Method and device for detecting solar radio burst information

Abstract

The application discloses a solar radio spectrum burst information detection method and device, and the method comprises the following steps: acquiring a solar radio spectrum image; performing feature extraction on the solar radio spectrum image to obtain a plurality of feature maps with different scales; mapping the plurality of feature maps with different scales to the solar radio spectrum image to obtain a plurality of regions of interest; pooling the plurality of regions of interest into a plurality of fixed-size feature maps; and determining information of each burst signal by using the plurality of fixed-size feature maps and a pre-established spectrum structure detection model. The application can effectively improve the accuracy of the detection result.

Description

Technical Field

[0001] This invention relates to the field of spectrum detection technology, specifically to a method and apparatus for detecting solar radio burst information. Background Technology

[0002] Solar radio bursts are radio emissions that increase dramatically in intensity when the Sun is subjected to strong disturbances. Currently, with the increasing number of ground-based and space-based solar radio observation instruments possessing high temporal, spatial, and frequency resolutions, numerous fine structures in the radio spectrum have been discovered. These fine structures are complex and diverse, and can be broadly categorized into Type I, Type II, Type III, Type IV, and Type V bursts. Within these bursts, complex fine structure phenomena such as spike bursts, zebra stripe structures, fibrous structures, rapid quasi-periodic pulsations, and lace-like structures are frequently observed. These often exhibit unique radiation brightness temperature, polarization, and rapid time-frequency structures, which are likely closely related to the non-thermal processes in the solar burst source region and the specific dynamic states of the plasma. Extracting characteristic parameters such as bandwidth, duration, frequency drift rate, relative bandwidth, and polarization of these burst phenomena is of significant value and importance for understanding the nature of solar bursts.

[0003] Currently, there is relatively little research on the detection of fine structures in the solar radio spectrum. The detection of fine structures in the solar radio spectrum is still at the stage of wavelet transform and traditional machine learning. Whether it is the morphological features extracted from the spectral image or the parametric features extracted through event analysis, the extraction process is very time-consuming because these features are manually designed, and they lack adaptability and robustness, which cannot meet the needs of massive data processing. Summary of the Invention

[0004] This invention provides a method and apparatus for detecting solar radio burst information, which can effectively improve the accuracy of detection results.

[0005] Therefore, the present invention provides the following technical solution:

[0006] A method for detecting solar radio burst information, the method comprising:

[0007] Acquire images of the solar radio spectrum;

[0008] Feature extraction was performed on the solar radio spectrum image to obtain multiple feature maps at different scales;

[0009] The feature maps of different scales are mapped onto the solar radio spectrum image to obtain multiple regions of interest;

[0010] The multiple regions of interest are aligned into multiple feature maps of fixed size;

[0011] By utilizing the multiple fixed-size feature maps and the pre-established spectral structure detection model, the information of each burst signal is determined.

[0012] Optionally, feature extraction is performed on the solar radio spectrum image to obtain multiple feature maps at different scales, including:

[0013] The solar radio spectrum image is input into a pre-constructed target detection network, and multiple feature maps of different scales are obtained based on the output of the target detection network.

[0014] Optionally, the object detection network includes a ResNet101 network and an FPN network; the ResNet101 network includes multiple convolutional layers, divided into multiple stages, and the size of the output feature map is reduced to half of its original size after each stage; the FPN network includes three parts: bottom-up, top-down, and lateral connections, which are used to process the C features at different scales obtained by the ResNet101 network separately, and then perform fusion processing to obtain P series features.

[0015] Optionally, mapping the multiple feature maps of different scales onto the solar radio spectrum image to obtain multiple regions of interest includes:

[0016] The feature maps of different scales are input into the RPN to determine the anchor boxes corresponding to each feature map. The anchor boxes are then subjected to binary classification and bounding box regression to obtain multiple regions of interest.

[0017] Optionally, aligning the plurality of regions of interest into a plurality of feature maps of fixed size includes:

[0018] The image values ​​of each region of interest are obtained by using bilinear interpolation with floating-point coordinates, resulting in multiple feature maps of fixed size.

[0019] Optionally, determining the information of each burst signal using the multiple fixed-size feature maps and the pre-established spectral structure detection model includes:

[0020] The burst signals within the multiple fixed-size feature maps are segmented at the pixel level using a pre-established shape contour detection model to obtain the shape contour of each burst signal; and / or

[0021] The multiple fixed-size feature maps are input into a pre-established category detection model to obtain candidate boxes and categories for each burst signal.

[0022] Optionally, the method further includes:

[0023] The burst parameters are calculated based on the coordinates of the candidate bounding box of the burst signal; the burst parameters include any one or more of the following: frequency bandwidth B, duration ΔT, and start time t. start End time t end Start frequency f start Stop frequency f end Center frequency f center Frequency drift rate B / Δt.

[0024] A solar radio burst information detection device, the device comprising:

[0025] The image acquisition module is used to acquire images of the solar radio spectrum.

[0026] The feature extraction module is used to extract features from the solar radio spectrum image to obtain feature maps of different scales;

[0027] The mapping module is used to map the multiple feature maps of different scales onto the solar radio spectrum image to obtain multiple regions of interest;

[0028] The alignment module is used to align the multiple regions of interest into multiple feature maps of fixed size;

[0029] The detection module is used to determine the information of each burst signal by utilizing the multiple fixed-size feature maps and the pre-established spectral structure detection model.

[0030] Optionally, the detection module includes: a contour detection unit and / or a category detection unit;

[0031] The contour detection unit is used to perform pixel-level segmentation of the burst signals in the multiple fixed-size feature maps using a pre-established shape contour detection model to obtain the shape contour of each burst signal.

[0032] The category detection unit is used to input the multiple fixed-size feature maps into a pre-established category detection model to obtain candidate boxes and categories for each burst signal.

[0033] Optionally, the device further includes:

[0034] The calculation module is used to calculate the burst parameters based on the coordinate values ​​of the candidate box of the burst signal.

[0035] The solar radio burst information detection method and apparatus provided in this invention extracts features from solar radio spectrum images to obtain multiple feature maps of different scales; maps these extracted feature maps onto the solar radio spectrum image to obtain multiple regions of interest; pools these regions of interest into multiple feature maps of fixed sizes; and uses these fixed-size feature maps and a pre-established spectral structure detection model to determine the information of each burst signal. This invention, starting from the statistical perspective of large amounts of data, achieves approximation of complex functions through layer-by-layer nonlinear transformation of a deep network structure, thereby conveniently and accurately obtaining solar radio burst information. Attached Figure Description

[0036] Figure 1 This is a flowchart of a method for detecting solar radio burst information provided in an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the target detection network composed of a ResNet101 network and an FPN network in an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of the RPN structure in an embodiment of the present invention;

[0039] Figure 4 This is a sampling diagram illustrating pooling of the region of interest using bilinear interpolation in an embodiment of the present invention;

[0040] Figure 5 Yes Figure 4 A schematic diagram of bilinear interpolation calculations performed on the sampling points in the diagram;

[0041] Figure 6 This is a schematic diagram of the Mask R-CNN deep learning network model in an embodiment of the present invention;

[0042] Figure 7 This is a flowchart illustrating the training process of the Mask R-CNN deep learning network model in an embodiment of the present invention.

[0043] Figure 8 This is a schematic diagram of a solar radio burst information detection device provided in an embodiment of the present invention. Detailed Implementation

[0044] To enable those skilled in the art to better understand the embodiments of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and implementation methods.

[0045] In existing technologies, there are two main types of methods for detecting the fine structure of the solar radio spectrum:

[0046] (1) Traditional detection methods mainly include the following:

[0047] a. Threshold-based segmentation method: The threshold-based segmentation method uses a contiguous search algorithm to set a threshold based on the actual noise level to determine all the fine structures in the entire burst region; the maximum inter-class variance method is used to realize the automatic detection of Type III radio bursts.

[0048] b. Image segmentation method based on wavelet analysis and wavelet transform: The image segmentation method based on wavelet analysis and wavelet transform accurately calculates the burst background of eight peak events and obtains their polarization profiles through wavelet transform and generalized interpolation; it uses feature detection and wavelet spectrum estimation to perform bandwidth analysis on the peak events; it separates the fiber structure in the solar radio spectrum through wavelet transform, calculates the frequency drift rate using linear regression, and calculates characteristic parameters such as instantaneous bandwidth and duration using numerical approximation and contiguous search algorithms; it uses the Ricker wavelet, the parent wavelet of continuous wavelet transform, to extract parameters such as peak flux density, spectral width, time span, and peak power of the spectrum, and detects the weakest feature to date.

[0049] c. Segmentation method based on active contour model: The Level set algorithm is used to extract the contour of fine structure and calculate the frequency drift rate; Hough transform is used to identify the line segments associated with radio type III bursts, and then an improved active contour model is used to track the bursts and estimate the frequency drift rate.

[0050] d. Edge detection-based segmentation method: By preprocessing, spectral intensity transformation, and 1 / f transformation of the spectrum to reduce the curvature of the Type II burst line, automatic detection of Type II bursts is achieved. Hough transform is then used to detect approximately straight lines to distinguish Type III bursts.

[0051] e. Statistical segmentation method: Gaussian model and codebook model are used for background modeling. The background subtraction idea in moving target detection is used to realize automatic burst detection. Then, binary morphological operation is used for smoothing and denoising to achieve accurate burst region localization. Finally, the binary burst region is automatically measured to obtain the burst feature parameter values.

[0052] (2) Machine learning-based methods

[0053] With the emergence of machine learning, algorithms can automatically learn features from large amounts of image data using network structures of varying sizes and various learning rules. Machine learning, especially deep learning, has achieved remarkable research results in many fields. Some studies have applied deep learning methods to the fine structure classification of solar radio spectra, achieving some progress. For example, several deep learning network models (including deep belief networks (DBN), multimodal networks, convolutional neural networks (CNN), and long short-term memory networks (LSTM)) have been used to perform binary classification of solar radio spectra for the presence or absence of bursts; CNN+capsule networks have been used for tri-class classification of spectra for Type III, Type IV, and other types, demonstrating that deep learning networks can automatically identify Type III and Type IV features. However, research on spectrum detection based on deep learning is relatively limited. Currently, only Type III bursts and spike bursts have been detected, and no research on the detection of other burst types has been found. The location coordinates of the outbreak event extracted by object detection networks such as Faster R-CNN, YOLOv2, and RetinaNet can be used to calculate the characteristic parameters of the outbreak, such as duration, frequency bandwidth, start and end frequency, and frequency drift rate. However, because the event edge is located using a rectangular bounding box, it is not possible to conduct in-depth research on the morphological contour feature parameters of the outbreak. Therefore, it is still insufficient in terms of localization and parameter extraction.

[0054] To this end, embodiments of the present invention provide a model, method and apparatus for detecting the fine structure of solar radio spectra. The automatic detection algorithm for the fine structure of solar radio spectra based on deep learning starts from the statistical perspective of a large amount of data and achieves the approximation of complex functions through layer-by-layer nonlinear transformation of deep network structure, and automatically labels burst parameters.

[0055] like Figure 1 The diagram shown is a flowchart of a method for detecting solar radio burst information provided in an embodiment of the present invention. The method includes the following steps:

[0056] Step 101: Obtain a solar radio spectrum image.

[0057] Step 102: Extract features from the solar radio spectrum image to obtain feature maps of different scales.

[0058] Specifically, an object detection network can be pre-built, and feature maps of multiple different scales can be obtained using this object detection network.

[0059] The target detection network is as follows Figure 2As shown, the object detection network includes a ResNet network and FPN (Feature Pyramid Networks). Of course, the ResNet network can also be replaced by networks such as VGG16, VGG19, GoogLeNet, and ResNet50, and this embodiment of the invention does not limit the specific network used.

[0060] In this embodiment of the invention, the ResNet network uses cross-layer connections, also known as residual connections, which can effectively prevent gradient vanishing.

[0061] The ResNet network includes multiple (e.g., 101) convolutional layers, divided into multiple stages (e.g., 5 stages). After each stage, the size of the output feature map is reduced to half of its original size.

[0062] FPN is a multi-scale detection network that avoids the problems of low accuracy and high false negative rate for small targets due to the limited feature size obtained during target detection.

[0063] In this embodiment of the invention, the FPN comprises three parts: bottom-up, top-down, and lateral connections. These parts are used to process C features at different scales obtained from the ResNet network separately, and then perform fusion processing to obtain P-series features. This structure of the FPN allows for the fusion of features from various levels output by the ResNet network to obtain P-series features, giving them both strong semantic and spatial information.

[0064] When extracting features from a solar radio spectrum image, the solar radio spectrum image is input into the target detection network, and multiple feature maps of different scales can be obtained based on the output of the target detection network.

[0065] It should be noted that FPN can obtain multi-layer features at different resolutions, and features at different resolutions have different uses. Low-resolution features are more suitable for detecting large targets, while small targets are better detected using low-level high-resolution features. The relationship between the selected feature level and the scale of the target to be detected is as follows:

[0066]

[0067] Where w and h represent the width and height of the RoI (Region of Interest) in the input image, respectively, and k0 represents the feature layer corresponding to the target region of size 256×256.

[0068] Step 103: Map the multiple feature maps of different scales onto the solar radio spectrum image to obtain multiple regions of interest.

[0069] In this embodiment of the invention, the Region Proposal Network (RPN) can be used to determine the region of interest.

[0070] The RPN is a lightweight neural network that uses a sliding window to scan an image and look for regions where a target exists.

[0071] like Figure 3 The diagram shown is a schematic representation of an RPN structure in an embodiment of the present invention. The RPN includes a 3×3 convolutional layer and two 1×1 convolutional layers.

[0072] On a feature map, each point can be mapped to a certain location in the original image. In the original image, anchor boxes with different scales and aspect ratios are predefined with this location as the center. These anchor boxes are then input into two network layers for binary classification (foreground or background) and bounding box regression to obtain multiple RoIs.

[0073] Step 104: Align the multiple regions of interest into multiple feature maps of fixed size.

[0074] Since the multiple RoIs generated by RPN have different scales and lengths, it is necessary to map them to produce feature maps of a fixed size. In Faster RCNN (Faster Regions with CNN features), RoI Pooling is commonly used to pool the corresponding regions in the feature map into fixed-size feature maps based on the position coordinates of the pre-selected boxes. Because the positions of the pre-selected boxes are usually obtained by model regression and are floating-point numbers, while the pooled feature maps require a fixed size, RoI Pooling involves two quantization processes. At this point, the candidate boxes have already deviated somewhat from the initially regressed positions, and this deviation affects the accuracy of detection or segmentation.

[0075] Therefore, in this embodiment of the invention, the quantization operation is cancelled, and the bilinear interpolation method is used to obtain the image values ​​of each region of interest with floating-point coordinates, thereby transforming the entire feature aggregation process into a continuous operation, and finally obtaining multiple feature maps of fixed size.

[0076] Figure 4 This is a sampling diagram illustrating the alignment of the region of interest using bilinear interpolation in an embodiment of the present invention. Figure 5 Yes Figure 4 A schematic diagram of bilinear interpolation calculations performed on the sampling points.

[0077] Step 105: Using the multiple fixed-size feature maps and the pre-established spectral structure detection model, determine the information of each burst signal.

[0078] The feature map obtained in step 104 above can be used for multiple tasks, such as any one or more of the following tasks:

[0079] (1) The burst signals in the multiple fixed-size feature maps are segmented at the pixel level using a pre-established outline detection model to obtain the outline (i.e., Mask) of each burst signal.

[0080] (2) Input the multiple fixed-size feature maps into the pre-established category detection model to obtain the candidate boxes and categories of each burst signal.

[0081] It should be noted that the above-mentioned shape contour detection model and category detection model can be trained independently or together, and this embodiment of the invention does not limit this.

[0082] The shape contour detection model can use FCN (Fully Convolutional Networks), while the category detection model can use a neural network with two convolutional layers.

[0083] The difference from existing neural network training methods lies in the fact that, during model training, a multi-level task loss function is defined for each sample's RoI:

[0084] L = L cls +L box +L mask (2)

[0085]

[0086]

[0087]

[0088]

[0089] Among them, L cls For classification loss, L box For the target bounding box regression loss, L mask The loss is due to the Mask branch.

[0090] Where i and j correspond to anchor points; if the anchor point is positive... on the contrary t i Represents the positive foreground box. This represents the ground truth (GT) box (background frame) corresponding to the positive Bbox.

[0091] In the Mask branch, the mask features predicted for class k are first obtained. Then, the mask region enclosed by the Bbox (candidate box) in the original image is mapped to an m*m mask region feature. Finally, the average binary cross-loss entropy of this m*m region is calculated. Sigmoid is used for classification to avoid inter-class loss, thus improving instance segmentation.

[0092] The solar radio burst information detection method provided in this invention extracts features from a solar radio spectrum image to obtain multiple feature maps of different scales; maps these extracted feature maps onto the solar radio spectrum image to obtain multiple regions of interest; pools these regions of interest into multiple feature maps of fixed sizes; and uses these fixed-size feature maps and a pre-established spectral structure detection model to determine the information of each burst signal. This invention, starting from the statistical perspective of large amounts of data, achieves approximation of complex functions through layer-by-layer nonlinear transformation of a deep network structure, thereby conveniently and accurately obtaining solar radio burst information.

[0093] It should be noted that in practical applications, the networks used in the above process can be combined to train a Mask R-CNN deep learning network. Using the Mask R-CNN deep learning network model, solar radio spectrum instance segmentation can be performed. That is, given any solar radio spectrum image, the Mask R-CNN deep learning network outputs the burst type, location information, and a binary mask for each category.

[0094] The structure of the Mask R-CNN deep learning network model is as follows: Figure 6 As shown.

[0095] Reference Figure 6 The Mask R-CNN deep learning network model includes: an object detection layer, a feature mapping layer, an alignment processing layer, a classification, bounding box regression, and a mask generation layer.

[0096] The object detection layer can use a ResNet101+FPN network, the feature mapping layer can use an RPN network, and the classification, bounding box regression, and mask generation layers can use neural networks. The network structure of each layer can be referred to the previous description, and will not be repeated here.

[0097] The training process of the Mask R-CNN deep learning network model is as follows: Figure 7 As shown, the specific process is as follows:

[0098] (1) Read the input data.

[0099] The input data can be a solar radio spectrum in .fits format.

[0100] (2) Data preprocessing

[0101] The main purpose is to remove RFI (radio frequency interference) from the spectrum to obtain a clear spectrum of solar radio bursts.

[0102] (3) Data labeling

[0103] Solar flare samples can be selected using a deep learning classification network. Then, the open-source image annotation software Labelme / Labellmg can be used to label the solar radio spectral map with rectangular or polygonal boxes, generating JSON-formatted labels. The `json_to_dataset.py` code is then run to convert the dataset into a Mask R-CNN neural network dataset. The dataset is then split into training and test sets according to a specific ratio.

[0104] (4) Model training and optimization

[0105] The Mask R-CNN network was used to learn the fine structure of the solar radio spectrum. The training process consisted of two parts: first, the backbone network was trained, mainly used to identify candidate regions; after the backbone network was trained, the complete network was trained. The number of image data samples fed into the network during training was determined by the memory of the running hardware platform, typically set to 2. n The specific standard for the number of training rounds in both parts is to stop when the loss function tends to stabilize.

[0106] Model optimization: Improve the network model and evaluation metrics based on actual conditions, adjust parameters to achieve the best recognition effect, and save the model parameters. Then consider using parallel computing to improve the model's detection speed.

[0107] (5) Identification of outbreak type and location

[0108] By inputting the test image into the trained Mask R-CNN network, mask information can be generated, which yields the burst type and the burst location coordinates.

[0109] The four coordinate values ​​are returned in the following order: x left y up x right y down These correspond to the top-left and bottom-right coordinates of the outbreak event detection box, respectively.

[0110] The burst types may include, but are not limited to: Type I burst, Type II burst, Type III burst, Type IV burst, Type V burst, spike burst, etc.

[0111] Using the Mask R-CNN deep learning network trained above, any solar radio spectrum image can be processed through the Mask R-CNN deep learning network to output the burst type, location information, and binary mask for each category in the image, thereby completing the solar radio spectrum instance segmentation through Mask R-CNN.

[0112] Furthermore, in a non-limiting embodiment, burst parameters can also be calculated based on the coordinate values ​​of the candidate bounding box of the burst signal. The burst parameters include, but are not limited to, any one or more of the following: frequency bandwidth B, duration ΔT, and start time t. start End time t end Start frequency f start Stop frequency f end Center frequency f center Frequency drift rate B / Δt. Here, the default input original spectrum graph increases from left to right on the time axis, while the frequency increases from top to bottom.

[0113] The formulas for calculating each burst parameter are as follows:

[0114] B = (y down -y up )×k f ×f resolution (7)

[0115] ΔT=(x right -x left )×k t ×t resolution (8)

[0116] t start =x left ×k t ×t resolution +T start (9)

[0117] t end =x right ×k t ×t resolution +T start (10)

[0118] f start =y down ×k f ×f resolution +F start (11)

[0119] f end =y up ×k f ×f resolution +F start (12)

[0120] f center =(f start +f end ) / 2 (13)

[0121]

[0122] Where, k t k represents the ratio of the actual number of sampling points on a single frequency channel to the number of pixels in the generated image length on the time axis. f This represents the ratio of the actual number of data points collected at a single time point to the number of pixels in the height of the generated image, on the frequency axis. resolution f resolution T represents the time resolution and frequency resolution of the input raw spectrogram, respectively. start F start These represent the start time and start frequency of the spectrum.

[0123] Accordingly, the present invention also provides a solar radio spectrum burst information detection device, such as... Figure 8 As shown, the solar radio burst information detection 800 includes the following modules:

[0124] Image acquisition module 801 is used to acquire solar radio spectrum images;

[0125] Feature extraction module 802 is used to extract features from the solar radio spectrum image to obtain multiple feature maps of different scales;

[0126] The mapping module 803 is used to map the multiple feature maps of different scales onto the solar radio spectrum image to obtain multiple regions of interest;

[0127] Alignment module 804 is used to align the plurality of regions of interest into a plurality of feature maps of fixed size;

[0128] The detection module 805 is used to determine the information of each burst signal by utilizing the multiple fixed-size feature maps and the pre-established spectral structure detection model.

[0129] The detection module may include a contour detection unit and / or a category detection unit.

[0130] The contour detection unit is used to perform pixel-level segmentation of the burst signals in the multiple fixed-size feature maps using a pre-established shape contour detection model to obtain the shape contour of each burst signal.

[0131] The category detection unit is used to input the multiple fixed-size feature maps into a pre-established category detection model to obtain candidate boxes and categories for each burst signal.

[0132] The solar radio burst information detection device provided in this invention extracts features from solar radio spectrum images to obtain multiple feature maps of different scales; maps these extracted feature maps onto the solar radio spectrum image to obtain multiple regions of interest; pools these regions of interest into multiple feature maps of fixed size; and uses these fixed-size feature maps and a pre-established spectral structure detection model to determine the information of each burst signal. This invention, starting from the statistical perspective of large amounts of data, achieves approximation of complex functions through layer-by-layer nonlinear transformation of a deep network structure, thereby conveniently and accurately obtaining solar radio burst information.

[0133] Furthermore, in another non-limiting embodiment of the device of the present invention, it may further include: a calculation module (not shown), used to calculate the burst parameters based on the coordinate values ​​of the candidate box of the burst signal.

[0134] The specific implementation methods of the above modules and units can be referred to the description in the previous embodiments of the present invention, and will not be repeated here.

[0135] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0136] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. Furthermore, the system embodiments described above are merely illustrative. The modules and units described as separate components may or may not be physically separate; that is, they may be located on a single network unit or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0137] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and apparatus of the present invention, and are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting solar radio burst information, characterized in that, The method includes: Acquire images of the solar radio spectrum; Feature extraction was performed on the solar radio spectrum image to obtain multiple feature maps at different scales; The feature maps of different scales are mapped onto the solar radio spectrum image to obtain multiple regions of interest; The multiple regions of interest are aligned into multiple feature maps of fixed size; Using the multiple fixed-size feature maps and the pre-established spectral structure detection model, the information of each burst signal is determined; The process of determining the information of each burst signal using the multiple fixed-size feature maps and the pre-established spectral structure detection model includes: The burst signals within the multiple fixed-size feature maps are segmented at the pixel level using a pre-established shape contour detection model to obtain the shape contour of each burst signal; the multiple fixed-size feature maps are then input into a pre-established category detection model to obtain the candidate boxes and categories of each burst signal. The burst parameters are calculated based on the coordinates of the candidate bounding box of the burst signal; the burst parameters include any one or more of the following: frequency bandwidth. Duration Start time End time Start frequency Stop frequency Center frequency Frequency drift rate ; The formulas for calculating each burst parameter are as follows: ； ； ； ； ； ； ； ； in, This represents the ratio of the actual number of sampling points on a single frequency channel to the number of pixels in the generated image length, on the time axis. This represents the ratio of the actual number of data points collected at a single time point to the number of pixels in the height of the generated image, expressed on the frequency axis. , These represent the time resolution and frequency resolution in the original input spectrogram, respectively. , These represent the start time and start frequency of the spectrum. Feature extraction was performed on the solar radio spectrum image to obtain multiple feature maps at different scales, including: The solar radio spectrum image is input into a pre-constructed target detection network, and multiple feature maps of different scales are obtained based on the output of the target detection network. The target detection network includes a ResNet101 network and an FPN network. The ResNet101 network includes multiple convolutional layers, divided into multiple stages. After each stage, the size of the output feature map is reduced to half of its original size. The FPN network includes three parts: bottom-up, top-down, and lateral connections. It is used to process the C features at different scales obtained by the ResNet101 network separately, and then perform fusion processing to obtain P-series features.

2. The method according to claim 1, characterized in that, The process of mapping the multiple feature maps of different scales onto the solar radio spectrum image to obtain multiple regions of interest includes: The feature maps of different scales are input into the RPN to determine the anchor boxes corresponding to each feature map. The anchor boxes are then subjected to binary classification and bounding box regression to obtain multiple regions of interest.

3. The method according to claim 1, characterized in that, The step of aligning the plurality of regions of interest into a plurality of feature maps of fixed size includes: The image values ​​of each region of interest are obtained by using bilinear interpolation with floating-point coordinates, resulting in multiple feature maps of fixed size.

4. A device for detecting solar radio burst information, characterized in that, The device includes: The image acquisition module is used to acquire images of the solar radio spectrum. The feature extraction module is used to extract features from the solar radio spectrum image to obtain feature maps of different scales; The mapping module is used to map the multiple feature maps of different scales onto the solar radio spectrum image to obtain multiple regions of interest; The alignment module is used to align the multiple regions of interest into multiple feature maps of fixed size; The detection module is used to determine the information of each burst signal by utilizing the multiple fixed-size feature maps and the pre-established spectral structure detection model; The process of determining the information of each burst signal using the multiple fixed-size feature maps and the pre-established spectral structure detection model includes: The burst signals within the multiple fixed-size feature maps are segmented at the pixel level using a pre-established shape contour detection model to obtain the shape contour of each burst signal; the multiple fixed-size feature maps are then input into a pre-established category detection model to obtain the candidate boxes and categories of each burst signal. The burst parameters are calculated based on the coordinates of the candidate bounding box of the burst signal; the burst parameters include any one or more of the following: frequency bandwidth. Duration Start time End time Start frequency Stop frequency Center frequency Frequency drift rate ; The formulas for calculating each burst parameter are as follows: ； ； ； ； ； ； ； ； in, This represents the ratio of the actual number of sampling points on a single frequency channel to the number of pixels in the generated image length, on the time axis. This represents the ratio of the actual number of data points collected at a single time point to the number of pixels in the height of the generated image, expressed on the frequency axis. , These represent the time resolution and frequency resolution in the original input spectrogram, respectively. , These represent the start time and start frequency of the spectrum.

Images