A self-supervised radio-interferometric imaging method and system

Abstract

The application belongs to the field of comprehensive aperture radio array imaging detection, and relates to a self-supervised radio interference imaging method and system. The method comprises the following steps: constructing a radio interference training set based on antenna-dependent complex gain uncertainty modeling; training a radio interference imaging model by using measurement consistency constraints; then inputting uncalibrated observation data into the trained imaging model to realize brightness temperature image reconstruction; the radio interference imaging model comprises a fast imaging processing module, a gradient descent affine module and an encoding and decoding module; the uncalibrated observation data sequentially passes through the above modules to complete high-quality imaging reconstruction. The application can complete high-quality imaging of uncalibrated observation data without relying on special calibration hardware and calibration processing flow, without estimating data error parameters, and without relying on multiple iterations to realize image cleaning processing pipeline. While ensuring the imaging quality, the imaging speed is improved, and the applicability of the method in actual radio observation scenarios is improved.

Description

Technical Field

[0001] This application belongs to the field of integrated aperture radio array imaging detection, specifically involving a self-supervised radio interferometry imaging method and system for high-quality and rapid imaging of uncalibrated radio interferometry data. Background Technology

[0002] Interferometric imaging is an important technique for obtaining high-resolution sky brightness temperature distribution using aperture radio arrays. Its basic principle is to obtain the visibility function in the spatial frequency domain through the antenna array, sample the sky signal, and obtain the brightness temperature image of the observation area through inversion and reconstruction.

[0003] An ideal radio array system is noise-free and error-free. However, in practical radio array systems, due to factors such as the discrete parameters of each antenna and its receiving link, and manufacturing errors, it is difficult to guarantee ideal consistency in amplitude and phase, resulting in fixed deviations in the amplitude and phase of a single antenna and receiving link. In actual observations, temperature drift of the electronic system, gravity and temperature deformation of the antenna, and environmental factors lead to time-varying errors in the amplitude and phase of a single antenna and receiving link. These fixed deviations and time-varying errors propagate to the visibility domain through the complex correlation combination of antenna pairs, causing the observed visibility data to deviate from the ideal model, thus significantly affecting the imaging results.

[0004] To mitigate the impact of these errors, existing techniques typically rely on precise calibration to correct the observation data, restoring ideal visibility by estimating and compensating for the antenna's complex gain parameters. However, such methods usually require high signal-to-noise ratio observation conditions or additional calibration source information, and the calibration process is complex and computationally expensive. Under certain observation conditions, it is difficult to obtain stable and reliable parameter estimation results, thus limiting imaging performance.

[0005] On the other hand, deep learning-based radio interferometry (RFI) imaging methods have gradually developed in recent years, accelerating the imaging process by constructing a mapping model from observation data to images. However, existing methods mostly rely on high-quality labeled data for supervised training, requiring known real-world sky images as a reference, which are often difficult to obtain in actual observations. Furthermore, most methods assume that the observation data has been calibrated; for uncalibrated data containing errors, their imaging performance remains significantly limited. Therefore, how to achieve stable and effective RFI reconstruction from uncalibrated observation data containing antenna complex gain errors without relying on real labeled data has become a pressing technical problem in this field. Summary of the Invention

[0006] The purpose of this invention is to overcome the defects of the prior art and provide a self-supervised radio interferometry imaging method and system, which can realize the imaging reconstruction of radio interferometry observation data without calibration and without reference true value constraints.

[0007] To achieve the above objectives, this application proposes a self-supervised radio interferometry imaging method, comprising: Uncalibrated observation data is input into a trained radio interferometry imaging model, and the observation data is used to reconstruct the image to obtain a brightness temperature image. The radio interferometry imaging model includes: a fast imaging processing module, a gradient descent affine module, and an encoding / decoding module; wherein... The input to the fast imaging processing module is the projected baseline and uncalibrated visibility data, and the output is a dirty image; The gradient descent affine module takes the dirty image as input and outputs a pre-optimized image. The input to the encoding / decoding module is the preliminary optimized image, and the output is the final imaging result.

[0008] As an improvement to the above method, the processing procedure of the fast imaging processing module includes: weighting the visibility data, mapping the weighted visibility data to the image domain through non-uniform discrete inverse Fourier transform, and generating the corresponding dirty image. The gradient descent affine module's processing includes: a forward feature propagation process and a backward update process, both of which are composed of the same set of learnable convolutional kernel parameters, wherein: The forward feature propagation process includes: performing convolution operations on the dirty graph using cascaded convolution operators to obtain an intermediate representation related to the observed structure; The reverse update process includes: constructing a residual term based on the difference between the intermediate representation and the dirty graph, and backpropagating the residual term through the transposed convolution operator corresponding to the convolution operator to obtain the update term; The dirty image is corrected and updated based on the update terms to obtain a preliminary optimized image; The encoding / decoding module includes an encoder and a decoder, and its processing includes: The pre-optimized image is input into the encoder for feature extraction to obtain a multi-scale hierarchical feature representation. The decoder reconstructs multi-scale features step by step, and fuses the multi-scale feature information output by the encoder during the reconstruction process to output the final imaging result.

[0009] As an improvement to the above method, the gradient descent affine module includes a forward feature map structure composed of multiple cascaded two-dimensional convolution operators, a cascaded transposed convolution operator that shares parameters with the two-dimensional convolution operators, and a residual connection structure.

[0010] As an improvement to the above method, the encoder includes multiple downsampling layers, each of which contains multiple two-dimensional convolution operators, self-attention modules, and max pooling layers; The decoder includes multiple upsampling layers, each of which includes multiple two-dimensional convolution operators, a self-attention module, and a bilinear interpolation unit. The decoder receives the feature representation output from the last layer of the encoder and fuses the feature representation of the encoder at the corresponding scale during the reconstruction process of each upsampling layer.

[0011] As an improvement to the above method, the training set construction process of the radio interferometry imaging model includes: ideal visibility data generation, antenna-dependent error injection, and training sample construction, wherein: Ideal visibility data generation: Based on any given observation array, perform forward modeling of the sky brightness temperature distribution to generate the visibility function of the ideal array system and the corresponding projected baseline; Antenna-dependent error injection: Amplitude and phase perturbations are applied to the time-domain signals output by each antenna in the array and its receiving and processing link, while a baseline-dependent error factor is introduced. The perturbations are then propagated to the visibility domain through the antenna pair combination, thereby perturbing the visibility data and generating uncalibrated visibility data. Training sample construction: Pair uncalibrated visibility data with projected baseline data to generate radio interferometry training set.

[0012] As an improvement to the above method, the training process of the radio interferometry imaging model does not depend on the reference ground truth image. Instead, it constructs measurement consistency constraints, maps the reconstructed image output by the model to the visibility domain, and applies consistency constraints with the input uncalibrated visibility data to guide the optimization of model parameters.

[0013] As an improvement to the above method, the loss function used during the training of the radio interferometry imaging model is expressed as: ; in, Represents the total loss function. and These represent the amplitude and phase loss of the visibility function, respectively. Represents image domain loss; The amplitude and phase loss of the visibility function are expressed as: ; ; in, and These represent the amplitude and phase of the uncalibrated visibility function, respectively. and These represent the magnitude and phase of the ideal visibility function obtained by forward modeling the final imaging result output by the model; Represents the cosine loss function; This represents the mean absolute error loss function; Image domain loss is expressed as: ; in, This represents a dirty graph obtained from an uncalibrated visibility function; This represents the dirty map obtained by forward modeling and imaging inversion of the final imaging result of the model. This represents the mean squared error loss function.

[0014] This application also provides a self-supervised radio interferometry imaging system, implemented based on the above method, the system comprising: An imaging module is used to input uncalibrated observation data into a trained radio interferometry imaging model, and to perform imaging reconstruction on the observation data to obtain a brightness temperature image. The radio interferometry imaging model includes a fast imaging processing module, a gradient descent affine module, and an encoding / decoding module. The fast imaging processing module takes the projected baseline and uncalibrated visibility data as input and outputs a dirty image. The gradient descent affine module takes the dirty image as input and outputs a preliminary optimized image. The encoding / decoding module takes the preliminary optimized image as input and outputs the final imaging result. The model training module is used to construct a radio interferometry training set based on antenna-dependent complex gain uncertainty modeling, generate training data that conforms to the characteristics of uncalibrated observation data, and train the radio interferometry imaging model based on measurement consistency constraints.

[0015] Compared with existing technologies, the advantages of this application are: 1. Radio interferometry training data is constructed based on antenna-dependent complex gain uncertainty modeling. By applying amplitude and phase perturbations to each antenna, the training data can characterize the error propagation characteristics of the interferometric imaging system. 2. A self-supervised radio interferometry imaging model was designed, which can achieve imaging reconstruction under the condition of visibility function amplitude and phase error, thus improving the method's adaptability to real observation data; 3. Self-supervised training is achieved based on measurement consistency constraints, which does not require the reference ground truth image, avoids the dependence on high-quality labeled data, and improves the generalization of the method in various real observation scenarios. 4. Compared with traditional imaging methods that rely on calibration processing, the present invention can complete imaging reconstruction without precise calibration, reducing the dependence on the calibration process and error parameter estimation, and improving the processing speed. Attached Figure Description

[0016] Figure 1 The flowchart shown is of the proposed self-supervised radio interferometry imaging method; Figure 2 The image shows the original sky brightness temperature distribution, the projected baseline distribution, and the resulting dirty map in the dataset. Figure 3 The diagram shows the structure of the gradient descent affine module and the encoder / decoder module. Figure 4 The image shown is a schematic diagram of the imaging results. Detailed Implementation

[0017] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0018] like Figure 1 As shown, this invention proposes a self-supervised radio interferometry imaging method, which includes: Construct a radio interferometry training set based on antenna-dependent complex gain uncertainty modeling to generate training data that conforms to the characteristics of uncalibrated observation data; The radio interferometric imaging model is trained based on measurement consistency constraints, enabling the model to learn the mapping relationship between observation data and sky brightness temperature distribution; The observation data is input into the trained radio interferometry imaging model, and the observation data is used to reconstruct the image to obtain a brightness temperature image.

[0019] Specifically 1. The steps for constructing a radio interferometry training set include: Ideal visibility data generation, antenna-dependent error injection, and training sample construction.

[0020] Ideal visibility data generation: Based on the radio observation array configuration, determine the spatial distribution information of the effective antenna elements, and calculate the corresponding projected baseline according to the observation elevation and azimuth angles. Based on this, a non-uniform discrete Fourier transform is used to perform forward modeling of the brightness temperature distribution in the observation area to obtain visibility data for an ideal observation system. Antenna-dependent error injection: Random amplitude and phase perturbations are applied to the time-domain signals output from each antenna and its receiving and processing link at the array antenna layer to simulate the complex gain error of the antenna in actual observations; antenna x The complex gain can be expressed as:

[0021] in, Indicates antenna x amplitude disturbance, This indicates the corresponding phase error; For a baseline formed by antennas x and y, the visibility data after introducing antenna complex gain error can be expressed as: in For the visibility function of an ideal observation system, Visibility data to incorporate antenna complex gain error; Indicates antenna y The complex gain; the superscript * indicates that the conjugate is used for calculation.

[0022] Furthermore, to simulate baseline correlation errors in actual observations, a baseline dependency error factor is introduced. For the cross-correlation data simulated by antenna x and antenna y, the simulated uncalibrated visibility function is:

[0023] Therefore, the antenna-dependent complex gain error propagates to the visibility domain through the antenna pair combination, and together with the baseline dependency error factor, it perturbs the visibility amplitude and phase, generating uncalibrated visibility data. .

[0024] Training sample construction: uncalibrated visibility data Paired with the corresponding projection baseline data, a radio interferometry training set is generated.

[0025] 2. Radio interferometry imaging models include: Fast imaging processing, gradient descent affine module, and encoding / decoding module; among which, Fast imaging processing includes visibility weighting, meshing, and inverse Fourier transform operations. These operations are performed using a non-uniform discrete inverse Fourier transform to achieve the desired projection baseline. and the corresponding uncalibrated visibility data Fast image domain mapping, output dirty map ; The gradient descent affine module processing flow includes a forward feature propagation process and a backward update process. It consists of multiple cascaded two-dimensional convolution operators, cascaded transposed convolution operators sharing parameters with the two-dimensional convolution operators, and a residual connection structure. The forward feature propagation process includes: transferring the dirty map... The input is fed into a forward feature map structure consisting of multiple cascaded two-dimensional convolution operators. After layer-by-layer convolution operations, an intermediate representation related to the observed structure is obtained. ; The reverse update process includes: based on the intermediate representation With dirty pictures The difference is used to construct the residual term. and the residual term Input to a cascaded transposed convolution operator that shares parameters with the convolution operator to obtain the update term. Update item Weighting, and for dirty graphs Updated to obtain a preliminary optimized image. .

[0026] The encoding and decoding module includes encoding and decoding operations, implemented by an encoder and a decoder, respectively. The encoder and decoder consist of multiple two-dimensional convolution operators, a leaky linear rectified function, a self-attention module, a max-pooling layer, a bilinear interpolation function, and a residual structure. The encoder consists of multiple downsampling layers, each containing multiple two-dimensional convolution operators, self-attention modules, and max-pooling layers, used for initial optimization of the input image. Perform stepwise feature extraction to output feature representations at different scales. ; The decoder consists of multiple upsampling layers, each including multiple two-dimensional convolution operators, self-attention modules, and bilinear interpolation units; the decoder receives the feature representation output from the last layer of the encoder. Furthermore, the feature representations of the encoder at the corresponding scale are fused during the reconstruction process of each upsampling layer. This enables step-by-step reconstruction and outputs the final imaging result. .

[0027] 3. The training process based on measurement consistency constraints includes: The radio interferometry training set is input into the radio interferometry imaging model, and the model parameters are optimized to obtain the trained radio interferometry imaging model.

[0028] The loss function used during training is expressed as follows:

[0029] in Represents the total loss function. and These represent the amplitude and phase loss of the visibility function, respectively. This represents the image domain loss.

[0030] The amplitude and phase loss of the visibility function are expressed as:

[0031]

[0032] in and These represent the amplitude and phase of the uncalibrated visibility function, respectively. and This indicates the final imaging result output by the model. The amplitude and phase of the ideal visibility function are obtained by performing forward modeling. Represents the cosine loss function; This represents the mean absolute error loss function.

[0033] Image domain loss is expressed as:

[0034] in The dirty graph obtained from the uncalibrated visibility function. The final imaging result of the representative model The dirty map is obtained by performing forward modeling and then imaging inversion. This represents the mean squared error loss function.

[0035] 4. The observation data imaging process includes, In practical applications, there is no need to calibrate the observation data or estimate the error parameters. The uncalibrated observation data can be directly input into the trained radio interferometry imaging model, and the model can complete the mapping from the visibility domain to the image domain and output the corresponding brightness temperature image.

[0036] Example 1 Embodiments of the present invention provide a self-supervised radio interferometry imaging method, which is implemented according to the above-mentioned technical framework and the following steps: Step 1: Construction of Radio Interferometry Training Set In this embodiment, the Circular Array Solar Radio Imaging Telescope (DART) is used as the observation array, and its antenna array contains 313 antenna elements.

[0037] Ideal visibility data generation: Based on the DART antenna array configuration, during each data generation process, at least 80% of the antenna elements are randomly selected as effective antennas, while the remaining antenna elements are considered ineffective, to simulate the situation of an incomplete array in actual observation. 1400 images from the Galaxy dataset are selected as the brightness temperature distribution of the observed sky region. In this embodiment, a uniform scale is set for the brightness temperature images, with the image size set to [size missing]. Pixels, with a single pixel angular resolution set to 30 arcsec, thus corresponding to a field of view of . arcsec. For each brightness temperature distribution image Randomly select the corresponding observation direction, and calculate the projected baseline under that observation direction based on the selected effective antenna element information. Based on this, the non-uniform discrete Fourier transform (NUFFT) is used to analyze the brightness temperature distribution. Perform forward observation modeling to obtain the projected baseline. Corresponding ideal visibility data .

[0038] Antenna-dependent error injection: For each sample, a random perturbation is applied to each antenna of its observation array to simulate uncalibrated observation conditions. The complex gain of the antenna after applying the perturbation can be expressed as:

[0039] in, Indicates antenna x amplitude disturbance, This indicates the corresponding phase error. In this embodiment, both amplitude perturbation and phase perturbation are randomly generated using a Gaussian distribution, and their noise levels are randomly selected within a preset range. The standard deviation range of amplitude perturbation is set to 0.1dB~1dB, and the standard deviation range of phase perturbation is set to 0.1rad~1rad. In this embodiment, a random baseline deviation is used to characterize the baseline dependency error factor to simulate the impact of array geometric model errors on visibility data. The randomness is applied to the antenna... x and antenna y The baseline is constructed, and a random baseline offset error is further introduced into its spatial coordinates to simulate the array geometric model error. The baseline offset is randomly generated within the range of 0~1m, and the baseline dependency error factor is further obtained. Therefore, the simulated uncalibrated visibility data can be represented as follows:

[0040] in For the visibility function of an ideal observation system, To introduce visibility data for antenna complex gain error and baseline dependency error factor; By using the above method, the antenna-dependent complex gain error is propagated to the visibility domain and, together with the baseline-dependent error factor, perturbs the visibility amplitude and phase, thus obtaining uncalibrated visibility data containing amplitude and phase uncertainties. .

[0041] Training sample construction: uncalibrated visibility data With the corresponding projection baseline data The dataset is paired to generate a radio interferometry training set, which serves as the input data for the radio interferometry imaging model. In this embodiment, the generated dataset is divided into a training set, a validation set, and a test set in a 6:2:2 ratio.

[0042] Step 2: Rapid Imaging Processing In this embodiment, the visibility function is subjected to natural weighting. Subsequently, the weighted visibility data is gridded, and the visibility function is mapped to the image domain using a non-uniform discrete Fourier inverse transform function to generate an image of size [size missing]. Dirty image with pixel count and single-pixel angular resolution of 30 arcsec This serves as the input for subsequent modules. Figure 2 Example data from this embodiment is shown, including the original brightness temperature image, the corresponding projected baseline distribution, and the dirty map generated from the visibility data.

[0043] Preferably, the dirty image After being converted to real-valued representation, it is input into the subsequent processing module. In this embodiment, the magnitude of the dirty map is taken as the network input, and its dimension is... This reduces the complexity of processing complex data and adapts to the computational requirements of convolutional neural networks.

[0044] After obtaining the dirty graph with the real-valued representation, it is input into the gradient descent affine module for further optimization.

[0045] Step 3: Gradient Descent Affine Module Design In this embodiment, multiple cascaded convolution operators and their adjoint operators are used in combination with a residual connection structure to realize an image update process based on data consistency. The overall structure is as follows: Figure 3 As shown on the left.

[0046] The forward feature propagation process consists of a cascade of multiple two-dimensional convolution operators. In this embodiment, 16 operators of size 16 are preferably used. The convolution operators are cascaded together to form a dirty graph. The input is fed into the forward feature propagation structure, and after layer-by-layer convolution operations, an intermediate representation related to the observed structure is obtained. Its dimensions are This process can be represented as

[0047] in This represents the forward propagation process of features.

[0048] Based on intermediate representation With input dirty graph The difference between them is used to construct the residual term. Its dimensions are ,Right now:

[0049] Further, the residual term Input into the reverse update structure; obtain the updated item. Its dimensions are The reverse update structure is implemented by a transposed convolution that shares weights with the forward connection operator. This process can be represented as:

[0050] By introducing a learnable step size parameter For the update item The image is weighted and then corrected and updated against the input dirty image to obtain a preliminary optimized image. Its dimensions are The process is represented as:

[0051] Step 4: Encoding / Decoding Module Design In this embodiment, the encoding and decoding module adopts an encoder-decoder structure, which has a symmetrical multi-scale feature extraction and reconstruction structure, as shown in the figure below. Figure 3 As shown on the right.

[0052] The encoding / decoding module includes encoding and decoding operations, implemented by the encoder and decoder respectively. The encoder is used for initial image optimization. Multi-scale feature extraction is performed. In this embodiment, the encoder preferably uses a VGG16 network as the basic feature extraction backbone framework. Self-attention modules are introduced into each downsampling layer, enabling the network to model global spatial dependencies while extracting local convolutional features, thereby enhancing feature representation capabilities. Through progressive downsampling, multi-scale feature representations are obtained. The dimensions of each feature are as follows: , , , , The process is represented as:

[0053] The decoder is used to reconstruct the multi-scale features obtained in the encoding stage step by step. In this embodiment, the decoder consists of four cascaded upsampling layers. In each upsampling layer, the input features are first upsampled, preferably using bilinear interpolation. Then, they are fused with the features of the corresponding scale from the encoder. In this embodiment, feature fusion is achieved by element-wise addition. The fused features are then processed by convolution and a leaky linear rectified function, and a self-attention module is introduced during the convolution process to further enhance the feature representation capability, thereby obtaining the final output imaging result. The process is represented as:

[0054] Step 5: Optimize the imaging neural network In this embodiment, the radio interferometry training set constructed in step one is input into the radio interferometry imaging model, and the model parameters are optimized and trained until the preset convergence condition is met, thus obtaining the trained radio interferometry imaging model.

[0055] Specifically, for each training sample, the uncalibrated visibility data is first input into the imaging model. After processing by the fast imaging processing, gradient descent affine module, and encoding / decoding module, the final imaging result is obtained. .

[0056] Subsequently, the imaging results A forward modeling process is performed to map the image to the visibility domain, obtaining the corresponding predicted visibility function. Furthermore, the corresponding dirty image is obtained through imaging inversion. , used to construct measurement consistency constraints.

[0057] During training, the following loss function is used to optimize the model parameters:

[0058] in Represents the total loss function. and These represent the amplitude and phase loss of the visibility function, respectively. Represents image domain loss:

[0059]

[0060]

[0061] in and These represent the magnitude and phase of the uncalibrated visibility function, respectively. and This indicates the final imaging result output by the model. The magnitude and phase of the ideal visibility function obtained by forward modeling. The dirty graph obtained from the uncalibrated visibility function. The final imaging result of the representative model The dirty map is obtained by performing forward modeling and then imaging inversion.

[0062] During training, a gradient descent-based optimization method is used to minimize the loss function. In this embodiment, the Adam optimizer is preferably used to update the model parameters. Through continuous optimization of the loss function, the model satisfies consistency constraints in both the visibility domain and the image domain, thereby learning the mapping relationship between uncalibrated observation data and the sky brightness temperature distribution.

[0063] When the loss function converges or reaches the preset number of training rounds, training stops, and the trained radio interferometry imaging model is obtained. Step Six: Imaging Operation of Observation Data In this embodiment, after the radio interferometry imaging model is trained, the uncalibrated visibility data obtained from actual observations is input into the trained radio interferometry imaging model for rapid imaging reconstruction.

[0064] Specifically, the corresponding projection baseline is first calculated based on the observation array configuration and observation direction information. And acquire uncalibrated visibility data corresponding to the projected baseline. .

[0065] Subsequently, the projection baseline and uncalibrated visibility data The input radio interferometry imaging model is processed step-by-step through fast imaging, gradient descent affine module, and encoding / decoding module to obtain the final imaging result. .

[0066] Figure 4 Partial imaging results are presented under antenna amplitude error levels of 1 dB and phase error levels of 1 rad. The first row shows the reference brightness temperature image of the observed sky area, the second row shows the imaging result obtained using the method of this invention, and the third row shows the residual image between the two. The results demonstrate that the method of this invention can effectively recover target structure information while effectively suppressing imaging errors.

[0067] Table 1 presents a statistical performance comparison between the proposed method and existing brightness temperature image reconstruction methods (Multiscale CLEAN method, SARA method) at different antenna amplitude and phase error levels. The evaluation metrics include root mean square error (RMSE) and structure similarity index (SSIM). The statistical sample consists of 100 simulated images. The statistical metrics show that the proposed method outperforms existing brightness temperature image reconstruction methods at higher antenna error levels.

[0068] Table 1. Comparison of quantitative evaluation indicators for different methods under different perturbation levels

[0069] The advantages of this invention are as follows: it constructs radio interferometric training data based on antenna-dependent complex gain uncertainty modeling; it designs a self-supervised radio interferometric imaging model that can still achieve stable imaging reconstruction under the condition of visibility function amplitude and phase errors; it achieves self-supervised training based on measurement consistency constraints, which reduces the dependence on high-quality labeled data and improves the generalization of the method in actual observation scenarios; compared with traditional imaging methods that rely on precise calibration, this invention can complete imaging reconstruction without precise calibration, reducing the dependence on calibration process and error parameter estimation, and improving processing speed.

[0070] Example 2 This application also provides a self-supervised radio interferometry imaging system, implemented based on the above method, including: An imaging module is used to input observation data into a trained radio interferometry imaging model, perform imaging reconstruction on the observation data, and obtain a brightness temperature image. The radio interferometry imaging model includes a fast imaging processing module, a gradient descent affine module, and an encoding / decoding module. The fast imaging processing module takes the projected baseline and uncalibrated visibility data as input and outputs a dirty image. The gradient descent affine module takes the dirty image as input and outputs a preliminary optimized image. The encoding / decoding module takes the preliminary optimized image as input and outputs the final imaging result. The model training module is used to construct a radio interferometry training set based on antenna-dependent complex gain uncertainty modeling, generate training data that conforms to the characteristics of uncalibrated observation data, and train the radio interferometry imaging model based on measurement consistency constraints.

[0071] This application may also provide a computer device, including: at least one processor, memory, at least one network interface, and a user interface. The various components in this device are coupled together via a bus system. It is understood that the bus system is used to implement communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0072] The user interface can include a display, keyboard, or clicking device. Examples include a mouse, trackball, touchpad, or touchscreen.

[0073] It is understood that the memory in the embodiments disclosed in this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0074] In some implementations, the memory stores elements such as executable modules or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0075] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. Programs implementing the methods of the embodiments of this disclosure can be included in the application programs.

[0076] In the above embodiments, the processor can also invoke programs or instructions stored in memory, specifically programs or instructions stored in an application program, for the following purposes: Follow the steps described above.

[0077] The above methods can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above methods can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic diagrams disclosed above. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the disclosed methods can be directly implemented by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0078] It is understood that the embodiments described in this application can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or combinations thereof.

[0079] For software implementation, the technology of this application can be implemented by executing the functional modules (e.g., procedures, functions, etc.) of this application. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0080] This application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, it can implement the steps in the above method embodiments.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. A self-supervised radio interferometry imaging method, comprising: Uncalibrated observation data is input into a trained radio interferometry imaging model, and the observation data is used to reconstruct the image to obtain a brightness temperature image. The radio interferometry imaging model includes: a fast imaging processing module, a gradient descent affine module, and an encoding / decoding module; wherein... The input to the fast imaging processing module is the projected baseline and uncalibrated visibility data, and the output is a dirty image; The gradient descent affine module takes the dirty image as input and outputs a pre-optimized image. The input to the encoding / decoding module is the preliminary optimized image, and the output is the final imaging result; The training process for the radio interferometry imaging model includes: generating ideal visibility data, injecting antenna-dependent errors, and constructing training samples, wherein: Ideal visibility data generation: Based on any given observation array, perform forward modeling of the sky brightness temperature distribution to generate the visibility function of the ideal array system and the corresponding projected baseline; Antenna-dependent error injection: Amplitude and phase perturbations are applied to the time-domain signals output by each antenna in the array and its receiving and processing link, while a baseline-dependent error factor is introduced. The perturbations are then propagated to the visibility domain through the antenna pair combination, thereby perturbing the visibility data and generating uncalibrated visibility data. Training sample construction: Pair uncalibrated visibility data with projected baseline data to generate a radio interferometry training set; The training process of the radio interferometry imaging model does not rely on the reference ground truth image. Instead, it constructs measurement consistency constraints, maps the reconstructed image output by the model to the visibility domain, and applies consistency constraints with the input uncalibrated visibility data to guide the optimization of model parameters. Specifically, for each training sample, the uncalibrated visibility data is first input into the imaging model. After processing by the fast imaging processing, gradient descent affine module, and encoding / decoding module, the final imaging result is obtained. ; Subsequently, the imaging results A forward modeling process is performed to map the image to the visibility domain, obtaining the corresponding predicted visibility function. Furthermore, the corresponding dirty image is obtained through imaging inversion. , used to construct measurement consistency constraints; The loss function used during the training of the radio interferometry imaging model is expressed as follows: ； in, Represents the total loss function. and These represent the amplitude and phase loss of the visibility function, respectively. Represents image domain loss; The amplitude and phase loss of the visibility function are expressed as: ； ； in, and These represent the amplitude and phase of the uncalibrated visibility function, respectively. and These represent the magnitude and phase of the ideal visibility function obtained by forward modeling the final imaging result output by the model; Represents the cosine loss function; This represents the mean absolute error loss function; Image domain loss is expressed as: ； in, This represents a dirty graph obtained from an uncalibrated visibility function; This represents the dirty map obtained by forward modeling and imaging inversion of the final imaging result of the model. This represents the mean squared error loss function.

2. The self-supervised radio interferometry imaging method according to claim 1, characterized in that, The processing steps of the fast imaging processing module include: weighting the visibility data, mapping the weighted visibility data to the image domain through non-uniform discrete inverse Fourier transform, and generating the corresponding dirty image. The gradient descent affine module's processing includes: a forward feature propagation process and a backward update process, both of which are composed of the same set of learnable convolutional kernel parameters, wherein: The forward feature propagation process includes: performing convolution operations on the dirty graph using cascaded convolution operators to obtain an intermediate representation related to the observed structure; The reverse update process includes: constructing a residual term based on the difference between the intermediate representation and the dirty graph, and backpropagating the residual term through the transpose convolution operator corresponding to the convolution operator to obtain the update term; The dirty image is corrected and updated based on the update terms to obtain a preliminary optimized image; The encoding / decoding module includes an encoder and a decoder, and its processing includes: The pre-optimized image is input into the encoder for feature extraction to obtain a multi-scale hierarchical feature representation. The decoder reconstructs multi-scale features step by step, and fuses the multi-scale feature information output by the encoder during the reconstruction process to output the final imaging result.

3. The self-supervised radio interferometry imaging method according to claim 2, characterized in that, The gradient descent affine module includes a forward feature map structure composed of multiple cascaded two-dimensional convolution operators, a cascaded transposed convolution operator that shares parameters with the two-dimensional convolution operators, and a residual connection structure.

4. The self-supervised radio interferometry imaging method according to claim 2, characterized in that, The encoder includes multiple downsampling layers, each of which contains multiple two-dimensional convolution operators, a self-attention module, and a max pooling layer. The decoder includes multiple upsampling layers, each of which includes multiple two-dimensional convolution operators, a self-attention module, and a bilinear interpolation unit. The decoder receives the feature representation output from the last layer of the encoder and fuses the feature representation of the encoder at the corresponding scale during the reconstruction process of each upsampling layer.

5. A self-supervised radio interferometry imaging system, implemented based on the method of any one of claims 1-4, characterized in that, The system includes: An imaging module is used to input uncalibrated observation data into a trained radio interferometry imaging model, perform imaging reconstruction on the observation data, and obtain a brightness temperature image. The radio interferometry imaging model includes a fast imaging processing module, a gradient descent affine module, and an encoding / decoding module. The fast imaging processing module takes the projected baseline and uncalibrated visibility data as input and outputs a dirty image. The gradient descent affine module takes the dirty image as input and outputs a preliminary optimized image. The encoding / decoding module takes the preliminary optimized image as input and outputs the final imaging result. The model training module is used to construct a radio interferometry training set based on antenna-dependent complex gain uncertainty modeling, generate training data that conforms to the characteristics of uncalibrated observation data, and train the radio interferometry imaging model based on measurement consistency constraints.

Images