Single pulse signal recognition system and method based on shenwei architecture

Abstract

The application belongs to the field of radio astronomy, and provides a single pulse signal identification system and method based on a Sunway architecture, which comprises the following steps: obtaining astronomical observation original data and performing preprocessing to obtain preprocessed astronomical observation data; generating an original candidate body pulse signal list from the preprocessed astronomical observation data, visualizing the original candidate body pulse signal list to obtain a pulse image, using a Sunway architecture to classify and identify a trained deep neural network model to obtain a classification result, screening a pulse image corresponding to a true classification result based on the classification result to obtain a true candidate body pulse signal list; assigning a drawing sub-process to each candidate body based on the true candidate body pulse signal list, combining a dynamic spectrum and a pulse fitting graph generated by each drawing sub-process to obtain a candidate body graph, and fusing the true candidate body pulse signal list and the candidate body graph to obtain a pulse signal identification result.

Description

Technical Field

[0001] This invention belongs to the field of radio astronomy technology, specifically relating to a single-pulse signal recognition system and method based on the Shenwei architecture. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] The current pulsar search in radio astronomy faces three major technical bottlenecks: First, the massive data processing pressure is immense. Taking the FAST telescope as an example, the amount of data generated in a single observation can reach the petabyte level, making it difficult for traditional serial processing methods to meet real-time requirements. Existing CPU-based acceleration solutions have significant shortcomings in energy efficiency and memory management. Second, signal recognition accuracy is low. Conventional threshold detection methods have an recognition rate of less than 60% for weak signals with a signal-to-noise ratio below 8σ, and the false alarm rate is as high as 30-40%. Although some studies have attempted to introduce machine learning, it has shown instability when processing signals in complex radio frequency interference (RFI) environments. Third, dynamic spectral analysis is inefficient. Generating high-resolution dynamic spectra involves complex dedispersion calculations, with single-node processing taking several hours, severely restricting the timeliness of pulsar discovery. Existing parallel solutions lack optimization on the Shenwei architecture and cannot fully utilize the computing potential of domestically produced processors.

[0004] To address the aforementioned technical bottlenecks, a series of methods for single-pulse signal identification have emerged in existing technologies, such as those utilizing the Shenwei processor. However, due to limited support for the Shenwei processor from mainstream international deep learning frameworks, particularly insufficient optimization in matrix operations and neural network inference, computational resource utilization is low, and architectural compatibility issues exist. Furthermore, traditional processing methods frequently experience memory overflows when loading extremely large amounts of data, lacking an effective hierarchical caching mechanism, which severely impacts system stability and leads to memory access bottlenecks. Therefore, current pulsar searches in radio astronomy face technical bottlenecks including massive data processing pressure, low signal identification accuracy, and inefficient dynamic spectral analysis. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a single-pulse signal identification system and method based on the Shenwei architecture. By optimizing the utilization of computing resources on the Shenwei architecture, this invention improves the accuracy and efficiency of signal identification, while also enhancing the performance of dynamic spectral analysis, thereby accelerating the pulsar discovery process.

[0006] According to some embodiments, the first solution of the present invention provides a single-pulse signal identification method based on the Shenwei architecture, which adopts the following technical solution: A single-pulse signal identification method based on the Shenwei architecture includes: The raw astronomical observation data is acquired and preprocessed to obtain preprocessed astronomical observation data. A list of original candidate pulse signals is generated based on the preprocessed astronomical observation data. The list of original candidate pulse signals is visualized to obtain pulse images. A deep neural network model optimized and trained using the Shenwei architecture is used for classification and recognition to obtain classification results. Based on the classification results, the pulse images corresponding to the real classification results are selected and analyzed to obtain a list of real candidate pulse signals. Based on the list of real candidate pulse signals, a drawing subprocess is assigned to each candidate. Each drawing subprocess generates a dynamic spectrum and a pulse fitting map, which are then combined to obtain a candidate map. The list of real candidate pulse signals and the candidate map are then fused to obtain the pulse signal recognition result.

[0007] Furthermore, the acquisition of raw astronomical observation data and its preprocessing to obtain preprocessed astronomical observation data specifically involves: The raw astronomical observation data is acquired, decoded, and converted into a standardized format to obtain astronomical observation data. Automatic gain control and bandpass correction are applied to standardized astronomical observation data to obtain corrected standardized astronomical observation data. Preliminary radio frequency interference marking is performed on the corrected and standardized astronomical observation data to obtain preprocessed astronomical observation data.

[0008] Furthermore, the automatic gain control and bandpass correction performed on the standardized astronomical observation data to obtain the corrected standardized astronomical observation data are specifically as follows: By using an automatic gain control mechanism, the gain is dynamically adjusted based on the statistical power of each channel in the standardized astronomical observation data to obtain the modulated standardized astronomical observation data. The spectral average power curve is extracted from the modulated standardized astronomical observation data, and the bandwidth gain is flattened according to the background noise level to obtain the corrected standardized astronomical observation data.

[0009] Furthermore, the generation of the original candidate body pulse signal list based on the preprocessed astronomical observation data specifically includes: The preprocessed astronomical observation data is then subjected to dedispersion, interference reduction, and data folding processes to obtain the original candidate pulse signal list. Extract the arrival time and increment modulation of each pulse from the original candidate pulse signal list, and plot the pulse image within a window around each pulse.

[0010] Furthermore, the deep neural network model is optimized and trained using the Shenwei architecture, specifically as follows: The matrix multiplication and addition kernel function is encapsulated into an optimized interface file using the Shenwei architecture; In each training cycle, the gradient magnitude of each channel of the convolutional layer in the deep neural network model is iteratively calculated as an importance indicator. When using a neural network framework to perform convolution operations during the training of a deep neural network model, the optimized interface file for the matrix multiplication and addition kernel function is called to accelerate training. After every N training cycles, the K% of channels with the lowest importance ranking are automatically pruned. Fine-tune the pruned model to compensate for the loss of accuracy; The process continues until the deep neural network model converges, resulting in a deep neural network model optimized and trained using the Shenwei architecture.

[0011] Furthermore, each plotting subprocess generates a dynamic spectrum and a pulse fitting map, which are then combined to obtain a candidate volume map, specifically: Dynamic spectra are plotted based on the data of each candidate in the real candidate pulse signal list; One-dimensional Gaussian fitting is performed on the temporal contour of each candidate in the real candidate pulse signal list to obtain the pulse fitting map; The dynamic spectrum and the pulse fitting graph are then combined to obtain the candidate graph for each candidate.

[0012] According to some embodiments, a second aspect of the present invention provides a single-pulse signal identification system based on the Shenwei architecture, employing the following technical solution: A single-pulse signal recognition system based on the Shenwei architecture includes: The front-end preprocessing module is configured to acquire raw astronomical observation data and perform preprocessing to obtain preprocessed astronomical observation data. The pulse classification module is configured to generate a list of original candidate pulse signals based on preprocessed astronomical observation data, visualize the list of original candidate pulse signals to obtain pulse images, use the Shenwei architecture to optimize and train a deep neural network model for classification and recognition to obtain classification results, and filter the pulse images corresponding to the real classification results based on the classification results to obtain a list of real candidate pulse signals. The parallel verification module is configured to assign a plotting subprocess to each candidate based on the real candidate pulse signal list. Each plotting subprocess generates a dynamic spectrum and a pulse fitting map, which are then combined to obtain a candidate map. The real candidate pulse signal list and the candidate map are then fused to obtain the pulse signal recognition result.

[0013] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the single-pulse signal recognition method based on the Shenwei architecture as described in the first embodiment above.

[0015] According to some embodiments, a fourth aspect of the present invention provides a computer device.

[0016] A computer device includes 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 steps in the single-pulse signal recognition method based on the Shenwei architecture as described in the first embodiment above.

[0017] According to some embodiments, a fifth aspect of the present invention provides a computer program product or computer program.

[0018] A computer program product or computer program includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the single-pulse signal recognition method based on the Shenwei architecture as described in the first embodiment above.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention represents a significant technological breakthrough in the field of radio single-pulsar signal processing, achieving a qualitative leap in data processing efficiency through an innovative three-stage pipeline architecture. Real-world testing data shows that the system's data processing capability on a single node of the Shenwei cluster is 19.5 times higher than traditional main-core solutions, while reducing energy consumption per unit of data processing by 40%, perfectly solving the real-time processing challenges faced by large radio telescopes such as FAST. This efficiency improvement is primarily attributed to the collaborative work of Shenwei's optimized deep neural network and innovative MPI parallel framework, reducing the dynamic spectrum generation task, which previously required several hours, to less than 3 minutes.

[0020] In terms of signal recognition accuracy, this invention achieves unprecedented performance improvement. The system employs an improved ResNet50 network combined with a multi-dimensional feature fusion algorithm, achieving a detection rate of 98.2% for pulsar signals with a signal-to-noise ratio (SNR) of 5σ or higher, while controlling the false alarm rate to below 3.8%, representing a revolutionary improvement compared to the 25-30% false alarm rate of traditional threshold methods. Particularly in millisecond pulsar detection, the system sensitivity is improved by 40%, primarily due to the combination of innovative DM-SNR feature enhancement technology and an adaptive radio frequency interference suppression algorithm, which significantly enhances the ability to extract weak periodic signals.

[0021] The system of this invention, deeply optimized on the Shenwei processor, significantly improves model training speed, achieving a qualitative leap in training efficiency through innovative computing architecture and algorithm optimization. Test results show that the system achieves a 3.8-fold speedup compared to traditional many-core solutions in standard ResNet50 network training tasks. This breakthrough is mainly attributed to three key technologies: First, the customized SW64 instruction set accelerator performs register-level optimization of matrix multiplication and addition operations, increasing the forward propagation speed of convolutional layers by 300%; second, the hybrid precision computing framework innovatively adopts a dynamic switching strategy of FP16 / FP32, reducing training time by 45% while maintaining model accuracy; these optimizations work together to not only significantly shorten the model convergence time but also control the average time per training epoch to within 85 seconds, providing strong computing power support for the rapid iteration of single-pulse signal recognition models. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a flowchart of a single-pulse signal recognition method based on the Shenwei architecture in an embodiment of the present invention; Figure 2 This is an execution architecture diagram of the single-pulse signal recognition method based on the Shenwei architecture in this embodiment of the invention; Figure 3 This is the processing flow of the preprocessing unit in this embodiment of the invention; Figure 4 This is the processing flow of the AI ​​intelligent recognition engine in the embodiments of the present invention; Figure 5 This is the processing flow of the parallel verification system in this embodiment of the invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0028] Example 1 like Figure 1 As shown, this embodiment provides a single-pulse signal identification method based on the Shenwei architecture. This embodiment uses the application of this method to a server as an example for illustration. It can be understood that this method can also be applied to a terminal, and can also be applied to a system including a terminal, a server, and a system, and can be implemented through the interaction between the terminal and the server. The server can be an independent physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communication, middleware services, domain name services, CDN security services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in this application. In this embodiment, the method includes the following steps: The raw astronomical observation data is acquired and preprocessed to obtain preprocessed astronomical observation data. A list of original candidate pulse signals is generated based on the preprocessed astronomical observation data. The list of original candidate pulse signals is visualized to obtain pulse images. A deep neural network model optimized and trained using the Shenwei architecture is used for classification and recognition to obtain classification results. Based on the classification results, the pulse images corresponding to the real classification results are selected and analyzed to obtain a list of real candidate pulse signals. Based on the list of real candidate pulse signals, a drawing subprocess is assigned to each candidate. Each drawing subprocess generates a dynamic spectrum and a pulse fitting map, which are then combined to obtain a candidate map. The list of real candidate pulse signals and the candidate map are then fused to obtain the pulse signal recognition result.

[0029] Specifically, such as Figure 2 As shown, the detailed process of the processing flow in this embodiment is as follows: Step S1: Obtain the raw astronomical observation data and perform preprocessing to obtain the preprocessed astronomical observation data, such as... Figure 3 As shown, specifically: Step S1.1: Obtain raw astronomical observation data, decode and convert the format to obtain standardized astronomical observation data, including: First, receive raw astronomical observation data from the FAST radio telescope or other radio equipment, typically in the format of a multi-channel PSRFITS file.

[0030] The data merging program completes data decoding and structure unification processing, supporting operations such as merging multiple files, channel normalization, and polarization averaging.

[0031] Extract key information headers (TBIN, NSBLK, OBSNCHAN) and construct standardized astronomical observation data (standardized FITS data), which is a cross-platform astronomical data exchange standard format, laying the foundation for subsequent candidate object extraction and image rendering.

[0032] This step is essentially a data input module that converts the format of the input graph data.

[0033] Step S1.2: Perform automatic gain control and bandpass correction on the standardized astronomical observation data to obtain the corrected standardized astronomical observation data, specifically as follows: By utilizing an automatic gain control (AGC) mechanism, the gain is dynamically adjusted based on the statistical power of each channel in the standardized astronomical observation data to obtain the modulated standardized astronomical observation data. In other words, embedding an automatic gain control (AGC) mechanism in the FITS decoding stage can correct for gain drift generated by the equipment during observation.

[0034] The average power curve of the spectrum is extracted based on the modulated standardized astronomical observation data, and the band gain is flattened according to the background noise level to obtain the corrected standardized astronomical observation data. In other words, bandpass correction: extract the average power curve of the spectrum and flatten the band gain according to the background noise level to ensure that signals in different frequency bands have a uniform energy scale and reduce modeling errors.

[0035] This step is essentially a channelization process for standardized astronomical observation data.

[0036] Step S1.3: Perform preliminary radio frequency interference (RFI) labeling on the corrected and standardized astronomical observation data to obtain preprocessed astronomical observation data, specifically as follows: The corrected, standardized FITS data is then subjected to a coarse RFI labeling method based on sliding window statistics. Channels with abrupt changes or constant intensities in the spectrum are masked and recorded in an RFI mask table for subsequent candidate analysis to perform "confidence weighting" or feature deduction. This step is essentially radio frequency interference suppression.

[0037] Step S1, through the front-end preprocessing unit, performs unified decoding, channelization, polarization information fusion, and preliminary radio frequency interference (RFI) suppression on the raw astronomical observation data (PSRFITS format). It reads the multi-channel data structure file by file, extracts header information, and performs frequency and time-dimension normalization on each sampling point. The merged data file retains key observation parameters such as OBSNCHAN, NSBLK, and TBIN, ensuring consistency and reproducibility in subsequent processing. This method is fast, highly compatible, and adaptable to FAST multi-beam and multi-polarization observation data.

[0038] Step S2: Generate an original candidate pulse signal list based on the preprocessed astronomical observation data. Visualize the original candidate pulse signal list to obtain pulse images. Use the Shenwei architecture to optimize and train a deep neural network model for classification and recognition to obtain classification results. Based on the classification results, filter the pulse images corresponding to the true classification results and analyze them to obtain the true candidate pulse signal list, such as... Figure 4 As shown, the details are as follows: Step S2.1: Generate a list of original candidate object pulse signals based on the preprocessed astronomical observation data, and visualize the list of original candidate object pulse signals to obtain a pulse image (DM-SNR feature map), specifically: 1) After preprocessing, the astronomical observation data undergoes dispersion reduction, interference removal, and data folding to obtain the original candidate pulse signal list, which includes dispersion, signal-to-noise ratio, pulse arrival time, and other necessary information. In other words, the original candidate pulse signal list is stored in a .cand or .singlepulse candidate record file; 2) Extract the arrival time and delta modulation of each pulse from the original candidate pulse signal list, and plot the pulse image within a window around each pulse. That is, in this embodiment, the DM-SNR_plot.py script is used to automatically extract the ToA (Time of Arrival) and DM (Delta Modulation) of each pulse, and plot a DM-SNR scatter plot within a window around it. The DM-SNR scatter plot is the final pulse image, thus realizing DM-SNR feature extraction.

[0039] Pulse images were standardized to 224×224 resolution, three-channel RGB format, and generated as PNG files using matplotlib. The naming convention included the ToA value for easy tracking and backtracking.

[0040] To improve the recognizability of weak fast radio burst candidate signals in the DM-SNR map, this invention designs a multi-scale feature enhancement module that integrates image processing and signal processing techniques to achieve weak signal enhancement, background removal, and frequency structure discrimination. First, for the signal records in the original candidate pulse signal list, before plotting the DM-SNR scatter plot, the system uses an adaptive histogram equalization algorithm to redistribute the image grayscale, enhancing the contrast of low-intensity pulse regions and improving the model's ability to perceive weak pulses. Second, a difference-of-Gaussian (DoG) filter bank is introduced in the image processing stage to extract the structural edge features of the signal in the DM and SNR (signal-to-noise ratio) dimensions in a multi-scale manner, effectively removing high-frequency noise and enhancing mesoscale features. Based on this, the system further constructs a three-dimensional feature tensor structure, using "DM × SNR × ToA" to form a multi-dimensional description space for the candidate, providing it as an additional input channel for the subsequent deep neural network. Meanwhile, to further enhance the model's ability to discriminate in the frequency space, a frequency attention mechanism is introduced. By adjusting the frequency channels with weights, high-confidence frequency bands are highlighted, effectively improving the model accuracy under multi-channel data.

[0041] Step S2.2: Based on the pulse image, use the Shenwei architecture to optimize and train the deep neural network model for classification and recognition, and obtain the classification result, specifically: Deep feature extraction and classification are performed. The generated pulse images are fed into a deep neural network model (using, but not limited to, the ResNet50 deep neural network model) optimized and trained using the Shenwei architecture for classification and recognition. This model has learned the distinguishing features between real and fake pulse images through a large number of manually labeled samples during the training phase.

[0042] After classification and recognition, the classification results are output, resulting in an info.csv file that stores the classification results. This file records the filename, ToA position, and discrimination result (real / fake) for each pulse image.

[0043] To meet the requirements of the domestic Shenwei platform for parallel computing, vectorization, and memory affinity, this system has undergone a comprehensive structural reconstruction and operator optimization based on the original ResNet50 network architecture, namely, Shenwei-optimized ResNet50, as detailed below: First, in the design of the convolutional layer, the traditional standard convolution is replaced with grouped convolution. The core idea is to divide the input feature map into G groups in the channel dimension, perform convolution operation independently in each group, and finally concatenate the results. While keeping the receptive field unchanged, the number of model parameters and computation are significantly reduced, and the overall weight parameter overhead is reduced by about 60%, which is beneficial for the model to run under the limited register resources of Shenwei.

[0044] Secondly, at the operator implementation level, a customized matrix multiply-add (GEMM) kernel function was designed for the Shenwei SW64 instruction set. By combining SIMD vectorized instructions with a multi-level pipeline scheduling mechanism, deep integration of data loading, computation, and storage processes was achieved. During optimization, register block partitioning technology was employed, splitting large matrices into 8x8 blocks and storing them entirely in registers. This allows each data block to be reused multiple times, reducing memory accesses by 60%. Simultaneously, loop unrolling technology was used to expand the originally multi-layered nested matrix computation loops into a continuous sequence of parallel instructions, significantly improving matrix computation throughput.

[0045] In deep learning, GEMM, whether in convolutional or fully connected layers, ultimately transforms or decomposes its core computation into large matrix multiplication operations. The SW64 instruction set certainly provides SIMD (Single Instruction Multiple Data) instructions. Customized GEMM kernel functions utilize these instructions, allowing a single instruction to process multiple data points simultaneously (e.g., performing a multiplication and addition of four floats at once), significantly improving computational throughput. GEMM optimization is an implementation detail of the underlying computation library, encapsulated in an interface `optimized_gemm()`. The upper-level neural network framework PyTorch calls this optimized function when performing convolution operations. In this embodiment, PyTorch is used to train the ResNet50 model. PyTorch calls the optimized EMM (Embedded Multi-Media Card) to accelerate training, and the trained ResNet50 model is then used for pulse image classification.

[0046] Furthermore, the model supports mixed-precision computation (FP16 / FP32) pipeline scheduling. The mixed-precision computation framework dynamically switches between FP16 and FP32 precision at different computation stages, significantly improving computational efficiency and resource utilization while maintaining model accuracy. Forward computation is accelerated using FP16, while FP32 is used in key stages such as loss function calculation, gradient accumulation, weight update, and BatchNorm to ensure numerical stability. Combined with adaptive loss scaling, it effectively balances training speed and numerical accuracy.

[0047] To address the bottleneck issue in activation function execution, this system implements a dedicated hardware acceleration unit interface for activation functions, and incorporates Shenwei-compatible acceleration functions such as swExp and swReLU to reduce the overhead of nonlinear operations.

[0048] Finally, a Dynamic Sparse Training strategy was introduced. During training, low-contribution channels are automatically pruned based on gradient information, significantly reducing inference computation and improving inference frame rate (FPS) and concurrency performance without sacrificing accuracy. Its core is model compression through structured pruning. The specific implementation involves three key steps: First, the gradient magnitude of each channel in the convolutional layer is calculated in real-time during training iterations as an importance indicator; second, after every N training epochs, the lowest K% of channels in terms of importance are automatically pruned; finally, the pruned model is fine-tuned to compensate for accuracy loss. These three steps are repeated until the model converges. This strategy forms a progressive relationship with the aforementioned grouped convolution modification and GEMM kernel function optimization: first, grouped convolution reduces the number of parameters; then, DST further compresses the model size; and finally, a customized GEMM kernel function fully leverages hardware performance. These three elements together constitute a complete optimization chain, ultimately achieving a significant improvement in inference efficiency on the Zhongwei platform.

[0049] Step S2.3: Based on the classification results, filter out the pulse images that are judged as true classification results, analyze them to obtain a list of true candidate pulse signals, specifically: The system automatically parses images identified as "real" in info.csv, extracts their corresponding ToA and DM information, and writes it to the DM-SNR_process.txt file, obtaining a list of real candidate pulse signals. This file serves as the input file for the next stage of the parallel verification system. Simultaneously, this stage supports multi-factor scoring of candidates (SNR, image quality, AI confidence), constructing a candidate credibility ranking list for human-assisted review, thus achieving multi-criteria decision-making.

[0050] The list of real candidate pulse signals includes dispersion, signal-to-noise ratio, pulse arrival time, classification results, and other necessary information.

[0051] This step utilizes an AI intelligent recognition engine to visualize the original candidate pulse signal list as a pulse image (DM-SNR scatter plot), and then automatically classifies and identifies it using a deep neural network model optimized and trained on the Shenwei architecture. First, the candidate .cand files are aggregated, grouping signals with similar Time of Arrival (ToA) into the same event, and selecting the one with the highest SNR as the representative. Subsequently, the system plots the corresponding DM-SNR scatter plot, normalizes the image to 224×224, and saves it as a three-channel color image. The deep neural network model is built using, but is not limited to, the ResNet50 architecture, trained using transfer learning, and ultimately outputs binary labels (real / fake). Both model training and inference can run on the Shenwei platform, and it supports conversion to ONNX or Caffe formats adapted to the architecture for performance optimization, greatly improving the efficiency of automatic candidate screening and significantly reducing human error and missed detection.

[0052] Step S3: Based on the list of real candidate pulse signals, assign a plotting subprocess to each candidate. Each plotting subprocess generates a dynamic spectrum and a pulse fitting map, which are then combined to obtain a candidate image. The list of real candidate pulse signals and the candidate image are then fused to obtain the pulse signal recognition result, such as... Figure 5 As shown, the details are as follows: Step S3.1: Based on the list of real candidate pulse signals, the main process assigns a plotting subprocess to each candidate. That is, it automatically reads DM-SNR_process.txt and runs the plotting subprocesses in parallel with ToA and DM for each candidate as parameters. For example, candidate 1 corresponds to plotting subprocess 1, which executes the plotting of the dynamic spectrum and pulse profile fitting map; candidate 2 corresponds to plotting subprocess 2, which executes the plotting of the dynamic spectrum and pulse profile fitting map. Simulating MPI scheduling, the main process is responsible for task distribution, that is, assigning a plotting subprocess to each candidate. The plotting subprocesses concurrently plot the dynamic spectrum and pulse fitting map, thus realizing MPI task scheduling.

[0053] Step S3.2: Draw a dynamic spectrum based on the data of each candidate in the real candidate pulse signal list, specifically as follows: Dynamic spectra (waterfall plots) are generated based on the data of each candidate in the real candidate pulse signal list. The system supports scalable task allocation strategies, and the plotting results are two-dimensional frequency-time images containing obvious signal sweep structures. The images are stored in a specified folder with a unified name to achieve dynamic spectrum generation.

[0054] Step S3.3: Perform one-dimensional Gaussian fitting on the temporal contour of each candidate in the real candidate pulse signal list to obtain the pulse fitting map, specifically: In the plotting subroutine, the system supports one-dimensional Gaussian fitting of the temporal pulse profile of each candidate in the real candidate pulse signal list to extract key parameters such as pulse width, symmetry, peak value and position for secondary screening and pulse verification.

[0055] Step S3.4: Combine the dynamic spectrum and the pulse fitting plot to obtain the candidate volume plot, specifically: After processing in steps S3.2 and S3.3, the input data is the DM-SNR_process.txt file - a list of real candidate pulse signals. Two sub-graphs are drawn from different perspectives. The dynamic spectrum and the pulse fitting graph are then combined to obtain the candidate graph.

[0056] Step S3.5: Fuse the list of real candidate pulse signals and the candidate image to obtain the pulse signal recognition result, specifically: The system fuses the results of candidates that have passed the two-stage screening (AI intelligent recognition and judgment + parallel verification and confirmation) to form a comprehensive output file (such as frb_candidates_final.csv) containing fields such as ToA, DM, SNR, image path, discrimination label, and confidence level – the pulse signal recognition results. All image, text, and candidate data are archived to a database or a designated path for manual review or subsequent astronomical analysis.

[0057] To further confirm the identified "true candidates," this step introduces a parallel verification system to generate high-precision dynamic spectrograms (Waterfall plots) for expert or automated algorithms to make image-level judgments. The system takes a .toa file as input, containing a list of candidates composed of parameters such as ToA and DM. A main control parallel script is invoked for multi-process scheduling, constructing a task pool structure similar to MPI, and plotting subroutines are called in parallel according to the number of cores. Each plotting task executes independently, outputting a frequency-time two-dimensional image containing signal drift characteristics and structural details, which is crucial for confirming the authenticity of astronomical signals. The plotting process fully utilizes multi-core resources, with no interference between tasks, making it suitable for large-scale image generation tasks in a cluster environment.

[0058] To adapt to the data scale and computing power of domestic high-performance platforms, this system is designed with a multi-level parallel processing framework, which fully leverages the synergistic effect of data-level parallelism, task-level parallelism and pipeline parallelism to achieve end-to-end computing acceleration.

[0059] In terms of data-level parallelism, the system divides the original FITS data into 128MB blocks according to the time dimension, which are processed in parallel by different processes, effectively utilizing the advantages of multi-core memory bandwidth.

[0060] In terms of task-level parallelism, a master-slave structure is adopted, with the master process dynamically allocating drawing subprocesses to each core. The drawing subprocesses execute spectrum drawing tasks independently, avoiding resource contention and thread blocking.

[0061] At the pipeline level, the entire candidate processing flow is divided into three stages: dispersion removal, integral sampling, and image rendering. Independent subprocesses are used for asynchronous processing, forming a three-stage pipeline of computation-storage-IO, which improves the utilization of data channels.

[0062] To further enhance system scalability, the system implements a computing and communication overlap technology, which interleaves data prefetching, drawing scheduling, and image writing processes. In actual tests, this can hide approximately 80% of the communication / IO wait time, achieving a near-linear scalability ratio.

[0063] In the overall data processing flow of this embodiment, the system first receives raw astronomical observation data from radio telescopes such as FAST, i.e., PSRFITS format data. This data undergoes unified decoding, channel normalization, polarization merging, and preliminary RFI masking to generate standardized data blocks. This stage accelerates the decoding task through inter-core parallel partitioning and constructs a unified metadata structure (preprocessed astronomical observation data) for subsequent processing. Subsequently, the system generates a DM-SNR scatter plot based on the candidate pulse signal .cand file (the original candidate pulse signal list) and completes deep learning classification of true and false pulses using a deep neural network model optimized by the Shenwei architecture (such as the ResNet50 network model). To adapt to the Shenwei platform, several optimization strategies are used, including SIMD accelerated convolution, GEMM kernel reconstruction, and mixed-precision inference, significantly improving inference performance and energy efficiency. Finally, a parallel verification process is initiated for the list of true candidate pulse signals classified as "true." A multi-core asynchronous scheduling mechanism is used to generate two-dimensional spectrograms in batches, supplemented by pulse contour fitting analysis, ultimately forming a complete single-pulse signal screening workflow.

[0064] Example 2 This embodiment provides a single-pulse signal recognition system based on the Shenwei architecture, including: The front-end preprocessing module is configured to acquire raw astronomical observation data and perform preprocessing to obtain preprocessed astronomical observation data. The pulse classification module is configured to generate a list of original candidate pulse signals based on preprocessed astronomical observation data, visualize the list of original candidate pulse signals to obtain pulse images, use the Shenwei architecture to optimize and train a deep neural network model for classification and recognition to obtain classification results, and filter the pulse images corresponding to the real classification results based on the classification results to obtain a list of real candidate pulse signals. The parallel verification module is configured to assign a plotting subprocess to each candidate based on the real candidate pulse signal list. Each plotting subprocess generates a dynamic spectrum and a pulse fitting map, which are then combined to obtain a candidate map. The real candidate pulse signal list and the candidate map are then fused to obtain the pulse signal recognition result.

[0065] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1 above. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0066] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0067] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0068] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the single-pulse signal recognition method based on the Shenwei architecture as described in Embodiment 1 above.

[0069] Example 4 This embodiment provides a computer 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 steps in the single-pulse signal recognition method based on the Shenwei architecture as described in Embodiment 1 above.

[0070] Example 5 This embodiment provides a computer program product or computer program, including computer instructions stored in a computer-readable storage medium. The processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the single-pulse signal recognition method based on the Shenwei architecture described in Embodiment 1 above.

[0071] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0072] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0073] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0076] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for identifying single-pulse signals based on the Shenwei architecture, characterized in that, include: The raw astronomical observation data is acquired and preprocessed to obtain preprocessed astronomical observation data. A list of original candidate pulse signals is generated based on the preprocessed astronomical observation data. The list of original candidate pulse signals is visualized to obtain pulse images. A deep neural network model optimized and trained using the Shenwei architecture is used for classification and recognition to obtain classification results. Based on the classification results, the pulse images corresponding to the real classification results are selected and analyzed to obtain a list of real candidate pulse signals. Based on the list of real candidate pulse signals, a drawing subprocess is assigned to each candidate. Each drawing subprocess generates a dynamic spectrum and a pulse fitting map, which are then combined to obtain a candidate map. The list of real candidate pulse signals and the candidate map are then fused to obtain the pulse signal recognition result. The Shenwei architecture is used to optimize the training of deep neural network models, specifically as follows: The matrix multiplication and addition kernel function is encapsulated into an optimized interface file using the Shenwei architecture; In each training cycle, the gradient magnitude of each channel of the convolutional layer in the deep neural network model is iteratively calculated as an importance indicator. When using a neural network framework to perform convolution operations during the training of a deep neural network model, the optimized interface file for the matrix multiplication and addition kernel function is called to accelerate training. After every N training cycles, the K% of channels with the lowest importance ranking are automatically pruned. Fine-tune the pruned model to compensate for the loss of accuracy; The process continues until the deep neural network model converges, resulting in a deep neural network model optimized and trained using the Shenwei architecture. Each plotting subprocess generates a dynamic spectrum and a pulse fitting plot, which are then combined to obtain a candidate volume plot. Specifically: Dynamic spectra are plotted based on the data of each candidate in the real candidate pulse signal list; One-dimensional Gaussian fitting is performed on the temporal contour of each candidate in the real candidate pulse signal list to obtain the pulse fitting map; The candidate map for each candidate is obtained by combining the dynamic spectrum and the pulse fitting map.

2. The single-pulse signal recognition method based on the Shenwei architecture as described in claim 1, characterized in that, The process of acquiring raw astronomical observation data and preprocessing it to obtain preprocessed astronomical observation data specifically involves: The raw astronomical observation data is acquired, decoded, and converted into a standardized format to obtain astronomical observation data. Automatic gain control and bandpass correction are applied to standardized astronomical observation data to obtain corrected standardized astronomical observation data. Preliminary radio frequency interference marking is performed on the corrected and standardized astronomical observation data to obtain preprocessed astronomical observation data.

3. The single-pulse signal recognition method based on the Shenwei architecture as described in claim 2, characterized in that, The process of performing automatic gain control and bandpass correction on the standardized astronomical observation data to obtain the corrected standardized astronomical observation data is as follows: By using an automatic gain control mechanism, the gain is dynamically adjusted based on the statistical power of each channel in the standardized astronomical observation data to obtain the modulated standardized astronomical observation data. The spectral average power curve is extracted from the modulated standardized astronomical observation data, and the bandwidth gain is flattened according to the background noise level to obtain the corrected standardized astronomical observation data.

4. The single-pulse signal recognition method based on the Shenwei architecture as described in claim 1, characterized in that, The process of generating a list of original candidate body pulse signals based on preprocessed astronomical observation data is as follows: The preprocessed astronomical observation data is then subjected to dedispersion, interference reduction, and data folding processes to obtain the original candidate pulse signal list. Extract the arrival time and increment modulation of each pulse from the original candidate pulse signal list, and plot the pulse image within a window around each pulse.

5. A single-pulse signal recognition system based on the Shenwei architecture, characterized in that, include: The front-end preprocessing module is configured to acquire raw astronomical observation data and perform preprocessing to obtain preprocessed astronomical observation data. The pulse classification module is configured to generate a list of original candidate pulse signals based on preprocessed astronomical observation data, visualize the list of original candidate pulse signals to obtain pulse images, use the Shenwei architecture to optimize and train a deep neural network model for classification and recognition to obtain classification results, and filter the pulse images corresponding to the real classification results based on the classification results to obtain a list of real candidate pulse signals. The parallel verification module is configured to assign a plotting subprocess to each candidate based on the real candidate pulse signal list. Each plotting subprocess generates a dynamic spectrum and a pulse fitting map, which are then combined to obtain a candidate map. The real candidate pulse signal list and the candidate map are then fused to obtain the pulse signal recognition result. The Shenwei architecture is used to optimize the training of deep neural network models, specifically as follows: The matrix multiplication and addition kernel function is encapsulated into an optimized interface file using the Shenwei architecture; In each training cycle, the gradient magnitude of each channel of the convolutional layer in the deep neural network model is iteratively calculated as an importance indicator. When using a neural network framework to perform convolution operations during the training of a deep neural network model, the optimized interface file for the matrix multiplication and addition kernel function is called to accelerate training. After every N training cycles, the K% of channels with the lowest importance ranking are automatically pruned. Fine-tune the pruned model to compensate for the loss of accuracy; The process continues until the deep neural network model converges, resulting in a deep neural network model optimized and trained using the Shenwei architecture. Each plotting subprocess generates a dynamic spectrum and a pulse fitting plot, which are then combined to obtain a candidate volume plot. Specifically: Dynamic spectra are plotted based on the data of each candidate in the real candidate pulse signal list; One-dimensional Gaussian fitting is performed on the temporal contour of each candidate in the real candidate pulse signal list to obtain the pulse fitting map; The candidate map for each candidate is obtained by combining the dynamic spectrum and the pulse fitting map.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the single-pulse signal recognition method based on the Shenwei architecture as described in any one of claims 1-4.

7. A computer 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 steps in the single-pulse signal recognition method based on the Shenwei architecture as described in any one of claims 1-4.

8. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps in the single-pulse signal recognition system and method based on the Shenwei architecture as described in any one of claims 1-4.

Images