CN115641441B - Magnetic layer system soft X-ray photon number maximum value detection method - Google Patents
Magnetic layer system soft X-ray photon number maximum value detection method Download PDFInfo
- Publication number
- CN115641441B CN115641441B CN202211432905.7A CN202211432905A CN115641441B CN 115641441 B CN115641441 B CN 115641441B CN 202211432905 A CN202211432905 A CN 202211432905A CN 115641441 B CN115641441 B CN 115641441B
- Authority
- CN
- China
- Prior art keywords
- magnetic layer
- layer system
- photon number
- soft
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 52
- 230000011218 segmentation Effects 0.000 claims abstract description 38
- 230000004927 fusion Effects 0.000 claims abstract description 21
- 230000006872 improvement Effects 0.000 claims abstract description 20
- 238000007781 pre-processing Methods 0.000 claims abstract description 5
- 239000011800 void material Substances 0.000 claims abstract description 5
- 230000007246 mechanism Effects 0.000 claims abstract description 4
- 238000004088 simulation Methods 0.000 claims description 49
- 238000000034 method Methods 0.000 claims description 47
- 238000012549 training Methods 0.000 claims description 40
- 230000010354 integration Effects 0.000 claims description 22
- 239000011159 matrix material Substances 0.000 claims description 21
- 230000005855 radiation Effects 0.000 claims description 19
- 238000012360 testing method Methods 0.000 claims description 16
- 230000008569 process Effects 0.000 claims description 15
- 230000006870 function Effects 0.000 claims description 14
- 238000012545 processing Methods 0.000 claims description 10
- 238000005070 sampling Methods 0.000 claims description 10
- 230000008014 freezing Effects 0.000 claims description 7
- 238000007710 freezing Methods 0.000 claims description 7
- 238000010257 thawing Methods 0.000 claims description 7
- 238000011176 pooling Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 238000010586 diagram Methods 0.000 description 16
- 230000000694 effects Effects 0.000 description 10
- 238000003384 imaging method Methods 0.000 description 10
- 238000013135 deep learning Methods 0.000 description 6
- 238000000605 extraction Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 3
- 238000013079 data visualisation Methods 0.000 description 3
- 230000003993 interaction Effects 0.000 description 3
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 239000012530 fluid Substances 0.000 description 2
- 239000005433 ionosphere Substances 0.000 description 2
- 150000002500 ions Chemical class 0.000 description 2
- 230000007935 neutral effect Effects 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 101100295091 Arabidopsis thaliana NUDT14 gene Proteins 0.000 description 1
- 241000282414 Homo sapiens Species 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000001364 causal effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000004907 flux Effects 0.000 description 1
- ZZUFCTLCJUWOSV-UHFFFAOYSA-N furosemide Chemical compound C1=C(Cl)C(S(=O)(=O)N)=CC(C(O)=O)=C1NCC1=CC=CO1 ZZUFCTLCJUWOSV-UHFFFAOYSA-N 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 238000011144 upstream manufacturing Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 239000013585 weight reducing agent Substances 0.000 description 1
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Image Analysis (AREA)
Abstract
The invention relates to the technical field of space, in particular to a magnetic layer system soft X-ray photon number maximum value detection method, which comprises the following steps: collecting soft X-ray photon number data of the magnetic layer system under different solar wind densities, inputting the data into a pre-established and trained semantic segmentation network model after preprocessing, obtaining the classification of the soft X-ray photon number of the magnetic layer system, and further extracting the position of the maximum value of the soft X-ray photon number of the magnetic layer system; the semantic segmentation network model is obtained by improving the deep LabV3+ network structure, and the improvement comprises the following steps: changing the backbone network Xattention of deep LabV3+ into Mobilenetv2, introducing a CA attention mechanism into the Mobilenetv2 network, changing the void ratio combination of the hollow convolution in the deep LabV3+ feature fusion module, and adding the deformable convolution branching degree into the feature fusion module.
Description
Technical Field
The invention relates to the technical field of space, in particular to a magnetic layer system soft X-ray photon number maximum detection method, and more particularly relates to a deep LabV3+ improved magnetic layer system soft X-ray photon number maximum detection method.
Background
The solar wind-magnetic layer interaction panoramic imaging satellite task (Solar wind magnetosphere Ionosphere Link Explorer, SMILE) is a space science exploration cooperation project initiated by the national academy of sciences and the European space agency. The SMILE satellite can realize panoramic real-time imaging of the top of the solar earth magnetic layer, the polar tip region and the polar light elliptical region for the first time, observe a complete event chain for driving the spatial weather change, reveal the association and causal relationship of the interaction of the solar wind and the earth magnetic layer and know the driving factors of the spatial weather change. The task provides important scientific support in the aspect that human beings further perfect the physical model of the earth magnetic layer and improve the space environment forecasting capability.
A key scientific goal of the SMILE task is to detect the large scale structure and fundamental mode of solar wind-magnetic layer interactions, including detection of magnetic layer tops by soft X-ray imaging.
The top of the magnetic layer is the boundary region between solar wind and magnetic layer, which distinguishes solar wind plasma and magnetic field from earth magnetic layer plasma and magnetic field. The change in the top position and shape of the magnetic layer may reflect the effect of upstream solar wind conditions on the earth's magnetic field. When heavy ions of high charge state (e.g., c6+, o7+, o8+, fe12+, si10+) from the solar wind, encounter neutral atoms or molecules that are ubiquitous in the geospatial environment (e.g., H, H) 2 0. OH), the two collide and interact, electrons are transferred from neutral atoms and molecules to heavy ions, a process called the solar-wind charge exchange SWXC process. Photons of one or more extreme ultraviolet or soft X-ray wave bands are radiated in the process of solar-wind charge exchange SWXC, and then detected by satellites. The SMILE-mounted soft X-ray imager (SXI) is based on the imaging principle and is expected to provide soft X-ray images of a large-scale magnetic layer system near the area of the subsurface point.
Satellites have not yet been launched, and there is no soft X-ray photon map of the magnetic layer system that is actually observed. But can generate a simulation image through SXI simulation software, the principle of which is as follows:
the X-ray Intensity (IX) is related to the solar wind flux. For a given Line Of Sight (LOS), the X-ray intensity Of the coronal SWCX radiation can be integrated from the 3-dimensional X-ray emissivity (PX) along the field Of view, see equation (1)
The parameter n required in the formula (1) is obtained by a global three-dimensional magnetohydrodynamic numerical simulation code of a solar wind-magnetic layer-ionosphere system, namely an MHD (Magneto Hydro Dynamic ) model sw 、u sw 、u th 。
Each data point can be input as a point light source, the radiation intensity of each position obtained by MHD integral data is input as SXI (Solar X-ray image) simulation software, and the number N of light rays corresponding to each position is calculated. An image of SXI is formed from a two-dimensional matrix of photon numbers. The number of rays N is known to be proportional to the X-ray intensity value under its MHD simulation.
The two-dimensional image of the magnetic layer system obtained through observation or simulation also needs to obtain the three-dimensional position information of the top of the magnetic layer through calculation. At present, a TFA tangential fitting method is mainly adopted, and the maximum direction of an X-ray signal is obtained by analyzing integrated data of the intensity of X-ray radiation or photon data collected by SXI, and is taken as the tangential direction of the top of a real magnetic layer; and then a series of magnetic layer top bit shapes are obtained through a magnetic layer system model containing variable parameters, tangential directions are calculated aiming at various possible bit shapes, and the magnetic layer top bit shape which is best matched with the true tangential directions is found and is used as an inversion result of the three-dimensional magnetic layer top.
The method needs to find the maximum value of the X-ray photon intensity, adopts a method for taking the maximum value of each row of an X-ray radiation intensity integral data matrix or a photon number two-dimensional data matrix collected by SXI at present, is feasible in the detection of an MHD integral image which is a theoretical condition, but faces a simulated X-ray imaging image and the detection of an actual image obtained after the future satellite emission, can be influenced by universe background noise and instrument noise, and can not realize effective detection when the solar wind density is smaller. And this method requires long time energy accumulation and is not efficient for detection.
The detection of photon number maxima locations in a soft X-ray image of a magnetic layer system can be considered as a semantic segmentation problem. Traditional semantic segmentation methods include thresholding, region growing, edge segmentation, and the like. The segmentation methods are suitable for pictures with obvious property differences among target classes, all the pictures need to be segmented manually from design features in information such as gray level, contrast, texture and the like of the images, the influence of noise is large, and the segmentation precision and accuracy cannot reach expectations.
With the development of deep learning, the semantic segmentation method based on machine learning can fully utilize the semantic information of the image to extract features in a high-dimensional space, so that the precision and accuracy of semantic segmentation are improved. By employing convolutional neural networks, a computer device can efficiently process a large number of magnetic layer system soft X-ray photon data. The semantic segmentation based on deep learning commonly used at present comprises FCN, U-Net, segNet, PSPNet, deepLab series and the like. The deep LabV3+ uses an Encoder-Decoder (the high-level features provide semantics, and the Decoder replies boundary information step by step), so that the segmentation effect is improved, meanwhile, the boundary information is focused, and the deep LabV3+ is widely applied to CT semantic segmentation in the medical field at present and has good detection effect. However, the deep LabV3+ is directly used for detecting the maximum value of soft X-ray photon number of the magnetic layer system, so that the effect is poor.
Disclosure of Invention
The invention aims to effectively solve the detection problem of the position of the maximum value of the soft X-ray photon number of the magnetic layer system when the solar wind density is small, and improve the overall detection speed of magnetic layer top detection.
In order to achieve the above purpose, the present invention is realized by the following technical scheme. The invention provides a deep LabV3+ improved magnetic layer system soft X-ray photon number maximum detection method, which considers the extraction problem of photon intensity maximum influenced by solar wind density, integration time, universe background noise and instrument noise as a semantic segmentation problem. The spatial characteristics of soft X-ray photon numbers of the magnetic layer system with noise under different solar wind densities and different integration times are learned, so that the classification of each element of the two-dimensional array matrix of soft X-ray photon numbers of the magnetic layer system is obtained, and the position of the maximum value of the required photon numbers is extracted.
The invention provides a magnetic layer system soft X-ray photon number maximum value detection method, which comprises the following steps:
collecting soft X-ray photon number data of the magnetic layer system under different solar wind densities, inputting the data into a pre-established and trained semantic segmentation network model after preprocessing, obtaining the classification of the soft X-ray photon number of the magnetic layer system, and further extracting the position of the maximum value of the soft X-ray photon number of the magnetic layer system;
the semantic segmentation network model is obtained by improving the deep LabV3+ network structure, and the improvement comprises the following steps: changing the backbone network Xattention of deep LabV3+ into Mobilenetv2, introducing a CA attention mechanism into the Mobilenetv2 network, changing the void ratio combination of the hollow convolution in the deep LabV3+ feature fusion module, and adding a deformable convolution branch into the feature fusion module.
As one of the improvements of the above technical solution, the semantic segmentation network model specifically includes: an encoder and a decoder;
the encoder is used for respectively extracting shallow semantic information and deep semantic information of input soft X-ray photon number two-dimensional data of a magnetic layer system, and comprises the following steps: an improved mobilenet 2 backbone network and an improved multi-branch feature fusion module; wherein, improved MobileNetv2 backbone network includes: 1 conv-2D and 7 InvertedRsblock modules and one CA attention module; the improved multi-branch feature fusion module is characterized in that the cavity rate of 3 cavity convolutions of a cavity space convolution pooling pyramid module in a deep LabV3+ network structure is changed into 4, 8 and 12 respectively, and deformable convolution branches are added at the same time;
the decoder is used for fusing the shallow semantic information and the deep semantic information extracted by the encoder, upsampling, and finally outputting the detection result of each coordinate of the soft X-ray photon number two-dimensional simulation data of the magnetic layer system.
As one of the improvements of the above technical solution, the processing procedure of the semantic segmentation network model specifically includes:
step 1) preprocessing a photon number two-dimensional data matrix, and inputting the preprocessed two-dimensional data matrix into an encoder; processing the two-dimensional simulation data of the input photon number by conv-2D convolution of 3*3 to obtain a feature map; the dimension of the input feature map is enlarged by using 1*1 convolution, then a 3*3 depth convolution mode is used for convolution operation, finally the dimension of the feature map is reduced by using 1*1 convolution operation and a linear activation function is used, and the feature map with the reduced dimension is sent to a CA attention module;
step 2) obtaining deep semantic information and shallow semantic information of an input feature map through a CA attention module, transmitting the deep semantic information to a multi-branch feature fusion module, and transmitting the shallow semantic information to a decoder;
step 3) sequentially processing the deep semantic information through each branch of the feature fusion module to obtain feature graphs output by each branch, stacking the feature graphs output by each branch, and convolving and integrating the stacked feature graphs through 1*1 to obtain feature graphs with specified scales; the multi-branch feature fusion module comprises 1 1*1 convolution branch, 3 cavity convolution branches, wherein the cavity rate of the three cavity convolutions is respectively 4, 8, 12,1 deformable convolution branch and 1 global average pooling branch;
step 4) sampling the feature map obtained in the step 3) by 4 times and combining the feature map with the low-level semantic features obtained in the step 3) to obtain a combined feature map;
step 5), convolving the combined feature map output in the step 4) by 3*3, and then up-sampling by 4 times to obtain a feature map;
and 6) adjusting the size of the feature map output in the step 5) to the original size of the input photon number two-dimensional data matrix, and outputting a final prediction result.
As an improvement of the foregoing technical solution, the method further includes: training the semantic segmentation network model, wherein the training process comprises the following steps of:
step A: simulating and constructing soft X-ray radiation intensity three-dimensional data of the magnetic layer system under different solar wind densities by using a magnetohydrodynamic MHD model, integrating the soft X-ray radiation intensity three-dimensional data of the magnetic layer system under a fixed viewing angle to obtain a soft X-ray radiation intensity two-dimensional data matrix of the magnetic layer system, namely MHD two-dimensional data, and importing the MHD two-dimensional data into SXI simulation software to obtain photon number simulation data with different solar wind densities and different integration times; selecting a radiation intensity maximum value from the MHD two-dimensional data matrix as a label of a photon intensity maximum value, and adding the label into simulation data; respectively constructing a training data set and a testing data set based on the simulation data added with the labels;
and (B) step (B): and setting a loss function and model training parameters, training the semantic segmentation network model by using a training data set, testing by using a testing data set, and finally obtaining the trained semantic segmentation network model.
As one of the improvements of the above technical solution, the step a specifically includes:
step A1: taking MHD simulation results under different solar wind density conditions as light source input, generating SXI photon number simulation data under corresponding solar wind density conditions through a SXI simulation program, wherein the integration time of SXI istThe same MHD simulation result input can obtain the incomplete SXI photon number simulation data through multiple outputs;
step A2: will be arbitrarynThe Zhang Tongyang MHD simulation result input obtained incomplete SXI photon number simulation data are superimposed to obtain integration time ofntSXI photon count simulation data of (a);
step A3: the pixels of the maximum gray value of each row in the MHD two-dimensional integral graph form an X-ray radiation intensity maximum value region, the maximum gray value region is used as a label on the top of a magnetic layer in SXI simulation data under the condition of corresponding solar wind density, and the label is added into SXI simulation data;
step A4: after the label is added, the simulation data is enhanced, and the sample data is expanded;
step A5: based on the sample data, multiple sets of training data sets and test data sets are made by setting different solar wind densities or integration times.
As one of the improvements of the above technical solution, in the step A4, enhancement processing is performed on the simulated two-dimensional data matrix, which specifically includes: and performing operations of rotating and adding Gaussian noise on the simulation two-dimensional data matrix.
As one of the improvements of the above technical solution, in the step B, an SGD optimizer is used to update model training parameters during the training process.
As one of the improvements of the above technical solution, in the step B, the loss functionLossUsing a binary cross entropy loss function, the expression is:
wherein ,Nthe number of representative samples is represented by the number of samples,irepresent the firstiA number of samples of the sample were taken,i=1,2,…N,y i represents the firstiA sample tag of a sample is used,p i represents the firstiAnd predicted values.
As an improvement of the foregoing technical solution, in step B, the training process sequentially includes: a freezing stage and a thawing stage;
in the freezing stage, the MobileNetv2 backbone network of the semantic segmentation network model is frozen and operated;
and in the thawing stage, the MobileNetv2 backbone network of the semantic segmentation network model is enabled to normally operate.
Compared with a detection method for directly taking the maximum value, the detection method for the soft X-ray photon number maximum value of the magnetic layer system based on the DeepLabV3+ improvement has the following advantages:
1. according to the method, the spatial characteristics of soft X-ray photon numbers of the magnetic layer system with noise under different solar wind densities and different integration times are learned, so that the classification of each element of the two-dimensional array matrix of the soft X-ray photon numbers of the magnetic layer system is obtained, the maximum detection of photon numbers is realized, meanwhile, the requirement on the integration time of input data is reduced, and the overall detection speed of magnetic layer top detection is improved.
2. Aiming at the difficult problem of detecting the maximum value of the soft X-ray photon number of the magnetic layer system with smaller solar wind density, a network model based on deep LabV3+ improvement is designed, and the detection of the top position of the magnetic layer with smaller solar wind density is indirectly realized.
3. CA attention mechanism is introduced in the Mobilenetv2 network, so that the expression capability of the mobile network learning characteristics is enhanced.
4. The feature fusion module is changed on the original ASPP module, and the void ratio combination of the void convolution is changed from 6, 12 and 18 to 4, 8 and 12, so that the network is more in line with the feature extraction of soft X-ray photon numbers of the magnetic layer system, and meanwhile, the deformable convolution is added, so that the local positioning of a target can be self-adapted, and the feature learning of a detection area of a network pair is enhanced.
Drawings
FIG. 1 is a schematic diagram of an SMILE satellite mission.
Fig. 2 is a two-dimensional integral data visualization image of soft X-ray radiation intensity at fixed viewing angle for a magnetic layer system MHD, where fig. 2 (a) is solar wind density n=5In the case of (b) in fig. 2, the solar wind density n=12 +.>In the case of (c) in fig. 2, the solar wind density n=20 +.>In the case of (d) in fig. 2, the solar wind density n=25 +.>Is the case in (2); />
Fig. 3 is a view of the simulation software for visualizing the generated photon count simulation data using fig. 2 (a), 2 (b), 2 (c) and 2 (d)) as input, wherein the integration time of fig. 3 (a 1), 3 (a 2), 3 (a 3) and 3 (a 4) is 30s, the integration time of fig. 3 (b 1), 3 (b 2), 3 (b 3) and 3 (b 4) is 60s, and the integration time of fig. 3 (c 1), 3 (c 2), 3 (c 3) and 3 (c 4) is 120s;
FIG. 4 is a schematic diagram of the structure of a magnetic layer system soft X-ray photon number maximum detection network based on DeepLabV3+ improvement of the method of the present invention;
FIG. 5 is a schematic diagram of a modified Mobilenv 2 network architecture for use in the method of the present invention;
FIG. 6 is a schematic diagram of the structure of InvertedRsBlock1 and InvertedRsBlock2 in the modified Mobilenv 2 network used in the method of the present invention, wherein FIG. 6 (a) is a structure diagram of InvertedRsBlock1 and FIG. 6 (b) is a structure diagram of InvertedRsBlock 2;
FIG. 7 is a schematic diagram of the CA_Block architecture in the modified Mobilenv 2 network used in the method of the present invention;
FIG. 8 is a graph showing the effect of deep LabV3+ on the segmentation of the magnetic layer system soft X image photon intensity maxima before and after improvement in an embodiment of the present invention.
Detailed Description
Therefore, the invention provides a deep LabV3+ improved magnetic layer system soft X-ray photon maximum detection method, which aims at the problems of inaccurate detection, low detection speed and the like of soft X-ray imaging image photon maximum when the solar wind density is smaller in the existing method. In a Decoder (Decoder) module, high-level semantic information and low-level semantic information are stacked, the size of a feature map is gradually restored through up-sampling and other operations, the extraction of space information is completed, the detection precision of the soft X-ray photon number maximum value of a magnetic layer system is improved, meanwhile, the integral time requirement on input data is reduced, and the integral detection speed of magnetic layer top detection is greatly improved.
Aiming at the problem of magnetic layer top detection based on soft X-ray detection imaging in an SMILE satellite engineering task, as shown in FIG. 1, which is a schematic diagram of the SMILE satellite task, the application provides a magnetic layer system soft X-ray photon number maximum detection method based on deep LabV3+ improvement, which comprises the following steps:
step A: and constructing soft X-ray radiation intensity two-dimensional integral data of the magnetic layer system under different solar wind densities by using an MHD fluid mechanics model, and importing SXI simulation software into the MHD integral data to obtain simulation images with different solar wind densities and different integration times. And selecting a radiation intensity maximum value from the MHD two-dimensional integral data as a label of the SXI photon intensity maximum value under the same condition, and constructing a training data set and a testing data set.
And (B) step (B): and constructing a semantic segmentation network model based on deep learning. The network model includes two phases: and the decoding stage fuses the shallow semantic information and the deep semantic information extracted in the encoding stage, and finally outputs the detection result of each pixel after up-sampling.
Step C: and setting a loss function and model training parameters, and training a semantic segmentation network model based on deep learning.
And D, testing the detection effect of the SXI photon intensity maximum value on the simulation data set by using the trained semantic segmentation network model based on deep learning. In order to make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention will be described in further detail with reference to the accompanying drawings and examples.
Examples
The scheme of the invention consists of four parts: firstly, simulating data by using an MHD model and SXI, and establishing a training data set and a testing data set; then establishing a semantic segmentation network model; setting a loss function and training parameters, and training a network model by using a training data set; and finally, verifying the detection effect of the network model on the maximum photon intensity of the soft X imaging graph SXI of the magnetic layer system by using a test data set.
The first step: creating a dataset
And A, constructing magnetic layer system images under different solar wind densities by using an MHD fluid mechanics model, and importing the MHD integral data into SXI simulation software to obtain simulation data of different solar wind densities and different integral times. And selecting a radiation intensity maximum value from the MHD two-dimensional integral graph as a label of the SXI photon intensity maximum value under the same condition, and constructing a training data set and a testing data set. The specific process comprises the following steps:
as shown in fig. 2, which is a two-dimensional integrated data visualization image of the magnetic layer system MHD at a fixed viewing angle, fig. 2 (a) is solar wind density n=5 cm -3 (A) In the case of (a), fig. 2 (b) shows a solar wind density n=12 cm -3 (A) In the case of (c) in fig. 2, the solar wind density n=20 cm -3 (A) In the case of (d) in fig. 2, the solar wind density n=25 cm -3 (A) Is the case in (a).
As shown in fig. 3, the SXI simulation software visualizes the generated photon number simulation data using fig. 2 (a), 2 (b), 2 (c) and 2 (d)) as input, wherein the integration time of fig. 3 (a 1), 3 (a 2), 3 (a 3) and 3 (a 4) is 30s, the integration time of fig. 3 (b 1), 3 (b 2), 3 (b 3) and 3 (b 4) is 60s, and the integration time of fig. 3 (c 1), 3 (c 2), 3 (c 3) and 3 (c 4) is 120s.
Step A1: taking MHD simulation results under different solar wind conditions as light source input, generating SXI simulation data visual images under corresponding conditions through a SXI simulation program, wherein the integration time of SXI is 30s, and the same MHD input can obtain non-identical SXI simulation data through multiple outputs.
Step A2: the SXI simulation data visualization image with the integration time of 60s can be obtained by superposing any two SXI simulation images obtained by inputting the same MHD. Similarly, SXI simulation data with integration times of 120s, 180s, 240s, 300s can be obtained.
Step A3: the elements of each row of maxima in the MHD two-dimensional data matrix form the X-ray radiation intensity maxima region as a label on top of the magnetic layer in the corresponding SXI emulation data. ( The entire X-ray image is irradiated by the magnetic layer system, including the magnetic sheath and the pole tip region. The maxima on the X-ray map correspond to the top of the magnetic layer. )
Step A4: and (5) data enhancement processing. And (5) performing operations of rotating and adding Gaussian noise on SXI simulation data, and expanding sample data. Because the satellite orbit is fixed, the angle of the shot image is limited, and the data enhancement processing modes such as turning, translation, scaling, length and width distortion and the like are not considered.
Step A5: by setting different parameters, multiple sets of training data sets and test data sets are produced. Each data set includes no less than 20000 pairs of training data. Randomly dividing the training set, the verification set and the test set according to the proportion of 8:1:1.
And a second step of: construction of network model
FIG. 4 is a schematic diagram of the structure of the magnetic layer system soft X-ray photon number maximum detection network based on DeepLabV3+ improvement according to the method of the present invention.
And (B) step (B): and constructing a deep LabV3+ improved magnetic layer system soft X-ray photon number maximum detection network model. The network model includes two phases: and the decoding stage fuses the shallow semantic information and the deep semantic information extracted in the encoding stage, and finally outputs the detection result of each pixel after up-sampling. The specific process comprises the following steps:
step B1: and a SXI photon number two-dimensional data matrix of 33 x 55 is input, and is adjusted to 512 x 512 size through the size, so that the afferent neural network can conveniently carry out convolution operation.
Step B2: the backbone network is modified by MobileNetv2 and includes 1 conv-2D and 7 InvertedRsblock modules and a CA attention module. As shown in fig. 5. The InvertedRsblock module consists of 3 conv-2D, 2 Relu, 1 residual connections, as shown in FIG. 6. Where the size of the convolution kernel is 3 x3, padding is 1, and the step size is 1. The input feature map dimension is enlarged by 1*1 convolution, then the convolution operation is carried out in 3*3 depthwise convolution mode, finally the dimension is reduced by 1*1 convolution operation, the ReLU activation function is not used any more, but a linear activation function is used, so that more characteristic information is reserved, and the expression capability of the model is guaranteed.
As shown in fig. 5, a schematic diagram of the improved mobiletv 2 network structure used in the method of the present invention is shown. As shown in fig. 6, the structure of the InvertedRsblock1 (inverted residual block 1) and the InvertedRsblock2 (inverted residual block 2) in the improved mobiletv 2 network used in the method of the present invention is schematically shown, wherein fig. 6 (a) is a structure diagram of the InvertedRsblock1 and fig. 6 (b) is a structure diagram of the InvertedRsblock 2.
Step B3: the feature map obtained in step B2 is sent to the CA attention module as shown in fig. 7. The feature is aggregated from two dimensions of height and width, enhancing the localization of the target-photon number maxima by the mobiletv 2 network. High-level and low-level semantic features of the input feature map are acquired. The deep semantic information (30×30×320) is transmitted to the feature fusion module, the shallow semantic information (128×128×24) is transmitted to the Encode stage, and more global information is reserved.
As shown in fig. 7, a schematic diagram of the structure of ca_block (CA attention module) in the modified mobiletv 2 network used in the method of the present invention is shown.
Step B4: and (3) sending the high-level semantic features of the image output in the step (B3) to a multi-branch feature fusion module. The method comprises 1 convolution of 1*1 and 3 convolution of holes with expansion rates of 4, 8 and 12 respectively, different receptive fields are obtained, multi-scale information is captured, and experiments show that the expansion rates of 4, 8 and 12 are more suitable for extracting the characteristics of SXI photon number maximum value under the condition of not changing other conditions. The 1 deformable convolution dynamically adjusts the bias of the sampling position of the feature map according to the content of the current feature map, so that the network is more sensitive to the position of the photon number maximum value and the 1 global average pooling module is adopted, and the problem of effective weight reduction under long distance is solved. The feature maps of the branches are stacked as shown in formulas (2) - (8). The features were then integrated by 1*1 convolutions to obtain a 32 x 256 feature map.
Wherein: x represents an input signal, and X is a feature map finally output by the multi-branch feature fusion module; x1 represents a characteristic diagram of 1*1 convolution branch output; x2 represents a feature map of the cavity convolution branch output with the cavity rate of 4; x3 represents a feature map of the cavity convolution branch output with the cavity rate of 8; a feature map of the cavity convolution branch output with the X4 cavity rate of 12;、/>3/>the hole convolution operations with hole ratios of 4, 8, and 12 are shown.
Step B5: the feature map obtained in the step B4 is convolved by a 1*1 convolution to adjust the channel number, then 4 times up-sampling is carried out, and the feature map is combined with the low-level semantic features obtained in the step B2 to obtain a combined feature map;
step B6: the output in the step B5 is convolved by 3*3 and then up-sampled again by 4 times to obtain a 512 x 512 size characteristic diagram, and then a SXI photon number maximum value segmentation result with the size of 55 x 33 is output through the resize;
and a third step of: training network
And C, setting a loss function and model training parameters, and training the network model, wherein the training is divided into two stages, namely a freezing stage and a thawing stage. The specific process comprises the following steps:
step C1, using a binary cross entropy loss function as the loss function, as shown in a formula (9).
wherein NThe number of representative samples is represented by the number of samples,y i representing either a label 0 or a label 1,p i representing the predicted value
And step C2, updating network model parameters by adopting an SGD optimizer commonly used for deep learning, wherein the learning rate lr is 7e-3. Initializing training parameters, loading a training set and a verification set, creating a network model, and training a network. The training process includes two phases: a freezing stage and a thawing stage. In the freezing stage, the trunk of the model is frozen, the feature extraction network is not changed, the occupied video memory is small, only the network is finely tuned, the batch size is 8, and the epoch is 50. In the thawing stage, the trunk of the model is not frozen, the feature extraction network is changed, and the batch size is 4, and the total Epoch is 100.
Fourth step: detection of
Step D: testing the detection effect of the magnetic layer system soft X imaging graph SXI photon intensity maximum value on a simulation data set by using a trained network model, wherein the specific process comprises the following steps of:
step D1: and creating a network model and loading pre-training weights. The data set is input into a network model, and classification of each pixel of the soft X imaging diagram of the input magnetic layer system is obtained through network model processing.
Step D2: and storing the detection result as a jpg file, and evaluating the performance of the detection by using the mIOU. The detection effect is shown in fig. 8, and is a graph of the segmentation effect of the deep labv3+ network and the deep labv3+ network modified by the method on the maximum photon intensity of the soft X-ray image of the magnetic layer system in the embodiment. By comparison, it can be seen that: when the solar wind density is low (N)<10cm -3 (A) Network failure before improvement; at solar wind density of n=10cm -3 (A) To n=20 cm -3 (A) In the case of the two-dimensional network, the detection result of the network before improvement sometimes deviates, for example, the density of the sun wind is N=12 cm -3 (A) Photon number detection result of (2) and positive wind density n=13 cm -3 (A) Closer together; when the solar wind density is high (N)>20cm -3 (A) The top position of the magnetic layer can be extracted, but the network is more accurate after improvement.
As can be seen from the above detailed description of the present invention, the method for detecting the maximum value of soft X-ray photon number of the magnetic layer system based on deep LabV < 3+ > provided by the present application uses deep LabV < 3+ > as a network basic structure, changes the backbone network of the Encoder part into a Mobilenet 2 network, and inputs the Mobilenet 2 network into the multi-branch feature fusion module through the CA attention module to extract high-level semantic information. In the Decoder module, the high-level semantic information and the low-level semantic information are stacked, the size of the feature map is gradually restored through up-sampling and other operations, the extraction of the space information is completed, the detection precision of the soft X-ray photon number maximum value of the magnetic layer system is improved, the integral time requirement on input data is reduced, and the integral detection speed of magnetic layer top detection is greatly improved.
Finally, it should be noted that the above embodiments are only for illustrating the technical solution of the present invention and are not limiting. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made thereto without departing from the spirit and scope of the present invention, which is intended to be covered by the appended claims.
Claims (6)
1. A method for detecting soft X-ray photon number maxima of a magnetic layer system, the method comprising:
collecting soft X-ray photon number data of the magnetic layer system under different solar wind densities, inputting the data into a pre-established and trained semantic segmentation network model after preprocessing, obtaining the classification of the soft X-ray photon number of the magnetic layer system, and further extracting the position of the maximum value of the soft X-ray photon number of the magnetic layer system;
the semantic segmentation network model is obtained by improving the deep LabV3+ network structure, and the improvement comprises the following steps: changing the backbone network Xattention of deep LabV3+ into Mobilenetv2, introducing a CA attention mechanism into the Mobilenetv2 network, changing the void ratio combination of the hollow convolution in the deep LabV3+ feature fusion module, and adding a deformable convolution branch into the feature fusion module;
the semantic segmentation network model specifically comprises: an encoder and a decoder;
the encoder is used for respectively extracting shallow semantic information and deep semantic information of input soft X-ray photon number two-dimensional data of a magnetic layer system, and comprises the following steps: an improved mobilenet 2 backbone network and an improved multi-branch feature fusion module; wherein, improved MobileNetv2 backbone network includes: 1 conv-2D and 7 InvertedRsblock modules and one CA attention module; the improved multi-branch feature fusion module is characterized in that the cavity rate of 3 cavity convolutions of a cavity space convolution pooling pyramid module in a deep LabV3+ network structure is changed into 4, 8 and 12 respectively, and deformable convolution branches are added at the same time;
the decoder is used for fusing the shallow semantic information and the deep semantic information extracted by the encoder, upsampling, and finally outputting the detection result of each coordinate of the soft X-ray photon number two-dimensional simulation data of the magnetic layer system;
the processing procedure of the semantic segmentation network model specifically comprises the following steps:
step 1) preprocessing a photon number two-dimensional data matrix, and inputting the preprocessed two-dimensional data matrix into an encoder; by passing throughconv-2D convolution pairs of (c)Inputting photon number two-dimensional simulation data to process to obtain a feature map; then use->Convolution enlarges the dimension of the input feature map and then uses +.>Performing convolution operation in a deep convolution mode, and finally using +.>The dimension of the feature map is reduced by convolution operation of the feature map, a linear activation function is used, and the feature map with the reduced dimension is sent to a CA attention module;
step 2) obtaining deep semantic information and shallow semantic information of an input feature map through a CA attention module, transmitting the deep semantic information to a multi-branch feature fusion module, and transmitting the shallow semantic information to a decoder;
step 3) sequentially processing the deep semantic information through each branch of the feature fusion module to obtain feature graphs output by each branch, stacking the feature graphs output by each branch, and then passing throughConvolving the integrated and stacked feature images to obtain a feature image with a specified scale; the multi-branch feature fusion module comprises 1 +.>The method comprises the steps of convoluting branches, namely 3 cavity convoluting branches, wherein the cavity rate of three cavity convoluting branches is respectively 4, 8, 12,1 deformable convoluting branch and 1 global average pooling branch;
step 4) sampling the feature map obtained in the step 3) by 4 times and combining the feature map with the low-level semantic features obtained in the step 3) to obtain a combined feature map;
step 5) passing the merged feature map output in step 4)Performing 4 times up-sampling again after convolution to obtain a feature map;
step 6) adjusting the size of the feature map output in the step 5) to the original size of the input photon number two-dimensional data matrix, and outputting a final prediction result;
the method further comprises the steps of: training the semantic segmentation network model, wherein the training process comprises the following steps of:
step A: simulating and constructing soft X-ray radiation intensity three-dimensional data of the magnetic layer system under different solar wind densities by using a magnetohydrodynamic MHD model, integrating the soft X-ray radiation intensity three-dimensional data of the magnetic layer system under a fixed viewing angle to obtain a soft X-ray radiation intensity two-dimensional data matrix of the magnetic layer system, namely MHD two-dimensional data, and importing the MHD two-dimensional data into SXI simulation software to obtain photon number simulation data with different solar wind densities and different integration times; selecting a radiation intensity maximum value from the MHD two-dimensional data matrix as a label of a photon intensity maximum value, and adding the label into simulation data; respectively constructing a training data set and a testing data set based on the simulation data added with the labels;
and (B) step (B): and setting a loss function and model training parameters, training the semantic segmentation network model by using a training data set, testing by using a testing data set, and finally obtaining the trained semantic segmentation network model.
2. The method for detecting soft X-ray photon number maxima of a magnetic layer system according to claim 1, wherein said step a specifically comprises:
step A1: taking MHD simulation results under different solar wind density conditions as light source input, generating SXI photon number simulation data under corresponding solar wind density conditions through a SXI simulation program, wherein the integration time of SXI istThe same MHD simulation result input obtains the incomplete SXI photon number simulation data through multiple outputs;
step A2: will be arbitrarynThe Zhang Tongyang MHD simulation result input obtained incomplete SXI photon number simulation data are superimposed to obtain integration time ofntSXI photon count simulation data of (a);
step A3: the pixels of the maximum gray value of each row in the MHD two-dimensional integral graph form an X-ray radiation intensity maximum value region, the maximum gray value region is used as a label on the top of a magnetic layer in SXI simulation data under the condition of corresponding solar wind density, and the label is added into SXI simulation data;
step A4: after the label is added, the simulation data is enhanced, and the sample data is expanded;
step A5: based on the sample data, multiple sets of training data sets and test data sets are made by setting different solar wind densities or integration times.
3. The method for detecting soft X-ray photon number maxima of a magnetic layer system according to claim 2, wherein in the step A4, enhancement processing is performed on the simulated two-dimensional data matrix, specifically comprising: and performing operations of rotating and adding Gaussian noise on the simulation two-dimensional data matrix.
4. The method for detecting soft X-ray photon number maxima of a magnetic layer system according to claim 1, wherein in step B, model training parameters are updated by using an SGD optimizer during training.
5. The method for detecting soft X-ray photon number maxima of a magnetic layer system according to claim 1, wherein in step B, a loss function is obtainedLossUsing a binary cross entropy loss function, the expression is:
6. The method for detecting soft X-ray photon number maxima of a magnetic layer system according to claim 1, wherein in step B, the training process sequentially comprises: a freezing stage and a thawing stage;
in the freezing stage, the MobileNetv2 backbone network of the semantic segmentation network model is frozen and operated;
and in the thawing stage, the MobileNetv2 backbone network of the semantic segmentation network model is enabled to normally operate.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211432905.7A CN115641441B (en) | 2022-11-16 | 2022-11-16 | Magnetic layer system soft X-ray photon number maximum value detection method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211432905.7A CN115641441B (en) | 2022-11-16 | 2022-11-16 | Magnetic layer system soft X-ray photon number maximum value detection method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN115641441A CN115641441A (en) | 2023-01-24 |
CN115641441B true CN115641441B (en) | 2023-05-02 |
Family
ID=84948793
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211432905.7A Active CN115641441B (en) | 2022-11-16 | 2022-11-16 | Magnetic layer system soft X-ray photon number maximum value detection method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115641441B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116206517B (en) * | 2023-03-14 | 2023-07-18 | 哈尔滨工业大学 | Ground simulation device and method for magnetic field structure of polar tip region of earth magnetic layer |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113128542A (en) * | 2020-01-15 | 2021-07-16 | 中国科学院国家空间科学中心 | All-sky aurora image classification method and system |
CN113382683A (en) * | 2018-09-14 | 2021-09-10 | 纽罗因恒思蒙特实验有限责任公司 | System and method for improving sleep |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114332601A (en) * | 2021-12-02 | 2022-04-12 | 重庆邮电大学 | Picking robot unstructured road identification method based on semantic segmentation |
CN114549555A (en) * | 2022-02-25 | 2022-05-27 | 北京科技大学 | Human ear image planning and division method based on semantic division network |
CN114549563A (en) * | 2022-02-26 | 2022-05-27 | 福建工程学院 | Real-time composite insulator segmentation method and system based on deep LabV3+ |
-
2022
- 2022-11-16 CN CN202211432905.7A patent/CN115641441B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113382683A (en) * | 2018-09-14 | 2021-09-10 | 纽罗因恒思蒙特实验有限责任公司 | System and method for improving sleep |
CN113128542A (en) * | 2020-01-15 | 2021-07-16 | 中国科学院国家空间科学中心 | All-sky aurora image classification method and system |
Also Published As
Publication number | Publication date |
---|---|
CN115641441A (en) | 2023-01-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112241766B (en) | Liver CT image multi-lesion classification method based on sample generation and transfer learning | |
CN111950453B (en) | Random shape text recognition method based on selective attention mechanism | |
CN113160234B (en) | Unsupervised remote sensing image semantic segmentation method based on super-resolution and domain self-adaptation | |
CN115641441B (en) | Magnetic layer system soft X-ray photon number maximum value detection method | |
CN116958163B (en) | Multi-organ and/or focus medical image segmentation method and device | |
CN116206185A (en) | Lightweight small target detection method based on improved YOLOv7 | |
CN115471754A (en) | Remote sensing image road extraction method based on multi-dimensional and multi-scale U-net network | |
Lin et al. | Semantic segmentation network with multi-path structure, attention reweighting and multi-scale encoding | |
CN114550014A (en) | Road segmentation method and computer device | |
CN112766381B (en) | Attribute-guided SAR image generation method under limited sample | |
Wu et al. | Depth dynamic center difference convolutions for monocular 3D object detection | |
Sebaq et al. | Rsdiff: Remote sensing image generation from text using diffusion model | |
CN113111740A (en) | Characteristic weaving method for remote sensing image target detection | |
CN113989612A (en) | Remote sensing image target detection method based on attention and generation countermeasure network | |
CN117036941A (en) | Building change detection method and system based on twin Unet model | |
CN115236606B (en) | Radar signal feature extraction method and complex field convolution network system | |
CN115327544A (en) | Little-sample space target ISAR defocus compensation method based on self-supervision learning | |
CN117593470B (en) | Street view reconstruction method and system based on AI model | |
Jia et al. | Siamese Meets Diffusion Network: SMDNet for Enhanced Change Detection in High-Resolution RS Imagery | |
CN116977747B (en) | Small sample hyperspectral classification method based on multipath multi-scale feature twin network | |
CN115205710B (en) | Double-time-phase remote sensing image change detection method combined with color correction | |
CN114998749B (en) | SAR data amplification method for target detection | |
CN116385881B (en) | Remote sensing image ground feature change detection method and device | |
Li et al. | Enhanced Deep Learning Super-Resolution for Bathymetry Data | |
Tian et al. | Deep iterative residual back-projection networks for single-image super-resolution |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |