CN112633227B - Automatic recognition method and system for data lightning whistle sound waves of Zhangheng first induction magnetometer - Google Patents
Automatic recognition method and system for data lightning whistle sound waves of Zhangheng first induction magnetometer Download PDFInfo
- Publication number
- CN112633227B CN112633227B CN202011623352.4A CN202011623352A CN112633227B CN 112633227 B CN112633227 B CN 112633227B CN 202011623352 A CN202011623352 A CN 202011623352A CN 112633227 B CN112633227 B CN 112633227B
- Authority
- CN
- China
- Prior art keywords
- lightning
- image
- data
- whistle sound
- convolution kernel
- 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
- 238000000034 method Methods 0.000 title claims abstract description 60
- 230000006698 induction Effects 0.000 title claims abstract description 20
- 238000012545 processing Methods 0.000 claims abstract description 62
- 238000001228 spectrum Methods 0.000 claims abstract description 23
- 230000009467 reduction Effects 0.000 claims abstract description 9
- 238000012706 support-vector machine Methods 0.000 claims description 41
- 238000012549 training Methods 0.000 claims description 33
- 230000008569 process Effects 0.000 claims description 19
- 238000011156 evaluation Methods 0.000 claims description 17
- 238000004364 calculation method Methods 0.000 claims description 14
- 238000011160 research Methods 0.000 abstract description 12
- 238000012163 sequencing technique Methods 0.000 abstract description 3
- 230000000694 effects Effects 0.000 description 41
- 238000012360 testing method Methods 0.000 description 17
- 238000000605 extraction Methods 0.000 description 16
- 238000002474 experimental method Methods 0.000 description 12
- 238000010586 diagram Methods 0.000 description 8
- 230000000877 morphologic effect Effects 0.000 description 8
- 239000006185 dispersion Substances 0.000 description 7
- 230000006870 function Effects 0.000 description 7
- 238000004458 analytical method Methods 0.000 description 4
- 238000013461 design Methods 0.000 description 4
- 239000005433 ionosphere Substances 0.000 description 4
- 238000013209 evaluation strategy Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 238000012935 Averaging Methods 0.000 description 2
- 241000287531 Psittacidae Species 0.000 description 2
- 230000001627 detrimental effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000001965 increasing effect Effects 0.000 description 2
- 230000001939 inductive effect Effects 0.000 description 2
- 238000013178 mathematical model Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000005855 radiation Effects 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 102100029469 WD repeat and HMG-box DNA-binding protein 1 Human genes 0.000 description 1
- 101710097421 WD repeat and HMG-box DNA-binding protein 1 Proteins 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 230000005672 electromagnetic field Effects 0.000 description 1
- 230000005674 electromagnetic induction Effects 0.000 description 1
- 230000005670 electromagnetic radiation Effects 0.000 description 1
- 230000003203 everyday effect Effects 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000010921 in-depth analysis Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- UQDJGEHQDNVPGU-UHFFFAOYSA-N serine phosphoethanolamine Chemical compound [NH3+]CCOP([O-])(=O)OCC([NH3+])C([O-])=O UQDJGEHQDNVPGU-UHFFFAOYSA-N 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000010998 test method Methods 0.000 description 1
- 230000003313 weakening effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
- G06F2218/10—Feature extraction by analysing the shape of a waveform, e.g. extracting parameters relating to peaks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- 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
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a data lightning whistle sound wave automatic identification method of a Zhangheng first induction magnetometer, which comprises the following steps: sample collection (SCM data is subjected to Fourier spectrogram), graying and scale reduction processing, fuzzy convolution processing, L-shaped convolution processing and SVM model classification and identification; the invention also discloses a device and a system for automatically identifying the data lightning whistle sound wave of the Zhangheng first induction magnetometer corresponding to the method; the invention develops the exploration and research of automatically identifying lightning whistle sound waves on SCM data of the Zhang He satellite I; according to the characteristics of spectrum data and the L-shaped characteristic of lightning whistle sound waves, a fuzzy convolution kernel and an L-shaped convolution kernel are designed, so that the lightning shape characteristic in a spectrogram is further enhanced, and the robustness and the identification performance of the characteristic are greatly improved; and the SVM classifier is adopted to classify the features, so that the recognition performance, recall rate and sequencing capability of the features are all over 94 percent.
Description
Technical Field
The invention relates to the field of lightning whistle sound wave identification in space physics research, in particular to a method and a system for automatically identifying data lightning whistle sound waves of a Zhangheng-Yi induction magnetometer.
Background
However, many factors that cause electromagnetic anomaly disturbance (Nirupal et al 2012), a number of research results show that some severe weather phenomena in the underlying atmospheric weather activity, such as typhoons, hurricanes, coldness, tornadoes, lightning, etc., can also cause ionospheric disturbances (Liu et al 2006; xiao et al 2007). Lightning is one of the most frequent natural phenomena in many meteorological activities, and Brooks (1925) uses thunderstorm day data to estimate 1800 thunderstorms at any one time in the world, so as to obtain a result (Brooks, 1925) that the global lightning frequency is 100 fls-1 (fl is flash shorthand, and represents the number of times lightning occurs). The lightning discharge process is easy to generate large current and strong electromagnetic radiation, the peak value of the radiation field can reach or even exceed 15V/m (Marshall et al 2010) at the height of an ionization layer, and the ionization and excitation light radiation threshold values of certain components in the atmosphere are greatly exceeded, so that the ionized layer is obviously influenced, and the electromagnetic space-time characteristics of the ionized layer are obviously changed (Xiao et al, 2013;Huang and Gu,2003;Davis and Lo,2008). Since the flashover can penetrate the ionosphere, researchers have gained information about geophysical physics by means of lightning. For example: zahlava et al analyzed the measurement of the longitudinal dependence of DEMETEER and RBSP on the whistle mode wave of the magnetic layer inside the earth, which indicated that the longitudinal dependence of whistle mode wave inside the plasma layer was strong (Z.hllava et al.2018). Bayupati et al analyze the dispersion of lightning whistle observed by AKEBONO, and the analysis discusses the relationship between the time of propagation of the lightning whistle along the track and the electron density distribution (Bayupati et al 2012), indicating that the dispersion trend of the lightning whistle is a powerful method of determining the overall electron density distribution in the plasma layer. The frequency of occurrence of lightning whistles detected by AKEBONO satellites was analyzed by Oike et al to determine spatial and temporal variations in ground-observed lightning activity (Oike et al, 2014), indicating that the occurrence of lightning whistles in the ionosphere is closely related to lightning activity and electron density distribution around the earth. Clilverd et al (2002) uses the worldwide lightning location network (WWLLN) to determine the location of the source of lightning and then remotely senses the electron density distribution frequency from the propagation time of the lightning whistle from the source point to the observation point, deriving a function of electron density along the propagation path by means of propagation theory (Oike et al, 2014). The physical parameters, the position and the like of the lightning whistle sound wave are known as the basis for developing and utilizing lightning, and the characteristics of space environments such as an ionosphere and the like are studied by means of the lightning whistle sound wave.
Most of research on lightning whistle sound waves at present uses limited data, and aiming at a large amount of lightning whistle sound wave data, the research on space-time distribution rules and related parameters is less, and mainly two reasons exist: 1. a sufficient lightning whistle sound wave event is required; 2. the acquisition of lightning whistle acoustic events from a large number of electromagnetic observations by means of manpower is a very challenging task. The magnetic field observation data carries disturbance signals such as complex and various space electromagnetism and the like, such as satellite platform disturbance, lightning whistle, VLF ground emission source, magnetic storm and the like. There are a large number of lightning whistle waves in satellite data, and the existing research is mainly focused on the aspects of obvious magnetic storm and disturbance of VLF ground emission sources on an ionosphere (Yang et al, 2020; liao et al, 2019; zhao et al, 2019), and less on how to automatically identify and locate lightning events from the data. With the formation of a large satellite data pool, how to automatically identify lightning whistle sound waves from a large amount of satellite observation data is also an important premise for in-depth analysis. There is a need to discuss a viable automatic recognition method of lightning whistle sound waves to break through the research bottleneck.
The lightning whistle originates from lightning discharges and has unique spectral characteristics, which are characterized by decreasing frequency and amplitude over time due to the attenuation differences of the high frequency and low frequency components. Such a dispersion spectrum is called "dispersion" and becomes broader when the path length is longer or the electron density along the transmission path is denser, so if these values can be derived along the propagation path, dispersion can be a very useful parameter to derive the electron density distribution in the plasma layer as shown in fig. 1, 2. Helliwell (1965) classifies the morphology of lightning whistle waves into 9 categories based on the degree of dispersion.
Current research on lightning whistle wave automatic identification is mainly directed to waveform data collected by AKEBONO satellites (Bayupati et al 2012), data collected by Arase satellites (Ali Ahmad et al 2019), and spectrum data collected by DEMETER satellites (part et al 2019).
According to the theorem of Eckersley's (Gurnett et al, 1990), the time and frequency of lightning whistle waves satisfy the formula (1):
wherein D is a lightning whistle wave dispersion constant; t is the time of arrival at frequency f and t0 is the time of lightning triggering. Bayupati applies this theorem to waveform data acquired by the AKEBONO satellite for lightning identification (2012). The conventional lightning whistle sound wave automatic identification method is realized by identifying straight lines in a power spectrum, as shown in fig. 1: t1 is the time at which the frequency f1 arrives, t2 is the time at which the frequency f2 arrives, and the calculation formula of D is shown in formula (2):
It is found from fig. 1 that the lightning whistle wave is a straight line with a certain slope, and that the formulas (1) (2) only represent a simple relationship between the frequency of the lightning whistle wave and time.
(Ali Ahmad et al, 2019) it is considered that the above method hardly detects other lightning types summarized in (Helliwel et al, 1965), and by referring to the definition of lightning types in (Helliwell et al, 1965), the types of lightning whistle waves of the Arase satellite are corrected again, and different mode features are formulated according to different types. The identification method comprises the following steps: firstly, FFT spectrum calculation is carried out, then, gaussian convolution kernel is adopted for fuzzy processing on a spectrogram, laplacian operator is used for edge extraction, and then, an image segmentation algorithm is used for segmenting an image, as shown in figure 3; and dividing the image into a plurality of grids on average, as shown in fig. 4; finally, the features (Bresenham et a l., 1977) are extracted by using a Bresenham's straight line detection algorithm and then are input into a decision tree for recognition, and the recognition accuracy is only 75%. The algorithm has the following problems:
(1) In the identification process, a large number of image preprocessing operations are needed, such as parameter selection of Gaussian convolution kernels, selection of Laplacian parameters, parameter selection of segmentation algorithms and the like, and the selection of the parameters has a large influence on feature extraction by adopting the Bresenham's algorithm.
(2) Parameters need to be set in Bresenham's algorithm, and the selection of the parameters also affects the robustness of the feature.
(3) The evaluation index of the algorithm only adopts precision as an evaluation index, and the index cannot evaluate recall capability and sequencing capability of the classifier.
The lightning whistle sound wave identification is based on data of AKEBONO satellite and Arase satellite respectively. Identification of lightning whistle waves (as shown in fig. 5) based on tensor-1 satellite data was not investigated.
The Zhangheng satellite number one is a satellite for observing electromagnetic information related to earthquake activities for the first time in China. The revisiting period is 5 days, about 15 orbit observations are carried out every day, the SCM is covered between the north and south latitude 65 degrees, the induction magnetic field data of the ionized layer is obtained through Faraday electromagnetic induction law, and only the power spectrum data is obtained in the inspection mode. The satellite observes in orbit for more than 2 years, and acquires a large amount of waveform and power spectrum data of global electromagnetic fields, wherein 3 components X/Y/Z of the SCM comprise 3 frequency bands ULF/ELF/VLF, the frequency ranges ULF:1Hz-200Hz, ELF:200Hz-2.2kHz, VLF:12.5Hz-25.6kHz, the sampling rate of the original data is 51.2kHz, the frequency point interval ULF of the power spectrum data is 0.25Hz, ELF:2.5Hz, VLF: the 12.5Hz, detail mode VLF waveform data 80ms contains 4096 points (Wang et a l.,2018; fan et a l.,2018; wang et a l., 2018), producing a data volume of about 10G per day, how to automatically identify lightning whistle events from such huge observations is particularly critical and urgent.
At present, research on lightning whistle sound wave automatic identification based on a Zhang He satellite is not developed yet, the robustness of a lightning whistle sound wave automatic identification algorithm based on other constellations is poor, and how to create a lightning whistle sound wave automatic identification method based on the Zhang He satellite so as to improve the robustness and the identification performance of characteristics is one of the important research and development subjects at present.
Disclosure of Invention
The invention aims to provide a lightning whistle sound wave automatic identification method based on a Zhang Heng satellite, which can improve the robustness and the identification performance of characteristics.
In order to solve the technical problems, the invention adopts the following technical scheme:
a Zhangheng first induction magnetometer data lightning whistle sound wave automatic identification method comprises the following steps:
sample collection: performing Fourier spectrogram on SCM data, dividing the spectrogram according to the obvious L-shaped characteristic of lightning whistle sound waves in the spectrogram to finish sample image collection, and collecting a lightning spectrogram and a non-lightning sample spectrogram;
graying and scale treatment: graying treatment and scale reduction are carried out on the frequency spectrum image so as to reduce the calculation dimension and strengthen the lightning characteristic;
fuzzy convolution processing: designing a fuzzy convolution kernel, and carrying out convolution calculation on the image to filter out the influence of a large amount of step edge information;
L-form convolution processing: designing an L-shaped convolution kernel based on the L-shaped characteristic of the lightning whistle sound wave, and carrying out convolution processing on the image to further enhance the L-shaped characteristic in the image;
SVM model classification and identification: and inputting the enhanced image into a Support Vector Machine (SVM) for training and classifying recognition to obtain a recognition result.
As a further improvement of the invention, the SCM data mainly come from waveform data of VLF wave bands of SCM loads of three orbits of a first satellite, a sliding window is designed in the collecting process, the width of the sliding window is 10 pieces of 2s, the step length is 2 pieces of 2s, and the sliding window is adopted to intercept data from 1000Hz-6000Hz from the power spectrum data.
Further, the graying process is as follows:
Gray＝RGB.R×0.3+RGB.G×0.59+RGB.B×0.11
where RGB represents the original sample spectrogram, rgb.r is the pixel value of its red channel, rgb.g and rgb.b are the values of the green and blue channels, respectively, and Gray is the graying spectrogram.
Further, the scale process is scaled down to 50×50. The method mainly considers that the detail information of the image reflected by the small scale of the image is richer and the overall shape information is weaker; the whole shape information of the large-scale image is richer and the detail information is weaker, which just accords with the image characteristics of lightning whistle sound waves, namely, the whole L-shaped morphological characteristics are presented.
Further, the fuzzy convolution kernel is a full 1 convolution template with a scale of 5×5, and the main function of the template is to perform noise reduction processing on the image.
Further, the L-shaped convolution kernel is a 9×9L-shaped convolution kernel.
Further, in the SVM model classification recognition, the SVM adopts a polynomial kernel function k (x i ，x j )：
Wherein x is j Feature vector, x representing the jth image i A feature vector representing an i-th image; gamma and r are parameters to be adjusted, the SVM is from SVC library of Python; wherein the optimal parameter d is 13, and both gamma and r use the default values under the libraryParameters.
Further, in the SVM model classification and identification, the adopted evaluation indexes are as follows: recognition accuracy, recall, F1 value, and AUC-ROC.
The invention also provides an automatic recognition device for the data lightning whistle sound wave of the Zhangheng first induction magnetometer, which comprises the following components:
sample collection module: the SCM data are used for making Fourier spectrograms, according to the obvious L-shaped characteristic of lightning whistle sound waves in the spectrograms, the spectrograms are segmented to finish sample image collection, and a lightning spectrogram and a non-lightning sample spectrogram are collected;
graying and scale processing module: the method is used for carrying out graying treatment and scale reduction on the frequency spectrum image so as to reduce the calculation dimension and strengthen the lightning characteristic at the same time;
And the fuzzy convolution processing module is used for: the method is used for designing a fuzzy convolution kernel, and carrying out convolution calculation on the image to filter out the influence of a large amount of step edge information;
l-shaped convolution processing module: the L-shaped convolution kernel is used for designing the L-shaped convolution kernel based on the L-shaped characteristics of the lightning whistle sound wave, and the L-shaped characteristics in the image are further enhanced by carrying out convolution processing on the image;
SVM model classification recognition module: the method is used for inputting the enhanced image into a Support Vector Machine (SVM) for training and classifying recognition, and obtaining a recognition result.
The invention also provides an automatic recognition system for the data lightning whistle wave of the first-order induction magnetometer, which is characterized by comprising the following steps: one or more processors; and the storage device is used for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors realize the automatic identification method of the data lightning whistle sound waves of the Zhang He Yi induction magnetometer.
By adopting the technical scheme, the invention has at least the following advantages:
1. according to the invention, on the basis of the Zhang He satellite SCM data, a lightning whistle sound wave automatic identification framework based on machine learning is explored, and a fuzzy convolution kernel and an L-shaped convolution kernel are designed according to the characteristics of frequency spectrum data and the L-shaped characteristics of the lightning whistle sound wave, so that the lightning shape characteristics in a spectrogram are further enhanced, and the robustness and the identification performance of the characteristics are greatly improved.
2. The invention is proved by carrying out a large number of experiments, and the results show that: the automatic identification method of the lightning event provided by the invention is effective, and the identification effect of the automatic identification method reaches more than 94% on the indexes of precision, recall rate, F1 value (F1 score) and receiver operation characteristic Curve Area (AUC).
Drawings
The foregoing is merely an overview of the present invention, and the present invention is further described in detail below with reference to the accompanying drawings and detailed description.
FIG. 1 is an AKEBONO satellite lightning morphology legend;
FIG. 2 is a DEETER satellite lightning morphology legend (Parrot et al 2015;Parrot et al, 2019);
FIG. 3 is a segmentation result graph (Ali Ahmad et al, 2019);
FIG. 4 is a ruled line graph (Ali Ahmad et al, 2019);
FIG. 5 is a graph of lightning morphology in a time-frequency plot of SCM power spectrum data for a Zhang He satellite number one;
FIG. 6 is a sample collection flow chart;
fig. 7 is a sample illustration: (a) lightning samples; (b) no lightning samples;
FIG. 8 is a recognition algorithm (solid lines are model training streams; dashed lines are model recognition streams);
FIG. 9 is a color spectrum graph (with lightning);
FIG. 10 is a graph of a grayscaling spectrum (with lightning);
FIG. 11 is a raw scale view;
FIG. 12 is a small scale view (and enlarged detail);
FIG. 13 is a fuzzy convolution diagram (a) fuzzy convolution kernel; (b) a fuzzy convolution effect;
FIG. 14 is an L-feature convolution diagram (a) an L-feature convolution kernel; (b) convolution effect;
fig. 15 is a lightning identification diagram: (a) correct identification; (b) error identification; (c) graying map;
fig. 16 is a non-lightning identification diagram: (a) correct identification; (b) error identification; (c) a grayscaled image (b); (d) a fuzzy convolution map; (e) an L-modality convolution kernel processing map;
FIG. 17 is a box plot of precision, recall, and F1: (a) lightning (b) non-lightning;
FIG. 18 is an AUC box plot;
FIG. 19 is an original gray scale map containing lightning information;
FIG. 20 is an image processed using a fuzzy convolution kernel;
FIG. 21 is a feature distribution trend graph; (a) Is the characteristic distribution of the original gray scale of the lightning and non-lightning samples; FIG. 21 (b) feature distribution of lightning and non-lightning samples after fuzzy convolution processing; wherein 1 represents a lightning sample, -1 is a non-lightning sample;
fig. 22 is a graph of image features (image containing lightning) extracted by different convolution kernels: (a) Is an original gray level image, and (b) is an image processed by a fuzzy convolution kernel; (c) is an image after the L convolution kernel processing b of the scale 3×3; (d) is an image after the L convolution kernel processing b of scale 5×5; (e) is an image after the L convolution kernel processing b of the scale 7×7; (f) is an image after the L convolution kernel processing b of the scale 9×9; (g) is an image after the L convolution kernel processing b of the scale 11×11; (h) is an image after the L convolution kernel processing b of scale 13×13; (i) is an image after the L convolution kernel processing b of a scale of 15×15; (j) is a partially enlarged image at the square in (c);
FIG. 23 is a graph of image features (including non-lightning images) extracted by different convolution kernels, where (a) is the original gray scale image and (b) is the blurred convolution kernel processed image; (c) is an image after the L convolution kernel processing b of the scale 3×3; (d) is an image after the L convolution kernel processing b of scale 5×5; (e) is an image after the L convolution kernel processing b of the scale 7×7; (f) is an image after the L convolution kernel processing b of the scale 9×9; (g) is an image after the L convolution kernel processing b of the scale 11×11; (h) is an image after the L convolution kernel processing b of scale 13×13; (i) is an image after the L convolution kernel processing b of a scale of 15×15;
FIG. 24 is a schematic diagram of image features falling in two dimensions: (a) is an image after the blurring convolution kernel processing; (b) is an image after the L convolution kernel processing a of the scale 3×3; (c) is an image after the L convolution kernel processing a of scale 5×5; (d) is an image after the L convolution kernel processing a of the scale 7×7; (e) is an image after the L convolution kernel processing a of scale 9×9; (f) is an image after the L convolution kernel processing a of the scale 11×11; (g) is an image after the L convolution kernel processing a of scale 13×13; (h) Is an image after L convolution kernel processing a with a dimension of 15×15
FIG. 25 is the AUC of a classifier trained on different features;
FIG. 26 is a graph of three index profiles for a classifier trained on different features: (a) identifying a lightning sample; (b) identifying a non-lightning sample.
Detailed Description
The Zhang Heng satellite No. one archives a large amount of inductive magnetometer (Search Coil Magnetometer, SCM) data, and the search of an algorithm for automatically identifying lightning whistle sound waves from the inductive magnetometer (Search Coil Magnetometer, SCM) data has important research significance for further summarizing the time-space change rule of space weather lightning events. The embodiment provides a data lightning whistle sound wave automatic identification method of a Zhangheng first induction magnetometer, which comprises the following steps:
sample collection: performing Fourier spectrogram on SCM data, dividing the spectrogram according to the obvious L-shaped characteristic of lightning whistle sound waves in the spectrogram to finish sample image collection, and collecting a lightning spectrogram and a non-lightning sample spectrogram;
graying and scale treatment: graying treatment and scale reduction are carried out on the frequency spectrum image so as to reduce the calculation dimension and strengthen the lightning characteristic;
fuzzy convolution processing: designing a fuzzy convolution kernel, and carrying out convolution calculation on the image to filter out the influence of a large amount of step edge information;
l-form convolution processing: designing an L-shaped convolution kernel based on the L-shaped characteristic of the lightning whistle sound wave, and carrying out convolution processing on the image to further enhance the L-shaped characteristic in the image;
SVM model classification and identification: and inputting the enhanced image into a Support Vector Machine (SVM) for training and classifying recognition to obtain a recognition result.
The following describes it in detail: 1 data collection (sample collection)
The data mainly comes from waveform data of VLF wave bands of SCM load of three orbits of a first satellite in 2019, the collection process is shown in figure 6, and the power spectrogram of the data is obtained by carrying out Fourier transform on the data. According to the data: most lightning whistle sound waves have an L shape in the region of 1000Hz-6000Hz and cannot exceed 20s (Fi ser et a.l., 2010) in duration, therefore, the embodiment designs a sliding window with the width of 10 pieces of 2s and the step length of 2 pieces of 2s, and the sliding window is adopted to intercept data of 1000Hz-6000Hz from power spectrum data, so that 8316 pieces of data are obtained in total: of these, 316 pieces of lightning whistle sound wave data, 8000 pieces of non-lightning whistle sound wave data, and a sample example is shown in fig. 7.
2 lightning whistle sound wave automatic identification algorithm
The deep learning model is a mainstream model of machine learning at present, but a large number of data samples are still needed as a basis. The lightning data volume collected at present in this embodiment is smaller, and based on this, the recognition scheme of lightning whistle sound waves is still explored by using the traditional recognition algorithm technology, and the scheme includes a training process and a recognition process, as shown in fig. 8: the solid line is the training flow and the dashed line is the recognition flow. The main purpose of the training process is to obtain a lightning identification model, and the main purpose of the identification process is to apply the lightning identification model. The training process comprises the following steps: graying treatment, scale treatment, fuzzy convolution treatment, L-shaped convolution treatment and SVM model training. The identification process comprises the following steps: graying treatment, scale treatment, fuzzy convolution treatment, L-shaped convolution treatment and lightning identification by using an SVM model.
2.1 Gray treatment
As shown in fig. 9, the primary spectrum of the original sample is that lightning is identified by observation, and is most based on the L-shape characteristic, so that the influence of color is eliminated by adopting graying treatment. The effect of graying the image of FIG. 9 according to equation (3) is shown in FIG. 10
Gray ＝ RGB.R×0.3 + RGB.G×0.59 + RGB.B×0.11 (3)
Where RGB represents the original sample spectrogram, rgb.r is the pixel value of its red channel, rgb.g and rgb.b are the values of the green and blue channels, respectively, and Gray is the graying spectrogram.
2.2 Scale processing
The larger the scale of the image, the more obvious the detail is, but the overall morphological characteristics are not enough outstanding, while the smaller the scale, the more easily the overall morphological characteristics are outstanding. As shown in fig. 11 and 12: fig. 11 is the original scale, fig. 12 is the effect of scaling it down to 1/4, and the L morphological feature is found to be more pronounced than fig. 11. Therefore, the embodiment chooses to scale down to 50×50, which not only can highlight morphological characteristics but also reduces data dimension to facilitate computation.
2.3 fuzzy convolution
The image is seen to be more heavily striped in texture, as shown in fig. 13, indicating that the image has distinct stepped edges, which are detrimental to identifying L-shaped features. Therefore, the present embodiment designs the fuzzy convolution check fig. 12 to perform convolution processing so as to achieve the purpose of weakening the step edge thereof. The convolution kernel design is shown in fig. 13 (a), and is an all-1 template with a scale of 5×5, and the convolution processing is performed on fig. 12 by using the same, and the processing result is shown in fig. 13 (b).
2.4 L feature extraction
The lightning whistle wave exhibits a remarkable L shape, and in order to further project the lightning characteristic, a convolution kernel of a scale of 9×9L is designed according to the characteristics of the L shape of lightning, as shown in fig. 14 (a) (as to why the scale of the convolution kernel is 9×9, detailed analysis will be performed in the discussion), and the result after processing using this convolution kernel fig. 13 (b) is shown in fig. 14 (b).
2.5 SVM classification
The convolved image is input to an SVM classifier for training and recognition. SVM is a supervised classifier based on statistical theory (Sanchez and David, 2003). The categories in this embodiment are mainly two: lightning and non-lightning. First, image data is divided into a training set and a test set.
The images of the training set are represented as: x is x 1 ，x 2 ，…,x n Its corresponding feature tag is y 1 ，y 2 ，…,y n ,y i ∈[-1，1]
-1 indicates that the sample is a non-lightning sample, 1 indicates that there is a lightning event in the sample.
The main objective of the linear SVM classifier is to find a hyperplane w= (w) 1 ，w 2 ，…,w n ) -1 and 1 can be classified. The mathematical model of the hyperplane is shown in formula (4):
wx+b＝0 (4)
where x is the sample point on the hyperplane and b is the offset.
If the lightning sample correctly falls in the classification area of lightning, then there is formula (5)
y i (wx i +b)≥+1 (5)
If the non-lightning sample correctly falls in the non-lightning classification zone, then there is formula (6)
y i (wx i +b)≤-1 (6)
Both equation (5) and equation (6) may be combined into one equation (7):
y i (wx i +b)-1≥0 (7)
there are many hyperplanes to complete the classification task, and the hyperplane found by the SVM maximizes the separation in feature space. The mathematical model thereof can be expressed as formula (8):
max 1/||w|| s.t. y i (wx i +b)-1≥0(i＝1，2，...，n) (8)
the purpose of the training model is to find w that satisfies the above equation.
Since in most cases the samples are not linearly separable, when a hyperplane satisfying the above conditions is not present, for this nonlinear separable problem, the SVM provides a kernel function k (x i ，x j ) Instead of equation (5), the data can be mapped to a high-dimensional space such that it is linearly separable in the high-dimensional space (Sanchez and David, 2003). The kernel function in this embodiment employs a polynomial kernel function (9):
wherein x is j Feature vector, x representing the jth image i A feature vector representing an i-th image;gamma and r are parameters to be adjusted, the SVM is from SVC library of Python; where the optimal parameters d are 13, γ and r are all other default parameters under the library.
3 experiment and analysis
3.1 Experimental protocol
Since lightning samples are unbalanced with non-lightning samples, the ratio of positive and negative samples cannot exceed 1 according to the conventional machine learning convention: 3. the number of lightning samples in this embodiment is 316, so 1000 non-lightning samples corresponding thereto should be selected. Whereas in practice the non-lightning samples obtained in this embodiment are 8848. In order to test whether the proposed lightning identification algorithm is effective, the following experimental scheme is formulated.
Data set: the data set consists of 316 lightning samples and 1000 non-lightning samples. The 1000 non-lightning samples were randomly selected from 8848 non-lightning samples, and the process resulted in 316 lightning sample sets and 1000 non-lightning sample sets.
Training set: the training set of lightning samples and the training set of non-lightning samples are obtained by taking 80% of lightning sample sets randomly as training samples and 80% of non-lightning sample sets randomly as training samples.
Feature extraction: and extracting the image features of the lightning sample training set and the non-lightning sample training set by three different feature extraction methods aiming at the training set. The embodiment provides three image feature extraction modes:
(1) The original Gray image features are denoted Gray.
(2) The feature of the original Gray image after the fuzzy convolution processing is represented by Gray_Blur.
(3) The method comprises the steps of performing fuzzy convolution processing on an original Gray image, and then adopting image features subjected to L feature convolution processing to represent the image features by Gray_Blur_L.
Training process: and respectively training the SVM models by using different characteristics to obtain three different SVM lightning recognition models.
Test set: and taking the remaining samples in the lightning sample set as test samples, and taking the remaining samples in the non-lightning sample set as training samples, wherein the process obtains the test set of the lightning sample and the test set of the non-lightning sample.
Feature extraction: in the same manner as the feature extraction above.
The identification process comprises the following steps: and respectively putting the three different features into different SVM models for recognition, and outputting a recognition result.
Evaluating the recognition effect: four indexes are adopted for evaluating the identification effect: precision (Precision), recall (Recall), F1 value, and ROC Area (AUC).
Number of experiments: 10000 times
3.2 SVM parameter selection
The SVM in the present algorithm is from the SVC library of Python. The selected kernel function is a polynomial kernel as shown in equation (9). The optimal parameter d obtained by cross-validation is 13, the others all use the default parameters under the library.
3.3 evaluation index
The effect of each identification is evaluated using four different evaluation indexes, assuming that lightning is a positive type sample and non-lightning is a negative type sample, the definition of the four indexes will be explained next.
First, symbol definitions are shown in Table 1
TABLE 1 symbol definition
Definition of recognition accuracy (Precision) is as shown in formula (10):
Precision＝TP/(FP+TP) (10)
TP+FP is the total lightning sample, is the number of pictures in the predicted picture that are positive types; TP is the number of pictures whose positive class is also predicted as positive class. The meaning is that the number of pictures with correct prediction is proportional to the total positive class prediction number, and the larger the index is, the better the index is.
The definition of Recall (Recall) is shown in formula (11):
Recall＝TP/(TP+FN) (11)
wherein TP+FN represents the number of pictures which completely meet the picture marking; TP is the number of pictures of which the positive class is predicted as the positive class; the meaning is that the positive type is predicted to be the number of the pictures of the positive type accounting for all the marked pictures, and the larger the index is, the better the index is.
The definition of the F1 value (F1-Score) is shown in formula (12):
F1＝2/(1/Precision+1/Recall) (12)
the meaning is as follows: typically, precision is high, recall is low, recall is high, and Precision is low. Index F1-score, taking into account the reconciliation values of Precision and Recall. The larger the Precision, the smaller the 1/Precision, and thus the larger the F1, when Recall is unchanged. And (3) the same principle: when Precision is unchanged, the larger the Recall, the smaller the 1/Recall, and thus the larger the F1. The larger the index 1, the better.
AUC-ROC: the higher the area under the ROC curve, the better the sorting performance of the classification.
3.4 evaluation strategy
On the training set and the test set of each experiment, three different feature extraction methods are used for extracting image features, three different classifiers are obtained through training, and the performance of each classifier is evaluated by adopting precision, recall, F1SCORE and ROC-AUC. Because each training set and each test set are different, the effect of the lightning identification algorithm provided by the embodiment is difficult to be fully evaluated by four single evaluation indexes, the experiment is carried out 10000 times, and the following evaluation strategies are formulated on the basis of the four evaluation indexes:
(1) And displaying part of the recognition results.
(2) Overall recognition accuracy evaluation: and (3) carrying out an evaluation strategy for averaging the evaluation indexes of 10000 experiments.
(3) Evaluation of recognition effect stability: and (3) evaluating the stability of 10000 times of classification by using a box-shaped graph on the evaluation index of 10000 times of experiments.
(4) Identification effect differentiation evaluation: in order to evaluate whether the classification effects caused by different features have obvious differences, T test is adopted to perform differential evaluation on the classification effects of the different features. The threshold p=0.05, i.e. less than 0.05 is a significant difference, if more than 0.05 indicates no significant difference.
3.4.1 partial recognition results display
The partial recognition results are shown in fig. 15 and 16: fig. 15 is a partial result of identifying a lightning image, (a) is a lightning image that can be correctly identified, and (b) is a lightning image that is incorrectly identified. All the identification errors in the step (b) are because the L-shaped feature in the step (b) is obviously weaker than the "roar" feature on the right, as shown in the figure (c), the main information of the figure is embodied on the roar feature, and the classifier has misjudgment; fig. 16 is a partial result of identifying a non-lightning image, (a) is a result that the non-lightning image can be correctly identified, and (b) is a result that a wrong non-lightning image is identified. The reason for the recognition error is that multiple "roar" appearing in the image causes the artifact of the L-shaped feature, and as shown in the diagrams (c) (d) (e), the classifier presents a false judgment.
3.4.2 evaluation of overall recognition accuracy
This section measures lightning identification effect under different image features mainly for 10000 times of Precision, recall, F score and AUC-ROC averaging. The results are shown in tables 2 and 3.
Table 2.10000 average effect after experiments
TABLE 3 mean value of ROC area AUC
By observing tables 2 and 3, it can be found that:
(1) The classification accuracy of the features extracted by the fuzzy convolution kernel is obviously superior to that of the original gray image. For example, gray features are identified with a Precision of 0.732, and classification with fuzzy convolution kernels with a Precision of 0.915; meanwhile, classification Precison after feature extraction by adopting L-shaped convolution kernels is as high as 0.945.
(2) The classification recall rate of the features extracted by the fuzzy convolution kernel is obviously superior to the original gray level image effect. For example, gray's Recall has a value of 0.610, while the Recall after fuzzy convolution has a value of 0.911, with the class Recall after feature extraction using the L convolution kernel being significantly highest, 0.974.
(3) The F1score of the features extracted by the fuzzy convolution kernel is also obviously superior to the original gray image effect. For example, gray has a value of F1score of 0.664, and the blurred convolutions have values of 0.912, with the L convolution kernel extracting features having the highest value of F1score of 0.958.
(4) The sorting capability of the classifier after feature extraction by using convolution kernel is better than that of the original gray image, as shown in table 3, the AUC of the classification of the original gray image is 0.882, the classification AUC after fuzzy convolution is 0.975, and the classification AUC of the l convolution kernel is 0.989.
3.4.3 evaluation of stability of recognition Effect
To evaluate whether the performance of the classifier was stable, 10000 data for each index was plotted with a box chart, and the results were shown in fig. 17 and 18, which were mainly qualitatively evaluated for the stability of the classification performance.
(1) The distribution of 10000 sets of data of lightning recognition accuracy (Precision) is as shown in the Precision diagram of fig. 17 (a): the horizontal axis is the different image features and the vertical axis is the precision. The discovery is as follows: the distribution of the non-lightning recognition accuracy of the Gray_Blur and Gray_Blur_L features makes up a box that is lower than the lightning recognition accuracy of the Gray features. This phenomenon illustrates that the Gray_Blur and Gray_Blur_L features make the classifier more stable in performance than the Gray features. The box corresponding to the Gray_Blur_L feature is the lowest, which indicates that the feature extraction method makes the classifier most stable in performance. The median line was also observed to further find: the median lines of the boxes corresponding to Gray_Blur_L and Gray_Blur are higher than the median line of the box of Gray, which shows that the accuracy of Gray_Blur_L and Gray_Blur is high in probability and higher than that of Gray characteristics; the accuracy of the lightning classifier for gray_blur_l feature learning is optimal. The same performance is also shown in the distribution of Recall and F1score indicators.
(2) By observing the distribution data of the non-lightning recognition accuracy in the above-described comparison manner, as shown in fig. 17 (b), it was found that: the data box body with the lightning identification Precision of the Gray characteristic is higher than the non-lightning identification Precision of Gray_Blur and Gray_Blue_L image characteristics, and meanwhile, the median line is higher than the median line of the box body with the Precision of the Gray level image, which indicates that the Gray_Blue_L characteristics have better Precision and smaller fluctuation than the original Gray level characteristics when the non-lightning identification is carried out.
(3) Also observed is a box plot of 10000 values of AUC indicators for the classifier, as shown in fig. 18: the horizontal axis is the different feature extraction methods, and the vertical axis is the AUC value. The discovery is as follows: the bin of AUC of the Gray feature is much higher than the bin of gray_blue feature and the bin of gray_blue_l feature convolution kernel feature, where the AUC bin of gray_blue_l feature is not only the smallest but also the highest number of bits. It is explained that the classifier trained by the gray_blue_l feature not only fluctuates less but also has the best ranking capability.
In summary, the classifier learned from gray_blur_l image features not only has a more reliable classification effect but also has more stable classification performance than the original Gray features.
3.4.4 evaluation of significance Difference of recognition Effect
In order to check whether the performance of the classifier under different image characteristics has obvious difference, a T-test method is adopted to quantitatively evaluate the saliency difference, wherein the higher the saliency P value is, the smaller the saliency P value is, the general threshold value is 0.05, and the meaning is that the difference is considered to be obvious if the difference is smaller than 0.05; if it is greater than 0.05, the two experiments are considered to have no significant difference. The results are shown in Table 4, table 5, table 6 and Table 7.
/>
(1) Focusing on Precision indexes for identifying lightning, performing T test on 10000 Precision of Gray feature classification, 10000 Precision of Gray_Blur feature classification and 10000 Precision of Gray_Blur_L feature classification two by two to obtain different test values, wherein the P value of the T test is 0, for example, the data of Gray and Gray_Blur are subjected to T test to obtain the P value of 0, and the data show that: there is a significant difference between the classification accuracy of the acquisition of the Gray feature and the Gray_Blur feature. For example, when Gray and Gray are T-checked, the P value is 1 and the P value is 1 indicates no significant difference between the data. And the recognition accuracy of Gray_Blur_L features and the recognition accuracy of other two features are subjected to T test to obtain P values of 0, so that the classifier trained by Gray_Blur_L features has obvious difference in lightning classification accuracy compared with the classifier trained by other two features.
(2) And observing a Recall index for identifying lightning, wherein the P value of a T test between the Recall rate of Gray characteristics and the Recall rate of Gray_Blur characteristics is 0, so that the classifier for identifying lightning learned by the two is obviously different in Recall rate. Similarly, the lightning identification effect of Gray_Blur_L features and the T test values of the other two features are 0, which also shows that three different features make the lightning classification effect have obvious difference in recall.
(3) By observing the F1score index in the above manner, a comparison result similar to the above can be obtained. There is a significant difference between the F1 values of the classifier learned by the three features. Also observing the index AUC, the significance test P between the AUC of the classifier learned by the three features is 0, which indicates that the sorting capability of the classifier learned by the three different features has obvious difference.
(4) The above conclusion also exists on the recognition performance of non-lightning pictures.
In a word, the Gray characteristic, gray_Blur characteristic and Gray_Blur_L characteristic are adopted for lightning identification, the identification effect of the Gray_Blur_L characteristic has obvious differences in precision, recall, F1 value and AUC value, and the classification effect of the Gray_Blur_L characteristic is optimal.
Discussion 4
The experiment shows that based on the Zhang Heng satellite SCM data, an algorithm research for automatically identifying lightning whistle sound waves is developed, and the method has a certain effect. The fuzzy convolution kernel and the L-feature convolution kernel in the algorithm scheme have very important influence on lightning identification. This section will carry out a deeper discussion and analysis of the effects that it has on.
4.1 fuzzy convolution kernel
Experimental results show that the effect of adopting a fuzzy convolution kernel to extract image features for lightning classification is better than the effect of adopting an original spectrogram as the features, and the part is discussed from a visual observation angle and a feature separable angle.
4.1.1 visual observation angle
The original spectrogram has a clear vertical information trace, as shown in fig. 19: the presence of a large amount of vertical information has a great influence on the recognition of the L-shape feature, and fig. 20 is an image obtained by blurring the vertical information by blurring convolution check, the vertical information of which is weakened while the L-shape feature thereof is highlighted.
4.1.2 characteristic separable angles
To explore whether the features of the image after the fuzzy convolution process have different partialities from the features of the original gray map. This section will use PCA dimension reduction to 2-dimension image data of 316 lightning samples and 1000 random non-lightning samples and draw its distribution in 2-dimensional space to observe the feature's separability trend, as shown in fig. 21. An observation of fig. 21 reveals that: the original image features have a large number of positive and negative samples gathered together, especially in the circular area in fig. 21 (a), and the overlapping positive and negative samples increase the difficulty of identifying lightning and non-lightning, so that the performance of the classifier for identifying lightning is reduced; in contrast, in the circular region of fig. 21 (b), the number of positive and negative samples gathered together is greatly reduced as compared with the former, which phenomenon increases the distinction between positive and negative samples, increases the degree of distinction between lightning and non-lightning image features, and enhances the ability of the classifier to recognize lightning.
4.2L feature convolution check lightning identification effect
Based on the experimental result, the L characteristic convolution kernel processing of 9 multiplied by 9 is adopted on the basis of the fuzzy convolution processing image, so that the lightning classification effect is better, and the recognition effect and the scale selection of the L characteristic convolution kernel are further discussed from two aspects.
4.2.1 visual observation angle
It is found from fig. 22 that the L feature extracted by the blur kernel is more pronounced as shown in fig. 22 (a) and 22 (b) compared to the original image; further observing that the L features extracted by using the 3×3L convolution kernel are shown in fig. 22 (c), because the L features are too dark, the rectangular area of fig. 22 (c) is enlarged as shown in fig. 22 (j), and most detailed information is found to be filtered, so that only the main contour of the L features is reserved; as the L convolution kernel scale becomes larger, more and more detail information around the L feature is available, as shown in fig. 22 (d, e, f); as the scale of the L convolution kernel increases further, not only is the intensity of detail around the L feature more and more apparent, but also more detail occurs where it is irrelevant to the L region.
Looking further at FIG. 23, the small scale L convolution kernel can filter out detailed information, as shown in FIG. 23 (c) (d) (e); with the increasing of the L convolution kernel scale, the detail information is more and more obvious, as shown in fig. 23 (f) (g); when the scale is further increased, the amplified interference information generates a morphology similar to that of the L feature, such as a red circular region of fig. 23 (i), that is, a large scale L convolution kernel easily amplifies noise of the region, which easily causes a certain trouble for recognition. Therefore, a small-scale L convolution kernel filters out too much detailed information around the L features, and a too large L convolution kernel is also easy to amplify noise in a non-lightning image, and how to automatically select an L convolution kernel with a suitable scale is a problem that needs further consideration.
4.2.2 characteristic separable angles
In order to qualitatively compare the separability of different features, firstly, extracting image features of 316 lightning samples and 1000 random non-lightning samples by adopting different feature extraction modes, reducing the feature dimension to 2 dimensions by adopting PCA technology, finally, drawing the feature dimension in a two-dimensional space to observe the distribution condition of the corresponding features, and analyzing the separability trend of the feature dimension, wherein +1 represents lightning sample data, and-1 represents non-lightning sample data, as shown in figure 24.
From fig. 24, it can be found that: as shown in fig. 24 (a), although most of positive and negative samples are separated as a whole, there is still a case where the positive and negative samples are superimposed in the middle; fig. 24 (b) shows that the distribution effect of the features obtained by processing the image by the fuzzy convolution first and processing the image by the L convolution kernel of 3*3 in the two-dimensional space, compared with fig. 24 (a), the distance between the positive and negative samples is pulled apart, and the separability between the positive and negative classes is better than that of fig. 24 (a) with the increase of the L convolution kernel scale, but the difference is not great compared with that of fig. 24 (b), as shown in fig. 24 (c) (d) (f); however, it is worth mentioning that the lightning samples accumulate more and more in the class, indicating that the difference in the class of the lightning samples is smaller and less, which is advantageous for the identification of lightning events, as shown in fig. 24 (f); with further increases in the scale of the L-feature convolution kernel, it was found that there was a phenomenon in which some of the lightning samples and the non-lightning samples overlapped with each other, as shown by the circular areas in fig. 24 (g) (h). And this overlap phenomenon becomes more serious as the scale increases, as shown in fig. 24 (h). In addition, the non-lightning samples are more and more evacuated in the class, which indicates that the non-lightning samples are more and more bad in the class, and the situation is unfavorable for identifying the non-lightning class.
In a word, the proper L convolution kernel scale can better realize the classification effect between the lightning sample and the non-lightning sample. Although the classification effect can be improved by a small scale, the lightning type internal difference is not favorable for the identification of a lightning sample; too large a scale is prone to overlap of non-lightning and lightning samples, resulting in smaller and smaller inter-class variances; also resulting in an increase in the intra-class variance of the non-lightning samples; a small inter-class variance and a large intra-class variance would be detrimental to the classifier for efficient classification.
4.2.3 statistics of angles of box graphs
Image features extracted by L convolution kernels with different scales are utilized, a classifier capable of identifying lightning is trained, an AUC value reflecting sorting capability of the classifier is calculated, and experiments are repeated 10000 times, so that distribution conditions of AUC data of the classifier under the features are obtained, and a box-shaped diagram is shown in figure 25: the horizontal axis of the method is different feature extraction methods, the leftmost image features extracted by adopting a fuzzy convolution kernel are sequentially 3×3L morphological convolution kernels, 5×5L morphological convolution kernels, … and 15×15L morphological convolution kernels. The bin of AUC for the original gray features was found to be much higher than the bin based on the L-shaped convolution kernel. Especially when the convolution kernel is 13 x 13, its AUC box is not only smallest but also has the highest median. The classifier obtained by training the features extracted by the L convolution kernel with the scale of 13 multiplied by 13 is explained, and the sorting capability of the classifier is not only the most stable but also the most optimal.
Similarly, training the classifier based on the features extracted by the L convolution kernels of different scales, obtaining three index values of lightning identification Precision (Precision), recall rate and F1-value of the classifier, repeating 10000 times of experiments, wherein the data distribution is shown in fig. 26 (a), blur in the figure represents the features extracted by the fuzzy convolution kernels, L3 represents the L-shaped convolution kernels of 3×3, and so on, and L15 represents the L-shaped convolution kernels of 15×15. Observations find that: in the process of recognizing the lightning image, the classifier for extracting the features by the L convolution kernel is not adopted, the boxes of the Recall index are higher than the boxes adopting the L-shaped convolution features, particularly, the position of L9 in the 2 nd graph of fig. 26 (a), the position means that the scale of the convolution kernel is 9×9, and meanwhile, the median line of the position is found to be higher than the median lines of the boxes at other positions, so that the feature training classifier extracted by the scale convolution kernel is shown to have the performance with the minimum fluctuation and the optimal precision in Recall rate. Meanwhile, the boxes of the Recall data are found at the L3 and L15 positions, the height of the boxes is high, and the median line is lowered. Meaning that an L-shaped convolution kernel of too small a scale, such as a scale of 3 x 3, and an L-shaped convolution kernel of too large a scale, such as a scale of 15 x 15, will degrade the performance of lightning identification. The same performance as above is also shown in the profile of the F1 score. But its accuracy does not vary much. Fig. 26 (b) shows the data of three indices in identifying a non-lightning image, respectively: in the precision index, the median line of the L9 position reaches the highest, which indicates that the convolution kernel effect of the scale is optimal at the position, and compared with the position, the minimum scale (L3) and the maximum scale (L15) have non-ideal effects; on the recall index, the convolution kernel scale has little influence on the recall index; at F1score, the least-scaled convolution kernel is less effective.
In summary, the image features extracted based on the convolution kernel not only have better classification performance, but also have smaller fluctuation than the original gray scale, and the classification is more stable.
In conclusion, the invention develops the exploration and research of automatically identifying lightning whistle sound waves on SCM data of Zhang He satellite I. According to the characteristics of spectrum data and L-shaped characteristics of lightning whistle sound waves, a fuzzy convolution kernel and an L-shaped convolution kernel are designed, so that the lightning-shaped characteristics in a spectrogram are further enhanced, and the characteristics are classified by adopting an SVM classifier, so that the identification performance of the lightning-shaped characteristics is more than 94% in precision, recall rate and sequencing capability.
In addition, the embodiment also provides a device for automatically identifying the data lightning whistle sound wave of the Zhang He-Yi induction magnetometer corresponding to the method, which comprises the following steps:
sample collection module: the SCM data are used for making Fourier spectrograms, according to the obvious L-shaped characteristic of lightning whistle sound waves in the spectrograms, the spectrograms are segmented to finish sample image collection, and a lightning spectrogram and a non-lightning sample spectrogram are collected;
graying and scale processing module: the method is used for carrying out graying treatment and scale reduction on the frequency spectrum image so as to reduce the calculation dimension and strengthen the lightning characteristic at the same time;
And the fuzzy convolution processing module is used for: the method is used for designing a fuzzy convolution kernel, and carrying out convolution calculation on the image to filter out the influence of a large amount of step edge information;
l-shaped convolution processing module: the L-shaped convolution kernel is used for designing the L-shaped convolution kernel based on the L-shaped characteristics of the lightning whistle sound wave, and the L-shaped characteristics in the image are further enhanced by carrying out convolution processing on the image;
SVM model classification recognition module: the method is used for inputting the enhanced image into a Support Vector Machine (SVM) for training and classifying recognition, and obtaining a recognition result.
The embodiment also provides an automatic recognition system for the data lightning whistle wave of the first induction magnetometer of the Zhangheng, which comprises: one or more processors; and the storage device is used for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors realize the automatic identification method of the data lightning whistle sound waves of the Zhang He Yi induction magnetometer. Since the hardware part is a conventional design in the art, it is not described in detail here.
The above description is only of the preferred embodiments of the present invention, and is not intended to limit the invention in any way, and some simple modifications, equivalent variations or modifications can be made by those skilled in the art using the teachings disclosed herein, which fall within the scope of the present invention.
Claims (10)
1. The automatic identification method for the data lightning whistle sound wave of the Zhangheng first induction magnetometer is characterized by comprising the following steps:
sample collection: performing Fourier spectrogram on SCM data, dividing the spectrogram according to the obvious L-shaped characteristic of lightning whistle sound waves in the spectrogram to finish sample image collection, and collecting a lightning spectrogram and a non-lightning sample spectrogram;
graying and scale treatment: graying treatment and scale reduction are carried out on the frequency spectrum image so as to reduce the calculation dimension and strengthen the lightning characteristic;
fuzzy convolution processing: designing a fuzzy convolution kernel, and carrying out convolution calculation on the image to filter out the influence of a large amount of step edge information;
l-form convolution processing: designing an L-shaped convolution kernel based on the L-shaped characteristic of the lightning whistle sound wave, and carrying out convolution processing on the image to further enhance the L-shaped characteristic in the image;
SVM model classification and identification: and inputting the enhanced image into a Support Vector Machine (SVM) for training and classifying recognition to obtain a recognition result.
2. The automatic identification method of lightning whistle sound waves of Zhang He-Yi induction magnetometer data according to claim 1, wherein the SCM data mainly come from waveform data of VLF wave bands of SCM loads of three orbits of Zhang He-Yi satellite, a sliding window is designed in the collection process, the width is 10 2s, the step length is 2s, and the sliding window is adopted to intercept data from 1000Hz-6000Hz from power spectrum data.
3. The automatic identification method of the Zhang Heng-one induction magnetometer data lightning whistle sound wave according to claim 1, wherein the graying treatment is as follows:
Gray＝RGB.R×0.3+RGB.G×0.59+RGB.B×0.11
where RGB represents the original sample spectrogram, rgb.r is the pixel value of its red channel, rgb.g and rgb.b are the values of the green and blue channels, respectively, and Gray is the graying spectrogram.
4. The method for automatically identifying data lightning whistle sound waves of Zhang He-Yi induced magnetometer according to claim 1, wherein the scale is processed to reduce the scale to 50 x 50.
5. The automatic recognition method of the Zhang He-Yi induced magnetometer data lightning whistle sound waves according to claim 1, wherein the fuzzy convolution kernel is an all-1 template with a scale of 5×5.
6. The automatic identification method of the Zhang He-Yi induced magnetometer data lightning whistle sound waves according to claim 1, wherein the L-shaped convolution kernel is a 9X 9L-shaped convolution kernel.
7. The automatic recognition method of the Zhang Heng one-number induction magnetometer data lightning whistle sound waves according to claim 1, wherein in the SVM model classification recognition, the SVM adopts a polynomial kernel function k (x i ，x j )：
Wherein x is j Representing the j-th figureFeature vector of image x i A feature vector representing an i-th image; gamma and r are parameters to be adjusted, the SVM is from SVC library of Python; where the optimal parameters d are 13, γ and r are all other default parameters under the library.
8. The automatic recognition method of the Zhang Heng one-number induction magnetometer data lightning whistle sound waves according to claim 1, wherein in the SVM model classification recognition, the adopted evaluation indexes are: recognition accuracy, recall, F1 value, and AUC-ROC.
9. Zhangheng first response magnetometer data lightning whistle sound wave automatic identification equipment, its characterized in that includes:
sample collection module: the SCM data are used for making Fourier spectrograms, according to the obvious L-shaped characteristic of lightning whistle sound waves in the spectrograms, the spectrograms are segmented to finish sample image collection, and a lightning spectrogram and a non-lightning sample spectrogram are collected;
graying and scale processing module: the method is used for carrying out graying treatment and scale reduction on the frequency spectrum image so as to reduce the calculation dimension and strengthen the lightning characteristic at the same time;
and the fuzzy convolution processing module is used for: the method is used for designing a fuzzy convolution kernel, and carrying out convolution calculation on the image to filter out the influence of a large amount of step edge information;
L-shaped convolution processing module: the L-shaped convolution kernel is used for designing the L-shaped convolution kernel based on the L-shaped characteristics of the lightning whistle sound wave, and the L-shaped characteristics in the image are further enhanced by carrying out convolution processing on the image;
SVM model classification recognition module: the method is used for inputting the enhanced image into a Support Vector Machine (SVM) for training and classifying recognition, and obtaining a recognition result.
10. Zhangheng first response magnetometer data lightning whistle wave automatic identification system, its characterized in that includes:
one or more processors;
storage means for storing one or more programs,
when the one or more programs are executed by the one or more processors, the one or more processors implement the tensegrity one induction magnetometer data lightning whistle sound wave automatic identification method according to any one of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011623352.4A CN112633227B (en) | 2020-12-30 | 2020-12-30 | Automatic recognition method and system for data lightning whistle sound waves of Zhangheng first induction magnetometer |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011623352.4A CN112633227B (en) | 2020-12-30 | 2020-12-30 | Automatic recognition method and system for data lightning whistle sound waves of Zhangheng first induction magnetometer |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112633227A CN112633227A (en) | 2021-04-09 |
| CN112633227B true CN112633227B (en) | 2024-02-23 |
Family
ID=75289740
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011623352.4A Active CN112633227B (en) | 2020-12-30 | 2020-12-30 | Automatic recognition method and system for data lightning whistle sound waves of Zhangheng first induction magnetometer |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112633227B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113361476B (en) * | 2021-07-02 | 2023-07-25 | 中国地震局地震预测研究所 | Zhang Heng one-number pre-earthquake abnormal signal identification method based on artificial intelligence technology |
| CN113782054B (en) * | 2021-09-22 | 2023-09-15 | 应急管理部国家自然灾害防治研究院 | Lightning whistle sound wave automatic identification method and system based on intelligent voice technology |
| CN114764845A (en) * | 2022-04-16 | 2022-07-19 | 中国人民解放军战略支援部队航天工程大学 | Imaging method for simulating event camera working in space |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN2232974Y (en) * | 1995-06-20 | 1996-08-14 | 中国科学院声学研究所 | Sound wave clog releasing device |
| US7987144B1 (en) * | 2000-11-14 | 2011-07-26 | International Business Machines Corporation | Methods and apparatus for generating a data classification model using an adaptive learning algorithm |
| CN103323026A (en) * | 2013-05-30 | 2013-09-25 | 北京控制工程研究所 | Attitude standard deviation estimation and correction method of star sensor and payload |
| CN106385756A (en) * | 2016-09-21 | 2017-02-08 | 北京机械设备研究所 | Electric arc heating type helicon wave plasma electric propulsion device |
| CN107545890A (en) * | 2017-08-31 | 2018-01-05 | 桂林电子科技大学 | A kind of sound event recognition method |
| CA2974031A1 (en) * | 2016-07-22 | 2018-01-22 | Appropolis Inc. | Method of tracking one or more mobile objects in a site and a system employing same |
| WO2019041629A1 (en) * | 2017-08-30 | 2019-03-07 | 哈尔滨工业大学深圳研究生院 | Method for classifying high-dimensional imbalanced data based on svm |
| CN109858450A (en) * | 2019-02-12 | 2019-06-07 | 中国科学院遥感与数字地球研究所 | Ten meter level spatial resolution remote sensing image cities and towns extracting methods of one kind and system |
| CN109977576A (en) * | 2019-04-03 | 2019-07-05 | 北京理工大学 | A kind of constellation system multidisciplinary design optimization based on agent model |
| CN111814699A (en) * | 2020-07-13 | 2020-10-23 | 中国地震局地震预测研究所 | Deep learning earthquake prediction method for SWARM electromagnetic satellite data |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| IN2012DE00578A (en) * | 2011-03-02 | 2015-08-21 | Bae Systems Information & Electornic Systems Integration Inc |
-
2020
- 2020-12-30 CN CN202011623352.4A patent/CN112633227B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN2232974Y (en) * | 1995-06-20 | 1996-08-14 | 中国科学院声学研究所 | Sound wave clog releasing device |
| US7987144B1 (en) * | 2000-11-14 | 2011-07-26 | International Business Machines Corporation | Methods and apparatus for generating a data classification model using an adaptive learning algorithm |
| CN103323026A (en) * | 2013-05-30 | 2013-09-25 | 北京控制工程研究所 | Attitude standard deviation estimation and correction method of star sensor and payload |
| CA2974031A1 (en) * | 2016-07-22 | 2018-01-22 | Appropolis Inc. | Method of tracking one or more mobile objects in a site and a system employing same |
| CN106385756A (en) * | 2016-09-21 | 2017-02-08 | 北京机械设备研究所 | Electric arc heating type helicon wave plasma electric propulsion device |
| WO2019041629A1 (en) * | 2017-08-30 | 2019-03-07 | 哈尔滨工业大学深圳研究生院 | Method for classifying high-dimensional imbalanced data based on svm |
| CN107545890A (en) * | 2017-08-31 | 2018-01-05 | 桂林电子科技大学 | A kind of sound event recognition method |
| CN109858450A (en) * | 2019-02-12 | 2019-06-07 | 中国科学院遥感与数字地球研究所 | Ten meter level spatial resolution remote sensing image cities and towns extracting methods of one kind and system |
| CN109977576A (en) * | 2019-04-03 | 2019-07-05 | 北京理工大学 | A kind of constellation system multidisciplinary design optimization based on agent model |
| CN111814699A (en) * | 2020-07-13 | 2020-10-23 | 中国地震局地震预测研究所 | Deep learning earthquake prediction method for SWARM electromagnetic satellite data |
Non-Patent Citations (3)
| Title |
|---|
| 小波计盒分形维数核爆电磁脉冲信号模糊识别;庞新良;李鹏;范江兵;常芸芬;李欣;;强激光与粒子束(09);全文 * |
| 张衡一号卫星记录的空间电磁波传播特征分析方法及算法实现;胡云鹏;泽仁志玛;黄建平;赵庶凡;郭峰;王桥;申旭辉;;地球物理学报(05);全文 * |
| 电离密度分布对中低纬非导管哨声波传播的影响;徐文耀;地球物理学报(03);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112633227A (en) | 2021-04-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112633227B (en) | Automatic recognition method and system for data lightning whistle sound waves of Zhangheng first induction magnetometer | |
| Huang et al. | A new building extraction postprocessing framework for high-spatial-resolution remote-sensing imagery | |
| US8977062B2 (en) | Reduction of CFAR false alarms via classification and segmentation of SAR image clutter | |
| Khazai et al. | An approach for subpixel anomaly detection in hyperspectral images | |
| CN105956612B (en) | Hyperspectral image classification method based on Active Learning and semi-supervised learning | |
| CN109657610A (en) | A kind of land use change survey detection method of high-resolution multi-source Remote Sensing Images | |
| CN111666854B (en) | High-resolution SAR image vehicle target detection method fusing statistical significance | |
| CN111428631B (en) | Visual identification and sorting method for unmanned aerial vehicle flight control signals | |
| CN104182985A (en) | Remote sensing image change detection method | |
| CN103208011A (en) | Hyperspectral image space-spectral domain classification method based on mean value drifting and group sparse coding | |
| Tang et al. | A multiple-point spatially weighted k-NN method for object-based classification | |
| CN102509120A (en) | Supervised image segmentation method for hyperspectral image based migration dictionary learning | |
| Li et al. | SDBD: A hierarchical region-of-interest detection approach in large-scale remote sensing image | |
| CN105139034B (en) | A kind of ship detecting method that combination spectrum filters out | |
| Zhang et al. | Isolation forest for anomaly detection in hyperspectral images | |
| Mesecan et al. | Searching the effects of image scaling for underground object detection using KMeans and KNN | |
| Zou et al. | A method of radar echo extrapolation based on TREC and Barnes filter | |
| CN108985357A (en) | The hyperspectral image classification method of set empirical mode decomposition based on characteristics of image | |
| Fu et al. | A noise-resistant superpixel segmentation algorithm for hyperspectral images | |
| CN110046639A (en) | A kind of Hyperspectral imaging noise label detection method based on super-pixel weight density | |
| CN108764145A (en) | One kind is towards Dragon Wet Soil remote sensing images density peaks clustering method | |
| CN112989940A (en) | Raft culture area extraction method based on high-resolution three-satellite SAR image | |
| Salem et al. | Anomaly detection in hyperspectral images based spatial spectral classification | |
| CN110175638B (en) | Raise dust source monitoring method | |
| Boursier et al. | Automatic detection of coronal mass ejections on LASCO-C2 synoptic maps |
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 | ||
| CB03 | Change of inventor or designer information |
Inventor after: Wang Qiao Inventor after: Zeren Zhima Inventor after: Shen Xuhui Inventor after: Yuan Jing Inventor after: Yang Dehe Inventor after: Zhou Xin Inventor after: Zhou Na Inventor before: Wang Qiao Inventor before: Yuan Jing |
|
| CB03 | Change of inventor or designer information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |