Low-frequency radio interferometer array real-time imaging method and system based on deep learning

Abstract

This invention discloses a real-time imaging method and system for low-frequency radio interferometer arrays based on deep learning. The method includes: constructing a solar radio burst simulation dataset for low-frequency interferometer arrays, wherein the solar radio burst simulation dataset includes a clean target image and a simulated dirty observation image; constructing a deep learning imaging model, and performing supervised training using the simulated dirty observation image from the solar radio burst simulation dataset to obtain a trained deep learning imaging model; and performing forward propagation on the real-time observational visibility data acquired by the target low-frequency interferometer array based on the trained deep learning imaging model to obtain a reconstructed high-fidelity radio image. This invention can shorten the imaging time and improve the fidelity of the imaging results. As a real-time imaging method and system for low-frequency radio interferometer arrays based on deep learning, this invention can be widely applied in the field of radio astronomy image processing technology.

Description

Technical Field

[0001] This invention relates to the field of radio astronomy image processing technology, and in particular to a real-time imaging method and system for low-frequency radio interferometer arrays based on deep learning. Background Technology

[0002] Solar radio bursts are important precursors to space weather events, and real-time, high-precision monitoring of them is crucial for space weather forecasting and early warning. Low-frequency radio radiation (e.g., 30-400 MHz) primarily originates from the solar corona and includes various types of radio bursts, from type I to type V. This radiation is closely related to physical processes in the corona, such as electron beams and shock waves, making it an ideal window for detecting coronal dynamics.

[0003] The MUSER-L low-frequency array, part of the related technology, operates in the 30-400 MHz range. Traditional imaging methods for MUSER-L and similar low-frequency interferometric arrays primarily rely on iterative deconvolution algorithms such as CLEAN and its variants. These methods generally suffer from the following inherent drawbacks: first, slow iterative convergence speed, high computational overhead, and long reconstruction time for a single image; second, sensitivity to radio frequency interference (RFI) and observation errors, easily producing artifacts and distortions in complex observation environments; and third, difficulty in achieving millisecond-level real-time imaging, failing to meet the real-time monitoring requirements of rapid transient phenomena such as solar flares. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a real-time imaging method and system for low-frequency radio interferometry arrays based on deep learning, which can shorten the imaging time and improve the fidelity of the imaging results.

[0005] The first technical solution adopted in this invention is: a real-time imaging method for low-frequency radio interferometry arrays based on deep learning, comprising the following steps:

[0006] A solar radio burst simulation dataset for low-frequency interferometric arrays is constructed, which includes a clean target map and a dirty simulated observation map.

[0007] A deep learning imaging model was constructed and supervised training was performed using simulated dirty maps from a solar radio burst simulation dataset to obtain the trained deep learning imaging model.

[0008] Based on the trained deep learning imaging model, the observation visibility data acquired in real time by the target low-frequency interferometric array is forward propagated to obtain the reconstructed high-fidelity radio image.

[0009] Furthermore, the step of constructing a solar radio burst simulation dataset for low-frequency interferometric arrays specifically includes:

[0010] Based on the layered structure model of the solar atmosphere, vertical distribution profiles are calculated and superimposed with burst source processing to generate ground truth images.

[0011] Convolve the ground truth image with the clean bundle to obtain the target clean image;

[0012] Based on the antenna layout of the target low-frequency interferometric array, the ground truth image is converted to the spatial frequency domain and noise simulation is applied to generate a simulated dirty image of observation.

[0013] A solar radio burst simulation dataset for low-frequency interferometric arrays is constructed by combining the target clean map and the simulated dirty map.

[0014] Furthermore, the step of calculating the vertical distribution profile and processing the superimposed burst sources based on the layered structure model of the solar atmosphere to generate a ground truth image specifically includes:

[0015] Based on the layered structure model of the solar atmosphere, the vertical distribution profile of brightness temperature with heliocentric distance in the low-frequency radio band is calculated to generate a two-dimensional tranquil solar background image.

[0016] Based on a two-dimensional tranquil solar background image, an explosion source modeled based on a parameterized two-dimensional elliptical Gaussian function is superimposed to simulate the brightness temperature distribution image of the explosion source with physical parameters.

[0017] The ground truth image is obtained by linearly superimposing a two-dimensional still solar background image with an image of the brightness temperature distribution of the eruption source.

[0018] Furthermore, the deep learning imaging model specifically includes an encoder, a decoder, an attention mechanism module, a multi-scale fusion module, and an output layer. The encoder introduces residual dense blocks, and the encoder and the decoder are connected by skip connections.

[0019] Furthermore, the step of constructing a deep learning imaging model and performing supervised training using simulated observation dirty maps from a solar radio burst simulation dataset to obtain a trained deep learning imaging model specifically includes:

[0020] Input the simulated dirty map of observations from the solar radio burst simulation dataset into the deep learning imaging model;

[0021] An encoder based on a deep learning imaging model extracts and reuses features from a simulated observation dirty map to obtain a simulated observation feature dirty map.

[0022] A decoder based on a deep learning imaging model performs channel stitching on the simulated observation feature dirty map to obtain the stitched simulated observation feature dirty map;

[0023] The attention mechanism module based on the deep learning imaging model performs a squeeze-excitation operation on the stitched simulated observation feature dirty map to obtain a weighted simulated observation feature dirty map.

[0024] A multi-scale fusion module based on a deep learning imaging model is used to perform multi-scale feature fusion processing on the weighted simulated observation dirty map to obtain a predicted clean map.

[0025] Based on the output layer of the deep learning imaging model, the predicted clean image is output, and it is evaluated against the target clean image using preset image quality evaluation metrics until the model converges and the trained deep learning imaging model is output.

[0026] Furthermore, the preset image quality evaluation indicators specifically include image generation time, peak signal-to-noise ratio, and structural similarity, wherein:

[0027] The mapping time refers to the time taken from simulating a dirty map to obtaining the target clean map;

[0028] The peak signal-to-noise ratio represents the ratio of signal power to noise power in the maximum power domain;

[0029] The structural similarity is a metric that measures the similarity between the simulated dirty map and the target clean map.

[0030] Furthermore, the expression for the loss function of the deep learning imaging model is as follows:

[0031]

[0032] In the above formula, This represents the loss function of the deep learning imaging model. This represents the mean squared error loss function. Represents the structural similarity loss function. Indicates the mean absolute error. , This represents the weighting coefficient.

[0033] The second technical solution adopted in this invention is: a real-time imaging system for low-frequency radio interferometry arrays based on deep learning, comprising:

[0034] The first module is used to construct a solar radio burst simulation dataset for low-frequency interferometric arrays. The solar radio burst simulation dataset includes a target clean map and a simulated observation dirty map.

[0035] The second module is used to build a deep learning imaging model. It combines simulated observation dirty maps from the solar radio burst simulation dataset for supervised training to obtain the trained deep learning imaging model.

[0036] The third module is used to forward propagate the real-time observation visibility data of the target low-frequency interferometric array based on the trained deep learning imaging model to obtain the reconstructed high-fidelity radio image.

[0037] The beneficial effects of the method and system of this invention are as follows: This invention constructs a solar radio burst simulation dataset for low-frequency interferometric arrays. The solar radio burst simulation dataset includes a clean target image and a simulated dirty image. The simulation dataset is constructed based on the observation characteristics of low-frequency interferometric arrays (such as MUSER-L) and the physical characteristics of solar radio bursts, making the model highly adaptable and robust to the target application scenario. Then, a deep learning imaging model is constructed and trained in a supervised manner using the simulated dirty image in the solar radio burst simulation dataset to obtain the trained deep learning imaging model. By training on a large-scale simulation dataset, the model can learn to effectively extract features from dirty images with complex noise and UV coverage undersampling, and suppress artifacts and sidelobes to reconstruct images with clear structure and high fidelity. Finally, based on the trained deep learning imaging model, the observation visibility data acquired in real time by the target low-frequency interferometric array is forward propagated to obtain the reconstructed high-fidelity radio image, which can shorten the imaging time and improve the fidelity of the imaging results. Attached Figure Description

[0038] Figure 1 This is a flowchart of the steps of the real-time imaging method for low-frequency radio interferometer arrays based on deep learning according to the present invention;

[0039] Figure 2 This is a structural block diagram of the low-frequency radio interferometer array real-time imaging system based on deep learning according to the present invention;

[0040] Figure 3 This is a schematic diagram illustrating the simulation process of the dirty graph in the simulation dataset provided in a specific embodiment of the present invention;

[0041] Figure 4 This is a schematic diagram illustrating the simulation process of a clean graph in a simulation dataset provided in a specific embodiment of the present invention;

[0042] Figure 5 This is a schematic diagram of a deep learning network structure provided in a specific embodiment of the present invention;

[0043] Figure 6 This is a schematic diagram of the imaging effect provided in a specific embodiment of the present invention. Detailed Implementation

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0046] Reference Figure 1 This invention provides a real-time imaging method for low-frequency radio interferometer arrays based on deep learning, the method comprising the following steps:

[0047] S100. Construct a solar radio burst simulation dataset for low-frequency interferometric arrays, wherein the solar radio burst simulation dataset includes a target clean map and a simulated observation dirty map;

[0048] S110. Based on the layered structure model of the solar atmosphere, vertical distribution profiles are calculated and superimposed burst source processing is performed to generate true image.

[0049] Specifically, based on the layered structure model of the solar atmosphere, the vertical distribution profile of brightness temperature with heliocentric distance in the low-frequency radio band is calculated to generate a two-dimensional quiet solar background image. Based on the two-dimensional quiet solar background image, an eruption source modeled based on a parameterized two-dimensional elliptical Gaussian function is superimposed to simulate an image of the brightness temperature distribution of the eruption source with physical parameters. The two-dimensional quiet solar background image and the image of the brightness temperature distribution of the eruption source are linearly superimposed to obtain the ground truth image.

[0050] It should be noted that the vertical distribution profile of the brightness temperature was calculated by combining the thermal bremsstrahlung mechanism. Additionally, the peak intensity of the burst source is... K to The spectral characteristics of the burst source, taking values ​​within the range of K, satisfy the power-law relationship between brightness temperature and frequency.

[0051] In this embodiment, a quiet solar brightness temperature distribution map based on a solar atmospheric physics model is constructed: Based on the layered structure model of the solar atmosphere, the vertical distribution profile of brightness temperature with heliocentric distance in the low-frequency radio band is calculated, thereby generating a two-dimensional quiet solar background image that conforms to physical laws. A solar radio burst source with physical parameters is simulated: On the quiet solar background, a burst source modeled based on a parameterized two-dimensional elliptical Gaussian function is superimposed. Its key parameters such as intensity, size, shape, and spectral index are set according to the statistical analysis of historical observation data and the radio burst physics model, thereby generating a high-fidelity true value image. The simulated burst source brightness temperature distribution is linearly superimposed with the quiet solar background brightness temperature distribution to form the final true value image.

[0052] In some specific embodiments, firstly, a brightness temperature distribution map of the quiet solar background is constructed: based on a layered structure model of the solar atmosphere, combined with the radio emission mechanism (mainly thermal bremsstrahlung) in the low-frequency radio band (150 MHz in this embodiment), the brightness temperature is calculated. With the distance of the sun's center Vertical distribution profile of the variation in distance from the center of the sun On a two-dimensional image plane (256x256 pixels in this embodiment), for each pixel... Calculate its heliocentric distance ,in The coordinates are the center coordinates of the solar disk. Substitute the outline This will give you the brightness temperature of the tranquil sun background corresponding to that pixel. This generates a physically consistent and tranquil image of the sun.

[0053] Furthermore, a simulated solar radio burst source was superimposed onto a quiescent solar background. The morphology of the burst source was modeled using a parameterized two-dimensional elliptic Gaussian function, the expression of which is:

[0054]

[0055] in, The eruption center is located at a random position on the solar disk; The peak brightness temperature increment is calculated based on observational statistics of different radio burst types (e.g., Type II and Type III bursts). K to Randomly selected from the range K; and These are the semi-major and semi-minor axes of the Gaussian ellipse, used to simulate the spatial expansion dimensions of the burst; The direction angle of the principal axis of the ellipse.

[0056] Finally, a ground truth image is generated, showing the simulated brightness temperature distribution of the burst source. Bright temperature distribution against a tranquil solar background Linear superposition is performed to generate a ground truth image containing the burst events. Its expression is:

[0057]

[0058] In the above formula, This represents a high-fidelity true value image.

[0059] like Figure 3As shown, based on the actual antenna coordinates of MUSER-L at 150 MHz, its UV coverage and dirty beam PSF are calculated, and a clean beam is obtained by Gaussian fitting of the PSF. The model visibility is obtained by projecting it onto the UV plane using a Fourier transform. ;exist Adding Gaussian system noise yields the simulated observation visibility. ;right Perform inverse Fourier transform to generate a dirty map. .

[0060] S120. Convolve the ground truth image with the clean bundle to obtain the target clean image;

[0061] In this embodiment, as Figure 4 As shown, the highest resolution image obtainable by the telescope is obtained by convolving the ground truth image with the clean bundle, which is called the clean image.

[0062] S130. Based on the antenna layout of the target low-frequency interferometric array, the true image is converted to the spatial frequency domain and noise simulation is applied to generate a simulated observation dirty map.

[0063] In this embodiment, based on the antenna layout and observation frequency of the target low-frequency interferometric array (such as MUSER-L), the generated ground truth (GT) image is converted to the spatial frequency domain (UV plane) to generate corresponding model visibility data. Simulated observation visibility data is generated by applying noise simulation and UV coverage undersampling. Finally, the simulated observation visibility data is subjected to inverse fast Fourier transform (IFFT) to generate a "dirty image" with sidelobes, artifacts, and noise, which serves as the network input data. Clean and dirty images are paired to form the training sample set.

[0064] S140. Construct a solar radio burst simulation dataset for low-frequency interferometric arrays by combining the target clean map and the simulated dirty map.

[0065] In this embodiment, based on the solar atmospheric physics model, a vertical distribution profile of brightness temperature varying with heliocentric distance in the low-frequency radio band is constructed, and a quiet solar background brightness temperature image is generated accordingly. A parameterized two-dimensional elliptic Gaussian function is used to simulate a solar radio burst source on the quiet solar background brightness temperature image, generating a burst source brightness temperature distribution image. Parameters include the burst center location, peak intensity, spatial spanning size, and azimuth angle. The burst source brightness temperature distribution image is superimposed with the quiet solar background brightness temperature image to generate a high-fidelity ground truth image. This ground truth image is then convolved with a clean beam to obtain the target clean image. Based on the antenna layout of the target low-frequency interferometric array, the ground truth image is converted to the spatial frequency domain and noise simulation is applied to generate a corresponding simulated observation dirty image. The target clean image and the simulated observation dirty image are paired to form a training sample set.

[0066] More specifically, generate 20,000 pairs ( Image pairs were divided into training, validation, and test sets in an 8:1:1 ratio. All images were normalized.

[0067] S200. Construct a deep learning imaging model and perform supervised training by combining simulated observation dirty maps in the solar radio burst simulation dataset to obtain the trained deep learning imaging model.

[0068] First, it should be noted that the deep learning imaging model specifically includes an encoder, a decoder, an attention mechanism module, a multi-scale fusion module, and an output layer. The encoder introduces residual dense blocks, and the encoder and the decoder are connected by skip connections.

[0069] The composite loss function of a deep learning imaging model is expressed as:

[0070]

[0071] in This is the mean squared error loss, used to constrain pixel-level accuracy; Structural similarity loss is used to preserve the structural information of the image; The mean absolute error is used to provide additional regularization constraints. and Weighting coefficients

[0072] In this embodiment, a deep neural network model based on an enhanced U-Net architecture is constructed. This model, building upon the classic U-Net encoder-decoder structure, introduces residual dense blocks to enhance feature extraction and reuse capabilities, incorporates a channel attention module to adaptively calibrate the importance of feature channels, and employs a multi-scale feature fusion strategy to comprehensively utilize information at different levels. The encoder progressively downsamples through convolution and pooling operations to extract multi-scale high-level semantic features of the image; the decoder progressively restores spatial resolution through upsampling and convolution operations. Skip connections are used to pass high-resolution detail features from the encoder to the decoder to reconstruct accurate image structure. The model output is the predicted "clean" image, where an adaptive moment estimation (ADAM) optimizer is used for end-to-end training until the model converges.

[0073] Specifically, simulated dirty images from a solar radio burst simulation dataset are input into a deep learning imaging model. The encoder of the deep learning imaging model extracts and reuses features from the simulated dirty images to obtain simulated dirty image features. The decoder of the deep learning imaging model performs channel stitching on the simulated dirty image features to obtain a stitched simulated dirty image features. The attention mechanism module of the deep learning imaging model performs a squeeze-excitation operation on the stitched simulated dirty image features to obtain a weighted simulated dirty image features. The multi-scale fusion module of the deep learning imaging model performs multi-scale feature fusion on the weighted simulated dirty image features to obtain a predicted clean image. The output layer of the deep learning imaging model outputs the predicted clean image and evaluates it against a target clean image using preset image quality evaluation metrics until the model converges, outputting the trained deep learning imaging model.

[0074] In some specific embodiments, a DA-UNet imaging network is constructed: [structures are described in the original text]. Figure 5 The network shown is enhanced based on the U-Net framework. Specifically, it includes:

[0075] 1) Encoder: Contains 4 downsampling stages. The core of each stage consists of a Residual Dense Block (RDB), which contains multiple densely connected 5x5 convolutional layers. The residual connections promote feature reuse and enhance the feature extraction capability for complex sidelobes and noisy patterns. Each RDB is followed by a 2x2 max pooling layer for downsampling.

[0076] 2) Decoder: Symmetrically contains 4 upsampling stages. Each stage first uses a 2x2 transposed convolution for upsampling, and then concatenates the channels with the feature map of the corresponding scale of the encoder through "skip connections".

[0077] 3) Attention Mechanism: A ChannelAttention Module is introduced on the feature fusion path after the skip connections. This module adaptively learns and weights the importance of different feature channels through a squeeze-excitation operation, making the network pay more attention to key features related to celestial structure reconstruction and suppressing noisy channels.

[0078] 4) Multi-scale fusion: At the end of the decoder, feature maps from different decoding stages are fused at multiple scales to make comprehensive use of information from global context to local details.

[0079] 5) Output layer: The last layer uses a 1x1 convolution to map the number of channels to 1, outputting a single-channel prediction clean map.

[0080] Further adjust the parameters of the composite loss function. , The optimizer used is ADAM, with a learning rate of 5e-4. The model is trained using the training set, with a batch size of 32 and epochs of 200. The loss function is monitored on the validation set, and training is stopped early when the loss stops decreasing. Finally, the weights of the best-performing model are saved.

[0081] It should also be noted that after training the neural network model, image quality is evaluated using metrics such as image generation time t, peak signal-to-noise ratio, and structural similarity.

[0082] 1) The time t for generating the image is the time taken from a dirty image to a clean image;

[0083] 2) Peak signal-to-noise ratio (PSNR) is the ratio of signal power to noise power in the maximum power domain, and its mathematical expression is:

[0084]

[0085]

[0086] In the above formula, A clean astronomical chart. Indicates a dirty image. Indicates the size of the astronomical chart;

[0087] 3) Structural similarity is an indicator that measures the similarity between two images. It is based on a comparison of brightness, contrast, and structure between clean and dirty images.

[0088]

[0089] In the above formula, Indicates the pixel position of the image. It measures the similarity in average intensity between the output image and the target image. The term represents a comparison function used to evaluate the similarity of pixel values ​​within a dynamic range. The term is a structural comparison function that measures the pattern similarity of pixel values ​​and is related to the spatial distribution of features in the image.

[0090] S300 uses a trained deep learning imaging model to forward propagate the real-time observation visibility data acquired by the target low-frequency interferometric array to obtain a reconstructed high-fidelity radio image.

[0091] In this embodiment, the real-time observation visibility data acquired by the target low-frequency interferometric array is preprocessed in the same way as the training data (such as format conversion and normalization) and then directly input into the trained deep learning imaging model. The model outputs the reconstructed high-fidelity radio image directly in a single forward inference in a forward propagation manner, thereby achieving millisecond-level real-time imaging.

[0092] It should be noted that the target low-frequency interferometric array is the low-frequency array of the Mingantu Radio Spectroheliograph, which operates in the 30-400 MHz frequency band.

[0093] In some specific embodiments, the trained DA-UNet model is deployed on a graphics processing unit (GPU) server. When MUSER-L performs observations, the data acquisition system packages the visibility data for each integration period and sends it to the imaging server. The server performs rapid format conversion, UV meshing, and image generation (generating a dirty map) on the received visibility data, and performs the same normalization operations as on the training data. The preprocessed dirty map is then input into the loaded U-Net model. The model performs one forward propagation on the GPU (taking approximately 2 milliseconds), directly outputting the reconstructed high-resolution solar radio image. Figure 6 As shown, the astronomical image obtained using the above method can be displayed or stored in real time for scientists to monitor and analyze solar flares.

[0094] On the independent test set, the method in this embodiment achieved an average PSNR of 37.54 dB and an average SSIM of 0.982, significantly higher than the results of the traditional CLEAN algorithm (PSNR ~22.46 dB, SSIM ~0.89). The total processing time for a single image is less than 5 milliseconds, fully meeting the real-time requirements.

[0095] In summary, the present invention has the following advantages compared with the prior art:

[0096] 1) Extremely fast imaging speed: By using a trained deep learning model for forward inference, the traditional CLEAN algorithm's multiple iterations are replaced, reducing imaging time from seconds to milliseconds, and truly realizing real-time monitoring capabilities for rapid phenomena such as solar flares.

[0097] 2) High imaging quality and strong anti-interference ability: By training on a large-scale simulation dataset, the model can learn to effectively extract features from dirty images with complex noise and undersampled UV coverage, and suppress artifacts and sidelobes to reconstruct images with clear structure and high fidelity. The design of the composite loss function further ensures a balance between pixel accuracy and structural integrity.

[0098] 3) Highly targeted and with good generalization ability: This invention specifically constructs a simulation dataset based on the observation characteristics of low-frequency interferometric arrays (such as MUSER-L) and the physical characteristics of solar radio bursts, making the model highly adaptable and robust to the target application scenarios.

[0099] 4) End-to-end automation: This method achieves end-to-end mapping from raw visibility data to the final scientific image without the need for complex parameter tuning by human intervention, which lowers the threshold for use and improves the automation level of the imaging process.

[0100] Reference Figure 2 A real-time imaging system for low-frequency radio interferometer arrays based on deep learning, including:

[0101] The first module 201 is used to construct a solar radio burst simulation dataset for low-frequency interferometric arrays. The solar radio burst simulation dataset includes a target clean map and a simulated observation dirty map.

[0102] The second module 202 is used to construct a deep learning imaging model, and to perform supervised training by combining simulated observation dirty maps in the solar radio burst simulation dataset to obtain the trained deep learning imaging model.

[0103] The third module 203 is used to forward propagate the real-time observation visibility data of the target low-frequency interferometric array based on the trained deep learning imaging model to obtain the reconstructed high-fidelity radio image.

[0104] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0105] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A real-time imaging method for low-frequency radio interferometry arrays based on deep learning, characterized in that, Includes the following steps: A solar radio burst simulation dataset for low-frequency interferometric arrays is constructed, which includes a clean target map and a dirty simulated observation map. A deep learning imaging model was constructed and supervised training was performed using simulated dirty maps from a solar radio burst simulation dataset to obtain the trained deep learning imaging model. Based on the trained deep learning imaging model, the observation visibility data acquired in real time by the target low-frequency interferometric array is forward propagated to obtain the reconstructed high-fidelity radio image.

2. The real-time imaging method for low-frequency radio interferometer arrays based on deep learning according to claim 1, characterized in that, The step of constructing a solar radio burst simulation dataset for low-frequency interferometry arrays specifically includes: Based on the layered structure model of the solar atmosphere, vertical distribution profiles are calculated and superimposed with burst source processing to generate ground truth images. Convolve the ground truth image with the clean bundle to obtain the target clean image; Based on the antenna layout of the target low-frequency interferometric array, the ground truth image is converted to the spatial frequency domain and noise simulation is applied to generate a simulated dirty image of observation. A solar radio burst simulation dataset for low-frequency interferometric arrays is constructed by combining the target clean map and the simulated dirty map.

3. The real-time imaging method for low-frequency radio interferometer arrays based on deep learning according to claim 2, characterized in that, The step of calculating the vertical distribution profile and processing the superimposed burst sources based on the layered structure model of the solar atmosphere to generate a ground truth image specifically includes: Based on the layered structure model of the solar atmosphere, the vertical distribution profile of brightness temperature as a function of heliocentric distance in the low-frequency radio band is calculated, and a two-dimensional tranquil solar background image is generated. Based on a two-dimensional tranquil solar background image, an explosion source modeled based on a parameterized two-dimensional elliptical Gaussian function is superimposed to simulate an image of the brightness temperature distribution of the explosion source with physical parameters. The ground truth image is obtained by linearly superimposing a two-dimensional still solar background image with an image of the brightness temperature distribution of the eruption source.

4. The real-time imaging method for low-frequency radio interferometer arrays based on deep learning according to claim 3, characterized in that, The deep learning imaging model specifically includes an encoder, a decoder, an attention mechanism module, a multi-scale fusion module, and an output layer. The encoder introduces residual dense blocks, and the encoder and the decoder are connected by skip connections.

5. The real-time imaging method for low-frequency radio interferometer arrays based on deep learning according to claim 4, characterized in that, The step of constructing a deep learning imaging model and performing supervised training using simulated dirty maps from a solar radio burst simulation dataset to obtain a trained deep learning imaging model specifically includes: Input the simulated dirty map of observations from the solar radio burst simulation dataset into the deep learning imaging model; An encoder based on a deep learning imaging model extracts and reuses features from a simulated observation dirty map to obtain a simulated observation feature dirty map. A decoder based on a deep learning imaging model performs channel stitching on the simulated observation feature dirty map to obtain the stitched simulated observation feature dirty map; The attention mechanism module based on the deep learning imaging model performs a squeeze-excitation operation on the stitched simulated observation feature dirty map to obtain a weighted simulated observation feature dirty map. A multi-scale fusion module based on a deep learning imaging model is used to perform multi-scale feature fusion processing on the weighted simulated observation dirty map to obtain a predicted clean map. Based on the output layer of the deep learning imaging model, the predicted clean image is output, and it is evaluated against the target clean image using preset image quality evaluation metrics until the model converges and the trained deep learning imaging model is output.

6. The real-time imaging method for low-frequency radio interferometer arrays based on deep learning according to claim 5, characterized in that, The preset image quality evaluation indicators specifically include image generation time, peak signal-to-noise ratio, and structural similarity, wherein: The mapping time refers to the time taken from simulating a dirty map to obtaining the target clean map; The peak signal-to-noise ratio represents the ratio of signal power to noise power in the maximum power domain; The structural similarity is a metric that measures the similarity between the simulated dirty map and the target clean map.

7. The real-time imaging method for low-frequency radio interferometer arrays based on deep learning according to claim 6, characterized in that, The expression for the loss function of the deep learning imaging model is as follows: In the above formula, This represents the loss function of the deep learning imaging model. This represents the mean squared error loss function. Represents the structural similarity loss function. Indicates the mean absolute error. , This represents the weighting coefficient.

8. A real-time imaging system for low-frequency radio interferometry arrays based on deep learning, characterized in that, Includes the following modules: The first module is used to construct a solar radio burst simulation dataset for low-frequency interferometric arrays. The solar radio burst simulation dataset includes a target clean map and a simulated observation dirty map. The second module is used to build a deep learning imaging model. It combines simulated observation dirty maps from the solar radio burst simulation dataset for supervised training to obtain the trained deep learning imaging model. The third module is used to forward propagate the real-time observation visibility data of the target low-frequency interferometric array based on the trained deep learning imaging model to obtain the reconstructed high-fidelity radio image.

Images