution reconstruction techniques based on neural networks, so as to improve the quality of ionograms.

[0007] Traditional interpolation methods are mainly used to complement the time-domain information of missing frames in the ionogram sequence to improve the time resolution. For example, the linear interpolation method directly calculates the linear relationship between adjacent frames to generate intermediate frames. Its calculation is simple and fast. However, traditional interpolation methods usually assume smooth image changes. In the case of rapid and non-linear changes in ionospheric reflection data, it is easy to produce blurred, ghosted or unnatural transition effects and is difficult to accurately restore the true dynamics; while spline interpolation (such as cubic spline interpolation) fits a higher-order polynomial function, making the generated intermediate frames have a smoother transition and better retaining the continuity of the image structure, but it may still introduce unnatural fluctuations or transition distortions when dealing with drastic changes. These deficiencies restrict the application effects of the above methods in high-precision ionospheric sounding. In addition, some techniques have also attempted to post-process ionograms using traditional image interpolation and super-resolution methods, but mostly focus on the optimization of a single domain and lack the collaborative processing of time-domain information and frequency-domain information.

[0008] The main functions of the ionosonde are to measure 1) the vertical profile of plasma density (electron density) and 2) the occurrence and development of density inhomogeneities, traveling ionospheric disturbances (TIDs), etc. These two natural phenomena have the following characteristics: (1) The change in the vertical profile of plasma density is usually continuous. Even during large events such as magnetic storms, the change rate is usually at the hourly level. (2) The duration of plasma density inhomogeneities is usually greater than 2 hours, and the moving speed is usually tens of meters per second. It is difficult to have obvious changes or move out of the observation area within 5 minutes. The period of TIDs is usually between 0.5 hours and 1.5 hours. Some studies have pointed out that the shortest period of mesoscale TIDs may be about 6 minutes. Therefore, the extrapolation of 15-minute data is sufficient to infer the changes in these phenomena. However, regardless of the time interval, there is room for improvement in the time-domain resolution and frequency-domain resolution of the ionogram actually collected by the ionosonde.

[0009] Combined with the heat dissipation, functions, etc. of the equipment and the characteristics of ionospheric phenomena, most ionosondes internationally are set to a 15-minute acquisition interval, and only in seasons with high occurrences of disturbance phenomena such as plasma density inhomogeneities, a 5-minute encrypted observation is selected to better observe the development process of disturbance phenomena. Some scientific research institutions (such as the Institute of Electrical Engineering, the Institute of Geology of the Chinese Academy of Sciences, etc.) are continuously developing "ultra-high time-resolution" ionosondes that can approach a 1-minute interval acquisition, aiming to capture scientific information such as high-frequency fluctuations and rapid changes in density inhomogeneous structures in the ionosphere. However, as mentioned above, focusing on improving the high time resolution may result in the loss of spectral details and the loss of the ionosonde's ability to observe the drift speed.

[0010] In addition, the existing technologies for enhancing the observation ability of ionosondes usually obtain more image details by increasing the transmission power and improve the time resolution by shortening the acquisition interval. These methods are improvements at the hardware level. However, increasing the transmission power will increase the equipment load and may also result in the loss of some information during the process of filtering and generating pictures. The method of shortening the acquisition interval usually uses a relatively large frequency step (such as 0.05 MHz), but this will result in the loss of spectral details and even make it difficult to calculate the plasma drift speed using the Doppler effect. Summary of the Invention

[0011] To solve the above problems, the present invention provides an ionogram enhancement method and system based on time-frequency collaboration, which performs collaborative processing on time-domain information and frequency-domain information. It can not only shorten the acquisition interval of the ionogram but also obtain more detailed ionogram information. Therefore, the present invention can significantly optimize the quality of the ionogram without changing the hardware structure of the ionosonde.

[0012] The ionogram enhancement method based on time-frequency collaboration includes the following steps: Step 1: Collect a sequence of ionograms and determine whether it is a first ionogram sequence or a second ionogram sequence according to a time threshold and a spectral resolution threshold; Step 2: Use a pre-trained encoder to preprocess the first ionogram sequence to generate first low-frequency spectral image features, a first latent encoding, and a first mask template corresponding to the first ionogram sequence; or preprocess the second ionogram sequence to generate second low-frequency spectral image features, a second latent encoding, and a second mask template corresponding to the second ionogram sequence; Step 3: Use an improved context modeling method to process the first low-frequency spectral image features to obtain a first time-frequency planning control signal corresponding to the first ionogram sequence; or use an improved context modeling method to process the second low-frequency spectral image features to obtain a second time-frequency planning control signal corresponding to the second ionogram sequence; Step 4: Under the guidance of the first time-frequency planning control signal, process the first ionogram sequence to obtain a low-frequency spectral interpolation image; Step 5: Under the guidance of the time-frequency planning control signal, process the low-frequency spectral interpolation image or the second ionogram sequence based on an improved residual transfer repair strategy to obtain a high-frequency spectral resolution ionogram sequence; Step 6: Based on the high-frequency spectral resolution ionogram sequence and the mask template, perform super-resolution reconstruction on the low-frequency spectral interpolation image or the second ionogram sequence to obtain a target ionogram sequence.

[0013] Preferably, step 2 specifically includes: Downsample the first ionogram sequence to generate a low-frequency spectral image and a first mask template, use a pre-trained encoder to process the low-frequency spectral image to generate first low-frequency spectral image features, and use the pre-trained encoder to compress the first low-frequency spectral image features into a low-dimensional continuous latent space to obtain a first latent encoding; Alternatively, upsample the second ionogram sequence to generate a second mask template, use a pre-trained encoder to process the second ionogram sequence to generate second low-frequency spectral image features, and use the pre-trained encoder to compress the second low-frequency spectral image features into a low-dimensional continuous latent space to obtain a second latent encoding.

[0014] Preferably, step 3 specifically includes: Perform context modeling on the first low-frequency spectral image features and the second low-frequency spectral image features respectively to capture the inter-frame dynamic changes and dependencies of the first low-frequency spectral image features and the second low-frequency spectral image features, and use the inter-frame dynamic changes and dependencies corresponding to the first low-frequency spectral image features to form a first time-frequency planning control signal, and use the inter-frame dynamic changes and dependencies corresponding to the second low-frequency spectral image features to form a second time-frequency planning control signal.

[0015] Preferably, step 4 specifically includes: Under the guidance of the first time-frequency planning control signal, add noise frames to the first frequency-height map sequence, and reconstruct the first frequency-height map sequence with added noise frames by gradually denoising to generate a low-frequency spectrum interpolation image.

[0016] Preferably, step 4 also specifically includes: Adopt the adaptive instance normalization technique to integrate the diffusion step information in the reconstruction process into a conditional signal, and recover the first frequency-height map sequence based on the conditional signal, so that the low-frequency spectrum interpolation image contains the detailed features in the first frequency-height map sequence, and adopt the cross-attention mechanism to fuse the global planning information to make each frame of the low-frequency spectrum interpolation image coherent in time and frequency spectrum.

[0017] Preferably, step 5 specifically includes: Under the guidance of the first time-frequency planning control signal, gradually predict the residual between the low-frequency spectrum interpolation image and the target frequency-height map sequence through the diffusion model, gradually correct the predicted residual through a multi-step iterative diffusion process to obtain the high-frequency details in the low-frequency spectrum interpolation image, and generate a high-frequency spectrum resolution frequency-height map sequence based on the low-frequency spectrum interpolation image and the high-frequency details; Alternatively, under the guidance of the second time-frequency planning control signal, gradually predict the residual between the second frequency-height map sequence and the target frequency-height map sequence through the diffusion model, gradually correct the predicted residual through a multi-step iterative diffusion process to obtain the high-frequency details in the second frequency-height map sequence, and generate a high-frequency spectrum resolution frequency-height map sequence based on the second frequency-height map sequence and the high-frequency details.

[0018] Preferably, in step 1, a frequency-height map sequence with a sampling interval not less than the time threshold and a spectral resolution less than the spectral resolution threshold is used as the first frequency-height map sequence, and a frequency-height map sequence with a sampling interval less than the time threshold and a spectral resolution not less than the spectral resolution threshold is used as the second frequency-height map sequence.

[0019] The ionospheric frequency-height map enhancement system based on time-frequency collaboration provided in the present invention realizes the ionospheric frequency-height map enhancement method based on time-frequency collaboration. The system includes: An acquisition module, configured to acquire a frequency-height map sequence and determine it as the first frequency-height map sequence or the second frequency-height map sequence according to the time threshold and the spectral resolution threshold; A preprocessing module, connected to the acquisition module, configured to use a pre-trained encoder to preprocess the first frequency-height map sequence to generate the first low-frequency spectrum image feature, the first latent encoding, and the first mask template corresponding to the first frequency-height map sequence; or preprocess the second frequency-height map sequence to generate the second low-frequency spectrum image feature, the second latent encoding, and the second mask template corresponding to the second frequency-height map sequence; A time series prediction module, connected to the preprocessing module, is used to process the first low-frequency spectral image features by using an improved context modeling method to obtain a first time-frequency planning control signal corresponding to the first frequency-height map sequence; or process the second low-frequency spectral image features by using an improved context modeling method to obtain a second time-frequency planning control signal corresponding to the second frequency-height map sequence. A low-frequency spectral resolution frequency-height map generation module, connected to the time series prediction module, is used to process the first frequency-height map sequence under the guidance of the first time-frequency planning control signal to obtain a low-frequency spectral interpolation image. A high-frequency spectral resolution frequency-height map generation module, connected to the time series prediction module and the high-frequency spectral resolution frequency-height map generation module, is used to process the low-frequency spectral interpolation image or the second frequency-height map sequence under the guidance of the time-frequency planning control signal based on an improved residual transfer repair strategy to obtain a high-frequency spectral resolution frequency-height map sequence. A frequency-height map fusion and reconstruction module, connected to the high-frequency spectral resolution frequency-height map generation module, is used to perform super-resolution reconstruction on the low-frequency spectral interpolation image or the second frequency-height map sequence based on the high-frequency spectral resolution frequency-height map sequence and the mask template to obtain a target frequency-height map sequence.

[0020] Preferably, the system further includes a pre-training module; the pre-training module is connected to the preprocessing module, the time series prediction module, the low-frequency spectral resolution frequency-height map generation module, and the high-frequency spectral resolution frequency-height map generation module, and is used to pre-train the preprocessing module, the time series prediction module, the low-frequency spectral resolution frequency-height map generation module, and the high-frequency spectral resolution frequency-height map generation module, and is also used to jointly train the time series prediction module and the low-frequency spectral resolution frequency-height map generation module, and jointly train the time series prediction module and the high-frequency spectral resolution frequency-height map generation module.

[0021] Preferably, the pre-training of the preprocessing module by the pre-training module includes: Set the preprocessing module to an 8-fold spatial compression rate and a 16-channel latent dimension, set the learning rate to 10 -4 , the batch size to 32, and the codebook size to 512; Set the reconstruction loss function of the preprocessing module to , where represents the first frequency-height map sequence or the second frequency-height map sequence, represents the frequency-height map sequence reconstructed and output by the preprocessing module; Set the codebook loss function of the preprocessing module to , where represents the latent representation generated by the preprocessing module for the first frequency-height map sequence or the second frequency-height map sequence, represents the nearest codebook vector, represents the stop gradient operation; Set the commitment loss function of the preprocessing module to ; The total loss function of the preprocessing module is expressed as , where represents the balance coefficient; When the curve of the total loss function of the preprocessing module converges, the pre-training of the preprocessing module is completed; The pre-training of the time series prediction module by the pre-training module includes: Set the reconstruction loss function of the time series prediction module to , where represents the binary function, indicating whether it is a mask, represents the prediction function adopted by the time series prediction module, represents the low-frequency spectral image features processed by the mask, represents the true time-frequency planning information corresponding to the low-frequency spectral image features. This reconstruction loss is used to encourage the difference between the output of the time series prediction module and the true time-frequency planning information not to exceed a preset threshold, improving the accuracy of the time series prediction module in capturing inter-frame dynamic changes and dependency relationships; When the curve of the reconstruction loss function of the time series prediction module converges, the pre-training of the time series prediction module is completed; The pre-training of the low-frequency spectral resolution frequency-height map generation module by the pre-training module includes: Set the diffusion loss function of the low-frequency spectral resolution frequency-height map generation module to , where represents the binary mask, represents the time step in the current diffusion process, represents the conditional signal, represents the time step of the low-frequency spectral part in the current diffusion process of the target residual, represents the first frequency-height map sequence with added noise, represents the low-frequency spectral resolution frequency-height map generation module, with its parameters being ; When the curve of the diffusion loss of the low-frequency spectral resolution frequency-height map generation module converges, the pre-training of the low-frequency spectral resolution frequency-height map generation module is completed; The pre-training of the high-frequency spectral resolution frequency-height map generation module by the pre-training module includes: Set the diffusion loss function of the high-frequency spectral resolution frequency-height map generation module to , where represents the binary mask, represents the time step in the current diffusion process, Represents a conditional signal, represents the time step of the high-frequency spectrum part in the current diffusion process of the target residual, represents the residual between the low-frequency spectrum interpolation image with added noise and the target frequency-height map sequence or the residual between the second frequency-height map sequence and the target frequency-height map sequence, represents a high-frequency spectrum resolution frequency-height map generation module, whose parameters are ; When the curve of the diffusion loss function of the high-frequency spectrum resolution frequency-height map generation module converges, the pre-training of the high-frequency spectrum resolution frequency-height map generation module is completed; The pre-training module jointly trains the time series prediction module and the low-frequency spectrum resolution frequency-height map generation module, including: Set the masked diffusion loss function for the joint training of the time series prediction module and the low-frequency spectrum resolution frequency-height map generation module as , where represents a binary mask, represents the time step in the current diffusion process, represents the time step of the low-frequency spectrum part in the current diffusion process of the target residual, represents a low-frequency spectrum resolution frequency-height map generation module, whose parameters are , represents the low-frequency spectrum resolution data with added noise, represents the time-frequency planning control signal output by the time series prediction module, , represents the prediction function adopted by the time series prediction module, represents the masked low-frequency spectrum image feature; When the curve of the masked diffusion loss function converges, the joint training of the time series prediction module and the low-frequency spectrum resolution frequency-height map generation module is completed; The pre-training module jointly trains the time series prediction module and the high-frequency spectrum resolution frequency-height map generation module, including: Set the loss function for the joint training of the time series prediction module and the high-frequency spectrum resolution frequency-height map generation module as , where represents a binary mask, represents the time step in the current diffusion process, represents the time step of the high-frequency spectrum part in the current diffusion process of the target residual, represents the residual between the low-frequency spectrum interpolation image with added noise and the target frequency-height map sequence or the residual between the second frequency-height map sequence and the target frequency-height map sequence, Denotes the high - frequency spectral resolution frequency - height map generation module, and its parameters are , Denotes the time - frequency planning control signal output by the time - series prediction module, , Denotes the prediction function adopted by the time - series prediction module, Denotes the low - frequency spectral image features after mask processing; When the curve of the loss function for the joint training of the time - series prediction module and the high - frequency spectral resolution frequency - height map generation module converges, the joint training of the time - series prediction module and the high - frequency spectral resolution frequency - height map generation module is completed.

[0022] According to the specific embodiments provided by the present invention, the following technical effects are disclosed: 1) The present invention can significantly improve the time - domain continuity and frequency - domain detail expression ability of the ionospheric frequency - height map. Specifically, by introducing a strategy combining time - series prediction and residual migration, accurate interpolation of the intermediate frames of the frequency - height map and super - resolution reconstruction of the low - frequency spectral resolution frequency - height map can be achieved, effectively improving the data accuracy of ionosondes in detection due to long acquisition intervals and low - frequency spectral resolution conditions. Adopting a technical path of multi - stage progressive generation and incremental correction, fully integrating the inter - frame timing information and spectral details, global coherence and local fine recovery of image reconstruction can be realized, and thus high - quality and reliable data support can be provided for the precise monitoring of the ionospheric structure and dynamic changes; 2) The present invention breaks through the limitation of only focusing on a single domain, can realize the collaborative processing of time - domain information and frequency - domain information, can improve the time resolution when the acquisition interval is long, and can restore the high - frequency details in the spectrum under limited sampling conditions, and can fundamentally solve the contradiction between time and frequency acquisition; adopting a multi - stage progressive generation and incremental correction strategy, effectively using the time - series prediction and residual migration mechanisms, global coherence of image generation and fine recovery of local details can be realized; 3) When optimizing the ionogram, prior art usually processes time-domain optimization and frequency-domain optimization separately. In terms of the time domain, traditional methods mainly use techniques such as linear interpolation and spline interpolation to fill in the missing intermediate frames within the acquisition interval. However, in the face of rapidly changing ionospheric data or non-linear dynamic scenarios, such methods are prone to problems such as blurring and ghosting. In terms of the frequency domain, a common approach is to use super-resolution reconstruction techniques based on deep learning, such as convolutional neural networks or generative adversarial networks, to restore high-frequency details by learning the mapping relationship between low-spectral-resolution images and high-spectral-resolution images. However, these methods often rely on large-scale training data and have deficiencies in real-time performance and noise suppression, making it difficult to balance overall performance. In contrast, the present invention can optimize time-domain data and frequency-domain data simultaneously. It can not only complete intermediate frames to improve time resolution when the acquisition interval is long, but also restore high-frequency details in the spectrum under low-sampling conditions, fundamentally solving the contradiction between time and frequency acquisition; 4) The present invention adopts a multi-stage progressive generation and incremental correction strategy, effectively utilizes time series prediction and residual migration mechanisms, can achieve global coherence and local fine restoration in image generation, and significantly improves the blurring and ghosting phenomena that occur in traditional methods; 5) The present invention can flexibly adapt to different acquisition conditions (such as an acquisition interval of 15 minutes and a spectral resolution of 0.025 MHz, and an acquisition interval of 5 minutes and a spectral resolution of 0.05 MHz). While ensuring high detection accuracy, it takes into account system real-time performance and wide applicability, and the overall performance is significantly better than traditional technologies; 6) The present invention does not modify the hardware of the existing ionosonde itself, not only will not cause loss of existing data, but also can effectively enhance the information content of scientific data; 7) Through the multi-stage progressive generation and incremental correction strategy (the present invention divides the process of time interpolation and spectrum refinement into several steps. In each step, a more realistic ionogram is generated based on the results of the previous stage; then, through a small amount of "correction" or "adjustment", possible deviations are gradually corrected. This can not only "fill" intermediate frames under the condition of insufficient sampling (missing some time frames), but also gradually improve the spectral resolution of a single image in the frequency domain, thus ensuring both time-domain coverage and obtaining more delicate and clear spectral details), the present invention can reconstruct missing intermediate frames and refine spectral resolution under limited sampling conditions. This can not only improve time resolution without changing the existing hardware, but also capture richer spectral details. Without changing the hardware conditions of the ionosonde, it can significantly improve the ability to accurately track and identify the dynamic changes of the ionosphere. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for use in the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0024] Figure 1 This is the schematic diagram of the ionogram enhancement method based on time-frequency collaboration in Embodiment 1 of the present invention; Figure 2 This is the schematic diagram of processing the first ionogram sequence using the ionogram enhancement method based on time-frequency collaboration in Embodiment 1 of the present invention; Figure 3 This is the schematic diagram of processing the second ionogram sequence using the ionogram enhancement method based on time-frequency collaboration in Embodiment 1 of the present invention; Figure 4 This is the schematic diagram of the network structure for processing the first ionogram sequence; Figure 5 This is the schematic diagram of the network structure for processing the second ionogram sequence; Figure 6 This is the schematic diagram of image fusion in Embodiment 1 of the present invention; Figure 7 This is the schematic diagram of the structure of the ionogram enhancement system based on time-frequency collaboration in Embodiment 2 of the present invention; Figure 8 This is the training schematic diagram of each module in the pre-training stage in Embodiment 2 of the present invention; Figure 9 This is the training schematic diagram of each module in the joint training stage in Embodiment 2 of the present invention. Specific Embodiments

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0026] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0027] Embodiment 1: As Figure 1As shown in the figure, it is the schematic diagram of the ionogram enhancement method based on time-frequency cooperation. Using this method, the first ionogram sequence (in this embodiment, it is the high-frequency spectral resolution ionogram sequence with a long acquisition interval) can be processed to improve the time-domain acquisition density by generating an interpolation image, or the second ionogram sequence (in this embodiment, it is the low-frequency spectral resolution ionogram sequence with a short acquisition interval) can be processed to improve the frequency-domain acquisition density by the frequency super-resolution method. In both cases, the quality of the ionogram can be optimized to obtain the target ionogram sequence (in this embodiment, it is the high-frequency spectral image sequence with a short acquisition interval).

[0028] It should be noted that the present invention distinguishes the first ionogram sequence and the second ionogram sequence according to the time threshold and the spectral resolution threshold. Specifically, the ionogram sequence with an acquisition interval not less than the time threshold and a spectral resolution less than the spectral resolution threshold is used as the first ionogram sequence, and the ionogram sequence with an acquisition interval less than the time threshold and a spectral resolution not less than the spectral resolution threshold is used as the second ionogram sequence. In the present invention, the time threshold is generally taken as 15 minutes, and the spectral resolution threshold is generally taken as 0.05 MHz. In actual operation, generally, the first ionogram sequence with an acquisition interval of 15 minutes and a spectral resolution of 0.025 MHz, or the second ionogram sequence with an acquisition interval of 5 minutes and a spectral resolution of 0.05 MHz is processed. It should be noted that the acquisition interval of the first ionogram sequence can also be set to 30 minutes according to actual needs, and the acquisition interval of the second ionogram sequence can be set to 10 minutes, which will not be listed one by one here.

[0029] In addition, as Figure 1 shown, the methods adopted for ionogram sequences with different acquisition intervals are different. Specifically, when inputting the high-frequency spectral resolution ionogram sequence with a long acquisition interval, it is necessary to downsample the ionogram sequence to generate a low-frequency spectral image and a mask template, then the encoder extracts the low-frequency spectral image features of the low-frequency spectral image, then generates a latent code through context modeling, then decodes the latent code to generate a low-frequency spectral interpolation image, and then performs super-resolution reconstruction on the low-frequency spectral interpolation image to obtain the high-frequency spectral resolution ionogram sequence with a short acquisition interval.

[0030] When inputting the low-frequency spectral resolution ionogram sequence with a short acquisition interval, directly generate an upsampled mask template of the low-frequency spectral resolution ionogram sequence, then extract the low-frequency spectral image features of the low-frequency spectral resolution ionogram sequence through the encoder, generate a latent code through context modeling, and then directly perform super-resolution reconstruction on the low-frequency spectral resolution ionogram sequence to obtain the high-frequency spectral resolution ionogram sequence with a short acquisition interval.

[0031] As Figure 2 and Figure 3As shown, it is a schematic flowchart of processing a high-frequency spectral resolution ionogram sequence with a long acquisition interval (corresponding to the first ionogram sequence) or a low-frequency spectral resolution ionogram sequence with a short acquisition interval (corresponding to the second ionogram sequence) using the ionogram enhancement method based on time-frequency collaboration in Embodiment 1 of the present invention.

[0032] As Figure 4 and Figure 5 shown, it is a schematic network structure diagram of processing the first ionogram sequence (as Figure 4 shown, a high-frequency spectral resolution ionogram sequence with a long acquisition interval) or the second ionogram sequence (shown in Figure 5 as a low-frequency spectral resolution ionogram sequence with a short acquisition interval) to obtain a target ionogram sequence (as Figure 5 shown, a high-frequency spectral resolution ionogram sequence with a short acquisition interval).

[0033] As Figure 4 shown, when processing the first ionogram sequence, there are three main steps (time series prediction , generation of low-frequency spectral resolution ionograms , and generation of high-frequency spectral resolution ionograms ). Specifically, in time series prediction , there are multiple root mean square normalization processes. Among them, the output result of the encoder is first subjected to root mean square normalization processing and spatial multi-head attention layer processing, and positional encoding is applied to it; then the processed result is subjected to a second root mean square normalization processing and spatio-temporal multi-head attention layer processing, and a second positional encoding is applied to it; finally, the processed result is subjected to a third root mean square normalization processing and processed using a fully connected layer to obtain time series prediction results, which are respectively used for the generation of low-frequency spectral resolution ionograms and high-frequency spectral resolution ionograms.

[0034] In the generation of low-frequency spectral resolution ionograms, noise is applied to the low-frequency spectral resolution ionogram sequence. Specifically, it passes through three processing layers in sequence (including the first layer composed of a spatial multi-head attention layer and a root mean square normalization layer, the second layer composed of a temporal multi-head attention layer and a root mean square normalization layer, and the third layer composed of a cross multi-head attention layer and a root mean square normalization layer) to process the noise to be applied, including the application of positional encoding and diffusion time steps, and then together with the result output in time series prediction, it is sent to the decoder for decoding to obtain a low-frequency spectral resolution ionogram sequence, which includes multiple low-frequency spectral resolution ionograms.

[0035] In the generation of high-frequency spectral resolution frequency-height maps, the low-frequency spectral resolution frequency-height maps obtained in the previous step are processed successively through three processing layers (including the first layer composed of a spatial multi-head attention layer and a root mean square normalization layer, the second layer composed of a temporal multi-head attention layer and a root mean square normalization layer, and the third layer composed of a cross multi-head attention layer and a root mean square normalization layer). During this process, position encoding and diffusion time steps are applied. Finally, the cross multi-head attention layer of the third layer outputs multiple high-frequency spectral resolution frequency-height maps with short acquisition intervals. These high-frequency spectral resolution frequency-height maps with short acquisition intervals form a sequence of high-frequency spectral resolution frequency-height maps with short acquisition intervals, that is, the target frequency-height map sequence is obtained.

[0036] However, when processing the second frequency-height map sequence, the low-frequency spectral resolution frequency-height map generation process as shown in Figure 4 is not required. The specific processing process is as shown in Figure 5 .

[0037] As shown in Figure 5 , for the sequence of low-frequency spectral resolution frequency-height maps with short acquisition intervals, it only contains two main steps (time series prediction and high-frequency spectral resolution frequency-height map generation ). As shown in Figure 5 , the network structure required to achieve time series prediction and the network structure required to generate high-frequency spectral resolution frequency-height maps are similar to the corresponding structures as shown in Figure 4 . These two structures will not be elaborated here. Figure 5

[0038] Specifically, as shown in Figure 5 , the encoder is used to encode multiple low-frequency spectral resolution frequency-height maps in the sequence of low-frequency spectral resolution frequency-height maps, and then processed through three layers in time series prediction to extract the context information in the low-frequency spectral resolution frequency-height maps, obtaining a time-frequency planning control signal. Then, under the guidance of this time-frequency planning control signal, the original low-frequency spectral resolution frequency-height maps are processed. Finally, the high-frequency details obtained by processing are fused onto the original low-frequency spectral resolution frequency-height maps using a mask template, thereby obtaining high-frequency spectral resolution frequency-height maps with short acquisition intervals. Multiple high-frequency spectral resolution frequency-height maps with short acquisition intervals form a sequence of high-frequency spectral resolution frequency-height maps with short acquisition intervals.

[0039] In the above, the present invention uses a vector quantized variational autoencoder (VQ-VAE) as a codec to compress the original frequency-height map into a low-dimensional continuous latent space, thereby improving the training and reasoning efficiency. Specifically, VQ-VAE consists of an encoder and a decoder, wherein the encoder uses an 8-fold spatial compression rate and a 16-channel latent dimension to map the video frame into a tile structure of shape N×16 (N represents the number of tokens after segmentation), effectively reducing the data dimension while retaining spatial details as much as possible; the decoder uses this low-dimensional representation to reconstruct high-quality video frames, providing auxiliary conditions and prior information for the subsequent generation step of the frequency-height map.

[0040] In addition, the above-mentioned three-layer network architecture required for time series prediction and generation of low spectral resolution high-frequency graph sequences and high spectral resolution high-frequency graph sequences, as well as the two-layer network architecture required for time series prediction and generation of high spectral resolution high-frequency graph sequences are all based on Transformer design, which focuses on the balance between efficiency and resource utilization, and adopts an efficient attention mechanism, optimized layer normalization, and carefully designed position encoding.

[0041] In order to enhance the stability of training, the input of each attention block adopts Root Mean Square Layer Normalization (RMS-Norm), and a spatiotemporal attention mechanism is introduced, which can effectively extract contextual information and focus on the correlation features of the previous and next frames in the frequency-height map sequence to be predicted, so as to better capture the temporal changes. In addition, the attention blocks use Rotary Position Embedding (RoPE) to encode the spatial position information and temporal position information of the token to capture the internal spatial structure of each frame token in the frequency-height map sequence to be predicted and the sequential relationship between frames. Specifically, a two-dimensional RoPE is used to encode the image of a single frequency-height map in the frequency-height map sequence to be predicted, retaining the two-dimensional spatial structure inside the frame; the image block (Patch) of each frame is flattened into a one-dimensional token sequence, and a learnable marker token is inserted between different frequency-height maps to clearly distinguish each independent frame. This design not only retains the spatial position information within each frame, but also clarifies the sequential relationship between frames through tokens, enabling the model to effectively process frame sequence data with different aspect ratios and resolutions.

[0042] To generate both low - frequency spectral resolution frequency - height maps and high - frequency spectral resolution frequency - height maps, not only a network architecture with Transformer as the core is adopted, but also the design idea of the Diffusion Transformer (DIT) is borrowed. The main differences between Transformer and DIT are as follows: In the generation of low - frequency spectral resolution frequency - height maps, an encoder - decoder is used to compress the frequency - height maps to better adapt to the characteristics of low - frequency data, while in the generation of high - frequency spectral resolution frequency - height maps, this compression strategy is not adopted, aiming to better retain high - frequency details. Additionally, during the generation of both low - frequency spectral resolution frequency - height maps and high - frequency spectral resolution frequency - height maps, the Adaptive Instance Normalization (AdaIN) technique is used. This technique can integrate diffusion step information into conditional signals and inject them into the spatial attention layer, enabling the generation processes of low - frequency spectral resolution frequency - height maps and high - frequency spectral resolution frequency - height maps to flexibly regulate these generation processes according to the diffusion progress, thereby more precisely restoring the detailed features in the frequency - height maps. Moreover, to make full use of the time - frequency planning control signals generated in time - series prediction, a cross - attention mechanism is added. This mechanism can ensure the effective integration of global planning information during the generation process and achieve the coherence in time and frequency among different frequency - height map frames. Such a design, while taking into account the characteristics of both low - frequency and high - frequency data, can achieve the fine repair and continuous generation of the inter - frame time - frequency relationship under the guidance of global planning information.

[0043] In addition, to achieve the smooth transition and detail completion from low - frequency spectral resolution frequency - height maps to high - frequency spectral resolution frequency - height maps, the present invention uses the ResShift correction method for repair during the generation of high - frequency spectral resolution frequency - height maps. Specifically, the low - frequency spectral resolution frequency - height map is used as the starting image, and the high - frequency spectral resolution frequency - height map is used as the target image. The diffusion model is used to gradually predict and correct the residuals between low - frequency and high - frequency. Specifically, first, the low - frequency spectral resolution frequency - height map is processed during the generation of high - frequency spectral resolution frequency - height maps. During the multi - step iterative diffusion process, the detail residuals are predicted according to the current diffusion progress, and high - frequency information is gradually fused, making the generated image gradually approach the detailed features of the target high - frequency spectral resolution frequency - height map. During this process, the model structure remains unchanged, and only the loss function and diffusion step design are adjusted to ensure the smooth connection and effective repair of low - frequency and high - frequency information. This can not only utilize the stability of the low - frequency spectral resolution frequency - height map but also achieve the restoration of high - frequency details through residual correction, thereby effectively improving the overall quality and resolution of the generated high - frequency spectral resolution frequency - height maps while taking into account efficiency and real - time performance.

[0044] Such as Figure 6As shown in the figure, it shows the workflow of fusing the generated low-frequency spectral resolution frequency-height map and the generated high-frequency spectral resolution frequency-height map. Each number in the figure corresponds to a specific frequency. The number 1 represents 5 MHz, the number 2 represents 5.0025 MHz, the number 3 represents 5.005 MHz, the number 4 represents 5.0075 MHz, the number 5 represents 5.01 MHz, the number 6 represents 5.0125 MHz, the number 7 represents 5.015 MHz, and the number 8 represents 5.0175 MHz.

[0045] In the fusion step of this fusion map, the high-frequency spectral resolution frequency-height map generated during the generation of the high-frequency spectral resolution frequency-height map is split into multiple parts according to frequency, and then the content in the low-frequency spectral resolution frequency-height map and the preset mask template are used to replace the corresponding areas (shadow areas) in the map, so as to achieve the organic fusion of high-frequency details and the original low-frequency information, and significantly improve the overall quality and detail performance of the frequency-height map.

[0046] Embodiment 2: As Figure 7 shown, it is a schematic structural diagram of an ionospheric frequency-height map enhancement system based on time-frequency collaboration provided by the present invention. The system includes an acquisition module, a preprocessing module, a time series prediction module, a low-frequency spectral frequency-height map generation module, a high-frequency spectral frequency-height map generation module, a frequency-height map fusion and reconstruction module, and a pre-training module. The system can implement the ionospheric frequency-height map enhancement method based on time-frequency collaboration described in Embodiment 1. Using the above system can effectively solve the problem of frame interpolation for a high-frequency spectral image sequence with a long acquisition interval.

[0047] In actual operation, the system first uses the acquisition module to obtain a first sequence of frequency-height diagrams with an acquisition interval not less than the time threshold and a spectral resolution less than the spectral resolution threshold (in this embodiment, an image sequence with a high spectral resolution with an acquisition interval of 15 minutes and a spectral resolution of 0.025 MHz can be selected) or a second sequence of frequency-height diagrams with an acquisition interval less than the time threshold and a spectral resolution not less than the spectral resolution threshold (in this embodiment, an image sequence with a low spectral resolution with an acquisition interval of 5 minutes and a spectral resolution of 0.05 MHz can be selected). Then, the preprocessing module is used to preprocess the above-mentioned sequences of frequency-height diagrams respectively to obtain low-frequency spectral image features and mask templates. Then, the time series prediction module processes the low-frequency spectral image features to generate potential codes and time-frequency planning control signals. Under the guidance of the time-frequency planning control signal, the low-frequency spectral resolution frequency-height diagram generation module generates a low-frequency spectral interpolation image of the first sequence of frequency-height diagrams. Under the guidance of the time-frequency planning control signal, the high-frequency spectral resolution frequency-height diagram generation module processes the low-frequency spectral interpolation image or the second sequence of frequency-height diagrams based on the residual transfer repair strategy to restore the high-frequency details in the low-frequency spectral interpolation image or the second sequence of frequency-height diagrams. The frequency-height diagram fusion and reconstruction module performs super-resolution reconstruction on the low-frequency spectral interpolation image or the second sequence of frequency-height diagrams based on the high-frequency spectral resolution frequency-height diagram and the mask template to obtain the target sequence of frequency-height diagrams.

[0048] To improve the processing accuracy of each module in the present invention, the system also trains the preprocessing module, the time series prediction module, the low-frequency spectral resolution frequency-height diagram generation module, and the high-frequency spectral resolution frequency-height diagram generation module through a pre-training module.

[0049] Specifically, the present invention adopts a two-stage progressive training strategy to improve the training stability and the final generation performance by gradually increasing the task difficulty. The main advantage of the progressive training strategy is that it helps to alleviate the problem of unstable training of the Transformer model under large-scale parameter settings. As Figure 8 and Figure 9 shown, it is the training principle diagram for pre-training and joint training of each module in this embodiment.

[0050] The training includes an initial training stage and a joint training stage. Among them, the initial training stage involves the training of the preprocessing module, the training of the time series prediction module, the training of the low-frequency spectral resolution frequency-height diagram generation module, the training of the high-frequency spectral resolution frequency-height diagram generation module, and the training of the vector quantization variational autoencoder (VQ-VAE) for the encoder. The VQ-VAE is used to encode the input frequency-height diagram data into a discrete latent space to provide auxiliary conditions and prior information for the above-mentioned multiple frequency-height diagram generation modules. The joint training stage involves the joint training of the time series prediction module and the low-frequency spectral resolution frequency-height diagram generation module, and the joint training of the time series prediction module and the high-frequency spectral resolution frequency-height diagram generation module.

[0051] In the initial training stage, the present invention pre-trains each submodule independently so that the parameters of each module reach a reasonable initialization state.

[0052] The training process for the preprocessing module aims to minimize the reconstruction error and quantization error. The loss function consists of three parts: reconstruction loss function, codebook loss function and commitment loss function.

[0053] The reconstruction loss function is , represents the original input frequency-height map (in this embodiment, represents the first frequency-height map sequence or the second frequency-height map sequence), Represents the frequency-height graph sequence reconstructed by the preprocessing module. The original reconstruction loss is calculated using the mean square error (MSE) to measure the frequency-height graph of the original input. The reconstructed output from the encoder and decoder in the preprocessing module The difference between.

[0054] The codebook loss function is , Indicates that the preprocessing module is for input The generated continuous latent representation, represents the most recent codebook vector, Indicates stopping the gradient operation and encouraging the feature vector output by the encoder to be close to the selected quantized embedding vector.

[0055] The commitment loss function is , to prevent the feature vector output by the encoder from deviating too far from the codebook.

[0056] The total loss function is expressed as follows, where is the balance coefficient: .

[0057] When the curve of the total loss function of the preprocessing module converges, the pre-training of the preprocessing module is completed.

[0058] The task of the time series prediction module is to independently train a network so that it can accurately capture the time-frequency planning information between frames in the input frequency-height graph sequence, thereby reflecting the temporal changes of each frame in the video. The time series prediction module uses a prediction function For the masked input (in this example, the low-frequency image features) Make a prediction to generate real planning information (in this embodiment, the real time-frequency planning information corresponding to the low-frequency image feature) Similar outputs. To this end, the time series prediction module uses the reconstruction loss (i.e., the MSE loss) to measure the difference between the prediction result and the true value. The reconstruction loss function is: ;

[0059] M is a binary function representing whether it is a mask. This reconstruction loss function encourages the time-frequency planning information output by the time series prediction module to be as close as possible to the true time-frequency planning information, thereby improving the accuracy of the time series prediction module in capturing the dynamic changes between frames and the temporal structure, and providing an effective conditional signal for the subsequent generation of the frequency-height map.

[0060] When the curve of the reconstruction loss function of the time series prediction module converges, the pre-training of the time series prediction module is completed.

[0061] The main task of the low-frequency spectral resolution frequency-height map generation module is to process the low-frequency spectral data using a network structure with an encoder-decoder compression and optimize it through the mask diffusion loss, thereby generating a high-quality low-frequency spectral resolution frequency-height map. Specifically, the low-frequency spectral resolution frequency-height map generation module repairs and refines the input low-frequency spectral data. The input of the low-frequency spectral resolution frequency-height map generation module includes the low-frequency spectral data with added noise (in this embodiment, it is the first frequency-height map sequence with added noise) , the time step in the current diffusion process , the conditional signal , and this conditional signal is usually the time-frequency planning control signal output by the time series prediction module (here, to maintain the consistency of the architecture during subsequent joint training, this parameter ) is retained. Specifically, the low-frequency spectral resolution frequency-height map generation module compresses the low-frequency spectral data through the encoder, enabling the module to perform efficient processing in a lower-dimensional space while retaining the necessary spatial detail information. The generation network gradually recovers and optimizes the detail features of the low-frequency spectral image during the diffusion process, and its training objective is achieved through the mask diffusion loss. The diffusion loss function is: ;

[0062] M is a binary mask used to mask the areas that are known or do not need to be repaired; represents the target residual of the low-frequency spectral part at the time step in the current diffusion process; represents the low-frequency spectral resolution frequency-height map generation module, and its parameter is .

[0063] When the curve of the diffusion loss function of the low-frequency spectral resolution frequency-height map generation module converges, the pre-training of the low-frequency spectral resolution frequency-height map generation module is completed.

[0064] The main task of the high - frequency spectral resolution frequency - height map generation module is to restore high - frequency details. Its core lies in using the ResShift repair strategy to perform noise modeling and correction on the difference between the low - frequency spectral resolution frequency - height map and the high - frequency spectral resolution frequency - height map, thereby complementing high - frequency details. Specifically, the high - frequency spectral resolution frequency - height map generation module first calculates the residual between the low - frequency spectral resolution frequency - height map and the high - frequency spectral resolution frequency - height map, then applies noise perturbation to this residual, and predicts and corrects this residual through the ResShift model within a small number of iteration steps (20 steps). In the high - frequency spectral resolution frequency - height map generation module, the inputs include the residual information with added noise (in this embodiment, the residual between the low - frequency spectral interpolation image with added noise and the target frequency - height map sequence or the residual between the second frequency - height map sequence and the target frequency - height map sequence) the time step in the current diffusion process and the conditional signal (such as the time - frequency planning control signal). The output of this module is the residual correction result generated by the ResShift mechanism. The difference between this residual correction result and the target residual is measured by the following diffusion loss function: ;

[0065] M is a binary mask used to mask those known or areas that do not need to be repaired; represents the target residual of the high - frequency spectral part at the current diffusion step ; represents the high - frequency spectral resolution frequency - height map generation module adopting the ResShift repair strategy, and its parameters are . In each iteration step, the predicted residual is superimposed on the low - frequency spectral resolution frequency - height map, thereby gradually approaching the detailed features of the target frequency - height map.

[0066] When the curve of the diffusion loss function of the high - frequency spectral resolution frequency - height map generation module converges, the pre - training of the high - frequency spectral resolution frequency - height map generation module is completed.

[0067] In the joint training stage, the present invention uses the above - mentioned pre - trained time - series prediction module to perform joint training on the low - frequency spectral resolution frequency - height map generation module and the high - frequency spectral resolution frequency - height map generation module respectively, so as to make full use of the time - frequency planning information corresponding to the time - frequency planning control signal. Each sub - task combines the unconditional guidance strategy and the cross - attention mechanism for conditional information fusion, thereby improving the quality and detailed performance of the generated frequency - height map sequence as a whole.

[0068] In the joint training of the time series prediction module and the low-frequency spectral resolution frequency-height map generation module, the time-frequency planning control signal output by the time series prediction module is used as a conditional signal to guide the low-frequency spectral resolution frequency-height map generation module to perform detail restoration and interpolation within the masked area, ensuring the temporal coherence of the generated low-frequency spectral resolution frequency-height map.

[0069] The input conditions include the time-frequency planning control signal generated by the time series prediction module , the low-frequency spectral resolution data with added noise perturbations and the time step in the current diffusion process , denotes the prediction function of the time series prediction module. denotes the low-frequency spectral image features after masking.

[0070] The objective of the joint training of the time series prediction module and the low-frequency spectral resolution frequency-height map generation module is to perform end-to-end training on the low-frequency spectral resolution frequency-height map generation module using a unified masked diffusion loss function. The masked diffusion loss function is: ; is a binary mask, denotes the target residual (or noise) at time step in the diffusion process of the low-frequency spectral part.

[0071] When the curve of the masked diffusion loss function converges, the joint training of the time series prediction module and the low-frequency spectral resolution frequency-height map generation module is completed.

[0072] In the joint training of the time series prediction module and the high-frequency spectral resolution frequency-height map generation module, the time-frequency planning signal output by the time series prediction module is used as a conditional signal to guide the high-frequency spectral resolution frequency-height map generation module (adopting the ResShift repair strategy) to perform noise modeling and correction on the difference between the low-frequency spectral resolution frequency-height map and the high-frequency spectral resolution frequency-height map, thereby achieving the restoration and enhancement of high-frequency details within fewer iteration steps (20 steps). The input conditions include the time-frequency planning control signal generated by the time series prediction module , denotes the prediction function of the time series prediction module; the residual between the low-frequency spectral interpolation image obtained after noise perturbation and the target frequency-height map sequence or the residual between the second frequency-height map sequence and the target frequency-height map sequence (or directly calculated by taking the difference of the high-frequency spectral data); and the time step in the current diffusion process . denotes the low-frequency spectral image features after masking.

[0073] An end-to-end joint training is carried out on the high-frequency spectral resolution frequency-height map generation module using a unified masked diffusion loss. The core lies in using the ResShift repair strategy to perform noise modeling and correction on the difference between the low-frequency spectral resolution frequency-height map and the high-frequency spectral resolution frequency-height map. Its loss function is: ; is a binary mask used to mask known or areas that do not need to be repaired; represents the time step of the high-frequency spectral part in the current diffusion process The target residual at, that is, the ideal difference between the low-frequency spectral resolution frequency-height map and the high-frequency spectral resolution frequency-height map; represents the high-frequency spectral resolution frequency-height map generation module adopting the ResShift repair strategy, and its parameter is , the high-frequency spectral resolution frequency-height map generation module predicts and corrects the residual between the low frequency and the high frequency within a small number of iteration steps (20 steps), and gradually corrects the low-frequency spectral resolution frequency-height map into an image containing more high-frequency details. To further enhance the robustness of the high-frequency spectral resolution frequency-height map generation module and support classifier-free guidance, a certain probability can be set during the above training process to randomly replace with an unconditional signal .

[0074] When the curve of the loss function for the joint training of the time series prediction module and the high-frequency spectral resolution frequency-height map generation module converges, the joint training of the time series prediction module and the high-frequency spectral resolution frequency-height map generation module is completed.

[0075] In a preferred embodiment of the present invention, the obtained dataset comes from the Hainan National Space Weather Scientific Observation and Research Station (its coordinates are 19.5°N, 109.1°E, magnetic 11°N), and this station is part of the Chinese Meridian Project. The dataset includes 469,213 ionospheric frequency-height maps generated during the period from 2010 to 2023. Among them, there are 387,668 frequency-height maps with a collection interval of 15 minutes and a spectral resolution of 0.025 MHz, and 44,358 frequency-height maps with a collection interval of 5 minutes and a spectral resolution of 0.05 MHz. These frequency-height maps are generated by the DPS-4D Digisonde ionosonde.

[0076] In addition, when training the VQ-VAE, its learning rate is set to 10 -4 , the batch size is set to 32, the codebook size is fixed at 512, the reconstruction loss uses MSE, and the balance coefficient is set to 0.25, and the number of training epochs is fixed at 20.

[0077] In the training of the time series prediction module, the time series prediction module adopts the Transformer structure and predicts the masked input through the prediction function to generate a time-frequency planning control signal, enabling the system to capture the dynamic changes and long-term dependencies between frames. The learning rate is set to 10 -4 , the batch size is set to 16, and the number of training epochs is set to 30.

[0078] In the training of the low-frequency spectral resolution ionogram generation module, a generative network with encoder-decoder compression is adopted to gradually denoise and reconstruct the low-frequency spectral data perturbed by noise to restore the basic structure and temporal coherence of the image. The learning rate is set to , the number of diffusion steps is set to 50, the masking ratio is set to 40%, the batch size is set to 16, and the number of training epochs is set to 50.

[0079] In the training of the high-frequency spectral resolution ionogram generation module, the residual between the low-frequency spectral resolution ionogram and the target high-frequency spectral resolution ionogram is calculated, noise perturbation is applied to the residual, and the residual is input into the of the ResShift module for residual prediction and correction to gradually complete the high-frequency details in the output. In this embodiment, the learning rate is set to , the number of diffusion steps is set to 20, the masking ratio is set to 40%, the batch size is set to 16, and the number of training epochs is set to 50.

[0080] In the joint training of the time series prediction module and the low-frequency spectral resolution ionogram generation module, the learning rate is set to , the number of diffusion steps is set to 50, the masking ratio is set to 40%, the batch size is set to 16, and the number of training epochs is set to 50.

[0081] In the joint training of the time series prediction module and the high-frequency spectral resolution ionogram generation module, the learning rate is set to , the number of diffusion steps is set to 20, the masking ratio is set to 40%, the batch size is set to 16, and the number of training epochs is set to 50.

[0082] The above specific settings are only for the preferred embodiments of the ionogram enhancement system based on time-frequency collaboration provided in the present invention. The present invention is not limited to the above parameter settings and can be appropriately adjusted according to specific situations, and all these are within the protection scope of the present invention.

[0083] In this article, specific examples are used to elaborate on the principles and implementation modes of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation modes and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. An ionogram enhancement method based on time-frequency collaboration, characterized in that Including the following steps: Step 1: Collect a sequence of ionograms, and determine whether it is a first ionogram sequence or a second ionogram sequence according to a time threshold and a spectral resolution threshold; Step 2: Use a pre-trained encoder to preprocess the first ionogram sequence to generate first low-frequency spectral image features, a first latent code, and a first mask template corresponding to the first ionogram sequence; or preprocess the second ionogram sequence to generate second low-frequency spectral image features, a second latent code, and a second mask template corresponding to the second ionogram sequence; Step 3: Use an improved context modeling method to process the first low-frequency spectral image features to obtain a first time-frequency planning control signal corresponding to the first ionogram sequence; or use an improved context modeling method to process the second low-frequency spectral image features to obtain a second time-frequency planning control signal corresponding to the second ionogram sequence; Step 4: Under the guidance of the first time-frequency planning control signal, process the first ionogram sequence to obtain a low-frequency spectral interpolation image; Step 5: Under the guidance of the time-frequency planning control signal, process the low-frequency spectral interpolation image or the second ionogram sequence based on an improved residual transfer repair strategy to obtain a high-frequency spectral resolution ionogram sequence; Step 6: Based on the high-frequency spectral resolution ionogram sequence and the mask template, perform super-resolution reconstruction on the low-frequency spectral interpolation image or the second ionogram sequence to obtain a target ionogram sequence.

2. The ionospheric frequency-height diagram enhancement method based on time-frequency cooperation according to claim 1, wherein Step 2 specifically includes: Downsample the first ionogram sequence to generate a low-frequency spectral image and a first mask template, use a pre-trained encoder to process the low-frequency spectral image to generate first low-frequency spectral image features, and use a pre-trained encoder to compress the first low-frequency spectral image features into a low-dimensional continuous latent space to obtain a first latent code; Or, upsample the second ionogram sequence to generate a second mask template, use a pre-trained encoder to process the second ionogram sequence to generate second low-frequency spectral image features, and use a pre-trained encoder to compress the second low-frequency spectral image features into a low-dimensional continuous latent space to obtain a second latent code.

3. The ionospheric frequency-height diagram enhancement method based on time-frequency cooperation according to claim 1, wherein Step 3 specifically includes: Perform context modeling on the first low-frequency spectral image features and the second low-frequency spectral image features respectively to capture the inter-frame dynamic changes and dependencies of the first low-frequency spectral image features and the second low-frequency spectral image features, and use the inter-frame dynamic changes and dependencies corresponding to the first low-frequency spectral image features to form a first time-frequency planning control signal, and use the inter-frame dynamic changes and dependencies corresponding to the second low-frequency spectral image features to form a second time-frequency planning control signal.

4. The ionospheric frequency-height diagram enhancement method based on time-frequency cooperation according to claim 1, characterized in that Step 4 specifically includes: Under the guidance of the first time-frequency planning control signal, apply noise frames to the first ionogram sequence, and reconstruct the first ionogram sequence with noise frames by gradually denoising to generate a low-frequency spectral interpolation image.

5. The ionospheric frequency-height diagram enhancement method based on time-frequency collaboration according to claim 4, characterized in that Step 4 also specifically includes: The adaptive instance normalization technique is used to integrate the diffusion step information in the reconstruction process into the conditional signal, and the first frequency-height map sequence is restored based on the conditional signal, so that the low-frequency spectrum interpolation image contains the detailed features in the first frequency-height map sequence. In addition, the cross-attention mechanism is used to fuse the global planning information, so that each frame of the low-frequency spectrum interpolation image is coherent in time and frequency spectrum.

6. The ionospheric frequency-height diagram enhancement method based on time-frequency cooperation according to claim 1, wherein Step 5 specifically includes: Under the guidance of the first time-frequency planning control signal, the diffusion model is used to gradually predict the residual between the low-frequency spectrum interpolation image and the target frequency-height map sequence, and the predicted residual is gradually corrected through a multi-step iterative diffusion process to obtain the high-frequency details in the low-frequency spectrum interpolation image. Based on the low-frequency spectrum interpolation image and the high-frequency details, a high-frequency spectrum resolution frequency-height map sequence is generated; Alternatively, under the guidance of the second time-frequency planning control signal, the diffusion model is used to gradually predict the residual between the second frequency-height map sequence and the target frequency-height map sequence, and the predicted residual is gradually corrected through a multi-step iterative diffusion process to obtain the high-frequency details in the second frequency-height map sequence. Based on the second frequency-height map sequence and the high-frequency details, a high-frequency spectrum resolution frequency-height map sequence is generated.

7. The ionospheric frequency-height diagram enhancement method based on time-frequency cooperation according to claim 1, characterized in that In step 1, a frequency-height map sequence with a sampling interval not less than the time threshold and a spectral resolution less than the spectral resolution threshold is used as the first frequency-height map sequence, and a frequency-height map sequence with a sampling interval less than the time threshold and a spectral resolution not less than the spectral resolution threshold is used as the second frequency-height map sequence.

8. An ionospheric frequency-height diagram enhancement system based on time-frequency collaboration, which implements an ionospheric frequency-height diagram enhancement method based on time-frequency collaboration as described in claim 1, is characterized in that, The system includes: An acquisition module, configured to acquire a frequency-height map sequence and determine it as the first frequency-height map sequence or the second frequency-height map sequence according to the time threshold and the spectral resolution threshold; A preprocessing module, connected to the acquisition module, configured to use a pre-trained encoder to preprocess the first frequency-height map sequence to generate the first low-frequency spectrum image feature, the first latent encoding, and the first mask template corresponding to the first frequency-height map sequence; or preprocess the second frequency-height map sequence to generate the second low-frequency spectrum image feature, the second latent encoding, and the second mask template corresponding to the second frequency-height map sequence; A time series prediction module, connected to the preprocessing module, configured to use an improved context modeling method to process the first low-frequency spectrum image feature to obtain the first time-frequency planning control signal corresponding to the first frequency-height map sequence; or use an improved context modeling method to process the second low-frequency spectrum image feature to obtain the second time-frequency planning control signal corresponding to the second frequency-height map sequence; A low-frequency spectrum resolution frequency-height map generation module, connected to the time series prediction module, configured to process the first frequency-height map sequence under the guidance of the first time-frequency planning control signal to obtain a low-frequency spectrum interpolation image; A high-frequency spectrum resolution frequency-height map generation module, connected to the time series prediction module and the high-frequency spectrum resolution frequency-height map generation module, configured to process the low-frequency spectrum interpolation image or the second frequency-height map sequence based on an improved residual transfer and repair strategy under the guidance of the time-frequency planning control signal to obtain a high-frequency spectrum resolution frequency-height map sequence; The frequency-height diagram fusion and reconstruction module, connected to the high-frequency spectral resolution frequency-height diagram generation module, is used to perform super-resolution reconstruction on the low-frequency spectral interpolation image or the second frequency-height diagram sequence based on the high-frequency spectral resolution frequency-height diagram sequence and the mask template to obtain the target frequency-height diagram sequence.

9. The ionospheric frequency-height diagram enhancement system based on time-frequency collaboration according to claim 8, characterized in that The system further includes a pre-training module; The pre-training module is connected to the preprocessing module, the time series prediction module, the low-frequency spectral resolution frequency-height diagram generation module, and the high-frequency spectral resolution frequency-height diagram generation module, and is used to pre-train the preprocessing module, the time series prediction module, the low-frequency spectral resolution frequency-height diagram generation module, and the high-frequency spectral resolution frequency-height diagram generation module, and is also used to jointly train the time series prediction module and the low-frequency spectral resolution frequency-height diagram generation module, as well as jointly train the time series prediction module and the high-frequency spectral resolution frequency-height diagram generation module.

10. The ionospheric frequency-height diagram enhancement system based on time-frequency collaboration according to claim 9, wherein The pre-training of the preprocessing module by the pre-training module includes: Set the preprocessing module to an 8-fold spatial compression rate and 16-channel latent dimension, and set the learning rate to 10 -4 , batch size of 32, and codebook size of 512; Set the reconstruction loss function of the preprocessing module as , where represents the first frequency-height map sequence or the second frequency-height map sequence, represents the frequency-height map sequence reconstructed and output by the preprocessing module; Set the codebook loss function of the preprocessing module to , where represents the latent representation generated by the preprocessing module for the first frequency-height map sequence or the second frequency-height map sequence, represents the nearest codebook vector, represents the stop gradient operation; Set the commitment loss function of the preprocessing module to ; Total loss function of the preprocessing module is expressed as , where represents the balance coefficient; When the curve of the total loss function of the preprocessing module converges, the pre-training of the preprocessing module is completed; The pre-training of the time series prediction module by the pre-training module includes: Set the reconstruction loss function of the time series prediction module as , where represents a binary function, indicating whether it is a mask, represents the prediction function adopted by the time series prediction module, represents the low-frequency spectral image features after mask processing, represents the true time-frequency planning information corresponding to the low-frequency spectral image features. The reconstruction loss is used to encourage the difference between the output of the time series prediction module and the true time-frequency planning information not to exceed a preset threshold, improving the accuracy of the time series prediction module in capturing inter-frame dynamic changes and dependencies; When the curve of the reconstruction loss function of the time series prediction module converges, the pre-training of the time series prediction module is completed; The pre-training of the low-frequency spectral resolution frequency-height diagram generation module by the pre-training module includes: Set the diffusion loss function of the low-frequency spectral resolution frequency-height map generation module to , where represents a binary mask represents the time step in the current diffusion process represents the conditional signal represents the time step of the low-frequency spectral part in the current diffusion process of the target residual represents the first sequence of frequency-height maps with added noise represents the low-frequency spectral resolution frequency-height map generation module, whose parameters are ; When the curve of the diffusion loss function of the low-frequency spectral resolution frequency-height diagram generation module converges, the pre-training of the low-frequency spectral resolution frequency-height diagram generation module is completed; The pre-training of the high-frequency spectral resolution frequency-height diagram generation module by the pre-training module includes: Set the diffusion loss function of the high-frequency spectral resolution frequency-height map generation module to , where represents a binary mask, represents the time step in the current diffusion process, represents the conditional signal, represents the time step of the high-frequency spectral part in the current diffusion process of the target residual, represents the residual between the interpolated low-frequency spectral image with added noise and the target frequency-height map sequence or the residual between the second frequency-height map sequence and the target frequency-height map sequence, represents the high-frequency spectral resolution frequency-height map generation module, whose parameters are ; When the curve of the diffusion loss function of the high-frequency spectral resolution frequency-height diagram generation module converges, the pre-training of the high-frequency spectral resolution frequency-height diagram generation module is completed; The joint training of the time series prediction module and the low-frequency spectral resolution frequency-height diagram generation module by the pre-training module includes: The masked diffusion loss function for the joint training of the time series prediction module and the low-frequency spectral resolution frequency-height map generation module is set to , where represents a binary mask, represents the time step in the current diffusion process, represents the time step of the low-frequency spectral part in the current diffusion process of the target residual, represents the low-frequency spectral resolution frequency-height map generation module, whose parameters are , represents the low-frequency spectral resolution data with added noise, represents the time-frequency planning control signal output by the time series prediction module, , represents the prediction function adopted by the time series prediction module, represents the masked low-frequency spectral image features; When the curve of the mask diffusion loss function converges, the joint training of the time series prediction module and the low-frequency spectral resolution frequency-height diagram generation module is completed; The joint training of the time series prediction module and the high-frequency spectral resolution frequency-height diagram generation module by the pre-training module includes: The loss function for the joint training of the time series prediction module and the high-frequency spectral resolution frequency-height map generation module is set as , where represents a binary mask, represents the time step in the current diffusion process, represents the time step of the high-frequency spectrum part in the current diffusion process of the target residual, represents the residual between the low-frequency spectrum interpolation image with added noise and the target frequency-height map sequence or the residual between the second frequency-height map sequence and the target frequency-height map sequence, represents the high-frequency spectral resolution frequency-height map generation module, whose parameters are , represents the time-frequency planning control signal output by the time series prediction module, , represents the prediction function used by the time series prediction module, represents the low-frequency spectrum image feature after mask processing; When the curve of the loss function of the joint training of the time series prediction module and the high-frequency spectral resolution frequency-height diagram generation module converges, the joint training of the time series prediction module and the high-frequency spectral resolution frequency-height diagram generation module is completed.

Images