A deep learning-based pulse parameter intelligent detection method and system

Abstract

The application discloses a kind of based on deep learning's pulse parameter intelligent detection method and system, the method is applied to the time-frequency data of transient astronomical signal is carried out automatic analysis, comprising: S1.image stage: the original FRB signal data observed by radio telescope is carried out physical correction, the original format of signal data is converted into the two-dimensional image of standardized time-frequency graph matrix;S2.detection stage: the time-frequency graph is input into deep learning model;Deep learning model utilizes its feature extraction capability, through a signal discrimination branch, automatically locates the time-frequency area where FRB pulse signal is located in background noise;S3.regression stage: after successfully locating signal area, network directly regresses each attribute of the target by two branches, so as to obtain the predicted value of each pulse parameter of FRB signal.The application integrates the extraction of multiple parameters into an end-to-end deep learning model, greatly improves the data processing speed, simplifies the processing procedure.

Description

Technical Field

[0001] This invention relates to the field of astrophysical data processing, and specifically to a method and system for extracting parameters from radio pulse signals using deep learning technology. Background Technology

[0002] Scientific research on radio pulse signals such as fast radio bursts (FRBs) requires precise measurement of a series of key physical parameters to reveal their physical origin, radiation mechanism, and propagation environment. Currently, the closest existing or commonly used technology in the industry is a semi-automated parameter extraction method based on traditional signal processing. The specific process of this type of method is typically as follows:

[0003] Pulse Width Measurement: After dedispersion processing of the signal, the two-dimensional time-frequency data is integrated along the frequency axis to obtain a one-dimensional time series (i.e., a pulse profile). Subsequently, the pulse width is usually determined by fitting the pulse profile with a Gaussian function or a multi-Gaussian function, or by calculating the ratio of the net signal integration area to the peak value to obtain the equivalent pulse width (Weff).

[0004] Frequency bandwidth measurement: Two-dimensional time-frequency data is integrated along the time axis to obtain a one-dimensional spectrum. The cumulative distribution function (CDF) of this spectrum is calculated, and an energy threshold (such as 5%-95%) is set manually to find the upper and lower limits of the signal frequency, thereby calculating the frequency bandwidth.

[0005] Signal-to-noise ratio (SNR) measurement: In the pulse profile, the statistical characteristics (such as standard deviation) of the background noise are estimated by manually selecting an "off-pulse" region that does not contain the pulse. Then, the ratio of the pulse peak value to the noise standard deviation is calculated to obtain the peak signal-to-noise ratio (SNRp).

[0006] These methods treat the extraction of different parameters as independent computational tasks, employing different algorithms to execute them step by step, and often require manual intervention from researchers to set thresholds, select fitting regions, or judge the fitting effect.

[0007] The drawbacks of existing technology are:

[0008] Traditional methods suffer from low processing efficiency and limited automation: They are cumbersome, requiring different fitting or calculation algorithms to be executed step-by-step for varying parameters such as pulse width, frequency range, and signal-to-noise ratio. The entire process is time-consuming (up to hundreds of milliseconds for a single sample) and relies on manual intervention, resulting in low automation. This approach is ill-suited to meeting the real-time processing needs of massive (PB-level) data generated by modern observational equipment like China's FAST telescope.

[0009] Results rely on subjective settings, resulting in poor accuracy and consistency: Traditional methods heavily depend on manually set thresholds (such as the energy threshold of CDF and the rejection criteria of RFI) and the selection of fitting parameters. For pulses with complex morphology or low signal-to-noise ratio, the processing results of different researchers may vary significantly, leading to subjective bias in the measurement results and making it difficult to guarantee the consistency and robustness of multiple measurements.

[0010] Fragmented processes and lack of integrated solutions: Existing technologies break down a complete signal analysis task into multiple independent sub-problems (width fitting, bandwidth calculation, SNR estimation), using different tools and models, lacking a unified and integrated framework. This fragmented approach not only increases operational complexity but also makes optimizing the entire processing flow difficult. Summary of the Invention

[0011] In view of the problems existing in the prior art, the purpose of this invention is to provide an end-to-end intelligent pulse parameter detection network and corresponding method based on deep learning, so as to overcome the fundamental defects of the prior art, such as low processing efficiency, strong subjectivity of results, limited degree of automation and cumbersome process.

[0012] To achieve the above objectives, this invention provides a deep learning-based intelligent pulse parameter detection method. This method is applied to the automated analysis of time-frequency data of transient astronomical signals. The method specifically includes the following steps:

[0013] S1. Imaging Stage: Physically correct the raw FRB signal data observed by the radio telescope, converting the original format of the signal data into a standardized two-dimensional image of the time-frequency diagram matrix; this step transforms the abstract physical signal data into a visual representation that deep learning models are good at processing.

[0014] S2. Detection Phase: The time-frequency graph is input into the pulse parameter intelligent detection network; the pulse parameter intelligent detection network uses its feature extraction capability to automatically locate the time-frequency region where the FRB pulse signal is located in the background noise through a signal discrimination branch; this process is conceptually equivalent to "finding the target object" in an image;

[0015] S3. Regression Phase: After successfully locating the signal area, the network directly regresses the various attributes of the target through two branches, thereby obtaining the various pulse parameters of the FRB signal.

[0016] Furthermore, the pulse parameters of the FRB signal include pulse width, frequency bandwidth, and signal-to-noise ratio.

[0017] Furthermore, in step S3, the two branches are as follows:

[0018] S3.1 Geometric Attribute Regression: A signal localization branch is responsible for regressing the bounding box coordinates of the target; the coordinates are directly mapped to the physical boundaries of the signal: the coordinates on the time axis correspond to the start and end times of the pulse, thereby determining the pulse width; the coordinates on the frequency axis directly correspond to the lower and upper frequency limits, thereby obtaining the frequency bandwidth;

[0019] S3.2 Physical Attribute Regression: Another independent signal-to-noise ratio estimation branch is responsible for regressing the signal-to-noise ratio of the signal.

[0020] Furthermore, in step S1, the original FRB signal data is physically corrected using a standardization preprocessing module, as follows:

[0021] Step S1.1: Metadata parsing and time correction; Use the astropy.io.fits professional library to read the raw FRB signal data PSRFITS file and extract the observation metadata, which includes: observation center frequency, total bandwidth, number of frequency channels, channel width, time sampling interval, and dispersion measurement value; According to the given pulse arrival time, use the dispersion time delay formula to correct the pulse arrival time;

[0022] Step S1.2: Time-frequency window truncation and downsampling: Based on the corrected pulse arrival time and the user-specified time window, calculate the range of time sampling points to be truncated from the original data;

[0023] Step S1.3: Perform physical correction to eliminate dispersion;

[0024] Step S1.4: Baseline subtraction and radio frequency interference removal;

[0025] Step S1.5: Normalization and dynamic range adjustment; All pixel values ​​of the two-dimensional time-frequency map matrix processed above are linearly normalized to the standard range of [0,1] or [0,255] by Min-Max scaling;

[0026] After the above five steps, the original FRB signal data is transformed into a two-dimensional time-frequency diagram.

[0027] Furthermore, in step S1.3, the dedispersion is specifically implemented as follows:

[0028] The system is based on each frequency channel Relative to reference frequency Dispersion time delay Perform correction:

[0029]

[0030] Calculated dispersion time delay Quantized as discrete time sampling point offsets:

[0031]

[0032] For each frequency channel For time-series data, perform a cyclic shift of an integer number of sampling points; the shift amount is... ;in, It is the dispersion constant. It is a dispersion measurement value. The time sampling interval;

[0033] After this operation, the pulse trajectory, which was originally tilted due to dispersion, will be "straightened" on the time-frequency diagram, appearing simultaneously on all frequency channels and forming a roughly rectangular energy cluster.

[0034] Furthermore, in step S1.4, the baseline subtraction and radio frequency interference removal are as follows:

[0035] Baseline subtraction: Calculate a baseline value independently for the time series of each frequency channel, and then subtract this baseline value from all data points of that channel;

[0036] RFI removal is achieved using a statistically based automated RFI removal strategy, including:

[0037] Narrowband RFI: Calculate the energy standard deviation of each frequency channel over the entire time window; channels with a standard deviation higher than a set first threshold are identified as RFI-contaminated, and their data will be zeroed or replaced by interpolation;

[0038] Wideband RFI: Calculate the total energy of each time sampling point across all frequency channels; time points where the total energy exceeds a set second threshold are identified as wideband RFI and are zeroed out.

[0039] Furthermore, in step S2, the pulse parameter intelligent detection network includes a region attention module, a feature extraction module, a feature fusion module, and an analysis head module.

[0040] Furthermore, in step S2, the data processing flow of the pulse parameter intelligent detection network is as follows:

[0041] S2.1 Feature extraction: The standardized time-frequency graph matrix two-dimensional image is input into the backbone network to extract multi-level deep features;

[0042] S2.2 Feature Fusion: Feature fusion is performed on features from different levels in a feature pyramid structure that integrates upsampling, stitching and R-ELAN modules.

[0043] S2.3 Multi-scale prediction feeds the fused features from each scale of the feature pyramid into an independent analysis head to generate initial prediction results, which include bounding boxes, categories, and signal-to-noise ratios.

[0044] S2.4 Output: All prediction results from different scale analysis heads are processed by a non-maximum suppression algorithm to filter out redundant detection boxes, and the final results regarding the signal-to-noise ratio, pulse width, and frequency bandwidth are output.

[0045] On the other hand, the present invention provides a pulse parameter intelligent detection system based on deep learning, which automatically analyzes the time-frequency data of transient astronomical signals and obtains the signal-to-noise ratio, pulse width and frequency bandwidth of FRB signals.

[0046] Furthermore, the system includes a standardized preprocessing module and a pulse parameter intelligent detection network; the standardized preprocessing module is used to physically correct the raw FRB signal data; the pulse parameter intelligent detection network is used to simultaneously output all target parameters end-to-end, including the signal-to-noise ratio, pulse width, and frequency bandwidth of the FRB signal.

[0047] Compared with the traditional parameter extraction methods described in the background art, the pulse parameter intelligent detection method and system based on deep learning proposed in this invention have the following significant advantages:

[0048] 1. Extremely High Processing Efficiency: This invention integrates the extraction of multiple parameters into an end-to-end deep learning model (neural network), achieving extremely high data processing speed. According to experimental results, the average processing time for a single pulse using this method is only 1.2 milliseconds, while the average processing time for traditional methods is 570 milliseconds. This represents a speed improvement of nearly 500 times, fully capable of meeting the real-time processing needs of massive amounts of data from future large-scale radio survey projects (such as the SKA).

[0049] 2. Superior Prediction Accuracy: This invention learns features directly from data using a deep learning model, avoiding systematic errors introduced by inaccurate model assumptions in traditional methods. The average relative errors of this invention's method in pulse width (PW), signal-to-noise ratio (SNR), and bandwidth (BW) are 2.8%, 5.0%, and 2.6%, respectively. In contrast, the errors of traditional methods are as high as 20.7%, 51.3%, and 28.4%, respectively, demonstrating the significant advantage of this invention in prediction accuracy.

[0050] 3. Complete Automation and Objectivity: This invention provides a fully automated solution, requiring no manual intervention from data input to parameter output. This completely eliminates the biases introduced by subjective operations such as manually selecting the fitting region and setting thresholds in traditional methods, ensuring the objectivity, consistency, and reproducibility of parameter measurement results.

[0051] 4. Integrated Solution: This invention innovatively integrates multiple parameter extraction tasks, which originally required different algorithms and were executed step by step, into a unified framework. This integrated design not only greatly simplifies the data processing flow but also provides an unprecedented, efficient, and intelligent analysis tool for scientific research on radio transient sources such as FRBs. Attached Figure Description

[0052] Figure 1 This is a comparative diagram of the complete attention mechanism (a) and the regional attention mechanism (b);

[0053] Figure 2 These are schematic diagrams of the architectures of (a) A2Block and (b) R-ELAN;

[0054] Figure 3 This is a schematic diagram of the analysis head module architecture;

[0055] Figure 4 This is a schematic diagram of the overall architecture of the intelligent pulse parameter detection network;

[0056] Figure 5 These are scatter plots of predicted values ​​versus true values. (a), (b), and (c) are scatter plots of predicted values ​​versus true values ​​of the intelligent detection network for the three parameters of pulse width, frequency bandwidth, and signal-to-noise ratio, respectively. (d), (e), and (f) are scatter plots of calculated values ​​versus true values ​​of the three parameters of pulse width, frequency bandwidth, and signal-to-noise ratio, respectively, by the traditional method.

[0057] Figure 6 This is a comparison chart of the running speed of intelligent detection networks and traditional methods. The results are displayed on logarithmic coordinates, with a line chart showing the comparison of total running time for different amounts of data, and a bar chart showing the comparison of average running time. Detailed Implementation

[0058] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0060] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0061] The following combination Figures 1-6 Specific embodiments of the present invention will be described in detail below. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the present invention.

[0062] This invention is particularly suitable for the automated analysis of time-frequency data of transient astronomical signals such as Fast Radio Bursts (FRBs) to efficiently and accurately obtain their key physical parameters, such as pulse width, frequency bandwidth, and signal-to-noise ratio.

[0063] The inventive concept of this invention lies in innovatively reconstructing the traditional, step-by-step parameter fitting problem into a direct mapping problem from a two-dimensional time-frequency diagram to multi-dimensional physical parameters (pulse width, frequency bandwidth, signal-to-noise ratio), and achieving fully automated processing through a single deep neural network model.

[0064] The system designed in this invention is a complete automated pipeline from raw data to final physical parameters. According to a deep learning-based intelligent pulse parameter detection method of this invention, its working principle can be summarized as three stages: "image generation - detection - regression," specifically including the following steps:

[0065] S1. Imageization Stage (also known as Input Stage): The physically corrected (e.g., achromatic) FRB signal data, regardless of its original format, is uniformly converted into a standardized two-dimensional image (i.e., a frequency map matrix). This step transforms the abstract physical signal data into a visual representation that deep learning models excel at processing.

[0066] S2. Detection Phase (Also known as Localization Phase): The time-frequency image is input into the deep learning model (Intelligent Impulse Parameter Detection Network). Leveraging its powerful feature extraction capabilities, the deep learning model uses a signal discrimination branch to automatically locate (detect) the time-frequency region where the FRB pulse signal is located amidst complex background noise. This process is conceptually equivalent to "finding the target object" in an image. The deep learning model and the Intelligent Impulse Parameter Detection Network are synonymous; the "Intelligent Impulse Parameter Detection Network" is a deep learning model with a specific architecture designed by this invention to solve a specific problem (FRB parameter detection).

[0067] S3. Regression Phase (also known as Measurement Phase): After successfully locating the signal region, the pulse parameter intelligent detection network directly regresses the various attributes of the target through two branches. This application includes three branches: one signal discrimination branch in the detection phase (S2), and two branches in the regression phase (S3). Specifically, it includes:

[0068] S3.1 Geometric Attribute Regression: A signal localization branch is responsible for regressing the bounding box coordinates (x1, y1, x2, y2) of the target. These coordinates are precisely designed to directly map to the physical boundaries of the signal: the coordinates (x1, x2) on the time axis correspond to the start and end times of the pulse, thus determining the equivalent pulse width; the coordinates (y1, y2) on the frequency axis directly correspond to the lower and upper frequency limits, thus obtaining the frequency bandwidth.

[0069] S3.2 Physical Attribute Regression: Another independent branch for signal-to-noise ratio estimation is responsible for regressing the non-geometric intrinsic properties of the signal, namely the signal-to-noise ratio (SNR).

[0070] Through the above methods, this invention integrates multiple tasks that originally required manual intervention, were executed step by step, and relied on different algorithms into a single automated process that can be completed in one forward propagation, fundamentally ensuring efficiency, objectivity, and integration.

[0071] To ensure the deep learning model can learn under diverse signal conditions, this invention generated 6000 FRB signals through computer simulation. During the simulation, key physical parameters were randomly distributed over a wide range. The signal-to-noise ratio (SNR) was uniformly distributed between 1 and 50. The pulse width (PW) was distributed between 0.5 and 50 milliseconds, primarily following a normal distribution with a mean of 6 ms and a standard deviation of 3 ms, but with some variation. The bandwidth followed a normal distribution with a mean of 250 MHz and a standard deviation of 100 MHz. The central frequency (CFR) was limited to between 1025 MHz and 1475 MHz, while ensuring that the highest signal frequency did not exceed 1500 MHz and the lowest frequency did not fall below 1000 MHz.

[0072] To ensure the effectiveness of the model in actual observation scenarios, this invention also uses real observation data from the Five Hundred Meter Aperture Spherical Radio Telescope (FAST), including 5,927 pulse signals from the repeating source FRB121102 and 600 positive sample pulse signals from the FAST-FREX dataset.

[0073] To train a deep learning model, accurate "ground truth" labels are needed for each signal. For simulated data, parameters such as pulse width, frequency bandwidth, and signal-to-noise ratio used during generation are directly used as the ground truth labels for the signal. For real data, since its "absolute truth" is unknown, this invention uses industry-recognized traditional methods to calculate its parameters and uses the calculation results as "surrogate truth." To train the pulse parameter intelligent detection network proposed in this invention, the above physical parameter labels need to be converted into pixel coordinates and regression values ​​of the image. The start and end times of the signal ( , ) and frequency start and end ( , ) is linearly mapped to the bounding box pixel coordinates on the time-frequency map ( , , , The calculated signal-to-noise ratio (SNR) value is directly used as the training ground value for the SNR regression branch.

[0074] In step S1, the system input is as follows: the standardized preprocessing procedure for time-frequency graph data is as follows:

[0075] To ensure the pulse parameter intelligent detection network receives high-quality, lossless, and scale-consistent input, this system includes a crucial, independent standardization preprocessing module (this module processes the data to obtain a standard time-frequency plot before inputting it into the network). This module is responsible for converting the raw data from the radio telescope (typically in PSRFITS format) into a two-dimensional tensor suitable for network processing. This process involves the following sophisticated steps:

[0076] Step S1.1: Metadata parsing and time correction

[0077] The PSRFITS file was read using specialized libraries such as astropy.io.fits, and key observation metadata was extracted, including: observation center frequency (OBSFREQ), total bandwidth (OBSBW), number of frequency channels (OBSNCHAN), channel width (CHAN_BW), and time sampling interval (…). ) and dispersion measurements ( )wait.

[0078] Based on the given pulse arrival time ( Using the dispersion time delay formula to Precise correction is performed. Due to dispersion, high-frequency signals arrive before low-frequency signals. The purpose of correction is to find the "true" arrival time of the signal, serving as a reference for subsequent data truncation and alignment. The correction formula is:

[0079]

[0080] in, It is the dispersion constant (approximately ), It is the highest frequency of observation. It is the reference frequency, and DM is the known signal source dispersion measurement value.

[0081] Step S1.2: Time-frequency window truncation and downsampling

[0082] According to the revised Based on a user-specified time window (e.g., ±50 milliseconds), the range of time sampling points to be extracted from the original data is calculated. To reduce data redundancy, improve the signal-to-noise ratio, and reduce computational burden, the extracted data blocks are downsampled in both time and frequency dimensions. This process is called integral downsampling, which involves... adjacent time sampling points or F The data from adjacent frequency channels are accumulated or averaged to form a new sampling point / channel.

[0083] Step S1.3: Core Physical Correction: De-dispersion

[0084] This step is crucial to ensuring that the time-frequency diagram accurately reflects the instantaneous characteristics of the pulse. The system determines the frequency based on each frequency channel. Relative to reference frequency Dispersion time delay Perform correction:

[0085]

[0086] Calculated dispersion time delay Quantized as discrete time sampling point offsets:

[0087]

[0088] For each frequency channel For time-series data, perform a cyclic shift (numpy.roll) of an integer number of sampling points; the shift amount is... .

[0089] After this operation, the pulse trajectory, which was originally tilted due to dispersion, will be "straightened" on the time-frequency diagram, appearing simultaneously on all frequency channels and forming an approximately rectangular energy cluster.

[0090] Step S1.4: Baseline subtraction and radio frequency interference (RFI) removal

[0091] Baseline subtraction: To eliminate the stable DC bias and slowly changing background introduced by the receiver system itself, a baseline value (usually the mean or median) is calculated independently for the time series of each frequency channel, and then this baseline value is subtracted from all data points of that channel.

[0092] RFI Removal: Radio frequency interference typically manifests as a persistent strong signal on a specific frequency channel, or a broadband signal affecting all frequencies at a specific point in time. This invention employs a statistically based automated RFI removal strategy:

[0093] Narrowband RFI: Calculates the energy standard deviation for each frequency channel over the entire time window. Channels with exceptionally high standard deviations are identified as RFI-contaminated, and their data are either zeroed out or replaced by interpolation (masking).

[0094] Wideband RFI: Calculate the sum of energy across all frequency channels at each time sampling point. Time points with abnormally high total energy are identified as wideband RFI and are also zeroed out.

[0095] Step S1.5: Normalization and Dynamic Range Adjustment

[0096] All pixel values ​​of the processed two-dimensional time-frequency graph matrix are linearly normalized to the standard range of [0,1] or [0,255] using methods such as Min-Max scaling. This ensures that the input data fed into the neural network has a consistent scale, which is beneficial to the stable convergence of model training.

[0097] After the above five steps, the original and complex observation data is transformed into a clean, standardized, and information-concentrated two-dimensional time-frequency graph, which serves as the final input to the neural network system of this invention.

[0098] In step S2, the pulse parameter intelligent detection network is implemented as follows:

[0099] The intelligent pulse parameter detection network proposed in this invention is based on a high-precision and high-efficiency target detection framework (YOLOv12). By employing a high-efficiency backbone network, a high-efficiency area attention mechanism, a high-efficiency feature aggregation mechanism (R-ELAN), and a specially designed analysis head, the network of this invention achieves efficient analysis of FRB signal parameters. The intelligent pulse parameter detection network of this invention includes a region attention module, a feature extraction module, a feature fusion module, and an analysis head module.

[0100] 1. Area Attention Module: such as Figure 1 As shown, compared to the traditional Full Attention mechanism, which requires calculating the correlation between all tokens globally, this mechanism is computationally intensive and inefficient. The Area Attention mechanism employed in this invention divides the input feature map into several (default 4) horizontal or vertical regions and performs attention calculations only within each region. This design maintains a broad receptive field while significantly reducing computational complexity, ultimately achieving efficient computation.

[0101] 2. Feature Extraction Module (A2 Block): The A2 Block is a fundamental functional module for efficient feature processing, and its structure is as follows: Figure 2 As shown in (a), the input features first pass through a region attention module, whose output is added to the original input (forming the first residual connection). Then, this result passes through two cascaded 1x1 convolutional layers, whose outputs are again added to the result of the first addition operation (forming the second residual connection), ultimately yielding the module's output. This dual residual design allows information to flow and fuse more effectively within the module, achieving a larger receptive field while maintaining computational efficiency.

[0102] 3. Feature Fusion Module (R-ELAN Enhanced High-Efficiency Layer Aggregation Network): R-ELAN is a powerful feature fusion module employing a dense connection approach. It is the core of the network neck structure of this invention, and its structure is as follows: Figure 2 As shown in (b), its working principle is as follows: the input features are split into two paths. One path passes through a 1x1 convolutional layer as the main path; the other path sequentially passes through multiple feature extraction modules (A2 Blocks) for deep feature extraction. The output features of each A2 Block and the initial input features are retained and then concatenated along the channel dimension. The massive concatenated features are then scaled and added to the features of the main path. This design, by densely fusing multi-level and multi-scale information, constructs a module with efficient parameters and extremely strong feature representation capabilities, while significantly improving gradient flow and making the training of deep models more stable.

[0103] 4. Analyze Head Module: The Analyze Head module is the core of this invention for achieving parallel output across multiple tasks. For example... Figure 3 As shown, in this analysis head module, independent branches are used for the three different tasks of signal localization, signal discrimination, and SNR estimation. For each location on the feature map input to the analysis head module, the network uses these branches to estimate the signal start and end times (t_start, t_end), start and end frequencies (f_start, f_end), signal category (cls), and signal-to-noise ratio (SNR).

[0104] The specific structural design of each branch is as follows: The signal localization branch consists of two 3x3 convolutional layers and one 1x1 convolutional layer for output. This branch is responsible for regressing the geometric coordinates of the signal bounding box. The signal discrimination branch consists of 3x3 depth-wise convolutions and 1x1 convolutions. This branch is responsible for determining whether the detected target is an FRB signal. The signal-to-noise ratio (SNR) estimation branch consists of three 3x3 convolutional layers and two 1x1 convolutional layers. This branch is responsible for regressing the physical property of the signal's SNR.

[0105] Based on the above modules, the intelligent pulse parameter detection network of the present invention is realized, as follows: Figure 4 As shown. Its data processing flow is as follows:

[0106] S2.1 begins with feature extraction. The signal (standardized time-frequency graph) is first input into a high-efficiency backbone network (for the entire pulse parameter intelligent detection network, the image first enters the backbone network, then enters the S2 feature fusion step consisting of multiple characteristic modules mentioned above, and then proceeds to S2.3 and S2.4) to extract multi-level deep features.

[0107] S2.2 Follows with feature fusion, which effectively fuses features from different levels within a feature pyramid-like structure that integrates up-sampling, concatenation, and R-ELAN feature fusion modules. The main structure of the feature fusion module is the feature extraction module, and the main structure of the feature extraction module is the region attention module.

[0108] S2.3 Next is multi-scale prediction. The fused features from each scale of the feature pyramid are fed into separate analysis heads to generate initial prediction results (including bounding boxes, categories, and SNR).

[0109] S2.4 Finally, the results are output. All prediction results from different scale analysis heads are processed by the Non-Maximum Suppression (NMS) algorithm to filter out redundant detection boxes and output the final results regarding the signal-to-noise ratio (SNR), pulse width (PW), and bandwidth (BW) of the signal.

[0110] Training strategies and multi-task loss functions:

[0111] In this invention, the deep learning model is trained end-to-end. The core of the training lies in a carefully designed multi-task composite loss function L, which consists of losses from signal localization, signal discrimination, and SNR regression, aiming to collaboratively optimize all prediction tasks of the network. Its overall loss function is defined as follows: .

[0112] During training, the loss functions for each part are applied to the output of the corresponding branch of the Analyze Head module. Their detailed definitions and reasons for selection are as follows: (These loss functions are used to optimize the output of the Analyze Head module during the model training phase.)

[0113] 1. Signal positioning loss This invention applies to the signal localization branch of the analysis head. To achieve accurate localization of the time-frequency region of pulse signals, it employs CIoU (Complete Intersection over Union) loss. Compared to traditional IoU loss, CIoU loss provides a more comprehensive geometric measurement by additionally considering the distance between the center points of the predicted box and the ground truth box, as well as the aspect ratio, thereby achieving faster convergence and higher localization accuracy during training.

[0114] 2. Signal discrimination loss This invention applies to the signal discrimination branch of the analysis head. To effectively address the class imbalance problem caused by the fact that background noise samples far outnumber valid FRB signal samples in the training data, Focal Loss is used as the signal discrimination loss function. Focal Loss reduces the weight of easily classifiable samples (such as obvious background noise) in the total loss, allowing the model to focus more on learning difficult-to-classify signals (such as low signal-to-noise ratio FRBs) during training, thereby improving the model's discrimination ability.

[0115] 3. SNR regression loss This invention applies to the signal-to-noise ratio (SNR) estimation branch of the analysis head. When predicting the key physical parameter SNR, this invention employs the MAPE (Mean Absolute Percentage Error) loss. This loss function, by measuring the relative error between the predicted and true values, can evenly optimize physical parameters of different orders of magnitude. Its calculation formula is as follows:

[0116]

[0117] Where y is the actual SNR value. It is the predicted SNR value of the model. Using MAPE loss can ensure that the model treats the prediction of signals of all scales (whether strong or weak) equally during optimization, and will not be biased towards optimizing parameters with larger values ​​due to the size of the absolute error, thereby obtaining a more balanced and robust regression performance across the entire parameter range.

[0118] 4. Optimizer and Training Parameter Settings

[0119] The network training in this invention uses momentum-based stochastic gradient descent (SGD) as the optimizer. The initial learning rate is set to 0.01, and a cosine annealing schedule is used to dynamically adjust it. This strategy smoothly reduces the learning rate in the later stages of training, helping the model converge to a better solution. The entire training process consists of 600 epochs. All training experiments are performed on a server equipped with four NVIDIA RTX 4090 Ti GPUs.

[0120] S4. Post-processing and output of prediction results

[0121] In this invention, the output of the pulse parameter intelligent detection network in step S3 is the initial prediction result. To improve prediction accuracy, step S4, post-processing and output of the prediction result, is preferably added to generate the final, highly reliable physical parameters. This process mainly includes two key steps: non-maximum suppression (NMS) and model ensemble.

[0122] S4.1. Non-Maximum Suppression (NMS): Because the network of this invention performs dense predictions at multiple scales of the feature map, a single FRB signal may generate multiple overlapping detection boxes. To solve this problem, the system of this invention employs the Non-Maximum Suppression (NMS) algorithm to process all prediction results from analysis heads at different scales. This algorithm iteratively filters and retains the prediction box with the highest confidence, while filtering out other redundant detection boxes whose overlap with this box exceeds a preset threshold. Through this step, it is ensured that each FRB signal corresponds to only one unique and optimal detection result.

[0123] S4.2. Model Ensemble and Error Estimation: To further improve the stability of the prediction results and provide reliable error estimation for the final output physical parameters, this invention adopts a model ensemble strategy. The specific implementation is as follows: By introducing a random factor during training, multiple deep learning models are trained independently. In the prediction phase, the standardized time-frequency plot of the same FRB signal is simultaneously input into multiple deep learning models, thereby obtaining multiple sets of independent prediction parameters (including bounding box coordinates and SNR). The final output parameter value is the average of the multiple prediction values. Simultaneously, the standard deviation of the multiple prediction values ​​is calculated and used as the prediction error of the parameter. Preferably, 10 deep learning models are used in this invention.

[0124] S4.3. Final Parameter Output: After the above post-processing steps, the system finally outputs the physical quantities of the signal-to-noise ratio (SNR), pulse width (PW), and bandwidth (BW) of the signal, along with their errors. These physical quantities are derived from the average predicted values ​​after model integration, ensuring the robustness and reliability of the results.

[0125] Experimental Design and Results:

[0126] To quantitatively evaluate the performance advantages of the technical solution of this invention, a series of comparative experiments were conducted. The experiments were carried out on an independent simulated data test set containing 1000 samples.

[0127] 1. Accuracy Comparison Experiment: This experiment aims to compare the accuracy of parameter prediction between the present invention and traditional methods. The "mean relative error" is used as the core evaluation index, defined as the absolute value of the difference between the predicted value and the true value divided by the true value. The experimental results are shown in Table 1. It can be seen that the prediction accuracy of the present invention is significantly better than that of the traditional method on all key parameters, with the error reduced by an order of magnitude. Figure 5 Plot a scatter plot of predicted values ​​versus true values. Figure 5 In the diagrams (a), (b), and (c), we see scatter plots showing the predicted and true values ​​of the intelligent detection network for the three parameters: pulse width, frequency bandwidth, and signal-to-noise ratio. Figure 5 In the diagrams (d), (e), and (f), we see scatter plots comparing the calculated values ​​of pulse width, frequency bandwidth, and signal-to-noise ratio (SNR) with the true values ​​obtained using the traditional method. It becomes clearer that the predicted points of the method in this invention are closely clustered around the diagonal of y=x, while the data points of the traditional method exhibit significant dispersion.

[0128] ;

[0129] Table 1. Comparison of the average relative error results between traditional methods and intelligent detection networks.

[0130] 2. Efficiency Comparison Experiment: This experiment aims to compare the computational efficiency of the two methods. We processed 1000 pulse signals using both methods. Each method processed the dataset five times, and the average time required to process a single sample was calculated. We also calculated the total data runtime for both methods for data ranging from 100 to 600 pulses. Furthermore, we compared the data runtime of the traditional method under five parallel thread counts: 1, 2, 4, 8, and 16. The results are as follows: Figure 6 As shown in the figure, the bar chart on the right represents the average running time. It reveals that the average processing speed of the intelligent detection network of this invention is approximately 475 times that of traditional methods, achieving a performance leap of nearly three orders of magnitude. The line chart on the left further shows that as the amount of data processed increases, the total processing time of the method of this invention increases gradually, while the processing time of the traditional method increases sharply. Crucially, the parallelization test of the traditional method, as shown in Table 2, reveals that even increasing the number of parallel threads from 1 to 16 did not significantly improve its running speed. This strongly demonstrates that the efficiency advantage of this invention stems from its end-to-end integrated algorithm flow, fundamentally solving the efficiency bottleneck of traditional methods.

[0131] ;

[0132] Table 2. Test results of computation time for parallelization of traditional methods with different numbers of threads

[0133] The core of this invention lies in proposing a novel paradigm for extracting parameters from radio pulse signals. Its technical advantages are as follows:

[0134] This invention reconstructs the parameter extraction problem into an end-to-end image target detection problem. Existing technologies treat parameter extraction as multiple independent fitting tasks based on traditional signal processing, resulting in cumbersome and fragmented processes. This invention, through a paradigm shift, reconstructs this problem for the first time into a unified direct mapping problem from a two-dimensional time-frequency graph to multi-dimensional physical parameters. Through a three-step process of "image-detection-regression," utilizing a single deep neural network model, it achieves full automation from raw data to final parameters, fundamentally solving the efficiency and objectivity bottlenecks of existing technologies.

[0135] This invention designs a multi-task parallel output network structure specifically for pulse signal analysis. To achieve the aforementioned end-to-end mapping, this invention designs a "deep learning-based intelligent pulse parameter detection network." Its key feature lies in the design of the analysis head at the network's end. This analysis head innovatively integrates three parallel, functionally independent prediction branches: a geometric localization branch for regressing the geometric coordinates of the signal bounding box (…). , , , The signal discrimination branch is used to determine whether the target is a valid signal. The physical attribute regression branch is used to directly regress the intrinsic physical attributes of the signal, i.e., the signal-to-noise ratio (SNR). This design, which processes geometric localization and physical attribute regression in parallel within a single network, is the core technological guarantee for achieving integrated and highly efficient analysis, and is the key structure that this invention aims to protect.

[0136] This invention constructs a composite loss function optimized for multidimensional heterogeneous parameters. To effectively train the aforementioned multi-task network, this invention designs a composite loss function consisting of three parts. The signal-to-noise ratio (SNR) is measured using the MAPE (mean absolute percentage error) loss function. Unlike conventional L1 or L2 losses, MAPE loss measures relative error, ensuring that the model treats SNR predictions of different magnitudes (strong or weak signals) equally during optimization. This avoids bias in optimizing parameters with larger values, thus significantly improving the regression accuracy and robustness of physical parameters across the entire dynamic range.

[0137] This invention proposes a method for simultaneous estimation of parameters and errors based on model ensemble. This invention not only provides predicted values ​​for parameters but also their corresponding prediction errors. It achieves this by employing a model ensemble strategy: multiple models are trained independently, and their outputs are statistically analyzed during prediction. The final average value of the parameter is used as the best estimate, while its standard deviation is used as the prediction error for that parameter. This technique, which combines deep learning with ensemble methods to simultaneously output physical quantities and their uncertainties, provides a direct quantitative basis for evaluating the reliability of model prediction results, and is a key technical aspect of this invention in enhancing its application value.

[0138] Any process or method described in the flowcharts of this invention or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, which can be implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device. The computer-readable medium can be any medium containing a program for storage, communication, propagation, or transmission for use by the execution system, apparatus, or device, including read-only memory, magnetic disks, or optical disks.

[0139] In the description of this specification, references to terms such as "embodiment," "example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, those skilled in the art can combine or combine the different embodiments or examples described in this specification and the features therein without causing contradiction.

[0140] While embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and alterations to the above embodiments within the scope of the present invention.

Claims

1. A deep learning-based intelligent pulse parameter detection method, characterized in that, The method is applied to the automated analysis of time-frequency data of transient astronomical signals, and specifically includes the following steps: S1. Imaging Stage: Physically correct the raw FRB signal data observed by the radio telescope, converting the original format of the signal data into a standardized two-dimensional image of the time-frequency diagram matrix; this step transforms the abstract physical signal data into a visual representation that deep learning models are good at processing. S2. Detection phase: Input the time-frequency map into the deep learning model; The deep learning model uses its feature extraction capability to automatically locate the time-frequency region where the FRB pulse signal is located in the background noise through a signal discrimination branch; S3. Regression Phase: After successfully locating the signal area, the network directly regresses the various attributes of the located signal through two branches, thereby obtaining the predicted values ​​of various pulse parameters of the FRB signal; In step S2, the deep learning model is an intelligent pulse parameter detection network, which includes a region attention module, a feature extraction module, a feature fusion module, and an analysis head module. The data processing flow of the pulse parameter intelligent detection network is as follows: S2.1 Feature extraction: The standardized time-frequency graph matrix two-dimensional image is input into the backbone network to extract multi-level deep features; S2.2 Feature Fusion: Feature fusion is performed on features from different levels in a feature pyramid structure that integrates upsampling, stitching and R-ELAN modules. S2.3 Multi-scale prediction feeds the fused features from each scale of the feature pyramid into an independent analysis head to generate initial prediction results, which include bounding boxes, categories, and signal-to-noise ratios. S2.4 Output: All prediction results from different scale analysis heads are processed by a non-maximum suppression algorithm to filter out redundant detection boxes, and the final results regarding the signal-to-noise ratio, pulse width, and frequency bandwidth are output.

2. The intelligent pulse parameter detection method based on deep learning according to claim 1, characterized in that, The pulse parameters of the FRB signal include pulse width, frequency bandwidth, and signal-to-noise ratio.

3. The intelligent pulse parameter detection method based on deep learning according to claim 1, characterized in that, In step S3, the two branches are as follows: S3.1 Geometric Attribute Regression: A signal localization branch is responsible for regressing the bounding box coordinates of the target; the coordinates are directly mapped to the physical boundaries of the signal: the coordinates on the time axis correspond to the start and end times of the pulse, thereby determining the pulse width; the coordinates on the frequency axis directly correspond to the lower and upper frequency limits, thereby obtaining the frequency bandwidth; S3.2 Physical Attribute Regression: Another independent signal-to-noise ratio estimation branch is responsible for regressing the signal-to-noise ratio of the signal.

4. The intelligent pulse parameter detection method based on deep learning according to claim 1, characterized in that, In step S1, the original FRB signal data is physically corrected using a standardization preprocessing module, as follows: Step S1.1: Metadata parsing and time correction; Use the astropy.io.fits professional library to read the raw FRB signal data PSRFITS file and extract the observation metadata, which includes: observation center frequency, total bandwidth, number of frequency channels, channel width, time sampling interval, and dispersion measurement value; According to the given pulse arrival time, use the dispersion time delay formula to correct the pulse arrival time; Step S1.2: Time-frequency window truncation and downsampling: Based on the corrected pulse arrival time and the user-specified time window, calculate the range of time sampling points to be truncated from the original data; Step S1.3: Perform physical correction to eliminate dispersion; Step S1.4: Baseline subtraction and radio frequency interference removal; Step S1.5: Normalization and dynamic range adjustment; All pixel values ​​of the two-dimensional time-frequency map matrix processed above are linearly normalized to the standard range of [0,1] or [0,255] by Min-Max scaling; After the above five steps, the original FRB signal data is transformed into a two-dimensional time-frequency diagram.

5. The intelligent pulse parameter detection method based on deep learning according to claim 4, characterized in that, In step S1.3, the dedispersion is specifically implemented as follows: The system is based on each frequency channel Relative to reference frequency Dispersion time delay Perform correction: Calculated dispersion time delay Quantized as discrete time sampling point offsets: For each frequency channel For time-series data, perform a cyclic shift of an integer number of sampling points; the shift amount is... ;in, It is the dispersion constant. It is a dispersion measurement value. The time sampling interval; After this operation, the pulse trajectory, which was originally tilted due to dispersion, will be "straightened" on the time-frequency diagram, appearing simultaneously on all frequency channels and forming a rectangular energy cluster.

6. The intelligent pulse parameter detection method based on deep learning according to claim 4, characterized in that, In step S1.4, the baseline subtraction and radio frequency interference removal are as follows: Baseline subtraction: Calculate a baseline value independently for the time series of each frequency channel, and then subtract this baseline value from all data points of that channel; RFI removal is achieved using a statistically based automated RFI removal strategy, including: Narrowband RFI: Calculate the energy standard deviation of each frequency channel over the entire time window; Channels with a standard deviation higher than a set first threshold are identified as RFI-contaminated, and their data will be zeroed or replaced by interpolation. Wideband RFI: Calculate the total energy of each time sampling point across all frequency channels; time points where the total energy exceeds a set second threshold are identified as wideband RFI and are zeroed out.

7. A pulse parameter intelligent detection system based on deep learning, characterized in that, The system is used to implement the method described in any one of claims 1-6, to automatically analyze the time-frequency data of transient astronomical signals, and to obtain the signal-to-noise ratio, pulse width, and frequency bandwidth of the FRB signal.

8. The pulse parameter intelligent detection system based on deep learning according to claim 7, characterized in that, The system includes a standardized preprocessing module and a deep learning model; the standardized preprocessing module is used to perform physical correction on the raw FRB signal data; the deep learning model is used to output all target parameters end-to-end simultaneously, including the signal-to-noise ratio, pulse width, and frequency bandwidth of the FRB signal.

Images