Baseline separation radio interferometric imaging method and system based on a heterogeneous architecture
By employing a baseline-separated radio interferometry imaging method with a heterogeneous architecture, the memory-computation contradiction in wide-field-of-view radio interferometry imaging is resolved, achieving load balancing and efficient data processing, improving imaging accuracy and robustness, and adapting to complex scenarios with multiple frequency bands and wide fields of view.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU UNIVERSITY
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies suffer from a memory-computation contradiction in wide-field radio interferometry imaging. GPUs struggle to accommodate large grid requirements, while CPUs become bottlenecks in handling high-throughput data streams, leading to load imbalances and hindering efficient data processing.
A baseline-separated radio interferometry imaging method based on heterogeneous architecture is adopted. By allocating CPU-GPU load, the data is segmented into short baselines and long baselines, and different processing steps are performed on the GPU and CPU respectively. Combined with the PSWF anti-aliasing model and W-Stacking algorithm, efficient image generation and merging are achieved.
It achieves stable GPU memory utilization and CPU computing power utilization, reduces imaging time difference, reduces memory usage, avoids memory overflow, improves imaging accuracy and robustness, and adapts to the complex scene requirements of multi-band and wide field of view.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the field of radio interferometry imaging technology, and in particular to a baseline separation radio interferometry imaging method and system based on a heterogeneous architecture. Background Technology
[0002] With the construction of next-generation radio interferometer arrays, the sensitivity and resolution of astronomical observation data have reached unprecedented levels, but this has also brought massive data processing challenges. The basic mathematical relationships of radio interferometric imaging are described by measurement equations, and recovering the sky brightness distribution from visibility data usually requires an inverse Fourier transform. Due to the excessive computational cost of direct Fourier transform, traditional imaging paradigms typically employ convolutional gridding to interpolate visibility onto a regular grid in order to use the Fast Fourier Transform (FFT).
[0003] For wide-field-of-view arrays, the W-term effect introduced by non-coplanar baselines makes simple two-dimensional inversion impossible, requiring algorithms such as W-Projection or W-Stacking. However, these wide-field-of-view imaging algorithms typically require huge grid sizes to cover the entire field of view and perform W-term corrections. This leads to a memory-computation contradiction: while GPUs are suitable for high-throughput parallel computing, their capacity is limited by high-bandwidth memory (HBM), making it difficult to accommodate the huge grids required for wide-field-of-view imaging; while CPUs can handle large memory tasks with high accuracy, their throughput becomes a bottleneck when faced with SKA-level data streams.
[0004] Therefore, there is an urgent need for an imaging framework that can effectively decouple the requirements of large grids from high-throughput data processing in order to achieve load balancing on heterogeneous architectures. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a baseline-separated radio interferometry imaging method and system based on a heterogeneous architecture, in order to achieve optimal CPU-GPU load distribution and improve imaging accuracy and computational efficiency in complex scenes with multiple frequency bands and wide field of view.
[0006] In a first aspect, the present invention provides a baseline-separated radio interferometry imaging method based on a heterogeneous architecture, comprising the following steps:
[0007] S1) Extract visibility data from the observation data of the radio interferometer array and extract observation feature parameters;
[0008] S2) Construct a heterogeneous load balancing model, input the observed feature parameters into the heterogeneous load balancing model, calculate the baseline length threshold through constraints, and divide the visibility data into short baseline datasets and long baseline datasets according to the baseline length threshold.
[0009] S3) Based on the spatial frequency characteristics of short baseline data, a three-dimensional mesh is dynamically allocated on the GPU, and convolutional meshing and inverse Fourier transform processing are performed. Combined with W-term correction, a short baseline image is generated.
[0010] S4) Based on the grid parameters synchronized by the GPU, a matching minimized 3D grid is allocated on the CPU, and the W-Stacking algorithm is used to perform gridding processing to generate a long baseline image that retains high-resolution details;
[0011] S5) Construct a multi-parameter PSWF anti-aliasing model, adjust the order and cutoff radius parameters of PSWF according to the observation frequency band and field of view size, perform regional anti-aliasing resampling on short and long baseline images, and achieve linear image superposition through frequency domain alignment and field of view cropping to obtain the target sky image;
[0012] S6) Based on the grid parameters and heterogeneous processing strategy of the imaging stage, the three-dimensional demeshing of short baseline data is performed on the GPU and the demeshing of long baseline data is performed on the CPU. The demeshing data is merged in real time through the heterogeneous data interaction channel to obtain the predicted visibility data.
[0013] Preferably, in step S2), the baseline length threshold is obtained by inputting the observed characteristic parameters into the heterogeneous load balancing model and applying constraints. The multivariate fitting function is expressed as:
[0014] ;
[0015] In the formula, These are the fitting coefficients; For observation frequency band; The size of the field of view; Baseline distribution density; GPU memory utilization; This refers to GPU computing power utilization. This represents the CPU computing power utilization rate.
[0016] Preferably, in step S3), the convolutional meshing is as follows:
[0017]
[0018] In the formula, The resulting 3D mesh is generated through convolutional meshing. For the first Visibility values of short baselines; For the first Baseline weights for short baselines; This is the convolution kernel function.
[0019] Preferably, in step S3), an inverse Fourier transform is performed on the convolutionally meshed 3D mesh data. The inverse Fourier transform is expressed as:
[0020] ;
[0021] In the formula, This is the original short baseline image after inverse transformation; These are the pixel coordinates of the original short baseline image;
[0022] Preferably, in step S3), the short baseline image is obtained by correcting the original short baseline image. :
[0023] ;
[0024] In the formula, These are the stratified correction coefficients; For W screen correction function;
[0025] Preferably, in step S4), the W-Stacking algorithm is used to perform high-precision gridding processing, and convolutional gridding operation is performed on the long baseline dataset to interpolate the visibility data onto the CPU 3D grid.
[0026]
[0027] In the formula, This is a 3D mesh generated by CPU meshing; This represents the amount of long baseline data; For the W-Stacking phase compensation term, for the th The long baseline, in the Performing on the W plane A proportional, precise phase rotation cancels out the phase error caused by the non-coplanar baseline; The imaginary unit; For the first Long baseline coordinate; This refers to the layer number of the current W-plane layer; This represents the total number of layers in the CPU 3D mesh along the w direction.
[0028] Preferably, in step S5), the long baseline image is used... and short baseline images The system is divided into a central region and an edge region, and a regional anti-aliasing filtering strategy is adopted, with different parameters of PSWF anti-aliasing filters applied to different regions.
[0029] The central region is filtered using a PSWF anti-aliasing filter with the maximum cutoff radius and the minimum order; the edge region is filtered using a PSWF anti-aliasing filter with the minimum cutoff radius and the maximum order.
[0030] ;
[0031] ;
[0032] In the formula, , These are the filtered short and long baseline images, respectively. , These are PSWF anti-aliasing filters for the center and edge regions, respectively.
[0033] Preferably, in step S6), the target sky image is transferred to the GPU memory, a full-size Fourier transform is performed on each W-plane layer, and small regions related to the short baseline are cropped according to the cropping region in the imaging stage, and stacked to form a three-dimensional mesh; three-dimensional demeshing is performed on the GPU to map the three-dimensional mesh data back to the spatial frequency domain of the short baseline, generating short baseline predicted visibility data. ;Right now:
[0034] ;
[0035] In the formula, For three-dimensional demeshing operators; The cropped area of the target sky image.
[0036] Preferably, in step S6), the target sky image is stored in CPU memory, and based on the grid parameters synchronized with the GPU, a W-Stacking algorithm matching the imaging stage is used to perform long baseline de-gridization to generate long baseline predicted visibility data. ;Right now:
[0037] ;
[0038] In the formula, This is the phase compensation term for W-Stacking.
[0039] In a second aspect, the present invention provides a baseline-separated radio interferometry imaging system based on a heterogeneous architecture, comprising:
[0040] The feature extraction module is used to extract visibility data from the observation data of the radio interferometer array and extract observation feature parameters in real time.
[0041] The baseline segmentation module inputs the observed feature parameters into the heterogeneous load balancing model, calculates the baseline length threshold through constraints, and segments the visibility data into short baseline datasets and long baseline datasets based on the baseline length threshold.
[0042] The GPU imaging module dynamically allocates a three-dimensional grid on the GPU based on the spatial frequency characteristics of short baseline data, performs convolutional meshing and inverse Fourier transform processing, and generates short baseline images by combining W-term correction.
[0043] The CPU imaging module allocates a matched, minimized 3D grid on the CPU based on the grid parameters synchronized with the GPU, and performs gridding processing using the W-Stacking algorithm to generate a long baseline image that retains high-resolution details.
[0044] The anti-aliasing imaging module is based on a multi-parameter PSWF anti-aliasing model. It adjusts the order and cutoff radius parameters of PSWF according to the observation frequency band and field of view size, performs regional anti-aliasing resampling on short and long baseline images, and achieves linear image superposition through frequency domain alignment and field of view cropping to obtain the target sky image.
[0045] The prediction module, based on the grid parameters and heterogeneous processing strategy of the imaging stage, performs 3D demeshing of short baseline data on the GPU and demeshing of long baseline data on the CPU. It achieves real-time merging of demeshing data through a heterogeneous data interaction channel to obtain predicted visibility data.
[0046] The monitoring module is used to monitor the memory and computing power utilization of GPU and CPU in real time. When the processor status exceeds the preset threshold, it automatically triggers dynamic adjustment instructions to recalculate the baseline threshold and grid parameters to achieve real-time balancing of heterogeneous loads.
[0047] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the baseline separation radio interferometry imaging method.
[0048] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when invoked by a processor, cause the processor to execute the baseline separation radio interferometry imaging method.
[0049] The beneficial technical effects of this invention are as follows:
[0050] 1. This invention dynamically generates baseline thresholds based on observation features and the real-time status of the processor, so that the GPU memory utilization rate is stabilized at 80%~90%, the CPU computing power utilization rate is stabilized at 60%~80%, and the processing time difference of heterogeneous processors is reduced to less than 10%, thus completely solving the load imbalance problem.
[0051] 2. This invention establishes a nonlinear mapping relationship between PSWF parameters and observation frequency band and field of view size, and adopts a regional anti-aliasing filtering strategy to take into account the imaging effect of the image center and edge, further reducing imaging aliasing error and improving detail retention rate.
[0052] 3. This invention reduces intermediate data storage through gridding and inverse Fourier transform. Combined with grid allocation and lossless data compression, it reduces GPU memory usage by more than 40%, completely avoiding memory overflow issues, while preserving the low-frequency characteristics of short baseline data.
[0053] 4. This invention reuses all parameters from the imaging stage 100% through gridding and degridring, avoiding redundant calculations and data redundancy, reducing the residual of predicted visibility data, and meeting the high-precision requirements for sky model verification.
[0054] 5. This invention achieves real-time balancing of heterogeneous loads by monitoring the utilization of GPU and CPU memory and computing power in real time, and automatically and dynamically adjusting the baseline threshold and grid parameters when the status is abnormal. The system can adapt to complex observation needs of multi-band and wide field of view, and its robustness and adaptability are significantly improved. Attached Figure Description
[0055] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention;
[0056] Figure 2 Flowchart for generating short baseline images according to embodiments of the present invention;
[0057] Figure 3 This is a schematic diagram illustrating the process of generating a target sky image according to an embodiment of the present invention;
[0058] Figure 4 This is a schematic diagram of the structural framework of the system according to an embodiment of the present invention. Detailed Implementation
[0059] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:
[0060] like Figure 1As shown, this embodiment provides a baseline-separated radio interferometry imaging method based on a heterogeneous architecture. This embodiment uses the SKA1-Low radio interferometer array for observation across multiple frequency bands (50-350MHz) and a wide field of view (30°×30°). The GPU is an NVIDIA A100, and the CPU is an Intel Xeon Platinum 8480H. The method includes the following steps:
[0061] S1) Extract visibility data from the observation data of the radio interferometer array and extract observation feature parameters in real time;
[0062] In this embodiment, the visibility data includes baseline spatial frequency coordinates. Baseline visibility value (VIS) and baseline weight.
[0063] The observation characteristic parameters include observation frequency band, field of view size, baseline distribution density, GPU memory remaining amount, and CPU computing power utilization rate.
[0064] S2) Construct a heterogeneous load balancing model, input the observed feature parameters into the heterogeneous load balancing model, calculate the baseline length threshold through constraints, and divide the visibility data into short baseline datasets and long baseline datasets according to the baseline length threshold.
[0065] In this embodiment, the constraint condition is:
[0066] GPU memory utilization constraint: 80% ;
[0067] GPU compute utilization constraint: 80% ;
[0068] CPU utilization limit: 60% ;
[0069] Heterogeneous processing time difference constraint: ,in Estimated time for GPU to process short baseline data; Estimated time for CPU to process long baseline data.
[0070] The baseline length threshold is obtained by inputting the observed characteristic parameters into the heterogeneous load balancing model and applying constraints. The multivariate fitting function is expressed as:
[0071] ;
[0072] In the formula, These are the fitting coefficients; For observation frequency band; The size of the field of view; Baseline distribution density; GPU memory utilization; This refers to GPU computing power utilization. This represents the CPU computing power utilization rate.
[0073] The loss function is to minimize the total time of heterogeneous processing. The fitting coefficients are iteratively updated using gradient descent until the loss function converges, thus obtaining the optimal baseline length threshold. .
[0074] And the two-dimensional spatial length of the baseline is calculated based on the spatial frequency coordinates of the baseline, that is:
[0075] ;
[0076] In the formula, For the first The two-dimensional spatial length of each baseline; , For the first Spatial frequency coordinates of each baseline;
[0077] Based on the optimal baseline length threshold The visibility data is split into a short baseline dataset and a long baseline dataset, i.e.:
[0078] like This constitutes a short baseline dataset;
[0079] like This constitutes a long baseline dataset;
[0080] in, For short baseline datasets, For long baseline datasets; This is visibility data.
[0081] S3) Based on the spatial frequency characteristics of short baseline data, a 3D mesh is dynamically allocated on the GPU, and convolutional meshing and inverse Fourier transform processing are performed, combined with W-term correction to generate a short baseline image; simultaneously, a real-time data interaction channel is established between the GPU and the CPU to synchronously transmit mesh parameters to the CPU; such as Figure 2 As shown, the details are as follows:
[0082] S31) Maximum spatial frequency range based on short baseline data By combining the Nyquist-Shannon sampling theorem to determine the frequency sampling interval of the GPU 3D mesh, the optimal size of the GPU 3D mesh is calculated. :
[0083]
[0084] In the formula, This is the memory redundancy factor; The frequency sampling interval; The maximum spatial frequency;
[0085] S32) Transfer the short baseline dataset to the GPU memory, allocate a 3D mesh of optimal size, and perform convolutional meshing and inverse Fourier transform processing;
[0086] In this embodiment, the convolutional meshing is as follows:
[0087]
[0088] In the formula, The resulting 3D mesh is generated through convolutional meshing. For the first Visibility values of short baselines; For the first Baseline weights for short baselines; The kernel function is the convolution kernel function. For the first Spatial frequency coordinates of a short baseline.
[0089] Perform inverse Fourier transform on convolutionally meshed 3D mesh data The inverse Fourier transform is expressed as:
[0090] ;
[0091] In the formula, This is the original short baseline image after inverse transformation; These are the pixel coordinates of the original short baseline image;
[0092] S33) Based on the field of view (FOV) and the W-axis coordinate range of visibility data, the W-plane layer is divided into several layers; among which, the number of layers... as follows:
[0093] ;
[0094] In the formula, The range of W-axis coordinates; This is the control factor.
[0095] A hierarchical W-term correction strategy was employed to correct different W-plane layers, resulting in short baseline images. ;Right now:
[0096] The short baseline image is obtained by correcting the original short baseline image. :
[0097] ;
[0098] In the formula, These are the stratified correction coefficients; For W screen correction function;
[0099] S34) The GPU grid size, frequency sampling interval, and W-term correction parameters are synchronized to the CPU in real time through the GPU-CPU high-speed data interaction channel.
[0100] S4) Based on the grid parameters synchronized by the GPU, a matching minimized 3D grid is allocated on the CPU, and the W-Stacking algorithm is used to perform gridding processing to generate a long baseline image that retains high-resolution details; specifically, the following steps are included:
[0101] S41) Receive grid parameters synchronized by the GPU, including the frequency sampling interval. , Directional sampling interval The matching size of the CPU 3D mesh is calculated based on the maximum spatial frequency of the long baseline data to ensure frequency sampling alignment with the GPU mesh.
[0102] S42) Long baseline dataset Retained in CPU memory;
[0103] S43) The W-Stacking algorithm is used to perform high-precision gridding processing. Convolutional gridding operation is performed on the long baseline dataset to interpolate the visibility data onto the CPU 3D grid.
[0104]
[0105] In the formula, This is a 3D mesh generated by CPU meshing; This represents the amount of long baseline data; For the W-Stacking phase compensation term, for the th The long baseline, in the Performing on the W plane A proportional, precise phase rotation cancels out the phase error caused by the non-coplanar baseline; The imaginary unit; For the first Long baseline coordinate; This refers to the layer number of the current W-plane layer; This represents the total number of layers in the CPU 3D mesh along the w direction.
[0106] S44) Reuse the W-term correction parameters synchronized by the GPU to perform W-term correction on the meshed 3D mesh data to eliminate the non-coplanar baseline effect, that is:
[0107] ;
[0108] In the formula, The corrected 3D mesh data; The W-screen correction function is defined in the spatial frequency domain.
[0109] S45) Corrected 3D mesh data A three-dimensional inverse fast Fourier transform is performed to convert the spatial frequency domain data to the image domain. After brightness normalization and resolution enhancement, a long baseline image is generated. ;
[0110] ;
[0111] In the formula, This is the long baseline image after inverse transformation; The coordinates are for the long baseline image.
[0112] S5) Construct a multi-parameter PSWF anti-aliasing model, adjust the PSWF order and cutoff radius parameters according to the observation frequency band and field of view size, perform regional anti-aliasing resampling on short and long baseline images, and achieve linear image superposition through frequency domain alignment and field of view cropping to obtain the final wide field of view, high-resolution target sky image; Figure 3 As shown, it includes the following steps:
[0113] S51) Construct a multi-parameter PSWF anti-aliasing model to establish the mapping relationship between the PSWF order, cutoff radius parameter, observation frequency band, and field of view size; that is:
[0114] ;
[0115] In the formula, Let be the order of the ellipsoidal wavefunction PSWF; Let be the cutoff radius of the ellipsoidal wavefunction PSWF; , These are the empirical fitting coefficients;
[0116] S52) Based on the current observation frequency band The optimal order and cutoff radius of the PSWF are obtained to generate the PSWF anti-aliasing filter; that is:
[0117]
[0118] In the formula, It is a PSWF anti-aliasing filter; The coordinates of the long baseline image or the short baseline image;
[0119] S53), long baseline image and short baseline images The system is divided into a central region and an edge region, and a regional anti-aliasing filtering strategy is adopted, with different parameters of PSWF anti-aliasing filters applied to different regions.
[0120] In this embodiment, the central area Represented as:
[0121] ;
[0122] The edge region Represented as:
[0123] ;
[0124] In this embodiment, the central region is filtered using a PSWF anti-aliasing filter with the largest cutoff radius and the smallest order to preserve the core brightness features; the edge region is filtered using a PSWF anti-aliasing filter with the smallest cutoff radius and the largest order to suppress edge aliasing, i.e.:
[0125] ;
[0126] ;
[0127] In the formula, , These are the filtered short and long baseline images, respectively. , These are PSWF anti-aliasing filters for the center and edge regions, respectively.
[0128] S54) Based on the grid sizes of the GPU and CPU, calculate the number of zero-padded pixels in the frequency domain to align grids of different resolutions, resulting in aligned short and long baseline images. , ;
[0129] In this embodiment, with target size Based on this, calculate the number of zero-padded pixels in the frequency domain; that is:
[0130] ;
[0131] In the formula, , These represent the number of zero-padded pixels in the frequency domain for short and long baseline images, respectively. , The GPU 3D mesh is respectively in Pixel count in direction, CPU 3D mesh in Number of pixels in the direction;
[0132] In the frequency domain, the filtered short and long baseline images , Zero-padding is applied to the edges to obtain dimensions that are all... Aligned short and long baseline images , ;
[0133] S55), for zero-padded short and long baseline images , Perform field-of-view cropping, automatically adjust the cropping factor based on the size of the observed field of view, extract the effective field of view region, remove invalid and redundant pixels, and obtain effective short and long baseline images. , ;
[0134] In this embodiment, the clipping factor for:
[0135] ;
[0136] In the formula, The target's field of view size; This refers to the actual size of the observed field of view;
[0137] And from the zero-padded short and long baseline images , The medium cutting size is Effective short and long baseline images , ;
[0138] S56) Assign a superposition weight to each pixel according to the weight distribution of the effective short and long baseline image data, and superimpose the resampled short and long baseline images according to the superposition weight to generate the final wide field of view, high resolution target sky image.
[0139] In this embodiment, a weight is assigned to each pixel based on the weight distribution of effective short and long baseline image data, that is:
[0140] ;
[0141] In the formula, , The first Weighting of short and long baselines for each pixel; , Effective short and long baseline image data in pixels, respectively. Normalized weights; , These are the resolutions of the effective short and long baseline images, respectively;
[0142] The resampled short and long baseline images are superimposed according to the superposition weights to generate the final wide-field-of-view, high-resolution target sky image; that is:
[0143] ;
[0144] In the formula, The first of the target sky images The radio emission brightness value of each pixel.
[0145] S6) Based on the mesh parameters and heterogeneous processing strategy of the imaging stage, 3D demeshing of short baseline data is performed on the GPU, and demeshing of long baseline data is performed on the CPU. Real-time merging of demeshed data is achieved through a heterogeneous data interaction channel to obtain redundancy-free predicted visibility data; specifically as follows:
[0146] S61) The target sky image is transferred to the GPU memory, a full-size Fourier transform is performed on each W plane layer, and small regions related to the short baseline are cropped according to the cropping region in the imaging stage and stacked to form a three-dimensional mesh.
[0147] S62) Perform 3D demeshing on the GPU to map the 3D mesh data back to the spatial frequency domain of the short baseline, generating short baseline prediction visibility data. ;Right now:
[0148] ;
[0149] In the formula, For three-dimensional demeshing operators; The cropped area for the target sky image;
[0150] (S63) The target sky image is stored in CPU memory. Based on the grid parameters synchronized with the GPU, the W-Stacking algorithm, which matches the imaging stage, is used to perform long baseline de-gridization and generate long baseline predicted visibility data. ;Right now:
[0151] ;
[0152] In the formula, For W-Stacking phase compensation terms;
[0153] S64) Through the GPU-CPU high-speed data interaction channel, the short baseline prediction visibility data and the long baseline prediction visibility data are merged in real time to obtain a complete prediction visibility dataset.
[0154] S65) Compare the predicted visibility dataset with the observed visibility dataset to verify and correct the target sky image.
[0155] S7) Real-time monitoring of GPU and CPU memory and computing power utilization. When the processor status exceeds the preset threshold, it automatically triggers dynamic adjustment instructions to recalculate the baseline threshold and grid parameters to achieve real-time balancing of heterogeneous loads.
[0156] In addition, this embodiment is compared with the traditional fixed baseline separation method, the pure CPU method, and the pure GPU method. The results are shown in Table 1:
[0157] Table 1. Comparison of results between this embodiment and traditional fixed baseline separation methods, pure CPU methods, and pure GPU methods.
[0158] Comparison indicators This embodiment Traditional fixed baseline separation method Pure CPU method Pure GPU method Imaging aliasing error ( ) Video memory overflow, failure Image detail retention ( ) 99.5% 92.0% 99 % Video memory overflow, failed. Overall processing efficiency (speed) 1 0.4 0.15 fail GPU memory usage (GB) 68 78 - fail CPU utilization rate (%) 75 88 95 - Forecast data residuals (ΔVIS) fail
[0159] As can be seen from Table 1, the method in this embodiment is significantly better than existing methods in terms of imaging accuracy, computational efficiency, GPU memory usage, and CPU load balancing, and fully meets the multi-band, wide field of view, and massive data imaging processing requirements of next-generation large radio interferometer arrays such as SKA.
[0160] Example 2
[0161] like Figure 4 As shown, this embodiment provides a baseline-separated radio interferometry imaging system based on a heterogeneous architecture, including:
[0162] The feature extraction module is used to extract visibility data from the observation data of the radio interferometer array and extract observation feature parameters in real time.
[0163] The baseline segmentation module inputs the observed feature parameters into the heterogeneous load balancing model, calculates the baseline length threshold through constraints, and segments the visibility data into short baseline datasets and long baseline datasets based on the baseline length threshold.
[0164] The GPU imaging module dynamically allocates a three-dimensional grid on the GPU based on the spatial frequency characteristics of short baseline data, performs convolutional meshing and inverse Fourier transform processing, and generates short baseline images by combining W-term correction.
[0165] The CPU imaging module allocates a matched, minimized 3D grid on the CPU based on the grid parameters synchronized with the GPU, and performs gridding processing using the W-Stacking algorithm to generate a long baseline image that retains high-resolution details.
[0166] The anti-aliasing imaging module is based on a multi-parameter PSWF anti-aliasing model. It adjusts the order and cutoff radius parameters of PSWF according to the observation frequency band and field of view size, performs regional anti-aliasing resampling on short and long baseline images, and achieves linear image superposition through frequency domain alignment and field of view cropping to obtain the target sky image.
[0167] The prediction module, based on the grid parameters and heterogeneous processing strategy of the imaging stage, performs 3D demeshing of short baseline data on the GPU and demeshing of long baseline data on the CPU. It achieves real-time merging of demeshing data through a heterogeneous data interaction channel to obtain predicted visibility data.
[0168] The monitoring module is used to monitor the memory and computing power utilization of GPU and CPU in real time. When the processor status exceeds the preset threshold, it automatically triggers dynamic adjustment instructions to recalculate the baseline threshold and grid parameters to achieve real-time balancing of heterogeneous loads.
[0169] Example 3
[0170] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the baseline separation radio interferometry imaging method.
[0171] In this embodiment, the memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. A processor, coupled to the memory, is used to execute computer programs stored in the memory.
[0172] The computer program includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form.
[0173] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.
Claims
1. A baseline-separated radio interferometry imaging method based on heterogeneous architecture, characterized in that, Includes the following steps: S1) Extract visibility data from the observation data of the radio interferometer array and extract observation feature parameters; S2) Construct a heterogeneous load balancing model, input the observed feature parameters into the heterogeneous load balancing model, calculate the baseline length threshold through constraints, and divide the visibility data into short baseline datasets and long baseline datasets according to the baseline length threshold. S3) Based on the spatial frequency characteristics of short baseline data, a three-dimensional mesh is dynamically allocated on the GPU, and convolutional meshing and inverse Fourier transform processing are performed. Combined with W-term correction, a short baseline image is generated. S4) Based on the grid parameters synchronized by the GPU, a matching minimized 3D grid is allocated on the CPU, and the W-Stacking algorithm is used to perform gridding processing to generate a long baseline image that retains high-resolution details; S5) Construct a multi-parameter PSWF anti-aliasing model, adjust the order and cutoff radius parameters of PSWF according to the observation frequency band and field of view size, perform regional anti-aliasing resampling on short and long baseline images, and achieve linear image superposition through frequency domain alignment and field of view cropping to obtain the target sky image; S6) Based on the grid parameters and heterogeneous processing strategy of the imaging stage, the three-dimensional demeshing of short baseline data is performed on the GPU and the demeshing of long baseline data is performed on the CPU. The demeshing data is merged in real time through the heterogeneous data interaction channel to obtain the predicted visibility data.
2. The baseline separation radio interferometry imaging method based on heterogeneous architecture according to claim 1, characterized in that: In step S2), the baseline length threshold is obtained by inputting the observed characteristic parameters into the heterogeneous load balancing model and applying constraints. The multivariate fitting function is expressed as: ； In the formula, These are the fitting coefficients; For observation frequency band; The size of the field of view; Baseline distribution density; GPU memory utilization; This refers to GPU computing power utilization. CPU computing power utilization; The loss function is to minimize the total time of heterogeneous processing. The fitting coefficients are iteratively updated using gradient descent until the loss function converges, thus obtaining the optimal baseline length threshold. .
3. The baseline separation radio interferometry imaging method based on heterogeneous architecture according to claim 2, characterized in that: In step S2), the two-dimensional spatial length of the baseline is calculated based on the spatial frequency coordinates of the baseline, and then the optimal baseline length threshold is applied. The visibility data is split into a short baseline dataset and a long baseline dataset; that is: ； In the formula, For the first The two-dimensional spatial length of each baseline; , For the first Spatial frequency coordinates of each baseline; like This constitutes a short baseline dataset; like This constitutes a long baseline dataset; in, For short baseline datasets, For long baseline datasets; This is visibility data.
4. The baseline separation radio interferometry imaging method based on heterogeneous architecture according to claim 3, characterized in that: In step S3), the specific details are as follows: S31) Maximum spatial frequency range based on short baseline data By combining the Nyquist-Shannon sampling theorem to determine the frequency sampling interval of the GPU 3D mesh, the optimal size of the GPU 3D mesh is calculated. : In the formula, This is the memory redundancy factor; The frequency sampling interval; The maximum spatial frequency; S32) Transfer the short baseline dataset to the GPU memory, allocate a 3D mesh of optimal size, and perform convolutional meshing and inverse Fourier transform processing; The convolutional meshing is represented as follows: In the formula, The resulting 3D mesh is generated through convolutional meshing. For the first Visibility values of short baselines; For the first Baseline weights for short baselines; The kernel function is the convolution kernel function. For the first Spatial frequency coordinates of a short baseline; Perform inverse Fourier transform on convolutionally meshed 3D mesh data The inverse Fourier transform is expressed as: ； In the formula, This is the original short baseline image after inverse transformation; These are the pixel coordinates of the original short baseline image; S33) Based on the field of view (FOV) and the W-axis coordinate range of visibility data, the W-plane layer is divided into several layers; and a layered W-term correction strategy is used to correct different W-plane layers to obtain a short baseline image; The short baseline image is obtained by correcting the original short baseline image. : ； In the formula, These are the stratified correction coefficients; For W screen correction function; S34) The GPU grid size, frequency sampling interval, and W-term correction parameters are synchronized to the CPU in real time through the GPU-CPU high-speed data interaction channel.
5. The baseline separation radio interferometry imaging method based on heterogeneous architecture according to claim 4, characterized in that: In step S4), a long baseline image that preserves high-resolution details is generated; specifically, the following steps are included: S41) Receive grid parameters synchronized by the GPU, including the frequency sampling interval. , Directional sampling interval The matching size of the CPU 3D mesh is calculated based on the maximum spatial frequency of the long baseline data to ensure frequency sampling alignment with the GPU mesh. S42) Long baseline dataset Retained in CPU memory; S43) The W-Stacking algorithm is used to perform high-precision gridding processing. Convolutional gridding operation is performed on the long baseline dataset to interpolate the visibility data onto the CPU 3D grid. In the formula, This is a 3D mesh generated by CPU meshing; This represents the amount of long baseline data; For the W-Stacking phase compensation term, for the th The long baseline, in the Performing on the W plane A proportional, precise phase rotation cancels out the phase error caused by the non-coplanar baseline; The imaginary unit; For the first Long baseline coordinate; This refers to the layer number of the current W-plane layer; This represents the total number of layers in the CPU 3D mesh along the w direction. S44) Reuse the W-term correction parameters synchronized by the GPU to perform W-term correction on the meshed 3D mesh data to eliminate the non-coplanar baseline effect, that is: ； In the formula, The corrected 3D mesh data; The W-screen correction function is defined in the spatial frequency domain. S45) Corrected 3D mesh data A three-dimensional inverse fast Fourier transform is performed to convert the spatial frequency domain data to the image domain. After brightness normalization and resolution enhancement, a long baseline image is generated. ; ； In the formula, This is the long baseline image after inverse transformation; The coordinates are for the long baseline image.
6. The baseline separation radio interferometry imaging method based on heterogeneous architecture according to claim 5, characterized in that: Step S5 includes the following steps: S51) Construct a multi-parameter PSWF anti-aliasing model to establish the mapping relationship between the PSWF order, cutoff radius parameter, observation frequency band, and field of view size; that is: ； In the formula, Let be the order of the ellipsoidal wavefunction PSWF; Let be the cutoff radius of the ellipsoidal wavefunction PSWF; , These are the empirical fitting coefficients; S52) Based on the current observation frequency band The optimal order and cutoff radius of the PSWF are obtained to generate the PSWF anti-aliasing filter; that is: In the formula, It is a PSWF anti-aliasing filter; The coordinates of the long baseline image or the short baseline image; S53), long baseline image and short baseline images The system is divided into a central region and an edge region, and a regional anti-aliasing filtering strategy is adopted, applying PSWF anti-aliasing filters with different parameters to different regions; that is: ； ； In the formula, , These are the filtered short and long baseline images, respectively. , These are PSWF anti-aliasing filters for the center and edge regions, respectively. S54) Based on the grid sizes of the GPU and CPU, calculate the number of zero-padded pixels in the frequency domain to align grids of different resolutions, resulting in aligned short and long baseline images. , ; S55), for zero-padded short and long baseline images , Perform field-of-view cropping, automatically adjust the cropping factor based on the size of the observed field of view, extract the effective field of view region, and obtain effective short and long baseline images. , ; S56) Assign a stacking weight to each pixel according to the weight distribution of the effective short and long baseline image data, and stack the resampled short and long baseline images according to the stacking weight to generate the final wide field of view, high resolution target sky image.
7. The baseline separation radio interferometry imaging method based on heterogeneous architecture according to claim 6, characterized in that: Weights are assigned to each pixel based on the weight distribution of effective short and long baseline image data, i.e.: ； In the formula, , The first Weighting of short and long baselines for each pixel; , Effective short and long baseline image data in pixels, respectively. Normalized weights; , These are the resolutions of the effective short and long baseline images, respectively; The resampled short and long baseline images are superimposed according to the superposition weight to generate the final wide field of view, high resolution target sky image; Right now: ； In the formula, The first of the target sky images The radio emission brightness value of each pixel.
8. The baseline separation radio interferometry imaging method based on heterogeneous architecture according to claim 7, characterized in that: In step S6), the specific details are as follows: S61) The target sky image is transferred to the GPU memory, a full-size Fourier transform is performed on each W plane layer, and small regions related to the short baseline are cropped according to the cropping region in the imaging stage and stacked to form a three-dimensional mesh. S62) Perform 3D demeshing on the GPU to map the 3D mesh data back to the spatial frequency domain of the short baseline, generating short baseline prediction visibility data. ;Right now: ； In the formula, For three-dimensional demeshing operators; The cropped area for the target sky image; (S63) The target sky image is stored in CPU memory. Based on the grid parameters synchronized with the GPU, the W-Stacking algorithm, which matches the imaging stage, is used to perform long baseline de-gridization and generate long baseline predicted visibility data. ;Right now: ； In the formula, For W-Stacking phase compensation terms; S64) Through the GPU-CPU high-speed data interaction channel, the short baseline prediction visibility data and the long baseline prediction visibility data are merged in real time to obtain a complete prediction visibility dataset. S65) Compare the predicted visibility dataset with the observed visibility dataset to verify and correct the target sky image.
9. A baseline-separated radio interferometry imaging system based on a heterogeneous architecture, characterized in that, include: The feature extraction module is used to extract visibility data from the observation data of the radio interferometer array and extract observation feature parameters in real time. The baseline segmentation module inputs the observed feature parameters into the heterogeneous load balancing model, calculates the baseline length threshold through constraints, and segments the visibility data into short baseline datasets and long baseline datasets based on the baseline length threshold. The GPU imaging module dynamically allocates a three-dimensional grid on the GPU based on the spatial frequency characteristics of short baseline data, performs convolutional meshing and inverse Fourier transform processing, and generates short baseline images by combining W-term correction. The CPU imaging module allocates a matched, minimized 3D grid on the CPU based on the grid parameters synchronized with the GPU, and performs gridding processing using the W-Stacking algorithm to generate a long baseline image that retains high-resolution details. The anti-aliasing imaging module is based on a multi-parameter PSWF anti-aliasing model. It adjusts the order and cutoff radius parameters of PSWF according to the observation frequency band and field of view size, performs regional anti-aliasing resampling on short and long baseline images, and achieves linear image superposition through frequency domain alignment and field of view cropping to obtain the target sky image. The prediction module, based on the grid parameters and heterogeneous processing strategy of the imaging stage, performs 3D demeshing of short baseline data on the GPU and demeshing of long baseline data on the CPU. It achieves real-time merging of demeshing data through a heterogeneous data interaction channel to obtain predicted visibility data.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the baseline separation radio interferometry imaging method according to any one of claims 1-8.